Artificial Intelligence Techniques for Industrial Predictive Maintenance: A Systematic Review of Recent Advances

Industrial assets, ranging from manufacturing machinery and power generation equipment to transportation systems like aircraft and vehicles, represent significant capital investments and are critical to operational continuity and economic productivity [1-3]. Maintaining these assets effectively is paramount, yet traditional maintenance strategies often fall short. Reactive maintenance, performed only after a failure occurs, leads to unplanned downtime, potential secondary damages, safety hazards, and significant economic losses [3-6]. Preventive maintenance (PvM), on the other hand, operates on predetermined schedules, such as fixed time intervals or usage counts, regardless of the actual condition of the equipment. While it aims to mitigate some failures by intervening before they are expected to occur based on these schedules, PvM often results in unnecessary interventions on healthy equipment or, conversely, fails to prevent unexpected breakdowns occurring between scheduled services [2,  3,  5,  7,  8]. The core limitation of PvM is its reliance on general statistical lifespans rather than real-time operational health.

These shortcomings of both reactive and traditional preventive strategies have spurred the transition to Predictive Maintenance (PdM). Distinct from PvM's schedule-based approach, PdM is a proactive strategy focused on anticipating potential failures by continuously monitoring the actual equipment condition and scrutinizing operational data [2, 4-6, 9, 10]. The emergence of Industry 4.0, characterized by the widespread adoption of sensors (Internet of Things - IoT), sophisticated data analysis techniques, and Artificial Intelligence (AI), has greatly enhanced the capabilities of PdM [4, 7-9, 11-15]. AI, notably Machine Learning (ML) and Deep Learning (DL), provides robust tools to handle immense volumes of intricate, high-dimensional sensor data, detect subtle indicators signaling degradation, determine fault categories, forecast the Remaining Useful Life (RUL) of parts, and refine maintenance timelines [6, 10, 13, 16]. By harnessing AI, organizations can transition to smarter, data-informed maintenance strategies, potentially realizing significant gains in asset uptime, operational effectiveness, safety, and cost savings [1-3, 11, 15, 16].

Despite the promise, the practical implementation of AI-driven PdM faces numerous challenges related to data acquisition and quality, algorithm selection and validation, model interpretability, system integration, cybersecurity, and organizational adoption [8, 9, 15, 17]. Understanding the current landscape of AI applications in PdM, identifying successful approaches, recognizing persistent challenges, and discerning future trends is crucial for both researchers developing new methods and practitioners seeking to implement these technologies effectively.

This systematic review endeavors to synthesize the peer-reviewed literature published within the past 10 years concerning the deployment of AI techniques for industrial predictive maintenance. Adhering to a defined search and screening process, 29 relevant publications were selected to explore the following research inquiries:

• Which principal AI methods (ML/DL algorithms and strategies) are presently utilized and under development for industrial PdM?
• What are the key industrial application areas and specific use scenarios (e.g., fault diagnosis, RUL prognosis, scheduling optimization) where AI-driven PdM is being implemented?
• What principal difficulties, constraints, and obstacles are faced during the creation and deployment of AI-powered PdM solutions?
• What developing patterns, models, and potential avenues for future study are highlighted in the existing research?

Through answering these inquiries, this review aims to deliver a thorough, empirically grounded summary of the current status of AI applications for enhancing PdM, yielding useful perspectives for both academic researchers and industry practitioners. The rest of this paper is organized as follows: Section 2 gives background information on PdM and relevant AI concepts. Section 3 describes the conceptual underpinnings for AI-driven PdM. Section 4 outlines the systematic review methodology applied. Section 5 presents the synthesized results from the literature analysis. Section 6 examines the implications, challenges, and future paths. Lastly, Section 7 brings the review to a close.

The integration of AI fundamentally reshapes the conceptual basis of PdM, shifting from reliance on predefined schedules or simple threshold rules towards a continuous, data-driven learning and adaptation cycle. The underlying principle is that historical and real-time operational data contain implicit information about the health state and degradation processes of industrial assets [5, 6, 8]. AI algorithms provide the means to extract this information, model complex system dynamics, and generate actionable insights for maintenance optimization [13, 16, 19, 26]. This AI-driven approach can be conceptualized as an iterative loop (See Figure 1), distinct from purely human-centric observation or basic condition monitoring:

• Data Acquisition & Processing: Continuous streams of data from diverse sources are collected, cleaned, integrated [17, 29], and prepared for analysis. Contextual information is crucial [17].
• AI Model Training & Validation: ML/DL models are trained on historical data to learn normal behavior, failure patterns, and degradation trajectories. Validation ensures model accuracy and generalization [5, 11, 12].
• AI Analysis: Trained models are deployed to monitor live data, detect anomalies [25-27], diagnose faults [4, 12, 29], and predict RUL [5, 16, 19, 29]. Handling uncertainty [5, 19] and providing explanations [8, 26] are increasingly important aspects.

1.png

Figure 1. Conceptual loop of AI-driven Predictive Maintenance

• Decision Support & Optimization: Insights from AI analysis are translated into optimal maintenance recommendations, considering economic factors [1, 16], operational constraints [16, 29], and risk assessments. DRL can automate policy generation [19, 3].
• Action & Feedback: Recommended maintenance tasks are scheduled and executed. Crucially, the outcomes of these actions provide feedback to refine data understanding and retrain AI models, enabling continuous learning and adaptation [21].

This conceptual model underscores the pivotal function of AI algorithms in converting unprocessed data into predictive understanding and optimized decisions, surpassing the constraints of conventional methods. The prevalent emphasis within the reviewed literatures [1-29] is largely placed upon enhancing and refining the elements inside this technological cycle, especially the phases involving AI modeling and analysis.

This part consolidates the principal discoveries derived from reviewing the 29 selected articles, tackling the research inquiries concerning AI techniques, applications, challenges, and frameworks within industrial PdM. The key results are detailed subsequently, accompanied by a certainty evaluation for each finding using the CERQual method.

5.1 AI techniques applied in predictive maintenance

It is important to note that a direct, universal comparison of performance (e.g., accuracy, efficiency) across all listed AI techniques is challenging to distill from the reviewed literature. This is due to the heterogeneity in datasets, specific PdM tasks (diagnosis, RUL estimation, scheduling), evaluation metrics, and implementation details across the primary studies. However, this section synthesizes the commonly employed AI techniques, highlighting their typical applications, and where possible, general characteristics related to their performance and computational considerations in the context of PdM as reported in the reviewed corpus [1-29].

The analyzed studies demonstrate a broad spectrum of AI techniques utilized for PdM, indicating a definite shift towards more advanced ML and particularly DL models. (CERQual Confidence: High) (Confidence is assessed as high due to the substantial volume, consistency, and depth of evidence across the body of work detailing specific AI methods employed [See Appendix A, Finding 1]).

5.1.1 Machine Learning algorithms

Traditional ML algorithms remain relevant, particularly for classification and baseline comparisons. Commonly cited techniques include:

• Support Vector Machines (SVM): Used for fault classification [11, 12] and anomaly detection [8]. SVMs are often chosen for their effectiveness in high-dimensional spaces and their ability to model non-linear decision boundaries with appropriate kernels. However, their training time can scale significantly with the size of the dataset, and performance is sensitive to kernel selection and parameter tuning.
• Random Forests (RF) & Decision Trees (DT): Applied for classification [4, 11, 12], feature importance analysis [4], and sometimes regression tasks [13]. RF is often favored for its robustness [21]. Decision Trees offer high interpretability, while Random Forests, by ensembling multiple trees, generally achieve higher accuracy and robustness against overfitting, often at the cost of some interpretability and increased computational load for training large ensembles. RFs are also valuable for their inherent ability to rank feature importance.
• k-Nearest Neighbors (k-NN): Employed for categorization [11] and outlier detection [27]. k-NN is a simple, non-parametric method that can be effective when local data structure is important. Its computational cost during inference can be high for large datasets as it requires computing distances to all training points, and its performance is sensitive to the choice of 'k' and the distance metric.
• Naive Bayes (NB): Employed for classification, though sometimes showing lower performance compared to others [12].
• Ensemble Methods: Techniques like AdaBoost [4], Gradient Boosting [11], and stacking ensembles [9] are used to combine multiple models, often improving accuracy and robustness, including resilience against adversarial attacks [9]. While often leading to superior predictive performance and robustness, ensemble methods inherently increase computational complexity both in training and inference compared to single models.
• Clustering Algorithms: Unsupervised methods like k-Means [21], DBSCAN [21], and Self-Organizing Maps (SOM) [27] are used for identifying operational states, detecting anomalies, or localizing faults [24, 27]. These unsupervised methods are crucial for discovering inherent structures or anomalies in unlabeled data. Their computational efficiency varies greatly depending on the algorithm (e.g., k-Means is relatively efficient, while density-based methods can be more demanding).
• Time Series Models: Traditional models like ARIMA are sometimes used as baselines or components in hybrid approaches [13].

To further summarize the characteristics of these traditional Machine Learning algorithms as applied within the reviewed PdM literature, Table 3 provides a comparative overview. This table highlights their typical tasks, common strengths and weaknesses, and general computational efficiency considerations.

Table 3. Summary of traditional Machine Learning techniques in PdM from reviewed literature

Note: This table provides a general synthesis. The actual performance and efficiency of these algorithms are highly dependent on the specific dataset, implementation details, and the nature of the PdM task. Computational efficiency refers to general trends; for instance, SVM training can be demanding for very large datasets, while k-NN inference can be slow without appropriate indexing structures.

5.1.2 Deep Learning architectures

DL models have become increasingly popular owing to their capacity for managing intricate, high-dimensional sensor data and automatically discerning relevant features [5, 6, 26]. Key architectures include:

• Convolutional Neural Networks (CNNs): Proficient at extracting spatial patterns from sensor data (like vibration signals viewed as images) or time-series sections; frequently employed for fault categorization [4, 6, 19] and identifying key features within more extensive models [16]. CNNs have shown strong performance in tasks involving spatial hierarchies in data, like processing vibration signals as images or spectrograms. They can be computationally intensive to train, especially with deep architectures, but can be very efficient at inference with optimized implementations.
• Recurrent Neural Networks (RNNs) - LSTM & GRU: Are particularly adept at modeling sequential information and temporal dependencies, rendering them highly suitable for RUL estimation [5, 6, 13, 16, 19] and health index forecasting [29]. LSTMs and GRUs are powerful for modeling long-term dependencies in sequential data, crucial for RUL estimation. Their training can be computationally demanding and slower than feedforward networks due to their recurrent nature, but they excel where temporal context is key.
• Autoencoders (AE): Primarily used in unsupervised anomaly detection by learning a compressed representation of normal data and identifying deviations [6, 7, 26]. Convolutional AEs are common [26]. Explainable variants are emerging [26]. AEs are effective for unsupervised anomaly detection and feature learning by learning compressed data representations. Their complexity and training time depend on the depth and width of the encoder/decoder networks.
• Hybrid Models: Combinations like CNN-LSTM [16], CNN-LSTM-Attention [16], LSTMs integrated with physical models [18], or generative models combined with health indicators [29] aim to leverage the strengths of different architectures for improved performance, particularly in complex RUL prediction or integrated planning tasks [16, 20, 29]. Hybrid models aim to combine the strengths of different architectures, potentially leading to better performance but also increased model complexity and training requirements.
• Generative Models: Used for data augmentation or directly for health prognostics [29].

Deep Reinforcement Learning (DRL): Applied to optimize maintenance scheduling and decision-making policies in dynamic and uncertain environments, learning through interaction [3, 19]. Multi-agent DRL addresses coordination in complex systems [3]. DRL shows promise for optimizing complex, sequential decision-making tasks like maintenance scheduling. However, DRL typically requires significant computational resources for training, often involving numerous simulation iterations, and careful environment design.

A summary of these Deep Learning architectures, detailing their application in PdM tasks, along with their common strengths, weaknesses, and computational aspects as observed in the reviewed corpus, is presented in Table 4.

Table 4. Summary of Deep Learning architectures in PdM from reviewed literature

Note: This table offers a general overview. Deep Learning models are generally data-hungry and computationally intensive to train, often requiring specialized hardware (e.g., GPUs). Their inference efficiency can vary. Performance is contingent upon architecture design, hyperparameter tuning, dataset size and quality.

5.2 Key application domains and use cases

AI-driven PdM is being applied across a diverse range of industrial sectors and equipment types, demonstrating its versatility. (CERQual Confidence: High) (Evidence for this finding is abundant across the corpus, covering numerous industries and equipment types consistently, indicating high coherence, adequacy, and relevance [See Appendix A, Finding 1]).

• Manufacturing: This domain is a prominent area for AI-driven PdM, with numerous studies leveraging empirical data from real enterprises. One such study presents a data-driven drift detection and diagnosis framework applied to a multiple tapping process in a real industrial setting [21]. It utilized sensor data (current, voltage, process parameters) and employed clustering algorithms (k-Means, DBSCAN) alongside Random Forests to identify and classify operational drifts, demonstrating the practical application of AI for improving the stability of heterogeneous production processes. Another study, conducted in the context of an automotive plant, linked product quality control from a machine tool to its predictive maintenance requirements [23]. It illustrated how machine degradation, monitored through sensor data and process parameters, could be correlated with part quality deviations, offering a framework for joint product–process–machine optimization using real-world manufacturing data. Further illustrating the application of AI in critical industrial infrastructure, a recent study developed an IoT and Machine Learning-based system for anomaly detection in electrical panels aimed at predictive maintenance and fire prevention [12]. The work involved an experimental setup using Arduino and Raspberry Pi platforms to collect real-time data (3,478 data points) including gas, temperature, humidity, current, and voltage from electrical panels under various simulated fault conditions leading to potential fire hazards. Several ML classifiers were evaluated (Decision Tree, Gaussian Naive Bayes, Support Vector Machine, and Gaussian Process Classifier). The study demonstrated that the Gaussian Process Classifier (GPC) achieved the highest performance, with an accuracy of 99.56% and an AUC of 0.99, in effectively distinguishing between normal and anomalous (pre-fire) conditions. This case study underscores the practical value of integrating IoT sensor data with robust ML algorithms for predictive maintenance of critical electrical infrastructure, enabling early anomaly detection and proactive safety interventions based on empirical evidence. Other applications within manufacturing include general systems [1, 14, 16, 17, 22, 23], specific equipment such as injection molding machines [1], industrial robots [18], power press motors [13], and textile knitting machines [4].
• Aerospace: Aircraft turbofan engines (widely used benchmark datasets) [5, 16, 19].
• Automotive: Vehicle components and systems [8].
• Energy & Utilities: Electrical distribution panels [12], electric motors [13, 20].
• Transportation: Railcar bearings [10].
• The primary use cases reported align with the core PdM tasks:

Fault Detection and Diagnosis: Identifying anomalous behavior and classifying specific fault types [4, 7, 9, 12, 13, 21, 24, 25, 27, 28].

• Prognosis and RUL Prediction: Estimating the remaining time until failure [4-6, 13, 16, 18, 19, 29]. This is often considered the most challenging but valuable task.
• Maintenance Scheduling and Optimization: Determining the optimal time and type of maintenance intervention [1, 14, 17, 19, 23, 29]. This increasingly involves integration with production schedules [16, 29].
• Condition Monitoring and Health Assessment: Tracking the overall health status using calculated indices or learned representations [18, 22, 26, 27].
• Linking Machine Health to Process Outcomes: Connecting machine degradation signals to impacts on product quality [23] or process efficiency [21].

5.3 Data aspects: Sources, processing, and challenges

Data is the cornerstone of AI-driven PdM. The reviewed articles highlight the use of diverse data sources and the critical importance of data processing, while consistently identifying data-related issues as major challenges. (CERQual Confidence: High) (Confidence is high, mirroring Finding 3 in Appendix A, due to strong coherence and adequacy of evidence across numerous studies detailing data sources, processing steps, and challenges).

• Data Sources: Vibration sensors [4, 13, 18, 19], temperature sensors [12, 13, 20], acoustic sensors [8], current/voltage sensors [12, 21], pressure sensors, process parameters (speeds, loads, flow rates) [21, 22, 24], event logs [22], quality measurements [23], thermal imaging [12, 20], and operational context data are commonly used. Multi-sensor fusion approaches are frequent [12, 28].
• Data Processing: Standard steps include cleaning, handling missing values, normalization/scaling [4], noise reduction [6], outlier detection/removal [13, 26], and time-series segmentation. Feature engineering (e.g., statistical features from time/frequency domains [8]) is still common, although DL aims to reduce this need [5]. Feature selection techniques (e.g., RFE [12], Permutation Importance [4]) are used to identify relevant inputs.
• Data Challenges: These remain a significant bottleneck:
  
  • Data Availability/Scarcity: A recurrent challenge is the lack of sufficient historical data, particularly comprehensive run-to-failure datasets, which critically hinders robust model training and validation [7, 8, 15, 18]. For instance, the scarcity of failure data in real-world automotive fleets often necessitates reliance on simulated data or limits the complexity of deployable ML models, potentially affecting their generalization to unseen failure modes, as highlighted in recent research [8]. Similarly, in the context of industrial robots, it has been noted that obtaining extensive degradation data can be time-consuming and costly, often leading to models being trained on limited operational histories which may not capture the full spectrum of wear patterns [18]. To mitigate such scarcity, the use of generative models for data augmentation is an emerging research avenue. One such effort aims to create synthetic but realistic operational data to supplement sparse real-world datasets [29].
  • Data Quality: The performance of AI models is also significantly impacted by issues such as noise, missing values, sensor drift, and inconsistent labeling in the available data [9, 15, 25]. For example, recent work on anomaly detection has emphasized that noisy sensor data or ambiguous labeling can lead to a high rate of false positives, thereby diminishing the trustworthiness and practical utility of PdM systems; a specific mitigation methodology to address this impact has been proposed [25]. In the context of resilient PdM model development, it has been acknowledged that adversarial noise—even slight perturbations—can drastically degrade model performance, illustrating the sensitivity of AI techniques to data integrity [9]. Furthermore, obtaining high‑quality, consistently labeled data remains a major hurdle for Small and Medium‑sized Enterprises looking to implement ML‑based PdM solutions [15].
  • Data Heterogeneity & Integration: Combining data from diverse sources with different formats, semantics, and sampling rates is complex [7, 11, 21, 28]. Semantic frameworks are proposed [17].
  • Imbalanced Data: Failure events are often rare compared to normal operation, requiring specific techniques during training [8, 25].
  • Data Confidentiality & Security: Protecting sensitive operational data is crucial [9, 17].

5.4 Integration approaches and frameworks

Recognizing that PdM does not operate in isolation, several articles focus on integration and systematic implementation frameworks, representing a key trend towards more holistic asset management. (CERQual Confidence: Moderate) (Confidence is moderate, mirroring Finding 2 in Appendix A, reflecting a clear trend but with varied approaches and fewer large-scale validated implementations reported in this corpus compared to core AI techniques).

• Integration with Production Planning: Optimizing maintenance schedules jointly with production lot-sizing to minimize overall costs and disruption [16, 29]. This requires predicting health based on operational parameters [29].
• Integration with Quality Control: Linking machine health degradation directly to potential impacts on manufactured part quality [23].
• Integration via Digital Twins (DTs): The use of DTs as comprehensive platforms for integrating real-time data, physics-based models, AI algorithms, and visualization capabilities is increasingly advocated to support holistic PdM [7, 14, 18]. DTs can facilitate advanced simulation and 'what-if' analysis for maintenance decisions [18], and offer a dynamic, up-to-date representation of asset health. Standardization roadmaps for DT-based PdM are being proposed to guide their development and interoperability [7]. However, the practical implementation of DTs, especially for Small and Medium-sized Enterprises (SMEs), faces significant cost and technical barriers. The development and maintenance of high-fidelity DTs can be resource-intensive, requiring substantial investment in sensorization, data infrastructure, modeling expertise, and computational power [7, 8]. Technical barriers include challenges in data integration from heterogeneous sources, ensuring real-time synchronization between the physical asset and its digital counterpart, and the complexity of validating and maintaining the twin over the asset's lifecycle [7, 17]. In the context of advanced ML solutions, some studies have emphasized that these challenges are often amplified for SMEs due to limited financial resources, a skills gap in areas like data science and complex system modeling, and the lack of standardized, cost-effective DT platforms tailored to their needs [15]. Therefore, while DTs offer transformative potential for PdM, a careful feasibility analysis considering these cost, technical, and organizational aspects is crucial, particularly for SMEs, before embarking on full-scale DT integration.

4-1.png

4-2.png

Figure 4. Levels of PdM integration approaches

• Semantic Frameworks: Utilizing ontologies and semantic technologies to integrate heterogeneous data sources and provide context for analysis in cloud environments [17].
• Implementation Methodologies: Structured methods are proposed to guide the implementation of ML-based PdM, especially for SMEs, covering data acquisition, model selection, validation, and deployment [15].
• Hybrid Intelligence Systems: Frameworks combining AI capabilities with human expertise through interfaces like Digital Intelligent Assistants (DIAs) for collaborative decision-making [14].

Figure 4 illustrates different levels of integration discussed in the literature, moving from standalone PdM to deeply integrated enterprise solutions.

5.5 Emerging trends and techniques

Beyond the established applications, the literature highlights several emerging research frontiers aimed at overcoming current limitations and enhancing PdM capabilities. (CERQual Confidence: Moderate) (Confidence is moderate, mirroring Finding 4 in Appendix A, as these are identifiable trends but represent newer research areas with a comparatively smaller or less mature evidence base within this specific corpus). These include a strong focus on:

• Uncertainty Quantification: Developing models that not only predict RUL but also provide associated confidence levels or probability distributions [5, 19].
• Deep Reinforcement Learning (DRL): Applying DRL for complex, adaptive scheduling policies that learn optimal actions through interaction [3, 19].
• Explainable AI (XAI): Developing approaches that render the predictions from AI models transparent and comprehensible to human experts [8, 26].
• Cybersecurity Resilience: Designing AI-based PdM systems robust against adversarial data manipulation or model attacks [9].
• Systematic Implementation: Devising practical methodologies and frameworks to guide successful adoption, particularly addressing the needs and constraints of SMEs [15].
• Hybrid Intelligence: Integrating human knowledge and reasoning capabilities with AI systems, potentially via DIAs [14].

This systematic review has synthesized the findings from 29 peer-reviewed articles published in the last 10 years on the application of Artificial Intelligence to improve Predictive Maintenance, identified through a targeted Scopus search. The analysis validates that AI, notably Deep Learning, is quickly reshaping the PdM field, facilitating higher accuracy in fault diagnosis, more precise Remaining Useful Life estimation, and better-planned maintenance scheduling throughout diverse industrial sectors.

There is a clear trend towards leveraging complex models like CNNs, LSTMs, and Autoencoders, often integrated within broader systems involving IoT data streams and Digital Twins. Key advancements focus not only on predictive accuracy but also on addressing critical aspects like model uncertainty, explainability (XAI), integration with production planning, and automated decision-making through techniques like Deep Reinforcement Learning.

However, significant challenges persist, primarily centered around data availability and quality, model interpretability and trust, system integration complexity, cybersecurity, and the practical hurdles of implementation, especially for SMEs. Despite these obstacles, the potential benefits of AI-driven PdM – improved operational efficiency, reduced costs, enhanced safety, and extended asset life – are substantial, driving continued research and development.

Future efforts should focus on creating more robust, trustworthy, integrated, and scalable AI solutions, alongside developing the necessary skills and methodologies for successful real-world deployment. The convergence of AI with other Industry 4.0 technologies holds the promise of enabling truly intelligent, adaptive, and ultimately autonomous maintenance systems, marking a fundamental shift in industrial asset management. This review provides a structured overview of this progress, highlighting the achievements while emphasizing the critical areas requiring continued research and innovation.