Potential, Challenges and Future Directions for Deep Learning in Prognostics and Health Management Applications

Data and feature challenges

• Complex assets generate heterogeneous, multiscale, noisy, and partly redundant sensor data, while different faults affect different signals, making raw-data PHM difficult.Monitoring may include temperature, pressure, flow, vibration, images, and video streams, with no generally one-to-one correspondence between faults and signals.Source: §1 · p. 2
• Successful PHM typically requires feature engineering, but feature selection must balance missed detections against false alarms while preserving useful degradation information.Feature engineering transforms and reduces raw data using statistical or signal-processing methods; selecting too few or too many features can impair detection, especially for previously unseen faults.Source: §1 · pp. 2–3
• Handcrafted features have limited generalization, transferability, and scalability because they depend on fault coverage, monitored-parameter volume, and domain-expert experience.Feature quality directly influences machine-learning performance, motivating automation or methods that circumvent feature engineering; deep learning can incorporate at least part of it end-to-end.Source: §1 · pp. 3–4

Prognostics challenges

• Prognostics remains difficult for deep learning because it estimates future health evolution and RUL, unlike fault detection and diagnosis, which have adopted deep-learning approaches more recently.RUL methods include model-based, data-driven, knowledge-based, and hybrid approaches, with hybrid approaches typically combining data-driven and physics-based techniques [11].Source: §1 · p. 4
• RUL prediction must accommodate either degradation prediction from onset or fault-onset detection followed by progression prediction under operating conditions.Earlier data-driven work emphasized statistical models and stochastic degradation trajectories, while prognostics research also evaluates algorithm performance and prediction uncertainty through key performance indicators.Source: §1 · pp. 4–5

• Deep learning could solve previously intractable PHM problems, improve traditional methods, automate model development, and make industrial deployment more robust and cost-effective.Source: §1 · p. 5
• Deep learning can automatically process massive condition-monitoring data and extract useful features from high-dimensional, heterogeneous sources.Source: §1 · pp. 5–6
• Key PHM challenges include handling outliers, evolving operating conditions and faults, unknown degradation types, highly varying conditions, cross-unit generalization, interpretability, and constrained decision support.These requirements reflect the specificities of complex industrial systems and their collected condition-monitoring data.Source: §1 · p. 6
• The review evaluates developments, drivers, challenges, and potential solutions for deep learning in PHM, focusing on promising directions that could address current industrial challenges.The paper’s content and structure are presented in Figure 1.Source: §1 · p. 7

2.1. Introduction to Deep Learning

• Deep neural networks use multiple computational layers and large parameterized models trained on large datasets, with architectures matched to data structure to identify relevant features.CNNs are typically used for images, whereas RNNs are often used for sequential data such as NLP; this structural match can reduce feature engineering.Source: §2.1 · pp. 7–8
• Deep learning’s growth was driven by large labeled datasets [27], fast affordable hardware, and freely available software libraries [31, 32, 33, 34, 35].Because labeled data and computation can be costly, transfer learning enables retraining on similar tasks with fewer resources.Source: §2.1 · pp. 8–9

2.2. Deep Learning Successes

• Deep learning successes in other fields provide inspiration for advancing PHM applications.The reviewed successes span speech recognition, machine translation, NLP, and image processing.Source: §2.2 · p. 9

2.2.1. Speech Recognition, Machine Translation and NLP

• Various types of RNNs were successfully applied to machine translation, achieving comparable performance and capturing aspects of writing style and syntax.RNNs were also successfully applied to classic NLP tasks such as part-of-speech tagging.Source: §2.2.1 · p. 10

2.2.2. Image Processing

• Deep CNNs revolutionized image processing, with AlexNet establishing a new top object-recognition score after training on the million-image ImageNet dataset.Subsequent work explored increasingly deeper and more intricate network topologies [25].Source: §2.2.2 · p. 11
• Image processing has progressed from object detection to pixel-level instance segmentation and multimodal image-caption generation using CNN–RNN systems.Source: §2.2.2 · p. 11

2.2.3. Recommendation Systems

• Recommendation systems learn user preferences from large, often vast collections of data and are well-suited to deep-learning technologies; DNNs have improved YouTube and Google Play recommendation systems [59, 60, 62].Source: §2.2.3 · p. 12

2.3. Interpretability

• Deep neural networks’ distributed nonlinear computations hinder interpretability, while post-hoc methods such as surrogates, perturbations, propagation, and LIME may not sufficiently explain physical systems [65–68].Trust concerns also include bias, unrepresentative training datasets, and adversarial attacks; explanations can target representations, individual predictions, or model behavior.Source: §2.3 · pp. 12–13

3.1. Natural Language Processing for maintenance reports

• NLP processes failure reports, maintenance records, and event logs to identify failure frequency, causes, remedies, and labels or targets despite free-form variability.These records require different processing from numerical sensor data and may contain ambiguity because technicians enter them manually.Source: §3.1 · p. 14
• Word embeddings perform automatic feature engineering, capture correlations between consecutive words and wider context, and are typically more compact than 1-of-N representations; RNNs enable end-to-end learning and NLP adoption in PHM.Source: §3.1 · p. 14
• NLP can identify keywords in maintenance records or failure reports and predict the failure type a maintenance record remedied, with applications including storage-device diagnosis and transformer failure analysis.Su et al. used LSTMs with self-monitoring logs, while Ravi et al. analyzed substation-transformer event logs.Source: §3.1 · pp. 14–15

3.2. Computer Vision for Images and Videos from Inspections and Quality Control

• Computer-vision PHM applications automate image processing for classification and quantitative description, including railway-part recognition and detection of cracked ties and broken or missing rail fasteners.Gilbert et al. used a multi-task CNN trained on manually labeled railway-infrastructure images.Source: §3.2 · p. 15
• Related vision systems use transfer learning, CNN ensembles, staged CNN inspection, and fully convolutional networks to detect coating corrosion, aircraft defects, engine parts, and blade-coating loss.These applications span marine structures, aircraft fuselage, jet-engine imagery, and turbine blades; blade imaging enabled accurate segmentation of lost protective coating as a condition indicator.Source: §3.2 · pp. 15–16

4.1. Supervised Learning Algorithms for PHM applications

• LSTM-based recurrent networks are widely used in PHM, with bidirectional LSTM, GRU, and attention mechanisms extending temporal modeling, although recent attention advances remain underexploited in PHM.Attention mechanisms have significantly improved LSTM performance on time-series data and achieved state-of-the-art results on several language tasks.Source: §4.1 · p. 16
• Supervised PHM models transform sensor time series into images for CNN processing, apply 1-D CNNs directly to signals, or combine CNN feature extraction with LSTM temporal modeling.GAF, MTF, and RP provide time-series image encodings; 1-D CNNs support classification, anomaly detection, and RUL prediction; hybrid models may embed convolution within LSTM cells for multivariate anomaly detection and fault diagnosis.Source: §4.1 · pp. 16–17
• The benefits of alternative time-series-to-image encodings across data and fault types remain an open research question.The unresolved comparison concerns which encoding is most useful for different types of time-series data and fault types.Source: §4.1 · p. 17

4.2. Unsupervised and semi-supervised learning for PHM applications

• Unsupervised and semi-supervised learning are needed because PHM labels can be expensive or infeasible, especially when faults are rare, lifetimes are truncated, and true health is difficult to assess.Labels may represent health indicators, fault types, or RUL values, but many PHM applications lack them.Source: §4.2 · p. 18
• Signal reconstruction with autoencoders learns normal system behavior and uses residuals to detect deviations without requiring large amounts of labeled data, while clustering can isolate faults when representations are separable.High-dimensional, noisy, correlated signals can make fault types inseparable; clustering on autoencoder-derived latent representations is therefore used to obtain more informative feature spaces.Source: §4.2 · pp. 18–19
• Semi-supervised learning leverages available unlabeled data to improve the performance of supervised learning, using generative models, graph-based methods, transductive methods, consistency regularization, and entropy minimization.Source: §4.2 · p. 19
• Active learning can achieve higher accuracy with less labelled data by selecting training samples from a large unlabelled pool for labelling by an oracle.The oracle may be a human expert or another method that labels only a subset of available data [149].Source: §4.2 · p. 20

4.3. Uncertainty quantification and decision support

• Uncertainty-aware prognostics must represent RUL with confidence intervals, probability distributions, or fuzzy representations because deterministic point predictions can be deceptive.Relevant uncertainties arise from model approximations, sensor measurement errors, and unknown future operating profiles and loads.Source: §4.3 · pp. 20–21
• Uncertainty propagation must account for model and asset-initial-state uncertainty, producing future-state and RUL distributions without unjustified distributional assumptions.Because RUL is generally nonlinear in the initial state, Monte Carlo simulation and uncertainty propagation are often used; computational approximations include response surfaces and extended Kalman filtering.Source: §4.3 · p. 21
• Prognostics supports decisions under uncertainty, including dynamic maintenance planning that combines RUL predictions with logistic and operational constraints.Examples include preventive-maintenance timing across fleets and spare-part or schedule-adherence constraints; RUL variability measured by a coefficient of variation decreases with the rate of ageing.Source: §4.3 · pp. 21–22

5.1. Transfer Learning

• Unsupervised domain adaptation can transfer knowledge across differently operated machines without target labels, reducing the need to collect, label, and retrain models for every fleet machine.Industrial source and target domains may differ in input sensor signals and output labels, creating transfer-learning challenges beyond standard domain-adaptation assumptions.Source: §5.1 · p. 23
• Directly converting raw signals into graphical images and fine-tuning ImageNet models produced promising fault-diagnosis results for gearboxes and induction motors.Other PHM transfer approaches align source and target distributions using AdaBN, adversarial training [174], and MMD minimization.Source: §5.1 · p. 24
• Transfer-learning methods still require refinement for nonidentical source-target spaces, transferable input encoding, and healthy-fault quantity imbalance in PHM applications.The section identifies these issues as open research questions after noting that industrial sensor and fault-label sets can differ across machines.Source: §5.1 · p. 24

5.2. DL for fleet approaches

• Fleet PHM is difficult because critical failures are rare and system configurations, operating conditions, materials, and manufacturing variability make fault patterns and lifetimes unit-specific.Consequently, models or degradation trajectories cannot be directly transferred across fleet units, and the common assumption of similar operating conditions is invalid for most complex industrial assets.Source: §5.2 · pp. 24–25
• Fleet approaches progress from parameter-based sub-fleet clustering, through time-series clustering and functional-behavior modeling, to feature-space domain alignment.Functional modeling does not depend on observation-period length, whereas feature alignment compensates for distribution shifts between units.Source: §5.2 · p. 25
• These fleet methods trade limitations for complexity: aggregation may miss relevant conditions, time-series clustering suffers dimensionality and evolving-regime problems, and alignment cannot guarantee representative future conditions.Functional-behavior methods require sufficient similarity among units, potentially demanding large fleets, while unsupervised alignment depends on future operating conditions resembling those aligned.Source: §5.2 · p. 26
• Fleet-oriented PHM remains largely unaddressed despite heterogeneous systems, sensors, locations, and sampling frequencies, and solving it could provide a major industrial advance.Fleets may contain functionally similar systems from different manufacturers and operating histories.Source: §5.2 · pp. 26–27

5.3. Deep Generative Algorithms

• Deep generative models learn data distributions without assuming a predefined family, using neural generators that map noise samples to ground-truth-like samples.GANs train a Generator against a Discriminator that distinguishes real from generated data, but can suffer mode collapse when generated examples concentrate on few modes.Source: §5.3 · p. 27
• W-GANs address GAN mode collapse by using the global, continuous 1-Wasserstein distance to discourage focusing on only a subset of the distribution.The W-GAN formulation uses the dual expression of the Wasserstein distance because the primal form can be intractable in high dimensions.Source: §5.3 · pp. 27–28
• Generative models support anomaly detection by learning an induced distribution and classifying samples through reconstruction or latent-space matching, but existing methods require fixed signal durations and difficult latent optimization.Adding an encoder during training may improve generated-sample quality.Source: §5.3 · pp. 28–29

5.4. Deep Reinforcement Learning

• Deep reinforcement learning can adapt maintenance decisions to new problems quickly by exploiting experience and knowledge from previously solved problems.This is valuable because corrective and predictive maintenance tasks are dynamic, complex, and time-constrained.Source: §5.4 · p. 29
• DRL uses deep learning as a function approximator to derive strategies for complex problems whose large state spaces exceed regular reinforcement learning lookup tables.DQN replaces Q-learning’s Q-table with a CNN-based approximation of the optimal action-value function.Source: §5.4 · p. 30
• DRL can support component maintenance decisions by using age and condition as state variables and choosing between maintenance and no maintenance, with rewards reflecting failure and repair costs.Source: §5.4 · p. 30
• Reward shaping is DRL’s major PHM challenge, because hand-crafted rewards can bias learning, while reinforcement learning may perform unstably without fixed datasets and ground-truth targets.Rewards require careful design by experienced experts, yet improperly shaped rewards can produce unexpected behaviors.Source: §5.4 · pp. 30–31

5.5. Physics-induced Machine Learning

• Physics-induced machine learning integrates prior knowledge into neural-network models, improving performance and interpretability, and is promising for applications outside image processing where visualizations may not be readily derived.Source: §5.5 · p. 31
• Proposed physics-integration methods include physics-informed and physics-guided neural networks, semantic regularization, and logic rules applied across multiple learning stages [198] [204].Knowledge can enter the training data, hypothesis space, training algorithm, or final hypothesis through simulations, statistical relations, symmetries, and constraints [200]. Hybrid and physics-informed PHM approaches remain subject to further research because physical integration affects interpretability and extrapolation [205, 11, 206].Source: §5.5 · p. 31

Methodological advancements

• Heterogeneous unsupervised domain adaptation for systems with different manufacturers, sensor types, locations, and numbers remains limited but highly impactful for industrial applications.Simulation-to-real adaptation is another promising direction because source-domain data are more likely to be sufficient.Source: §6 · p. 32
• Generative models require research on physical-system coherence, controllable fault generation, transfer to complex time-series datasets, and effects of generated faults on algorithm performance.Existing work has focused largely on vibration signals preprocessed into image-like representations.Source: §6 · pp. 33–34
• Physics-induced machine learning needs consolidation for industrial transfer, while combining deep learning with expert knowledge could enable adaptive digital twins supporting maintenance decisions.Further development may also improve model and method interpretability.Source: §6 · p. 33

Data needs

• In PHM, a lack of representative datasets obstructs broad industrial use and adaptation of deep-learning approaches; insufficient data impair generalization and increase overfitting risk, while augmentation mitigates that risk [207].Research on data augmentation for time-series data remains limited.Source: §6 · pp. 33–34

Data and uncertainty management

• Future systems must continuously select and compose training data across changing operating conditions and fleets, including decisions about newly measured data, redundancy, updates, and active-learning value.Uncertainty propagation also requires accurate, computationally reasonable Monte Carlo methods for nonlinear, non-Gaussian, non-stationary stochastic processes, with stronger reliability-engineering synergies.Source: §6 · p. 34

Industrial implementation

• Broad PHM implementation requires automated and standardized data collection, maintenance-report sharing, accepted data-quality assessment, and business and legal frameworks that support stakeholder collaboration.These requirements are prerequisites for realizing deep learning’s potential to improve complex-asset availability, safety, and cost efficiency.Source: §6 · pp. 34–35