Predictive Maintenance – An Overview

1.1 Background and motivation

Traditional maintenance strategies fall into two primary categories: reactive maintenance, which addresses failures after they occur, and preventive maintenance, which follows fixed schedules regardless of actual equipment condition. Both approaches have significant limitations. Reactive maintenance leads to unexpected downtime, production losses, and potentially catastrophic failures, while preventive maintenance often results in unnecessary interventions and premature component replacement [1]. In the industrial era, any equipment requires periodic maintenance of individual components, critical components, or general maintenance. These maintenance processes are prescribed by the manufacturer or based on the experience of the owner. Nowadays, large technical and industrial systems still rely on these two types of maintenance; however, PdM is a new and robust paradigm that is increasingly being adopted [3]. The aim of PdM is to better estimate the optimal moment to perform maintenance, thanks to the collected, stored, and processed data.

The advent of Industry 4.0 technologies, particularly the Internet of Things (IoT), Big Data analytics, Cloud Computing, and Artificial Intelligence, has enabled a more sophisticated approach with a huge volume of collected data, powerful treatments, and also a better manner to deploy complete solutions. PdM leverages continuous monitoring of equipment conditions through sensors, real-time data analysis using machine learning algorithms, and predictive models that forecast potential failures with increasing accuracy [2, 4].

1.2 Scope and objectives

This literature review provides a comprehensive analysis of recent PdM research to address the following research objectives: identify and categorize the main research themes and areas in PdM, analyze key methodologies and techniques, including machine learning, deep learning, and IoT approaches, examine applications across different industrial domains, highlight recent advances and innovations in the field, identify persistent challenges and limitations, and propose future research directions.

The remainder of this review is organized as follows: Section 2 presents the framework and a brief evolution of PdM. Section 3 identifies and discusses the main research themes. Section 4 analyzes key methodologies and techniques. Section 4 also discusses challenges and future research directions. Finally, Section 5 provides concluding remarks and a synthesis of key findings.

3.1 Data-driven condition monitoring

Data-driven condition monitoring represents the foundational theme in modern PdM research. The field has undergone a significant transformation from knowledge-based rule systems to large-scale analytics that combine IoT sensing with machine learning for remaining useful life (RUL) estimation and quality prediction in manufacturing environments [1].

This transition reflects a broader shift in industrial practice. Traditional condition monitoring relied on periodic manual inspections and expert interpretation of equipment behavior. Modern approaches continuously collect sensor data – including vibration, temperature, acoustic emissions, current consumption, and pressure – and apply sophisticated algorithms to detect subtle patterns indicative of degradation or impending failure [2, 4]. Key aspects of data-driven condition monitoring include continuous sensing, feature engineering, condition indicators, and threshold-based and model-based detection. Continuous sensing involves real-time data collection from multiple sensor modalities [12–14]. Feature engineering consists of the extraction of time-domain, frequency-domain, and time-frequency domain features from raw sensor signals [15, 16]. Condition indicators pertain to the development of health indices and degradation metrics that quantify equipment condition. Threshold-based and model-based detection refers to hybrid approaches combining statistical thresholds with predictive models [17].

3.2 Deep learning for fault diagnosis and anomaly detection

Deep learning has emerged as the dominant paradigm for fault diagnosis and anomaly detection in PdM. Comprehensive surveys document the widespread adoption of NNs, CNNs, autoencoders, generative adversarial networks (GANs), and transformer-based models to detect anomalies and model temporal patterns in multisensor industrial data [4].

The appeal of deep learning lies in its ability to automatically learn hierarchical feature representations from raw or minimally processed data, eliminating the need for manual feature engineering that characterized earlier approaches. This is particularly valuable in industrial settings where sensor data is high-dimensional and exhibits complex temporal dependencies, multiple failure modes may manifest with subtle, overlapping signatures, labeled failure data is scarce – requiring transfer learning or semi-supervised approaches – and real-time processing demands efficient inference.

Common deep learning architectures for PdM include classical NNs, CNNs, RNNs and Long Short-Term Memory (LSTM), autoencoders, GANs, and transformers.

Classical NNs are used to build regression models or classification models. In [18], the resulting model is then reduced, and transfer learning is realized. The CNNs are effectively used for spatial pattern recognition in vibration spectrograms and image-based monitoring. For example, in [19], after cleaning and normalizing the data, the 1-D data is transformed into 2-D data by matrix multiplication, making it a good data format for CNN. In [20], the IoT gathers the vibrations, followed by the construction of a spectrogram, which is ultimately used as the input for the CNN to detect defect frequencies.

The RNNs and LSTM aim to capture temporal dependencies in time-series sensor data. When we assimilate all measures at the same moment as a point in history, all points can be treated by an LSTM network in an aim to predict the RUL of a system; see De Vita and Bruneo [21].

The autoencoders aim at unsupervised anomaly detection by learning normal operation patterns, like [22] with an LSTM autoencoder [23], with deep convolutional autoencoders [24], proposed a Wasserstein Autoencoder regularized with GANs [25], and used a distributed stacked sparse autoencoder.

GANs rely on data augmentation to address class imbalance in failure datasets or to reduce the space of data [24, 26, 27].

The transformers are attention mechanisms for long-range temporal dependencies. In [28], the transformers are used in order to model relationships and patterns in sequential data. In [29], the authors adapted a transformer-KDE model; the found kernel is much more adapted.

3.3 IoT/IIoT architectures and sensing systems

The integration of IoT and Industrial Internet of Things (IIoT) technologies forms a critical research theme, as these systems provide the sensing infrastructure that enables data-driven PdM. Research in this area treats PdM as an integrated technology stack encompassing edge devices, cloud infrastructure, dashboards, and digital twins [1]. Examples of such architecture can be found in [25, 30, 31]. In [32], even though the architecture framework is less complicated, the author highlights the edge computing layer.

Modern IIoT architectures for PdM typically follow a hierarchical structure. These architectures contain, firstly, a sensor layer (physical sensors like accelerometers, thermocouples, and current sensors attached to equipment), an edge layer with local processing units that perform initial data filtering, feature extraction, and lightweight inference, the network layer for data transmission with communication protocols like MQTT, OPC-UA, and HTTP, the cloud/server layer for centralized storage, advanced analytics, and model training, and, finally, the application layer with user interfaces, dashboards, and DSSs.

Key research challenges in this area include optimizing the edge-cloud computation split to balance latency, bandwidth, and accuracy; ensuring interoperability across heterogeneous sensor systems and industrial protocols; implementing secure data transmission and storage; and managing the volume, velocity, and variety of streaming sensor data.

3.4 RUL estimation and prognostics

RUL estimation – predicting how much longer a component or system will operate before failure – represents a core objective of PdM. Unlike fault detection, which identifies current anomalies, prognostics provide forward-looking predictions that enable proactive maintenance planning.

Research in RUL estimation employs three main approaches: model-based methods, such as physics-based degradation models that simulate failure progression, like [33] working with degradation curves; data-driven methods, such as statistical and machine learning models trained on historical run-to-failure data, as in [1], where a comparison of performance was conducted; and hybrid methods, which combine physics-based and data-driven approaches, like [34].

Recent work emphasizes data-driven and hybrid approaches due to their ability to handle complex systems where physical models are difficult to formulate. Common techniques include survival analysis and proportional hazards models [35, 36], regression models (support vector regression, Gaussian process regression) [14], RNNs for sequential RUL prediction [37], and uncertainty quantification to provide confidence intervals on predictions [38, 39].

3.5 Integrated DSSs

The final major theme concerns the integration of PdM capabilities into comprehensive DSS that serve operators, maintenance technicians, and plant managers. Research recognizes that accurate predictions alone are insufficient; the system must translate technical outputs into actionable recommendations [1, 2]. DSSs for PdM could include maintenance scheduling, root cause analysis, visualization, and dashboards. Scheduling is a hard-to-solve classical problem in our case when we know the RUL for each piece of equipment. Another approach, like [40], highlights the RUL of a single piece of equipment as a function varying in probability depending on the risks associated with each component.

Ideally, this DSS is fully integrated into enterprise information systems and also is able to incorporate expert knowledge and enable operators to provide feedback.

3.6 Synthesis

To represent the new trends in PdM with respect to the initial schema in Figure 1, we can see the IoT Layer, the Edge Computing Layer, and the integration with the Decision Support System. Figure 2 attempts to synthesize the new trends in PdM.

The development of Industry 4.0 and the use of IoT and IIoT are increasing the volume of data captured, which is essential to implement PdM. This leads to the need for more powerful machine learning methods, and NNs are one of them. The increased need for computing and storage requires a different way to organize these calculations: put an Edge Computing layer near the equipment and tools under surveillance, and then transmit the reduced data to a server capable of setting up prediction calculations for both risk and RUL. PdM is no longer an isolated application; it becomes part of the DSS, and the captured data, as well as the calculated data, can also be integrated into other modules such as DSS reporting and DSS visualization.

4.1 Machine learning classification and regression

Traditional machine learning methods remain widely used in PdM, particularly when interpretability, computational efficiency, or limited training data are priorities. These methods typically operate on engineered features extracted from raw sensor signals.

4.2 Deep learning time series models

Deep learning has become the dominant approach for PdM due to its ability to automatically learn representations from raw or minimally processed sensor data.

4.3 Multimodal fusion techniques

Industrial equipment generates data from multiple modalities: vibration sensors, thermal cameras, acoustic sensors, current measurements, and visual inspection. Multimodal fusion techniques combine these complementary information sources for improved predictions.

Common fusion approaches include early fusion, which involves concatenating raw features from different modalities; late fusion, which entails training separate models per modality and combining their predictions; intermediate fusion, which combines learned representations at intermediate network layers; and attention-based fusion, which uses attention mechanisms to dynamically weight modality contributions.

A significant recent advance is OmniFuser [61], which introduces contamination-free crossmodal fusion and recursive refinement to combine images and force signals for robust tool-state classification and forecasting in milling operations. This framework addresses challenges of modal imbalance, where one modality dominates; temporal misalignment between different sensors; and contamination, where noise in one modality degrades overall performance.

4.4 Edge, federated, and decentralized learning

The need for low-latency predictions, bandwidth efficiency, and data privacy has driven research into edge and federated learning approaches for PdM.

4.5 Computer vision for process monitoring

Computer vision techniques enable PdM through visual inspection of equipment and processes. We can cite a wide panel of applications: defect detection (identifying surface defects, cracks, corrosion, and wear), alignment monitoring (detecting misalignment in rotating machinery), process quality (monitoring manufacturing processes for anomalies), and root cause localization (pinpointing failure locations visually).

In the OminiFuser framework [48], a notable long-term deployment used real-time computer vision plus sensor fusion for early failure prediction on a steel rolling mill, demonstrating practical root-cause insights and reduced breakdowns. The system processed live video streams using CNNs to detect visual indicators of impending failures.

4.6 Anomaly detection and unsupervised methods

When labeled failure data is scarce, unsupervised and semi-supervised methods become essential. In Machine Learning, the most used methods are: One-Class SVM (which learns a boundary around normal operation data), Isolation Forest (which identifies anomalies as data points that are easily isolated), Local Outlier Factor (able to detect anomalies based on local density), and statistical methods (like control charts, CUSUM (cumulative sum control chart), and exponential smoothing).

Classifier auditing, improved by One-Class SVM and ensemble anomaly filters, has been demonstrated in packaging equipment PdM [62]. In [63], the authors have proved that the One-Class SVM with a radial function kernel had an excellent average recall score of 0.95. In [8], the conclusion is also that One-Class SVM is highly effective at pinpointing anomalies. In [47], the One-Class SVM is slightly less performant than the Local Outlier Factor, with a recall of 0.913.

4.7 Evaluation of a PdM tool

PdM algorithms are typically evaluated along two complementary dimensions: their accuracy and their impact on maintenance decision-making. Prognostic performance focuses on how well an algorithm predicts RUL or failure probability, using metrics such as mean absolute error (MAE), root mean squared error (RMSE), mean squared error (MSE), prediction interval coverage, and timeliness indices designed specifically for RUL estimation. However, pure statistical accuracy is often insufficient because the ultimate goal of PdM is to reduce life-cycle cost and risk; recent work therefore proposes decision-oriented metrics that estimate the long-run expected maintenance cost per unit time under a given PdM policy, directly linking algorithm quality to operational performance [64].

In industrial PdM for Industry 4.0, evaluation frameworks frequently combine classical machine learning metrics (accuracy, precision, recall, F1‑score, ROC curves) with time‑series error measures (MSE, RMSE, MAE) to compare forecasting and classification models such as Random Forests, Gradient Boosting, ARIMA, and Prophet on real sensor data. For example, in [65] the authors evaluate multiple forecasting models for equipment monitoring using blocked cross‑validation and rolling‑window validation to mimic online deployment, and they report RMSE and MAE across different training windows to assess robustness. In [66], multi‑criteria decision‑making methods (e.g., CRITIC–AROMAN) have also been adopted to rank PdM models by combining accuracy, robustness, scalability, interpretability, and computational efficiency, reflecting the fact that deployability and transparency are as important as raw predictive power in industrial environments.

4.8 Synthesis

Table 1 summarizes the key methodologies, their typical algorithms, strengths, and representative applications.

This wide range of methods responds to different uses (predictions, outlier detection, time series predictions), for different formats (numerical values, images, data streams), with different volumes depending on the nature, number, and type of sensors used.

The choice of methodology depends on factors including data availability, data volume, computational resources, interpretability requirements, real-time constraints, and the specific industrial application. If a large dataset is available, the Machine Learning methods could be tuned with the aim of achieving the best performance. If the volume of input data is huge, feature engineering becomes an imperative step, as well as the choice of a neural network approach. If the computational resources are local, this is an excellent choice to be able to respond to real-time constraints. If these computational resources are limited, Edge Computing and Cloud Computing become the solution, but the use of these technologies makes it more difficult to comply with real-time constraints. Modern systems often combine multiple approaches, for example, using edge-based anomaly detection for rapid response and cloud-based deep learning for detailed diagnostics.

The interpretability of the results, which is opaque for neural network methods, can be problematic when users trust the results. On the other hand, highly interpretable classical methods, such as k-nearest neighbors or decision trees, have poorer performance.

Finally, the integration of PdM into complex production monitoring and decision-making systems becomes an important technical problem to be solved.

5.1 Data quality and availability

The most frequently cited challenge in PdM is the scarcity of labeled failure data. Failure events are rare because modern equipment is highly reliable, so failures are infrequent. The run-to-failure requirements impose that supervised learning ideally requires equipment to fail, which is economically unacceptable. The labeling cost is another problem because expert annotation of sensor data is time-consuming and expensive. Even when the labeling can be done, ambiguous labels can arise because failure modes may be unclear or misdiagnosed.

Some proposed solutions could be: semi-supervised and self-supervised learning to leverage unlabeled data, transfer learning from similar equipment or operating conditions, active learning to prioritize labeling of informative examples, and simulation-based synthetic data generation.

Real industrial data presents numerous quality challenges: missing data (as sensor failures, communication interruptions, and maintenance periods create gaps), noise and artifacts (in the case of electromagnetic interference, sensor drift, and environmental factors introducing noise), temporal misalignment (different sensors may have different sampling rates or timestamps), and concept drift (as operating conditions, equipment configurations, and degradation patterns change over time).

Papers call for strategies to automate annotation and fuse upper/lower shop-floor topics to mitigate labeling bottlenecks [1, 4].

The research community lacks high-quality, publicly available benchmark datasets:

• Most industrial datasets are proprietary and cannot be shared, for example, [23, 28]

• Existing public datasets (e.g., NASA bearing data [74], CWRU bearing data [75]) are small and do not reflect modern industrial complexity

• Lack of standardization makes cross-study comparisons difficult

• Multisensor, multimodal datasets are particularly scarce

Surveys note the shortage of high-quality multisensor industrial datasets and recommend public benchmarks and multisensor datasets to validate deep learning methods.

5.2 Model interpretability and trust

Deep learning models, while accurate, are often criticized as black boxes, so maintenance personnel may not trust predictions they cannot understand. Regulatory compliance may require explainable decisions. Debugging model failures is difficult without interpretability. Integration with human expertise requires transparent reasoning.

Papers call for interpretable models and decision-support tightly integrated with maintenance workflows to ensure actionable outputs for operators [1, 2].

On the other hand, the technical predictions must be translated into maintenance actions. Knowing that a failure is likely is insufficient; operators need to know what to do. Predictions must be integrated with spare parts availability, technician schedules, and production plans. False positives erode trust and waste resources. Uncertainty quantification is needed to communicate prediction confidence.

Some future directions could be: explainable prognostics that combine RUL predictions with explanations mapping sensor patterns to likely failure modes; integration with computerized maintenance management systems (CMMS); human-in-the-loop approaches that enable operator feedback; and counterfactual explanations showing what conditions would prevent predicted failures.

5.3 Security, privacy, and trust

Industrial IoT systems face significant security risks: sensor data can be intercepted or manipulated, model poisoning attacks can degrade prediction accuracy, adversarial examples can fool deep learning models, and unauthorized access to PdM systems could enable sabotage.

Centralized training and storage pose risks; decentralized architectures using blockchain/IPFS are proposed to improve data integrity and model trustworthiness.

Operational data is often considered proprietary: sensor data reveals production rates, process parameters, and equipment configurations. Sharing data for collaborative model development risks competitive advantage. Regulatory requirements (e.g., GDPR) may restrict data handling. Multi-party scenarios (manufacturers, operators, service providers) complicate data governance.

As future directions, we could suggest: federated and privacy-preserving PdM protocols to train cross-site models while preserving industrial privacy and heterogeneity; differential privacy techniques to protect individual data points; secure multi-party computation for collaborative analytics; and security-by-design PdM that integrates immutable logging and secure storage (blockchain/IPFS) where provenance and auditability are required.

5.4 Operational deployment constraints

Industrial edge devices have limited computational capabilities: real-time inference requires low latency (milliseconds to seconds), edge hardware has limited memory, processing power, and energy budget; model compression may sacrifice accuracy, and balancing edge versus cloud processing is application-specific.

Studies demonstrate edge deployment of CNN-LSTM and server-based inference for low PLC load as practical mitigations.

Most industrial facilities have existing control and information systems. Various problems could arise: legacy equipment may lack sensors or connectivity, proprietary protocols complicate data integration, operational technology (OT) and information technology (IT) systems are often isolated, and retrofitting existing equipment is costly.

Technical systems must fit into existing organizational processes. Maintenance teams have established procedures and responsibilities. PdM changes job roles and skill requirements. Organizational change management is often underestimated. Demonstrating return on investment (ROI) is necessary for continued support.

5.5 Generalization and transfer learning

Models trained on one piece of equipment often fail on similar equipment. Some causes could be: equipment variations (manufacturer, model, age) that affect sensor signatures, operating conditions (load, speed, environment) that differ across installations, degradation patterns that vary due to maintenance history and usage, and domain adaptation techniques that are needed but underdeveloped for industrial applications.

Transferring models across facilities faces additional challenges: different sensor types, placements, and sampling rates; facility-specific processes and operating procedures; varying data quality and availability; and legal and organizational barriers to data sharing.

Some future directions could be: domain adaptation and transfer learning techniques tailored to industrial scenarios, meta-learning approaches that learn how to adapt quickly to new equipment, standardized sensor configurations and data formats, and federated learning to leverage cross-facility diversity.

5.6 Evaluation and benchmarking

The evaluation of the performances of each PdM system or method is done based on some criteria that are not unified. Also, it is quite difficult, and sometimes impossible, to reproduce the results of a detailed research paper.

Different studies use different evaluation metrics, making comparison difficult: classification metrics (accuracy, precision, recall, F1) do not capture cost-benefit trade-offs; RUL estimation metrics (RMSE, MAE) do not reflect maintenance decision quality; time-to-detection and false alarm rates are often unreported; business metrics (downtime reduction, maintenance cost savings) are rarely quantified.

Many studies lack sufficient detail for reproduction: proprietary datasets cannot be shared, hyperparameters and training details are often omitted, code availability is inconsistent, and long-term deployment results are rarely published.

We can suggest standardizing evaluation protocols with common metrics and reporting standards. We also need examples of long-term deployments with common metrics and open sharing of anonymized datasets to improve reproducibility and cross-site generalization, benchmark datasets that reflect modern industrial complexity, and robust multimodal benchmarks that curate multisensor, multimodal datasets (visual + force + telemetry) to empirically evaluate fusion models.

5.7 Specific technical challenges

Work on a complete solution for PdM is not easy; a sum of challenges needs to be solved.

Class imbalance is the first challenge to solve. Failure events are rare compared to normal operation. As a result, some drawbacks are brought into focus: standard classification algorithms are biased toward the majority class, oversampling and undersampling have limitations, cost-sensitive learning requires accurate cost estimates, and evaluation metrics must account for the imbalance.

Equipment degradation unfolds over multiple time scales: short-term fluctuations versus long-term trends, seasonal and cyclic patterns, sudden versus gradual failures, and variable-length sequences.

Industrial systems involve complex interactions: sensor readings are correlated, failure modes affect multiple sensors, causal relationships are difficult to infer, and confounding factors complicate analysis.