---
title: "13"
source: "13.docx"
converted: "2026-10-01 10:23:20 +0800"
converter: "MarkForge (mdconvert)"
engine: "MarkItDown"
---

A Systematic Review of Deep Learning and Quantum Reinforcement Learning for Secure Predictive Maintenance in Upstream Oil and Gas Systems

Philip Adebayo[[1]](#footnote-1)\*, Afolayan Obiyini2, Victoria Yemi-Peters3, Joshua Aggobun4

1Department of Computer Engineering. Kogi State Polytechnic, Lokoja, Nigeria.

2,3,4Department of Computer Science, Federal University Lokoja, Kogi State, Nigeria.

|  |  |  |
| --- | --- | --- |
| ARTICLE INFO |  | ABSTRACT |
| *Article history:*  Received 18 February 2026  Revised 12 May 2026  Accepted 8 June 2026  Online first  Published 1 September 2026 |  | Predictive maintenance (PdM) plays a crucial role in maintaining reliability and operational efficiency of upstream oil and gas systems. Recent developments in artificial intelligence and specifically deep learning (DL), reinforcement learning (RL), and the new quantum reinforcement learning (QRL) have increased PdM capabilities. Nonetheless, cybersecurity vulnerabilities arise with sensor-driven model integration and can undermine system integrity. This study conducts a systematic literature review following PRISMA guidelines to examine AI-based PdM in upstream oil and gas. Major scientific databases were used to retrieve relevant studies that were published between 2015 and 2025. The rigor and relevance of the studies were evaluated using a hybrid quality appraisal framework, which is a combination of methodological and machine learning-specific criteria. The findings show that DL-based approaches, especially hybrid models such as CNN-LSTM and CNN-Transformer, demonstrate high predictive accuracy and strong validation using real-world sensor data. RL methods exhibit potential for adaptive maintenance decision-making but are largely limited to simulation-based environments. QRL remains an emerging paradigm with limited empirical validation. Additionally, while cybersecurity threats such as data spoofing and adversarial attacks are increasingly recognised, few studies integrate security mechanisms into PdM frameworks. Quality assessment indicates that high-quality studies are predominantly DL-based, whereas RL and QRL studies are generally of moderate quality. The review identifies a very important gap in the designing of integrated, secure and scalable PdM systems. Future studies must aim at integrating the real-life validation, adaptive learning models, and cybersecurity resilience to facilitate strong AI-based maintenance solutions in upstream oil and gas. |
| *Keywords:*  PdM  Machine Learning  Deep Learning  Reinforcement Learning  Quantum Reinforcement Learning Upstream Oil and Gas  Industrial Analytics  Knowledge Extraction  DOI:  [10.24191/jcrinn.v11i2.601](https://doi.org/10.24191/jcrinn.v11i2.601) |  |

1. INTRODUCTION

The oil and gas upstream industry business is in a very complex and safety-sensitive environment where equipment reliability is required to maintain operational efficiency and reduce risks. Predictive maintenance (PdM) is a new initiative that is being propelled by real-time sensor-based data and advanced analytics to predict equipment failures and optimise maintenance according to the equipment. Deep learning (DL) methods have proved to be very effective in the last few years in the modelling of nonlinear degradation trends, whereas reinforcement learning (RL) and emerging quantum reinforcement learning (QRL) are adaptive decision-making models to optimise maintenance. Nevertheless, the growing interconnection of sensors and AI models creates cybersecurity risks that will result in undermining the quality of maintenance decisions.

Most of the upstream oil and gas activities are based on scheduled inspections or reactive repairs, which are usually expensive and clearly inefficient (Gowekar, 2024). Meanwhile, the move towards digital oilfields, remote sensors, Internet of Things (IoT) connectivity and cloud analytics has thrown down a challenge to cybersecurity that is too serious to ignore. Cyber-attacks have become a new and common method of intrusion into sensor networks and SCADA systems of oil and gas platforms, and even the most intelligent model may be confused by an infected data stream. Practically, cases where the upstream operations are interfered with by equipment failure or a cyber-intrusion (or both) have been observed; these then highlight the necessity of a pragmatic or proactive maintenance paradigm, which should be both predictive and secure.

A lot has been achieved in all these fronts individually. A significant literature on Machine Learning (ML) applies to equipment health monitoring on the PdM side. As an example, Gowekar (2024) points to the Predictive Maintenance (PdM) analysis of sensor data using AI, predicting the malfunction of machines to reduce cases of unexpected downtime and prolong their useful life. Deep neural networks, e.g., convolutional networks on vibration spectrograms or LSTM recurrent networks on time-series readings will be able to find more complicated patterns that are not detected by simpler techniques. The application of such models is widespread, according to reviews of pipeline and drilling applications. As an example, (Magdin, 2025) survey of AI in gas infrastructure observes that neural networks, support vector machines, and Bayesian networks prevail when it comes to failure prediction. Such works show remarkable improvements in fault detection. However, as Magdin (2025) also explained, there are still practical challenges, such as a lack of high-quality data for model training, high cost of implementation, and the existent of regulatory barriers. In other words, even properly designed AI maintenance may be delicate or difficult to scale.

At the cybersecurity level, oil and gas companies have spent a lot of money on the protection of IT/OT systems, yet primarily using traditional (firewalls, intrusion detection, and rule-based monitoring) methods. What has not been well investigated is the application of predictive concepts to security. Some revolutionary works have been embarked upon to fill this gap. (Hafsi et al., 2024) proposed a predictive maintenance-inspired cybersecurity framework in which network components are modelled using anomaly detection algorithms to forecast potential failures or attacks. Their approach demonstrated that early anomaly signals in network traffic can improve intrusion detection responsiveness. Similarly, (Amato et al., 2024) employed a TimeGAN-based data augmentation model combined with LSTM networks to detect adversarial attacks in IoT-enabled PdM systems, achieving detection accuracy in the range of 80 to 90% under simulated attack conditions. They demonstrated that the lack of independence between interrelated sensors and machine learning models exposes vulnerabilities that can be exploited by means of adversarial inputs or spoofed data, thus affecting the integrity of maintenance decisions. This is why it is imperative that the smart maintenance systems must be strongly safeguarded against malicious attacks as well as unintentional errors.

The gap identified is evident: PdM, which integrates deep learning, new approaches based on quantum, and oil and gas security, is not systematically studied. Surveys have so far analysed subsets of this space. As an example, some existing reviews include both traditional and deep learning-based maintenance in manufacturing and energy systems, while others review quantum ML in general energy applications. They do not all specialise in upstream oil and gas, and do not take into account cybersecure maintenance practices. Similarly, the fast-developing discipline of quantum reinforcement learning (QRL), which is expected to make decision-making faster and more efficient, has not received much attention in this regard.

Strata (2025) gives a generalised "scoping review" of Quantum Machine Learning (QML) in energy, citing that the existing hybrid quantum-classical agents already perform better than their classical counterparts in power system control tasks. Indicatively, Ajagekar and You (2024) trained a quantum-enhanced RL controller to achieve 13.6% lower energy consumption than a classical controller, and Andrés et al. (2022) trained a hybrid quantum agent to find optimal policies compared to a classical neural agent. Such encouraging outcomes indicate that one day QRL will be able to enhance PdM scheduling or intrusion response, although they have not been linked to oilfield systems. As a matter of fact, reviews such as Magdin (2025) or Gowekar (2024) do not mention quantum computing at all, and QRL papers often have little to say about industrial maintenance or security.

Therefore, the niche is clear: we require an all-encompassing, rigorous overview of how the deep learning and QRL methods may be jointly used in assisting secure PdM of upstream oil and gas. Gowekar (2024) and Magdin (2025) emphasise AI usage to reduce downtime; Hafsi et al. (2024) and Amato et al. (2024) raised the importance of predicting cyber-attacks in maintenance; and Strata (2025) features quantum models that could serve to learn in such systems faster. These insights will be incorporated into a more general conceptual framework of a secure PdM pipeline, starting with physical sensors, via AI analytics, to quantum-sensitive control.

This systematic review is important both theoretically and in practice. In academia, it creates a link between the research on machine learning for maintenance and cybersecurity/quantum computing literature, two domains that have been confined too frequently in recent times. It presents a conceptual framework and taxonomy that can be used to inform future experiments and designs. In practice, it provides upstream operators and engineers with a critical view on implementing the state-of-the-art AI and quantum methods. This knowledge of the joint landscape enables decision-makers to make their investments and risks more informed (e.g., is a quantum-enhanced planner worth developing, and how to certify the security).

* 1. Research objectives

The aim of the systematic review is to critically synthesise and discuss the role of deep learning (DL), quantum reinforcement learning (QRL), and cybersecurity in facilitating secure predictive maintenance (PdM) in upstream oil and gas systems. To fulfil this purpose and address the aforementioned gaps, this study defines clear research objectives and contributions, which guide the systematic review process:

1. To examine the state-of-the-art deep learning methods currently used for predictive maintenance with upstream oil and gas systems.
2. To analyse and compare the quantum reinforcement learning strategies with classical reinforcement strategies in maintenance decision-making and cybersecurity applications.
3. To evaluate the cybersecurity risks to AI-based predictive maintenance systems and the efficiency of current DL and QRL-based mitigation strategies.
4. To identify and synthesise the research gaps, limitations, and integration issues that are crucial in secure PdM frameworks in the context of DL, QRL, and cybersecurity.
5. To propose a systematic research agenda and conceptual direction for future development of secure, smart, and scalable PdM systems.
   1. Research questions

The study tries to answer the following research questions (RQs):

**RQ1:** What deep learning methods are used for PdM in upstream oil and gas systems, and what are their reported strengths and limitations?

**RQ2:** Which quantum reinforcement learning approaches have been proposed for maintenance decision-making or cybersecurity, and how do they compare with classical reinforcement learning methods?

**RQ3:** What cybersecurity threats affect AI-based PdM systems in upstream oil and gas, and how are these threats addressed in existing DL and QRL approaches?

**RQ4:** What gaps remain in the integration of deep learning, quantum reinforcement learning, and cybersecurity for secure PdM?

* 1. Distinct contribution of this review

This systematic review makes the following key contributions:

1. Comprehensive synthesis: It is the systematic synthesis and appraisal of deep learning, reinforcement learning and novel quantum reinforcement learning methods in the context of upstream oil and gas predictive maintenance
2. Comparative analytical structure: It compares and contrasts these three paradigms of DL, classic RL and QRL with respect to predictive, decision making, scalability and applicability to cybersecurity-aware PdM systems
3. Cybersecurity integration outlook: It analyzes and unites cyber threats and mitigation measures in AI-based PdM and the fact that the intersection between predictive analytics and cyber resilience is frequently overlooked.
4. Critical research gaps identification: It classifies and characterises the main technical, methodological, and implementation issues, such as data heterogeneity, interpretability of models, adversarial resistance, and quantum hardware constraints.
5. Future research agenda: It suggests a systematic and practice-oriented research agenda on hybrid DL -QRL architectures, secure-by-design PdM systems, and quantum-enhanced decision intelligence.

The rest of the paper will adhere to the PRISMA 2020 systematic review framework for transparency and rigour. Our methodology is outlined in Section 2: search terms, inclusion criteria and screening procedure. In section 3, the findings of our review are provided, where the most important deep learning strategies in oilfield PdM are arranged, and the applicable quantum reinforcement methods are analysed. Section 4, the discussion section, evaluates the intersection of these approaches with cybersecurity, we examine vulnerabilities that are described in the literature and defensive mechanisms suggested. We bring out repetitive themes and unresolved questions along the way. Lastly, Section 5, the conclusion summarizes our results, directly stating what we have not learned and making recommendations on the future agenda. At this point, the reader will get a clear understanding of what has been accomplished and what there is to accomplish at the intersection of deep learning, QRL and the secure PdM of upstream oil and gas.

1. METHODOLOGY

This review was carried out in compliance with PRISMA-2020 systematic reviews (Haddaway et al., 2022). We have put together a clear protocol of the objectives, inclusion and exclusion criteria and how to analyse it prior to searching. The theoretical scope contains several overlapping subfields: upstream oil and gas (O&G) systems, deep learning (DL), and reinforcement learning (RL) algorithms (including quantum reinforcement learning (QRL), Predictive Maintenance (PdM) tasks, and cybersecurity considerations. PdM finds application in upstream O&G, which focuses on exploration and production, where sensor data and AI are used to predict equipment failures as described in recent reviews (Betha, 2021; Saputelli et al., 2022). The hybrid quantum/classical RL paradigm, also known as QRL, is a new strategy at this intersection (Kim, 2024). This review intends to find out how the DL/RL/QRL methods can be used to contribute to safe PdM in upstream O&G by integrating these fields. The PRISMA framework steered the protocol and reporting to facilitate the transparency of the objectives, methods, and findings (Haddaway et al., 2022).

* 1. Search strategy and query formulation

To ensure transparency and reproducibility, a structured search was formulated with Boolean operators using a combination of keywords and the syntax of respective databases. The search was carried out according to four main dimensions, namely, predictive maintenance, upstream oil and gas, artificial intelligence methods (DL/RL/QRL), and cybersecurity.

The following generic search string was adapted for each database:

(“predictive maintenance” OR “PdM” OR “condition-based maintenance”)

AND (“upstream oil and gas” OR “oilfield” OR “petroleum production”)

AND (“deep learning” OR “neural network” OR “reinforcement learning” OR “quantum reinforcement learning”)

AND (“cybersecurity” OR “intrusion detection” OR “adversarial attack” OR “data integrity”)

The search strings were customised as follows:

**Scopus**

TITLE-ABS-KEY (“predictive maintenance” OR “condition-based maintenance”) AND TITLE-ABS-KEY (“oil and gas” OR “upstream petroleum”) AND TITLE-ABS-KEY (“deep learning” OR “reinforcement learning” OR “quantum reinforcement learning”) AND TITLE-ABS-KEY (“cybersecurity” OR “adversarial attack”)

**Web of Science**

TS = (“predictive maintenance” OR “condition-based maintenance”) AND TS = (“oil and gas” OR “upstream petroleum”) AND TS = (“deep learning” OR “reinforcement learning” OR “quantum reinforcement learning”) AND TS = (“cybersecurity” OR “data integrity”)

**IEEE Xplore**

(“All Metadata”:“predictive maintenance” OR “condition-based maintenance”) AND (“oil and gas”) AND (“deep learning” OR “reinforcement learning”) AND (“cybersecurity” OR “intrusion detection”)

**Search Period and Dates**

The literature search was performed between January 10, 2026 and February 15, 2026. Only studies published between 2015 and 2025 were taken into consideration to reflect the current trends in deep learning, reinforcement learning, and cybersecurity applications in predictive maintenance.

The number of searches was restricted to articles published from 2018 to 2025. The seven-year window was selected as this specific window includes more recent developments of fast-moving fields such as DL and QRL and excludes older work that might be out-of-date. This was theoretically justified because the time limit was suggested by systematic review guidelines. Peer-reviewed journal and conference articles that were in English were retained. Reference lists of included papers were also screened (this is also known as reference chaining), in order to find any other relevant studies (Vizváry & Grigas, 2025).

* 1. Eligibility Criteria

Our inclusion/exclusion criteria were clearly defined before screening and included the following dimensions. Studies were selected based on the following inclusion and exclusion criteria:

**Inclusion Criteria:**

1. Peer-reviewed journal articles and conference papers
2. Studies that focus on predictive maintenance in upstream oil and gas
3. Use of DL, RL, or QRL
4. Empirical or simulation-based validation using sensor or operational data
5. Journal articles published between 2015 and 2025

**Exclusion Criteria:**

1. Studies having nothing to do with upstream oil and gas operations
2. Studies that are purely theoretical works without validation
3. Studies lacking clear methodology or results
4. Non-English publications
5. Duplicate studies or articles with inaccessible full-text
   1. Study Selection and Screening Process

The study selection process aligned with the PRISMA guidelines and consisted of four distinct phases or stages; they are:

1. **Identification**

300 records were extracted from the sampled databases. 3 records were eliminated, and 297 unique studies were obtained.

**2. Screening**

The rest of the studies were screened based on predetermined inclusion criteria (titles and abstracts), and N₄ records that were irrelevant to either predictive maintenance, upstream oil and gas, or AI-based solutions were excluded.

**3. Eligibility:**

65 articles were evaluated in terms of eligibility. Other studies were eliminated at this point (n = 40) because of:

1. Not dependent on upstream oil and gas (n = 14)
2. No cybersecurity component (n = 17)
3. No DL/RL/QRL methodology (n = 9)

**4. Inclusion:**

A final set of 25 studies was included in the qualitative and comparative analysis. The study selection process is illustrated in the PRISMA flow diagram of Fig. 1, showing the number of records at each stage and reasons for exclusion.

Records identified from:

Databases (n = 300)

Registers (n = 0)

Records removed *before screening*:

Duplicate records removed

(n = 3)

Records removed for other reasons (n = 0)

Records screened

(n = 297)

Records excluded (n = 232)

Reports sought for retrieval

(n = 65)

Reports not retrieved:

Publisher paywall (n = 0)

Conference abstract only (n = 0)

Reports assessed for eligibility

(n = 65)

Reports excluded: (n = 40)

Not Upstream oil & gas (n = 14)

No cybersecurity component (n = 17)

No DL/RL/QRL methodology (n = 9)

Studies included in review

(n = 25)

Reports of included studies

(n = 25)

**Identification of studies via databases**

**Identification**

**Screening**

**Included**

Fig. The PRISMA 2020 flow diagram of study selection

Source: Haddaway et al. (2022)

All search results were pooled, and then duplicates were removed following PRISMA guidelines. Two independent reviewers screened the titles and abstracts of unique records based on compliance with the inclusion criteria. To ensure consistency, a screening form was utilised with the above criteria. Full text of the records was obtained, which were considered likely to be eligible. Two reviewers then independently assessed and evaluated the full texts on all criteria. All issues of disagreement were solved by discussion, and if the need arises, by a third reviewer. For the sake of transparency, reasons given at the full-text stage (e.g. wrong domain, non-ML technique, no PdM focus, or insufficient detail) were recorded. All the included studies had their reference lists searched by hand in order to find any more articles that were not included by the initial query.

In full-text assessment, a subset of papers that were originally considered relevant at the abstract stage were filtered out based on the inadequacy with the upstream oil and gas operations, not satisfying a cybersecurity focus, or not covering a substantive deep learning or reinforcement learning methodology. Figure 1 shows the number of records at each stage that were identified, screened, excluded and included in the final review. It provides a clear overview of how the filtering of the studies was done. Any omissions (with reasons) were recorded in compliance with PRISMA guidelines.

The systematic review methodology and screening process are summarised in Table 1.

Table 1. Summary of systematic review methodology

|  |  |  |
| --- | --- | --- |
| **Methodological Component** | **Approach Adopted** | **Alignment with Research Questions** |
| Review Design | Systematic Literature Review | Supports all RQs through structured evidence synthesis |
| Study Selection | RQ-driven inclusion/exclusion criteria | RQ1–RQ4 |
| Quality Assessment | A hybrid evidence synthesis approach | Ensures methodological and computational rigor |
| Data Extraction | Manual thematic extraction with validation | Enables consistent cross-study comparison |
| Evidence Synthesis | Narrative, comparative, and thematic synthesis | RQ1–RQ4 |
| Analytical Focus | Applied ML and optimisation | Emphasises real-world  relevance |
| Output | Taxonomies, challenges, and research agenda | Informs both research and practice |

* 1. Data Extraction Strategy

Based on every included study, the following fields were extracted in a standardized format: authors and year; O&G context (equipment or process); type and architecture of the DL/RL/QRL model; source of data (type and size of sensor data); preprocessing of data; type of PdM activity (e.g. RUL regression, fault classification, anomaly detection); cybersecurity implementation (e.g. encryption, intrusion detection, secure architectures); performance metrics and results; and limitations were noted. Experimental settings and hyperparameters, if any, that were reported as well as code and model, publicly available, were also noted. These details were collated in a summary table of all the studies.

Due to the heterogeneity of approaches and measures, a qualitative synthesis of the results was used. A narrative thematic synthesis was performed, which clustered studies according to major themes: e.g. DL-based vs RL-based methods, traditional vs quantum methods, and the kind of security measures that were implemented. Potential common patterns (e.g. common architectures, common datasets, common accuracy ranges) were found and catalogued. A taxonomy of approaches was also created by classifying them based on the type of model (CNN, RNN, classical RL, QRL, etc.) and PdM task. In this case, narrative synthesis is relevant because there could not be a meta-analysis (Melendez-Torres et al., 2015). Data and quality ratings extracted from studies were compared in order to make a point of the state-of-the-art techniques, performance, and gaps.

* 1. Evidence Synthesis Method

A hybrid evidence synthesis approach was used, which entailed the synthesis of narrative, thematic, and comparative synthesis. Overall methodological trends and spheres of application were described with the help of narrative synthesis. Thematic synthesis also made it possible to find common patterns in the learning architectures, optimisation strategies and deployment issues. Where enough empirical detail existed, comparative synthesis was used to enable the performance properties and empirical trade-offs between deep learning and reinforcement learning methods. Such a multi-level synthesis choice was chosen to strike a balance between the level of analysis and practical application, as it is with the applied nature of computing research.

* 1. Quality Appraisal

In order to increase the reliability and objectivity of the quality assessment (Purssell et al., 2024), all the studies included were assessed by two independent reviewers (Reviewer A and Reviewer B) with the help of the specified hybrid appraisal framework. The studies were individually evaluated by each reviewer, taking into consideration the level of methodology, the criteria of machine learning evaluation, and the relevance to the domain. The reviewers were not aware of one another's scores in the initial assessment period. The scores were cross-examined after the independent evaluation, and any divergence was resolved through a discussion. An average score was calculated where there was no agreement. The Inter-rater agreement was determined by means of percentage agreement which ensured uniformity in assessment.

* 1. Bias Management

Numerous precautions were taken to reduce bias. Such a wide, synonym-expanded search of multiple databases reduced the potential danger of failing to identify pertinent studies, and the inclusion of "quantum reinforcement learning" directly avoided the domination of DL-focused terminology. The search strategy was created a priori, and it was peer-reviewed; screening and extraction were conducted independently, and disputes were solved through consensus or arbitration by a third reviewer. All the exclusion reasons were documented, and the PRISMA diagram documents the selection process for transparency. Technological bias was also put into consideration by the inclusion of journals and conferences, and the cross-checking of references and author names, so that the QRL and cybersecurity works should not be neglected.

Research was also excluded when it did not have a noticeable computational approach, only addressed the conventional statistical maintenance models, and was too sparse to supply adequate technical detail that would assist in reproducibility or comparative research. This made sure that the result of the corpus of studies was closely aligned to the goals and the analytical scope of the review.

* 1. Generative Artificial Intelligence use

To write the present manuscript, generative AI (ChatGPT, OpenAI) was used to assist with refining the clarity of narratives, structuring the parts of the systematic review, and creating tables to visually summarize the methodological procedures (e.g., quality assessment models like GRADE and CASP). The authors worked alone in the development of all conceptual decisions and interpretations, methodological designs, and data analyses, and all output produced by AI had to be subjected to critical review and revision to guarantee accuracy and adherence to the objectives of the research.

1. RESULTS AND INTERPRETATION

The results of the study are now presented within the confines of the research questions that guided the execution of the systematic literature review. The quality appraisal findings (Table 2) were included in the analysis process to make sure that the conclusions are made with a proper weighting, according to the methodological rigour. The high-quality studies were given priority in the synthesis process, especially those that used real-world sensor data and strong validation methods, and moderate-quality studies found were taken cautiously.

In order to give an overview of the methodological distribution of the included studies, Fig. 2 presents the categorisation of the studies based on their dominant techniques.

![C:\Users\MAIN USER\Pictures\Screenshots\method_distribution_bar (3).jpg](data:image/jpeg;base64...)

Fig. 2. Distribution of studies by method category

Fig. 3 demonstrates that deep learning-related approaches take up the majority of the literature; next are cybersecurity-related and reinforcement learning-related studies. Quantum reinforcement learning is a relatively new field that has fewer contributions.

![C:\Users\MAIN USER\Pictures\Screenshots\Screenshot 2026-04-08 134149.png](data:image/png;base64...)

Fig. 3. Proportion of studies by method category

Each study was given a final score as an average of the two scores of the reviewers. Quality level was allocated as follows:

i. High: greater than or equal to 8.0

ii. Moderate: 6.0 to 7.9

iii. Low: 4.0 to 5.9

iv. Very Low: less than 4.0

Table 2. The results of quality assessment

|  |  |  |  |  |  |  |  |
| --- | --- | --- | --- | --- | --- | --- | --- |
| **Study** | **Model** | **Data Type** | **Validation** | **Cybersecurity Considered** | **Score** | **Quality Level** |  |
| Aranha et al. (2024) | Unsupervised ML | Real sensor data | Field validation | No | 8.4 | High |  |
| Du et al. (2024) | Deep Learning | Real production data | Train/test split | No | 7.8 | Moderate |  |
| Sun et al. (2025) | CNN–Transformer | Real field data | Benchmarking | No | 8.7 | High |  |
| Lu et al. (2025) | CNN–BiLSTM–AdaBoost | Real pipeline data | Cross-validation | No | 8.8 | High |  |
| Wang et al. (2025) | Ensemble Learning | Real equipment data | Comparative validation | No | 8.3 | High |  |
| Putro et al. (2024) | Deep Learning | Gas pipeline data | Train/test split | No | 7.5 | Moderate |  |
| Zhang et al. (2022) | GRU | Industrial/simulated | Experimental | No | 7.0 | Moderate |  |
| Varalakshmi & Kumar (2025) | DRL + Ensemble | IoT streaming data | Simulation + benchmarking | Yes | 8.1 | High |  |
| Mesadieu et al. (2024) | Deep RL | SCADA (simulated) | Experimental | Yes | 7.9 | Moderate |  |
| Sangoleye et al. (2024) | DRL | Industrial control data | Benchmark testing | Yes | 8.2 | High |  |
| Amato et al. (2024) | LSTM + TimeGAN | IoT PdM data | Adversarial testing | Yes | 8.5 | High |  |
| Dionisopoulos et al. (2023) | ML robustness | Simulated adversarial | Experimental | Yes | 7.6 | Moderate |  |
| Chen et al. (2022) | QRL (VQC) | Simulation | Experimental | No | 6.8 | Moderate |  |
| Neumann et al. (2023) | QRL vs DRL | Synthetic | Comparative | No | 6.5 | Moderate |  |
| Skolik et al. (2022) | Quantum DQN | Simulation | Controlled experiment | No | 6.7 | Moderate |  |
| Said et al. (2024) | RL + Quantum entropy | Smart grid (simulated) | Experimental | Yes | 7.7 | Moderate |  |
| Hohenfeld et al. (2022) | QRL | Simulation (robotics) | Experimental | No | 6.4 | Moderate |  |
| Jerbi et al. (2021) | Quantum RL | Synthetic | Theoretical + experimental | No | 6.6 | Moderate |  |
| Saggio et al. (2021) | Quantum RL | Experimental physics | Lab validation | No | 7.2 | Moderate |  |
| Moll & Kunczik (2021) | Hybrid QRL | Simulation | Comparative | No | 6.3 | Moderate |  |
| Zhao et al. (2024) | Transformer-DRL | Simulation | Benchmarking | No | 7.8 | Moderate |  |
| Shingne et al. (2025) | DRL + VAE | Synthetic/security data | Experimental | Yes | 7.9 | Moderate |  |
| Khanzadeh et al. (2025) | DL (Cyber defense) | Security datasets | Comparative validation | Yes | 8.0 | High |  |
| Progoulakis et al. (2021) | Cybersecurity models | Real offshore data | Case-based | Yes | 7.4 | Moderate |  |
| Stergiopoulos et al. (2020) | Cyber-attack analysis | Incident datasets | Analytical validation | Yes | 7.6 | Moderate |  |

The distribution of study quality levels is illustrated in Fig. 4.

![C:\Users\MAIN USER\Desktop\IN-COUNTRY\Publication 2 Systematic Review\Screenshot 2026-04-08 081012.png](data:image/png;base64...)

Fig. 4. A bar chart showing quality distribution for publication figures

Any difference in rating of more than 1.0 points between reviewers led to re-evaluation and consensus discussion. Table 4 shows the results of quality assessment, in which every study was analysed on the hybrid appraisal framework. It is noted that research that used deep learning models that utilized real-world sensor data and effective validation approaches had more quality scores, whereas quantum reinforcement learning research was usually constrained by simulation-based validation. This highlights a critical gap in the integration of quantum reinforcement learning with real-world predictive maintenance and cybersecurity applications in upstream oil and gas systems.

**RQ1**: What deep learning methods are used for PdM in upstream oil and gas systems, and what are their reported strengths and limitations?

To provide a structured overview of the included studies, the distribution of methodologies and model architectures is first examined.

Among 300 papers that were screened, 25 articles 25 met the inclusion criteria, focusing on deep learning (DL) and quantum reinforcement learning (QRL) to predictive maintenance in upstream oil and gas systems with little cybersecurity integration. The methodological space was dominated by convolutional neural networks (CNNs) and recurrent architectures (LSTM/GRU), with CNNs being used in about a quarter to half of the studies and RNN variants about a third. The most common models were transformer, attention-based, and autoencoder models that represent about 10 to 20 per cent of the literature. Empirical results also emphasize the benefit of hybrid models, as in one study, Sun et al. (2025) found that a CNNTransformer model achieved an R² of approximately 0.72 for shale gas flowback prediction, outperforming CNN-LSTM (R² ≈ 0.49) and CNN-GRU (R² ≈ 0.57). Also, hybrid and ensemble methods are becoming increasingly popular, and some studies combine CNNs and expert rules or meta-learning methods to boost robustness. In general, a more in-depth, multi-block architecture, with CNNs, LSTMs, and attention, frequently enhanced with physics-based or heuristic components, is the current trend in methodologies.

Deep learning is transforming predictive maintenance in the upstream oil and gas industry into a direction where the industry is no longer reactive and time-based but is anticipatory in all its decision-making (Alabadi and Habbal, 2023; Babayeju et al., 2024). These models exploit the dense, continuous sensor channels produced by Industrial Internet of Things (IIoT) systems in each drilling rig, pump, turbine, and subsea system, by which subtle degradation patterns can be observed before they become a failure (Shil, 2025). In such contexts, with degradation being subtle and failure being expensive, deep learning offers a way to transform unfiltered sensor data into early warning messages as opposed to failure explanations.

RNNs based on LSTM and GRU models are the most popular architectures, which can process multivariate time series data, including pressure, temperature, and vibration, to estimate remaining useful life (Gowekar, 2024; Azmi et al., 2024). CNNs models are also increasingly becoming more important since they can detect faulty signatures in vibration spectrograms and other high-frequency sensor channels (Khan et al., 2025; Elijah et al., 2021). The unsupervised methods, such as stacked autoencoders and deep belief networks, are extensively deployed in the detection of anomalies, where the labelled data may be insufficient, or patterns are strongly irregular (Gong et al., 2021). There are new physics-informed neural networks, which impose physical constraints on the learning process, which are based on equipment behaviour and which enhance stability in the presence of noisy or incomplete sensor data (Fan et al., 2025). In the meantime, deep reinforcement learning helps to make downstream decisions by learning maintenance strategies through interacting with the operational environment via rewards (Du et al., 2024).

Although there are these strengths, there are several limitations (Du et al., 2024). Deep learning models require extensive amounts of labelled failure information that upstream operations do not have, and thus the difficulty in training predictable models (Lawal et al., 2024). They are also opaque and black box, which also limits interpretability in safety-critical environments, where the reasons behind predicted failures are frequently demanded (Serradilla et al., 2021). Most architectures are computationally intensive, which makes them less suitable to be deployed on edge devices in real time (Putro et al., 2024). Besides, geological and operational variability causes models to be hard to generalise across fields and rigs and adds strong dependence on domains (Tariq et al., 2021). Overfitting can always be a significant issue in small or imbalanced datasets, even in case of high performance, when the models can pick up noise rather than actual degradation patterns (Zhang et al., 2022).

Table 3. Overview of deep learning techniques in PdM in upstream oil and gas

|  |  |  |  |  |  |  |  |
| --- | --- | --- | --- | --- | --- | --- | --- |
| **DL Category** | **Representative Methods** |  | **PdM Tasks** | **Sensor Data** | **Key Strengths** | **Key Limitations** | **Core References** |
| Temporal DL Models | RNN, LSTM, GRU |  | RUL prediction, failure forecasting | Pressure, temperature, vibration | Capture long-term temporal dependencies; strong performance in degradation modelling | Data-hungry; limited interpretability; domain dependence | Gowekar (2024); Azmi et al. (2024); Tariq et al. (2021) |
| Convolutional Models | CNN, 1D-CNN, CNN-spectrogram |  | Fault classification, bearing diagnostics | High-frequency vibration, acoustic | Effective pattern extraction from raw signals; robust to noise | High computational cost; sensitive to data representation | Khan et al. (2025); Elijah et al. (2021); Putro et al. (2024) |
| Unsupervised DL | Stacked Autoencoders |  | Anomaly detection | Unlabelled multivariate sensors | Suitable for rare-failure scenarios; reduces labelling burden | Threshold sensitivity; limited fault interpretability | Gong et al. (2021); Lawal et al. (2024) |
| Probabilistic DL | Deep Belief Networks |  | RUL estimation, failure probability | High-dimensional sensor data | Handles uncertainty and noisy inputs | Training instability; declining adoption | Li et al. (2024) |
| Physics-Integrated DL | Physics-Informed Neural Networks |  | PdM under sparse or noisy data | Sensor + physical constraints | Improved generalisation and physical consistency | Model formulation complexity | Fan et al. (2025) |
| Decision-Oriented DL | Deep Reinforcement Learning |  | Maintenance scheduling, policy optimisation | System-level operational data | Enables proactive maintenance decisions | Computationally intensive; slow convergence | Du et al. (2024) |
| General DL Frameworks | Hybrid and comparative DL models |  | Multi-task PdM | IIoT sensor streams | Models nonlinear, high-dimensional systems; automated feature learning | Overfitting risk; black-box behaviour | Alabadi & Habbal (2023); Zhang et al. (2022) |

The quality analysis shows that the deep learning-based predictive maintenance research is the most numerous high-quality contributions in this field. It is worth noting that researchers like Sun et al. (2025) and Lu et al. (2025) scored high-quality ratings because it uses real-world operational data, clearly defined hybrid architecture (i.e. CNN-Transformer and CNN-BiLSTM-AdaBoost), and controlled validation approaches (i.e. cross-validation and comparative benchmarking). Likewise, Wang et al. (2025) were able to provide high methodological strength with ensemble learning using real equipment data. Conversely, the research by Zhang et al. (2022) and Putro et al. (2024) was classified as medium-quality because of constraints in the scale of the dataset, the use of biased validation methods, or low external validity. Altogether, the pre-eminence of high-quality DL studies represents their maturity and usability in the upstream oil and gas predictive maintenance. A Comparison of the model-type is as shown in Table 4.

Table 4. Model-type comparison

|  |  |  |  |
| --- | --- | --- | --- |
| **Model Type** | **Strengths** | **Limitations** | **Typical Use Case** |
| CNN | Feature extraction from sensor signals | Limited temporal modeling | Fault detection |
| LSTM/GRU | Temporal dependency modeling | Data-intensive | RUL prediction |
| Hybrid (CNN-LSTM, CNN-Transformer) | High accuracy, captures spatial + temporal features | Computational cost | Complex PdM systems |
| Model Type | Strengths | Limitations | Typical Use Case |

**RQ2:** Which quantum reinforcement learning approaches have been proposed for maintenance decision-making or cybersecurity, and how do they compare with classical reinforcement learning methods?

Quantum reinforcement learning is a promising and yet to be explored technique in solving complex decision-making problems in maintenance optimisation and cybersecurity (Said et al., 2024). There are three methodological families that are recurrently used throughout the literature: variational quantum circuit-based learning, quantum multi-agent reinforcement learning, and quantum annealing-based optimisation. These are usually developed with high-dimensional and hostile conditions, resembling the required conditions in upstream oil and gas operations, where uncertainty is the new reality and sensor-rich systems and cyber-physical dependency. QRL can frequently be found as an adaptative defence layer in applications with any cybersecurity focus. Hossain et al. (2024) suggests a quantum policy-gradient autonomous incident response, and Olutimehin (2025) presents quantum multi-agent reinforcement learning in financial networks and provides structural analogies to distributed industrial control. The literature on variational quantum deep Q-networks by Chen et al. (2022) on cognitive radio supports the relationship between quantum policy learning and the problem of secure communication in the industrial context.

QRL applications are not as widespread in maintenance decision-making, but they are conceptually similar. Chen et al. (2022) came up with a variational quantum circuit-based deep Q-network (VQ-DQN), in which the value function approximation uses parameterised quantum circuits instead of classical neural networks. Their model had a higher learning efficiency in high-dimensional environments and required less training episodes than classical DQN. Likewise, Neumann et al. (2023) compared the quantum annealing and gate-based QRL methods of solving optimisation problems, demonstrating that quantum-enhanced agents more efficiently searched solutions in stochastic problems. The surveys give several benefits over classical reinforcement learning, such as a smaller number of parameters because of amplitude encoding (Hohenfeld et al., 2022),and in VQ-DQN systems up to a hundred times fewer training episodes (Moll and Kunczik, 2021). Quantum-classical hybrid agents can also achieve quadratic speed-ups through amplitude amplification (Saggio et al., 2021), This is converted into practical operational advantages in applied cases of cybersecurity simulation, where QRL-based systems cut the incident response time by up to 28 percent compared to rule-based baselines (Hossain et al., 2024).

Despite these advantages, there are also major constraints that lessen the expectations of QRL. The majority of reported results are based on classical simulations of quantum systems, which bring high computational costs, and large-scale classical reinforcement learning is more reliable in safety-critical industrial settings. Though deep quantum models have the capability to express multimodal state-action distributions in a more expressive way (Jerbi et al., 2021), a mature quantum hardware is missing, as well as the large quantum circuits are not stable, limiting their real-world application. In the final analysis, the existing evidence makes QRL a complementary and not a substitutive method of maintenance decision-making and cybersecurity, and its future application will require hardware development, hybrid algorithm creation, and strict benchmarking with classical models. Table 5 shows Quantum Reinforcement Learning Solutions to Maintenance Decision-Making and Cybersecurity

Table 5. Quantum reinforcement learning solutions to maintenance decision-making and cybersecurity

|  |  |  |  |  |  |
| --- | --- | --- | --- | --- | --- |
| **QRL Category** | **Representative**  **Methods** | **Application Context** | **Key Advantages over Classical RL** | **Key Limitations** | **Core References** |
| Variational Quantum RL | VQC, VQ-DQN | Cybersecurity response, cognitive radio, secure networking | Higher parameter efficiency; faster convergence in low-episode regimes | Training instability; quantum simulation overhead | Chen et al. (2022); Moll & Kunczik (2021); Hohenfeld et al. (2022) |
| Quantum Policy Optimisation | Quantum policy gradients | Autonomous incident response | Faster adaptation to dynamic threats; reduced response time | Limited scalability; evaluated mainly in simulations | Hossain et al. (2024); Saggio et al. (2021) |
| Quantum Multi-Agent RL | QMARL | Distributed cyber defence | Improved coordination under adversarial dynamics | High communication complexity; limited industrial validation | Olutimehin (2025) |
| Annealing-Based QRL | Quantum annealing, QBM | Maintenance routing, asset traversal | Efficient cost-path optimisation; global search capability | Hardware dependency; task-specific formulations | Neumann et al. (2023) |
| Hybrid Quantum–Classical RL | TN-VQC, hybrid agents | Robotic maintenance, complex control | Improved expressiveness for high-dimensional states | Requires careful partitioning between classical and quantum layers | Chen et al. (2022); Jerbi et al. (2021) |
| Quantum-Enhanced Security RL | Quantum entropy Q-learning | DDoS mitigation, adaptive defence | Faster adaptation; lower false-positive rates | Early-stage validation; limited benchmarks | Said et al. (2024) |
| Trust and Secure Communication RL | Distributed Q-learning + QKD | Device trust, secure IIoT communication | Dynamic trust management; non-static cryptographic defence | Integration complexity; edge deployment challenges | Moudoud & Cherkaoui (2023); Shingne et al. (2025) |

The quality appraisal shows that reinforcement learning (RL) and quantum reinforcement learning (QRL) research is of average quality. Although RL-based methodologies, including Varalakshmi and Kumar (2025), and Mesadieu et al. (2024) show good performance in the tasks of optimisation and intrusion detection, their use of simulated or semi-synthetic data restricts the application to real-world scenarios. The literature on QRL-based models, such as Chen et al. (2022), Neumann et al. (2023), and Skolik et al. (2022), was all evaluated as of moderate quality because of the experimental and simulation-based design. Even though these studies present useful theoretical and computational information, they are not validated with actual upstream oil and gas data. It points out a serious mismatch between the development of algorithms and their use in industry. Moreover, the lack of quality QRL research is an indicator of the modern limitations of technology, specifically scalability of quantum hardware and the integration of datasets, which are hindrances to the practical implementation of a predictive maintenance system. The comparison of deep learning, reinforcement learning, and quantum reinforcement learning is as shown in Table 6.

Table 6. Comparison of DL, RL, and QRL

|  |  |  |  |
| --- | --- | --- | --- |
| **Technique** | **Strengths** | **Limitations** | **Maturity Level** |
| DL | High predictive accuracy | Static decision-making | High |
| RL | Adaptive decision-making | Requires environment modeling | Moderate |
| QRL | High-dimensional optimization potential | Hardware + simulation limitations | Emerging |
| Technique | Strengths | Limitations | Maturity Level |

**RQ3:** What cybersecurity threats affect AI-based PdM systems in upstream oil and gas, and how are these threats addressed in existing DL and QRL approaches?

The existence of AI-based predictive maintenance applications in the upstream oil and gas industry is vulnerable to a broad array of cybersecurity risks that demand simultaneous attacks on both operation technology and information technology levels. Distant offshore platforms, longer communications routes, and integrative systems of control provide the environment that a failure spreads rapidly and attacks do not give a chance to salvage the situation (Progoulakis et al., 2021). As the number of Industrial Internet of Things infrastructures grows, data flows, which are constantly supplied by pressure, temperature, flow, and vibration sensors, are new points of vulnerability to adversaries (Akinniyi, 2021). One of the most frequent threats is the data poisoning, specifically manipulated sensor inputs that can confuse predictive models into unsafe, expensive, or misleading decisions and frequently cover up early signs of degradation (Ayemere et al., 2024; Kochkodan and Petryna, 2025). Advanced persistent threats, ransomware, and other threats that are more aggressive, like Stuxnet and Triton, are capable of bypassing signature-based defences and compromising SCADA and safety instrumented systems, and they are a direct physical threat (Mesadieu et al., 2024; Stergiopoulos et al., 2020). There is even more vulnerability with insider attacks and the vulnerability of centralised AI architectures (Alabadi & Habbal, 2023).

These threats are increasingly detected and acted upon by using deep learning and reinforcement learning. ConvLSTM3D is a spatiotemporal architecture that learns coordinated or changing anomalies on interconnected offshore subsystems, as opposed to individual sensor faults (Ogunmolu, 2025). An additional defence mechanism is physics-informed hybrid models which incorporate flow and pressure constraints into the learning mechanisms, and thus, allow them to signal data trends which may seem statistically normal but inconsistent with the physical laws that are indicative of poisoning (Sangoleye et al., 2024; Odili et al., 2024). Blockchain-based and IPFS-based decentralised AI systems ensure that single points of failure are minimised and model integrity is ensured despite the compromise of individual nodes (Alabadi and Habbal, 2023). Quantum-enhanced reinforcement models, such as Quantum Entropy Q-Learning, are more adjustable in changing environment network traffic and harsh denial-of-service (Said et al., 2024).

Complementary defence is achieved through trust management and secure communication. In distributed Q-learning, IIoT devices are assigned a dynamic reputation score, which limits bandwidth to those with suspicious behaviour (Moudoud and Cherkaoui, 2023). Quantum key distribution with enhanced reinforcement-based protocols to optimize the secure communications can be used to generate the cryptographic parameters continuously by updating them to account for noise and any intrusion attempts that might happen to mitigate the weaknesses of fixed security settings (Shingne et al., 2025). Altogether, the literature portrays cybersecurity threats in AI-based predictive maintenance as being complex and closely related to system architecture. Deep learning provides fine-resolution detection and structural resilience, whereas reinforcement learning especially quantum-enhanced versions provides flexibility and quick autonomous reaction. However, the vast majority of methods consider detection, response, and trust as independent layers indicating that complex frameworks are required to implement security in the entire predictive maintenance lifecycle. A summary of these threats and mitigation mechanisms is summarised in Table 7.

Table 7. Summary of cybersecurity threats in AI-driven PdM and mitigation mechanisms

|  |  |  |  |  |
| --- | --- | --- | --- | --- |
| **Threat Category** | **Manifestation in Upstream PdM** | **DL-Based Mitigation** | **RL / QRL-Based Mitigation** | **Key References** |
| **Data Poisoning & False Data Injection** | Manipulated sensor readings (pressure, temperature, vibration) causing false alarms or hidden degradation | Spatiotemporal models (ConvLSTM3D), physics-informed DL to detect physically inconsistent patterns | RL policies adapt maintenance actions under uncertainty; QRL improves robustness under high traffic variability | Ayemere et al. (2024); Kochkodan & Petryna (2025); Ogunmolu (2025); Said et al. (2024) |
| **Advanced Persistent Threats (APT) & Malware** | Covert intrusion into SCADA/SIS (e.g., Stuxnet, Triton), manipulation of safety logic | DL-based anomaly and intrusion detection across cyber–physical layers | RL agents optimise isolation, rerouting and recovery actions post-detection | Stergiopoulos et al. (2020); Mesadieu et al. (2024); Ahmad et al. (2025) |
| **Denial-of-Service (DoS/DDoS)** | Loss of visibility during critical operational windows, delayed maintenance decisions | Traffic-pattern learning for early DoS detection | Quantum Entropy Q-Learning for faster adaptation and reduced false positives | Said et al. (2024); Ayemere et al. (2024) |
| **Insider Threats** | Authorized misuse or coercion compromising monitoring pipelines | Behavioural DL models to detect deviations from normal operator patterns | Reputation-based distributed Q-learning to dynamically restrict malicious nodes | Progoulakis et al. (2021); Alabadi & Habbal (2023); Moudoud & Cherkaoui (2023) |
| **Single Point of Failure (Centralised AI)** | Compromise of one node cascades across PdM pipeline | Distributed DL architectures for decentralised detection | RL-assisted decentralised coordination; blockchain-backed integrity assurance | Alabadi & Habbal (2023); Stergiopoulos et al. (2020) |
| **Secure Communication & Trust Management** | Key compromise, eavesdropping, unreliable IIoT nodes | DL-assisted traffic and trust anomaly detection | Adaptive RL/QRL for dynamic cryptographic tuning and reputation scoring | Moudoud & Cherkaoui (2023); Shingne et al. (2025) |
|  |  |  |  |  |

Studies with a cybersecurity focus show comparatively high methodological rigour from a quality standpoint, especially those that incorporate reinforcement learning approaches. For instance, by using adversarial testing frameworks, realistic data environments, and reliable assessment criteria, Amato et al. (2024) and Sangoleye et al. (2024) were able to obtain high-quality ratings. Similarly, Khanzadeh et al. (2025) showed excellent methodological quality by thoroughly validating cyber defence methods based on deep learning. However, several studies, such as Dionisopoulos et al. (2023) and Progoulakis et al. (2021), received a moderate rating since they relied on simulated environments or had little empirical support. Significantly, the quality assessment reveals that although cybersecurity research in AI systems is progressing, there are still few studies that cover both domains concurrently, and its integration with predictive maintenance in upstream oil and gas is still fragmented. Table 8 shows the cybersecurity threat mapping.

Table 8. Cybersecurity Threat Mapping

|  |  |  |
| --- | --- | --- |
| **Threat Type** | **Impact on PdM** | **Mitigation Approach** |
| Data Spoofing | Incorrect predictions | Data validation, anomaly detection |
| Adversarial Attacks | Model manipulation | Robust training, adversarial defense |
| Network Intrusion | System compromise | RL-based intrusion detection |

**RQ4:** What gaps remain in the integration of deep learning, quantum reinforcement learning, and cybersecurity for secure PdM?

Even though the combination of DL, QRL and cybersecurity to PdM is becoming an increasingly popular topic, the body of literature is still scarce and technically underdeveloped. The vast majority of works consider the individual elements but do not provide end-to-end systems, and the theoretical basis of hybrid DL-QRL systems does not always coincide with the practical constraints of deployment (Bellante et al., 2025; Cerezo et al., 2022). The major problems of algorithms remain, such as barren plateaus, the disappearance of the gradient with the depth of the circuit and its dimensionality, which contribute to unstable optimisation and diminished application in industries (Andres et al., 2022; Melnikov et al., 2023). To make matters worse, there is no standardised encoding of quantum data and observable design; even high-dimensional PdM sensor streams do not have agreed-upon strategies to map quantum circuits, and invalid output ranges are common with poor encodings, which make it difficult to make control decisions (Andres et al., 2022; Skolik et al., 2022). Contextual awareness is also poorly developed: DL and DRL models are still mostly statistical recognisers that are incapable of differentiating between real degradation and natural operational deviation, and which propagate false alarms and lower the confidence in safety-critical settings (Khanzadeh et al., 2025; Fink et al., 2020; Li et al., 2024; Zhao et al., 2024).

An even greater gap in integrations is brought about by cybersecurity problems. Adversarial manipulation is especially a significant problem in deep learning and reinforcement learning systems, as well as false-data injection and poisoning attacks that manipulate sensor streams as the core components of PdM processes (Nguyen and Reddi, 2023; Ozkan-Okay et al., 2024). These types of attacks are able to subdue early warnings or create unwarranted shutdowns, but strong and industry-prepared defences are rare (Olakunle et al., 2024). Simultaneously, although quantum technologies may have better cryptography guarantees, they have a paradox of their own, they allow quantum capable attackers to compromise classical encryption schemes, including RSA and elliptic-curve cryptography, leaving a security gap between classical and quantum security in the transition phase, where neither technics can guarantee security (Ali et al., 2025; Hossain et al., 2024; Michael et al., 2024).

Privacy ensuring technologies such as federated learning have the advantage of reducing the risk of exposure of data, but at the cost of latency and additional communication overheads not compatible with real-time industrial maintenance (Fink et al., 2020; Ravi, 2025; Shingne et al., 2025). These challenges are compounded by hardware and infrastructural constraints: the current quantum processors are still based on noisy intermediate-scale quantum computers as they do not have the number of qubits and stability to support large-scale PdM, hardware noise also causes additional barren plateaus, and both DL and QRL models can be well beyond the compute, memory, and energy capacity of edge devices necessary to deploy upstream PdM (Melnikov et al., 2023; Nguyen and Reddi, 2023; Ali et al., 2025; Varalakshmi and Kumar, 2025; Bharati and Podder, 2022).

Other than technical constraints, there are practical and organisational ones that slow down progress. The nature of industrial data streams is non-stationary, and there are changing operational regimes and long-lasting equipment decay that cause concept drift that many existing models do not account for; ad hoc adaptation approaches tend to introduce concept drift and catastrophic forgetting with disastrous results (Theissler et al., 2021). Interpretability has been lacking between DL and hybrid quantum systems and hinders the implementation in systems where the maintenance decision may have a huge financial and safety impact (Olakunle et al., 2024; Li et al., 2024). Regulatory preparedness and workforce are even further behind schedule: the industries are not adequately prepared regarding expertise in the realms of AI, quantum computing, and cybersecurity, and regulatory frameworks have not yet kept pace with the ethical and privacy implications of quantum-enhanced systems (Ali et al., 2025; Callahan et al., 2025; Michael et al., 2024; Fink et al., 2020; Hossain et al., 2024). All these gaps show that to get to secure, quantum-enhanced PdM, algorithmic innovation is not enough, but a concerted effort on hardware, system design that believes in cybersecurity, interpretability, regulatory preparation, and workforce capability is all needed. These gaps in the research of deep learning, quantum reinforcement learning, and cybersecurity in upstream oil and gas PdM are summarised in table 9.

Table 9. Highlights of research gaps in the development of deep learning, QRL, and cybersecurity

|  |  |  |  |  |
| --- | --- | --- | --- | --- |
| **Gap Category** | **Key Gap Identified** | **Underlying Cause** | **Implications for Secure PdM** | **Representative Studies** |
| **Algorithmic & Learning Stability** | Barren plateaus and vanishing gradients in quantum–DL models | High-dimensional parameterised quantum circuits and poor trainability | Limits scalability of QRL for complex industrial environments | Andrés et al. (2022); Melnikov et al. (2023)) |
| **State Representation & Encoding** | Lack of standardised quantum data encoding and observable selection | Mismatch between quantum outputs and continuous industrial state spaces | Learning instability and poor policy convergence | Andrés et al. (2022); Skolik et al. (2022) |
| **Context Awareness** | Limited incorporation of domain and operational context | Data-driven models trained without physics or process constraints | Inability to distinguish faults from operational variability | Fink et al. (2020); Li et al. (2024); Zhao et al. (2024) |
| **Adversarial Robustness** | High susceptibility to poisoning and spoofing attacks | Absence of adversarial training and robust validation pipelines | False alarms, missed failures, and unsafe maintenance actions | Amato et al. (2024); Nguyen & Reddi (2023); Ajala et al. (2024) |
| **Cryptographic Resilience** | Dependence on quantum-vulnerable classical encryption | Slow adoption of post-quantum and quantum-secure protocols | Future exposure of IIoT/PdM data to quantum-enabled attacks | Ali et al. (2025); Hossain et al. (2024) |
| **Privacy vs. Latency Trade-off** | High communication and computation overhead of privacy-preserving AI | Federated and secure learning not optimised for real-time PdM | Reduced responsiveness in time-critical fault detection | Fink et al. (2020); Shingne et al. (2025) |
| **Quantum Hardware Limitations** | Noise, limited qubit count, and short coherence times | Reliance on NISQ-era devices | Constrains real-world deployment of QRL-enhanced PdM | Cerezo et al. (2022); Melnikov et al. (2023) |
| **Edge Deployment & Scalability** | DL/QRL models exceed edge resource constraints | High model complexity and energy demands | Bottlenecks real-time, on-site PdM decision-making | Bharati & Podder (2022); Nguyen & Reddi (2023) |
| **Concept Drift Handling** | Poor adaptation to non-stationary industrial data | Static training and catastrophic forgetting | Degradation of PdM accuracy over system lifetime | Theissler et al. (2021) |
| **Interpretability & Trust** | Black-box nature of DL and QRL models | Limited integration of XAI techniques | Low operator trust and resistance to AI-driven maintenance actions | Li et al. (2024); Bellante et al. (2025) |
| **Human & Regulatory Readiness** | Skills shortage and immature governance frameworks | Fragmented expertise across AI, quantum, and cybersecurity | Slowed adoption and unclear accountability | Ali et al. (2025); Callahan et al. (2025); Michael et al. (2024) |

Table 10. Comparison between deep learning, quantum reinforcement learning, and cybersecurity

|  |  |  |  |
| --- | --- | --- | --- |
| **Dimension** | **Deep Learning (DL)** | **Quantum Reinforcement Learning (QRL)** | **Cybersecurity Mechanisms** |
| **Primary Role** | Fault detection and failure prediction | Dynamic optimization of maintenance scheduling | Protection of data, models, and system infrastructure |
| **Core Function** | Learns patterns from historical sensor data | Learns optimal actions through environment interaction | Detects and prevents cyber threats and attacks |
| **Input Data** | Sensor data (vibration, pressure, temperature) | System state (equipment health, operational context) | Network logs, system access data, anomaly signals |
| **Output** | Failure prediction, anomaly detection, RUL estimation | Optimal maintenance decisions (repair, replace, delay) | Alerts, intrusion detection, secure communication |
| **Strengths** | High accuracy in pattern recognition | Efficient decision-making in complex environments | Ensures system integrity and data confidentiality |
| **Limitations** | Requires large labeled datasets | Still emerging; limited real-world deployment | May introduce computational overhead |
| **Computational Complexity** | High (especially deep architectures) | Potentially reduced via quantum parallelism | Moderate to high depending on security layers |
| **Adaptability** | Static after training (unless retrained) | Highly adaptive through continuous learning | Adaptive via real-time threat monitoring |
| **Application in PdM** | Predicts when equipment will fail | Decides when maintenance should occur | Secures PdM system from cyber threats |
| **Integration Role** | Provides predictive insights | Uses predictions to optimize actions | Ensures safe and reliable system operation |
|  |  |  |  |

Table 9 gives a distinct comparison of the functions of Deep Learning, Quantum Reinforcement Learning, and Cybersecurity in predictive maintenance systems, whereas Table 10 provides a comparison between Deep Learning, Quantum Reinforcement Learning and Cybersecurity. Deep learning concentrates on the proper prediction of failures, QRL provides a better approach to decision-making when it comes to maintenance scheduling, and cybersecurity guarantees the integrity and protection of the systems in the event of possible risks. Such a multiple-level integration is fundamental to the construction of powerful and intelligent predictive maintenance systems.

1. DISCUSSION

The sensory form of data is a major concern in PdM models: the time series of subsea well pressure/temperature is, in a study by Aranha et al. (2023) and Lu et al. (2025) use wellhead pressure, pump current, and load. Practically, mechanical wear is detected using pressure and vibration sensors, whereas fluid anomalies are monitored by the use of flowmeters and acoustic monitors. Conversely, the hybrid models combining steady-state (pressure, flow) and dynamic (vibration, acoustic) characteristics are always better than the single-modality ones. Wang et al. (2025) show that time-domain and frequency-domain data obtained on several sensors is more effective in enhancing the fault detection accuracy of CNN. On the other hand, single-mode studies usually have lower detection rates. In this way, simultaneous multimodal measurements (e.g., vibration, pressure and electric measurements) are generally viewed to be critical in trustworthy PdM in upstream oil and gas. Fig. 5 shows a typical upstream well instrumentation; pressure and temperature sensors at the wet end at the top of the Christmas Tree (TPT) and downhole gauge (PDG), flow measurements, and electrical measurements are real time data feeds into PdM models.

![C:\Users\MAIN USER\Pictures\Screenshots\Screenshot 2025-12-29 180810.png](data:image/png;base64...)

Fig.5. A typical upstream well instrumentation

Source: Aranha et al. (2023)

* 1. Model performance and evaluation metrics

The metrics of performance used in PdM studies are varied, making it difficult to compare them to each other directly. Classification tasks usually report accuracy, precision, recall or F1, but regression and forecasting tasks report such metrics as RMSE, MAE or R2. New parameters are typically compared to classical baselines, for instance, (Lu et al., 2025) reported approximately 98% accuracy of a CNN+AdaBoost model (9-25% above baselines), and quantum classifiers reach the same level of accuracy, reaching 99% (Zhang et al., 2022). Ablation experiments support such improvements: Lu et al. (2025) demonstrated that their CNN+AdaBoost model is more than twice as good as a plain CNN or SVM, and Sun et al. (2025) also assert that CNN +Transformer has reduced the prediction error of a CNN-LSTM by more than twofold. This tendency of gradual improvements is stable across research. But since every study works with another set of data and measures, the aggregate results of papers are hard to obtain. However, it is generally agreed that deep models are more effective than shallow models (e.g., SVM, random forest) in cases where enough labelled data is provided.

* 1. Cybersecurity and robustness

The effects of cybersecurity concerns on PdM algorithm design are currently trending. (Amato et al., 2024) apply GAN-based data augmentation to identify malicious sensor injections: their TimeGAN model used the synthetic failure log which allowed an LSTM to reach an accuracy of about 80-90% when distinguishing between attacked sequences. In the same breath, (Dionisopoulos et al., 2023) test standard ML models with focused adversarial attacks (e.g. ZOO, HopSkipJump) and determine that the accuracy of a strong random forest decreases to 83.6% from 96% when being attacked. The implication of these studies is that standard PdM models can be easily perturbed, and adversarial defenses (training or anomaly detection) can also significantly increase robustness.

Nevertheless, in addition to these PdM-specific initiatives, general AI and energy audits underline the need to address such problems as sensor noise, quantum decoherence and cryptographic vulnerability in order to improve the security of the entire system. Thus, despite the field of mainstream PdM research always being concentrated on predictive accuracy, there is a growing body of research that incorporates security measures (e.g. adversarial filtering or input hardening). In practice, it implies that PdM solutions will need to include anomaly detectors and data-validation pipelines in addition to prognostic models, instead of using raw AI predictions.

* 1. Classical vs. Quantum reinforcement learning

The comparisons of the classical deep RL and quantum-enhanced RL remain largely conceptual (Kim, 2024). Classical agents (e.g. DQN, A3C) have been only experimented on proxy tasks in a handful of studies, and are used as benchmarks, but QRL methods include variational quantum circuits (VQCs) in the agent policy or value function (Chen et al., 2020). This is supported by early trials (in other fields) showing that VQC-based agents can be as efficient as, or more efficient than, their classical equivalents with the same number of parameters counts: a VQC-REINFORCE model was able to achieve comparable performance with fewer trainable weights, and a quantum asynchronous Actor-Critic (Q-A3C) trained faster than an equivalent classical A3C. This is said to be due to quantum superposition, which enables exploration of state-action spaces in a more efficient way. Yet, a large-scale application of a quantum processor has not been used in oilfield applications, and the QRL studies are still only proofs-of-concept.

To date, QRL agents provide similar stability and even greater accuracy compared to their classical counterparts of equal size. Less complex classical ensemble techniques (e.g. boosted decision trees or CNN ensembles) are often still used in practice (simply because they are easier to understand and interpret). In short, it can be concluded that QRL has not yet shown any apparent production advantages to PdM, but it suggests a direction toward faster and more effective learning when the practical limitations (hardware noise, data encoding) are eliminated.

* 1. Gaps and Biases in the literature

It has been identified that the literature has several long-standing gaps, such as an acute lack of publicly accessible upstream PdM datasets, inconsistent reporting of sensor configurations and pre-processing procedures, and without comparable performance measures in the literature. The findings are also biased by the publication bias whereby complex models are always published as superior to simpler models, and by the almost lack of cybersecurity assessment, where most PdM studies do not consider adversarial robustness. Besides, quantum reinforcement learning studies are mostly theoretical or simulated with no field implementations at all, and its applicability is yet to be tested.

* 1. Implications for upstream oil and gas predictive maintenance

Summing everything up, there are a number of practical and theoretical implications. In practice, it is important for upstream operators to understand that state of art AI/ML tools can actually improve PdM, but only under a careful implementation. CNNs and LSTMs are deep models with demonstrated capability of identifying subtle fault patterns in the vibration, pressure, and electrical signals. The new Transformer-based hybrids can provide additional gains when data volumes justify their complexity. But such models are also to be combined with powerful data practices: sensor fusion (multimodal inputs) and intense verification against anomalies. Most importantly, cyber-physical security has to be implemented into the design. According to our survey, the GANs or other methods of generating an attack would greatly enhance the possibility of detecting malicious data, as was observed by Amato et al. (2024). That is, PdM systems cannot expect clean data; either implementing anomaly detectors or using adversarial training might cause the AI systems to become significantly more resilient to both random noise and targeted attacks.

Hypothetically, these results suggest that AI research can fill the actual restrictions of the energy industry. As an example, the behavior of pressure and vibration sensors can differ significantly, which implies that the representation learning will have to be customised (e.g. spectrogram as the input in acoustic data, flow rates as the time-dependent embedding). Hybrid models that incorporate domain rules, as indicated by their success, indicate that physics knowledge together with data-driven DL is an opportunity. Based on the QRL viewpoint, theory offers a chance in the future where quantum models may gain acceleration in learning and enhance resilience to uncertainty. The fact that small-parameter QRL policies are able to be equivalent to classical networks should inspire additional research on quantum-safe PdM. But until then, quantum methods still have to conquer hardware noise and data encoding limitations before they can be used upstream.

In all, our review creates a dynamic image: on the one hand, the techniques of mature deep learning (CNNs, LSTMs, autoencoders, even GANs) have continued to enhance the fault prediction in oil and gas machinery. Alternatively, developing quantum-enhanced RL and resistance to cyber threats are the future directions. The direction to follow is obvious: introduce rich sensor data and hybrid DL systems to achieve accuracy and interweave security to make the solutions resistant. It is only at this point that PdM in oil and gas will be informed, as well as credible.

1. CONCLUSION

This article provided a systematic review of artificial intelligence-based predictive maintenance (PdM) in the upstream oil and gas industry, with specific attention to deep learning (DL), reinforcement learning (RL), new quantum reinforcement learning (QRL), and cybersecurity considerations. The review employed a PRISMA-based approach and a hybrid quality appraisal model to synthesise evidence that empirical and sensor-based studies have produced to come up with a critical and well-structured perception of the existing research trends. The results indicate that deep learning methods are the most developed and practically proven methods, particularly when used on real-life sensor data with solid validation plans. The hybrid CNN-LSTM and CNN-Transformer architectures are consistently shown to be more effective in fault detection and remaining useful life (RUL) prediction since they can both resolve spatial and temporal dependencies. Conversely, reinforcement learning methods have promising potential to adaptive maintenance decision-making, although their use is still highly limited to the scope of simulations.

Moreover, quantum reinforcement learning (QRL) has been proven as a new paradigm that has enormous theoretical prospects, especially in high-dimensional optimisation tasks. Its effective application in upstream oil and gas predictive maintenance is, however, still constrained due to hardware limitations and the unavailability of real-world test data. From a cybersecurity perspective, the review notes that there is an increasing awareness of the risks of data spoofing, adversarial attacks, and system intrusion, but it shows that integrated frameworks that are based on PdM, advanced learning models and cybersecurity mechanisms are still rare. These findings are further supported by the quality assessment results, which indicate that high-quality studies are mostly related to deep learning models that have been validated by real operational data, whereas RL and QRL studies are, on average, of moderate quality because of methodological and deployment challenges. This lack of balance highlights an urgent research gap in the creation of comprehensive, reliable and scalable predictive maintenance.

Considering these results, future studies need to address:

1. The combination of RL and QRL models with real sensor-based PdM systems
2. The creation of resilient cybersecurity-sensitive maintenance systems, and
3. The provision of experimental validation outside of the simulation setting.

The solutions on how to handle such challenges will be critical in facilitating the development of reliable, intelligent and secure predictive maintenance solutions in the upstream oil and gas industry.

* 1. Future Research Agenda

According to the results of this review, a number of important directions are outlined to develop AI-inspired predictive maintenance (PdM) in upstream oil and gas systems. To begin with, in future studies, the real-world application of reinforcement learning (RL) and quantum reinforcement learning (QRL) models should be a priority and leave the simulation-based research, as the latter has less practical use. Second, it is necessary to have unified frameworks integrating predictive maintenance, adaptive decision-making, and cybersecurity to consider the interdependencies between system reliability and security.

More so, it is necessary to develop hybrid and multi-model architectures that include deep learning, attention mechanisms, and domain knowledge to enhance robustness and performance. Since exposure of PdM systems to cyber threats is on the rise, maintenance frameworks should incorporate cybersecurity-conscious design solutions. In addition, standardised datasets and evaluation protocols should be established to enhance the reproducibility and to compare studies fairly.

On a final note, the scalable quantum-enhanced solutions and the integration of explainable AI methods should be considered in the future to provide transparency and trust in safety-critical settings. These points will be vital in the development of the secure, intelligent and deployable PdM systems in the upstream oil and gas sector.

1. Acknowledgements/FUNDING

The authors would like to acknowledge the support of the Petroleum Technology Development Fund (PTDF) Nigeria, for providing the financial support for this research.

1. CONFLICT OF INTEREST STATEMENT

The authors declare that there are no conflicts of interest.

1. AUTHORS’ CONTRIBUTIONS

**Philip Adebayo**: Conceptualisation, methodology, formal analysis, investigation and writing-original draft; **Victoria Yemi-Peters**: Conceptualisation, methodology, and formal analysis; **Joshua Agbogun**: Conceptualisation, formal analysis, and validation; **Afolayan Obiniyi**: Conceptualisation, supervision, writing- review and editing, and validation.

1. REFERENCES

Ajagekar, A., and You, F. (2024). Variational quantum circuit based demand response in buildings leveraging a hybrid quantum-classical strategy. *Appl. Energy 364*, 123244. <https://doi:10.1016/j.apenergy.2024.123244>.

Akinniyi J. S. (2021). Cloud-Native AI solutions for PdM in the energy sector: A security perspective. *World Journal of Advanced Research and Reviews*, *9*(3), 409–428. <https://doi.org/10.30574/wjarr.2021.9.3.0052>.

Alabadi, M., & Habbal, A. (2023). Next-generation PdM: Leveraging blockchain and dynamic deep learning in a domain-independent system. *PeerJ Computer Science*, *9*, e1712. <https://doi.org/10.7717/peerj-cs.1712>.

Ali, G., Samuel, A., Kabiito, S. P., Morish, Z., Thomas, A., Robert, W., Denis, A., Sallam, M., Mijwil, M. M., Ayad, J., Salau, A. O., & Dhoska, K. (2025). Integration of Artificial Intelligence, blockchain, and quantum cryptography for securing the Industrial Internet of Things (IIoT): Recent advancements and future trends. applied data science and analysis, 2025, 19–82. <https://doi.org/10.58496/ADSA/2025/004>.

Amato, F., Cirillo, E., Fonisto, M., & Moccardi, A. (2024). Detecting adversarial attacks in IoT-Enabled PdM with time-series data augmentation. *Information*, *15*(11), 740. <https://doi.org/10.3390/info15110740>.

Andrés, E., Cuéllar, M. P., & Navarro, G. (2022). On the use of quantum reinforcement learning in energy-efficiency scenarios. *Energies*, *15*(16), 6034. <https://doi.org/10.3390/en15166034>.

Aranha, P.E., Policarpo, N. A. & Sampaio, M.A. (2024). Unsupervised ML model for predicting anomalies in subsurface safety valves and application in offshore wells during oil production. *J Petrol Explor Prod Technol 14*, 567–581. <https://doi.org/10.1007/s13202-023-01720-4>.

Ayemere U., Oludayo, O. O., Dazok D. J. & Obinna, J. O. (2024). Optimizing maintenance logistics on offshore platforms with AI: Current strategies and future innovations. *World Journal of Advanced Research and Reviews*, *22*(1), 1920–1929. <https://doi.org/10.30574/wjarr.2024.22.1.1315>.

Azmi, P.A.; Yusoff, M.; Mohd Sallehud-din, M.T. (2024). A review of predictive analytics models in the oil and gas industries. *Sensors 2024*, 24, 4013. <https://doi.org/10.3390/s24124013>.

Babayeju O.A., Adefemi A., Ekemezie I.O., Sofoluwe O.O. (2024). Advancements in PdM for aging oil and gas infrastructure. *World Journal of Advanced Research and Reviews*, 22(03), 252–266. <https://doi.org/10.30574/wjarr.2024.22.3.1669>.

Bellante, A., Fioravanti, T., Carminati, M., Zanero, S., & Luongo, A. (2025). Evaluating the potential of quantum ML in Cybersecurity: A Case-Study on PCA-based intrusion detection systems. *Computers & Security*, *154*, 104341. <https://doi.org/10.1016/j.cose.2025.104341>.

Betha, R. (2021). AI in oil and gas: Predicting equipment failures and maximizing uptime. *IJSAT-International Journal on Science and Technology*, 12(1).

Bharati, S., & Podder, P. (2022). Machine and deep learning for IoT security and privacy: Applications, challenges, and future directions. *Security and Communication Networks*, 2022, 1–41. <https://doi.org/10.1155/2022/8951961>.

Callahan, B., Schilp, K., Colognato, Q., Goldman, E., Sugerman, S., Mehta, A., Imanuel, A., Kaii, K., & Rose, H. (2025). Multidisciplinary quantum cybersecurity research for the undergraduate laboratory. *Journal of The Colloquium for Information Systems Security Education*, *12*(1). <https://doi.org/10.53735/cisse.v12i1.206>.

Cerezo, M., Verdon, G., Huang, H.-Y., Cincio, L., & Coles, P. J. (2022). Challenges and opportunities in quantum machine learning. *Nature Computational Science*, *2*(9), 567–576. <https://doi.org/10.1038/s43588-022-00311-3>.

Chen, S. Y. C., Huang, C. M., Hsing, C. W., Goan, H. S., & Kao, Y. J. (2022). Variational quantum reinforcement learning via evolutionary optimization. *Machine Learning: Science and Technology*, *3*(1), 015025. <https://doi.org/10.1088/2632-2153/ac41b6>.

Chen, S. Y. C., Yang, C. H. H., Qi, J., Chen, P. Y., Ma, X., & Goan, H. S. (2020). Variational quantum circuits for deep reinforcement learning. *IEEE Access*, *8*, 141007-141024. <https://doi.org/10.1109/ACCESS.2020.3010470>.

Dionisopoulos, N., Vrochidou, E., & Papakostas, G. A. (2023, March). Machine learning robustness in predictive maintenance under adversarial attacks. In *Congress on Control, Robotics, and Mechatronics* (pp. 245-254). Singapore: Springer Nature Singapore.

Du, J., Zheng, J., Liang, Y., Ma, Y., Wang, B., Liao, Q., Xu, N., Ali, A. M., Rashid, M. I., & Shahzad, K. (2024). A deep learning-based approach for predicting oil production: A case study in the United States. *Energy*, 288, 129688. <https://doi.org/10.1016/J.ENERGY.2023.129688>.

Elijah, O., Ling, P. A., Rahim, S. K. A., Geok, T. K., Arsad, A., Kadir, E. A., ... & Abdulfatah, M. Y. (2021). A survey on Industry 4.0 for the oil and gas industry: Upstream sector. *IEEE Access*, *9*, 144438-144468. <https://doi.org/10.1109/ACCESS.2021.3123507>.

Fan, D.; Lai, S.; Sun, H.; Yang, Y.; Yang, C.; Fan, N.;Wang, M. (2025). Review of ML methods for steady state capacity and transient production forecasting in oil and gas reservoir. *Energies 2025*, 18, 842. <https://doi.org/10.3390/en18040842>.

Fink, O., Wang, Q., Svensén, M., Dersin, P., Lee, W., & Ducoffe, M. (2020). Potential, challenges and future directions for deep learning in prognostics and health management applications. *Engineering Applications of Artificial Intelligence*, *92*, 103678. <https://doi.org/10.1016/j.engappai.2020.103678>.

Gong, F., Ji, X., Gong, W., Yuan, X., & Gong, C. (2021). Deep learning based protective equipment detection on offshore drilling platform. *Symmetry*, *13*(6), 954. <https://doi.org/10.3390/sym13060954>.

Gowekar, G. S. (2024). Artificial intelligence for PdM in oil and gas operations. *World Journal of Advanced Research and Reviews*, *23*(3), 1228–1233. <https://10.30574/wjarr.2024.23.3.2721>.

Haddaway, N. R., Page, M. J., Pritchard, C. C., & McGuinness, L. A. (2022). PRISMA2020: An R package and Shiny app for producing PRISMA 2020-compliant flow diagrams, with interactivity for optimised digital transparency and Open Synthesis. *Campbell Systematic Reviews*, *18*, e1230. <https://doi.org/10.1002/cl2.1230>.

Hafsi, M., Agoun, J., Hamour, N., Ouchani, S., & Abuarqoub, A. (2024). Strategic PdM for internet system security and risk management: A roadmap. In *Proceedings of the 8th International Conference on Future Networks and Distributed Systems* (pp. 472 - 480). <https://doi.org/10.1145/3726122.3726190>.

Hohenfeld, H., Heimann, D., Wiebe, F., & Kirchner, F. (2024). Quantum deep reinforcement learning for robot navigation tasks. *IEEE Access*, *12*, 87217-87236. <https://doi.org/10.1109/ACCESS.2024.3417795>.

Hossain, F., Hasan, K., Amin, A., & Mahmud, S. (2024). Quantum ML for enhanced cybersecurity: Proposing a hypothetical framework for next-generation security solutions. *Journal of Technologies Information and Communication*, *4*(1), 32222. <https://doi.org/10.55267/rtic/15824>.

Jerbi, S., Trenkwalder, L. M., Poulsen Nautrup, H., Briegel, H. J., & Dunjko, V. (2021). Quantum enhancements for deep reinforcement learning in large spaces. *PRX Quantum*, *2*(1), 010328. <https://doi.org/10.1103/PRXQuantum.2.010328>.

Khan, U., Cheng, D., Setti, F., Fummi, F., Cristani, M., & Capogrosso, L. (2025). A comprehensive survey on deep learning-based PdM. *ACM Transactions on Embedded Computing Systems, 25*(2), 1-43. <https://doi.org/10.1145/3732287>.

Khanzadeh, S., Neto, E. C. P., Iqbal, S., Alalfi, M., & Buffett, S. (2025). An exploratory study on domain knowledge infusion in deep learning for automated threat defense. *International Journal of Information Security*, *24*(1), 71. <https://doi.org/10.1007/s10207-025-00987-4>.

Kim, J. (2024). Quantum reinforcement learning: Concepts, models, and applications. In N.-N. Dao, Q.-D. Pham, S. Cho, & N. T. Nguyen (Eds.), *Intelligence of Things: Technologies and Applications* (Vol. 229, pp. 3–11). Springer Nature Switzerland. <https://doi.org/10.1007/978-3-031-75593-4_1>.

Kochkodan, V. B., & Petryna, M. Yu. (2025). Risks of implementing and using artificial intelligence by oil and gas sector enterprises. *The Actual Problems of Regional Economy Development*, *2*(21), 333–343. <https://doi.org/10.15330/apred.2.21.333-343>.

Lawal, A., Yang, Y., He, H., & Baisa, N. L. (2024). MLin oil and gas exploration: A review. *IEEE Access*, *12*, 19035-19058. <https://doi.org/10.1109/ACCESS.2023.3349216>.

Li, Z., He, Q., & Li, J. (2024). A survey of deep learning-driven architecture for predictive maintenance. *Engineering Applications of Artificial Intelligence*, *133*, 108285. <https://doi.org/10.1016/j.engappai.2024.108285>.

Lu, Q., Wang, Y., Gu, C., Guo, Y., Yang, J., Xiao, H., & Yang, Z. (2025). An Integrated CNN-BiLSTM-Adaboost Framework for Accurate Pipeline Residual Strength Prediction. *Applied Sciences*, *15*(16), 9059. <https://doi.org/10.3390/app15169059>.

Magdin, M. (2025). Current trends and advances in artificial intelligence for ensuring the safety and efficiency of gas pipeline operations: A literature review. *Results in Engineering*, *28*, 107309. <https://doi.org/10.1016/j.rineng.2025.107309>.

Melendez-Torres, G. J., Thomas, J., Richardson, M., Felix, L., Lorenc, T., Thomas, S., & Petticrew, M. (2015). Lessons from comparing narrative synthesis and meta-analysis in a systematic review. *The Lancet*, *386*, S9. [https://doi.org/10.1016/S0140-6736(15)00847-8](https://doi.org/10.1016/S0140-6736%2815%2900847-8).

Melnikov, A., Kordzanganeh, M., Alodjants, A., & Lee, R.-K. (2023). Quantum machine learning: From physics to software engineering. *Advances in Physics: X*, *8*(1), 2165452. <https://doi.org/10.1080/23746149.2023.2165452>.

Mesadieu, F., Torre, D., & Chennamaneni, A. (2024). Leveraging deep reinforcement learning technique for intrusion detection in SCADA infrastructure. *IEEE Access*, *12*, 63381–63399. <https://doi.org/10.1109/ACCESS.2024.3390722>.

Michael, C. I., Campbell, T.-A., Idoko, I. P., Bemologi, O. U., Anyebe, A. P., & Odeh, I. I. (2024). Enhancing cybersecurity protocols in financial networks through reinforcement learning. International *Journal of Scientific Research and Modern Technology (IJSRMT)*, *3*(9), 44–59. <https://doi.org/10.38124/ijsrmt.v3i9.58>.

Moll, M., & Kunczik, L. (2021). Comparing quantum hybrid reinforcement learning to classical methods. Human-Intelligent Systems Integration, 3(1), 15-23. <https://doi.org/10.1007/s42454-021-00025-3>.

Moudoud, H., & Cherkaoui, S. (2023). Empowering security and trust in 5G and beyond: A deep reinforcement learning approach. *IEEE Open Journal of the Communications Society*, *4*, 2410–2420. <https://doi.org/10.1109/OJCOMS.2023.3313352>.

Neumann, N. M., de Heer, P. B., & Phillipson, F. (2023). Quantum reinforcement learning: Comparing quantum annealing and gate-based quantum computing with classical deep reinforcement learning. Quantum Information Processing, 22(2), 125. <https://doi.org/10.1007/s11128-023-03867-9>.

Nguyen, T. T., & Reddi, V. J. (2023). Deep reinforcement learning for cyber security. *IEEE Transactions on Neural Networks and Learning Systems*, *34*(8), 3779–3795. <https://doi.org/10.1109/TNNLS.2021.3121870>.

Odili, P. O. Daudu, C.D. Adefemi, A., Ekemezie, I. O. & Usiagu, G. S. (2024). Integrating advanced technologies in corrosion and inspection management for oil and gas operations. *Engineering Science & Technology Journal*, *5*(2), 597–611. <https://doi.org/10.51594/estj.v5i2.835>.

Ogunmolu, A. M. (2025). A multiscale approach to cyber-mechanical threat modeling for predicting and preventing failures in critical energy infrastructure. *Journal of Engineering Research and Reports*, *27*(6), 1–20. <https://doi.org/10.9734/jerr/2025/v27i61523>.

Olakunle A. A., Chinwe C. O., Onyeka C. O., Chuka A. A., & Obinna D. D. (2024). Review of AI and ML applications to predict and Thwart cyber-attacks in real-time. *Magna Scientia Advanced Research and Reviews*, *10*(1), 312–320. <https://doi.org/10.30574/msarr.2024.10.1.0037>.

Olutimehin, Abayomi Titilola. 2025. The synergistic role of machine learning, deep learning, and reinforcement learning in strengthening cyber security measures for crypto currency platforms. *Asian Journal of Research in Computer Science*, *18*(3), 190-212. <https://doi.org/10.9734/ajrcos/2025/v18i3586>.

OpenAI. (2026). ChatGPT (GPT-5.3) [Large language model]. https://chat.openai.com/

Ozkan-Okay, M., Akin, E., Aslan, Ö., Kosunalp, S., Iliev, T., Stoyanov, I., & Beloev, I. (2024). A comprehensive survey: Evaluating the efficiency of Artificial Intelligence and ML techniques on cyber security solutions. *IEEE Access*, *12*, 12229–12256. <https://doi.org/10.1109/ACCESS.2024.3355547>.

Progoulakis, I., Nikitakos, N., Rohmeyer, P., Bunin, B., Dalaklis, D., & Karamperidis, S. (2021). Perspectives on cyber security for offshore oil and gas assets. *Journal of Marine Science and Engineering*, *9*(2), 112. <https://doi.org/10.3390/jmse9020112>.

Purssell, E. (2020). Can the critical appraisal skills programme check‐lists be used alongside grading of recommendations assessment, development and evaluation to improve transparency and decision‐making?. *Journal of Advanced Nursing*, *76*(4), 1082-1089. <https://doi.org/10.1111/jan.14303>.

Putro, E. C., Zulfikri, H., & Sommeng, A. N. (2024). Deep learning application as a method for classification equipment failure on natural gas pipeline. *Cognizance Journal of Multidisciplinary Studies*, *4*(6), 87–97. <https://doi.org/10.47760/cognizance.2024.v04i06.009>.

Saggio, V., Asenbeck, B. E., Hamann, A., Strömberg, T., Schiansky, P., Dunjko, V., ... & Walther, P. (2021). Experimental quantum speed-up in reinforcement learning agents. Nature, 591(7849), 229-233. <https://doi.org/10.1038/s41586-021-03242-7>.

Said, D., Bagaa, M., Oukaira, A., & Lakhssassi, A. (2024). Quantum entropy and reinforcement learning for distributed denial of service attack detection in smart grid. *IEEE Access*, *12*, 129858–129869. <https://doi.org/10.1109/ACCESS.2024.3441931>.

Sangoleye, F., Johnson, J., & Eleni Tsiropoulou, E. (2024). Intrusion detection in industrial control systems based on deep reinforcement learning. *IEEE Access*, *12*, 151444–151459. <https://doi.org/10.1109/ACCESS.2024.3477415>.

Saputelli, L., Palacios, C., & Bravo, C. (2022). Case studies involving ML for PdM in oil and gas production operations. In *ML Applications in Subsurface Energy Resource Management* (pp. 313-336). CRC Press.

Serradilla, O., Zugasti, E., Ramirez de Okariz, J., Rodriguez, J., & Zurutuza, U. (2021). Adaptable and explainable predictive maintenance: Semi-supervised deep learning for anomaly detection and diagnosis in press machine data. *Applied Sciences*, *11*(16), 7376. <https://doi.org/10.3390/app11167376>.

Shil, S. K. (2025). AI driven PdM in petroleum and power systems using random forest regression model for reliability engineering framework. *American Journal of Scholarly Research and Innovation*, *4*(1), 363-391. <https://doi.org/10.63125/477x5t65>.

Shingne, H., Chikmurge, D., Parkhi, P., & Agrawal, P. (2025). Design of an integrated model using deep reinforcement learning and Variational Autoencoders for enhanced quantum security. *MethodsX*, *15*, 103445. <https://doi.org/10.1016/j.mex.2025.103445>.

Skolik, A., Jerbi, S., & Dunjko, V. (2022). Quantum agents in the Gym: A variational quantum algorithm for deep Q-learning. *Quantum*, *6*, 720. <https://doi.org/10.22331/q-2022-05-24-720>.

Stergiopoulos, G., Gritzalis, D. A., & Limnaios, E. (2020). Cyber-Attacks on the Oil & Gas Sector: A survey on incident assessment and attack patterns. *IEEE Access*, *8*, 128440–128475. <https://doi.org/10.1109/ACCESS.2020.3007960>.

Strata, F. (2025). Quantum machine learning: early opportunities for the energy industry – A scoping review. *Frontiers in Quantum Science and Technology*, *4*, Article 1653104. <https://doi.org/10.3389/frqst.2025.1653104>.

Sun, S., Bai, W., Shao, Z. et al. Exploring the predictive performance of deep learning for fracturing fluid flowback and shale gas production. *Sci Rep 15*, 42748 (2025). <https://doi.org/10.1038/s41598-025-26761-z>.

Tariq, Z., Aljawad, M. S., Hasan, A., Murtaza, M., Mohammed, E., El-Husseiny, A., ... & Abdulraheem, A. (2021). A systematic review of data science and ML applications to the oil and gas industry. *Journal of Petroleum Exploration and Production Technology*, *11*(12), 4339-4374. https://doi.org/10.1007/s13202-021-01302-2.

Theissler, A., Pérez-Velázquez, J., Kettelgerdes, M., & Elger, G. (2021). PdM enabled by machine learning: Use cases and challenges in the automotive industry. *Reliability Engineering & System Safety*, *215*, 107864. <https://doi.org/10.1016/j.ress.2021.107864>.

Varalakshmi, K., & Kumar, J. (2025). Optimized PdM for streaming data in industrial IoT networks using deep reinforcement learning and ensemble techniques. *Scientific Reports*, *15*(1), 27201. <https://doi.org/10.1038/s41598-025-10268-8>.

Vizváry, P., & Grigas, V. (2025). Unravelling citation rules: A comparative analysis of referencing instruction patterns in Scopus‐indexed journals. *Learned Publishing*, *38*(2), e1661. <https://doi.org/10.1002/leap.1661>.

Wang, M., Su, X., Song, H., Wang, Y., & Yang, X. (2025). Enhancing predictive maintenance strategies for oil and gas equipment through ensemble learning modeling. *Journal of Petroleum Exploration and Production Technology*, *15*(3), Article 46. <https://doi.org/10.1007/s13202-025-01931-x>.

Zhang, J., Zhang, F., Guo, X., Yang, X., & Mu, P. (2022). A pdm model of equipment system based on GRU model. In *2022 International Conference on Virtual Reality, Human-Computer Interaction and Artificial Intelligence (VRHCIAI)* (pp. 100-105). IEEE. <https://doi.org/10.1109/VRHCIAI57205.2022.00024>.

Zhao, Y., Yang, J., Wang, W., Yang, H., & Niyato, D. (2024). TranDRL: A transformer-driven deep reinforcement learning enabled prescriptive maintenance framework. *IEEE Internet of Things Journal*, *11*(21), 35432–35444. [https://doi.org/10.1109/JIOT.2024.3436110](https://www.google.com/search?q=https://doi.org/10.1109/JIOT.2024.3436110)

|  |  |
| --- | --- |
| ![A picture containing text, clipart  Description automatically generated](data:image/png;base64...) | © 2026 by the authors. Submitted for possible open access publication under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/). |

1. \* Corresponding author. *E-mail address*: philipadebayo41@gmail.com [↑](#footnote-ref-1)