Source-linked AI summary
Machine Learning for Reliability Engineering and Safety Applications: Review of Current Status and Future Opportunities
Zhaoyi Xu, Joseph Homer Saleh
TL;DR
Reliability engineering and safety face a large, fragmented ML literature as these fields undergo growing ML influence. This paper provides a panoramic synthesis, reviews prior applications, and identifies future opportunities, arguing that ML can yield more accurate accident-dataset insights and support better-informed decisions and accident prevention.
Problem
Reliability engineering and safety are increasingly influenced by ML, but their large literature is fragmented and difficult to integrate.
Method
The paper provides a panoramic synthesis and roadmap, reviews ML use in reliability engineering and safety, and looks ahead to future opportunities.
Results
ML can provide novel insights and opportunities for reliability and safety applications and more accurate insights from accident datasets than traditional analysis tools.
Takeaways & Limitations
These ML capabilities can support better-informed decision-making and more effective accident prevention.
Takeaways & Limitations
The paper does not address safe reinforcement learning beyond a brief discussion, while AI safety deserves careful attention.
Abstract
from arXiv · showhide
Machine learning (ML) pervades an increasing number of academic disciplines and industries. Its impact is profound, and several fields have been fundamentally altered by it, autonomy and computer vision for example; reliability engineering and safety will undoubtedly follow suit. There is already a large but fragmented literature on ML for reliability and safety applications, and it can be overwhelming to navigate and integrate into a coherent whole. In this work, we facilitate this task by providing a synthesis of, and a roadmap to this ever-expanding analytical landscape and highlighting its major landmarks and pathways. We first provide an overview of the different ML categories and sub-categories or tasks, and we note several of the corresponding models and algorithms. We then look back and review the use of ML in reliability and safety applications. We examine several publications in each category/sub-category, and we include a short discussion on the use of Deep Learning to highlight its growing popularity and distinctive advantages. Finally, we look ahead and outline several promising future opportunities for leveraging ML in service of advancing reliability and safety considerations. Overall, we argue that ML is capable of providing novel insights and opportunities to solve important challenges in reliability and safety applications. It is also capable of teasing out more accurate insights from accident datasets than with traditional analysis tools, and this in turn can lead to better informed decision-making and more effective accident prevention.
1. Introduction
Machine learning is reshaping many fields, while reliability and safety applications face a large, fragmented literature and increasing demands for precise analysis. This work synthesizes that landscape, reviews current ML use, and outlines future opportunities for reliability and safety.
- The paper provides a synthesis and roadmap of current ML use in reliability engineering and safety, while outlining future opportunities.
- ML has fundamentally altered fields such as autonomy and computer vision, with reliability engineering and safety expected to follow.
- Reliability and safety ML research is large and fragmented, making it difficult to navigate and integrate coherently.
- Traditional reliability tools face challenges modeling increasingly complex systems, improving predictions, propagating uncertainty, and meeting industry demands for precision and accuracy.
- ML can provide new insights, extract more accurate information from accident and degradation datasets, and support better engineering decisions and accident prevention.
2. Machine learning: a brief overview
Machine learning organizes data-driven analysis around supervised, unsupervised, semi-supervised, and reinforcement learning, with tasks and models suited to different data and questions.
- Learning categories: ML categories are selected according to the available data and the questions being asked about it.The paper identifies supervised, unsupervised, semi-supervised, and reinforcement learning as major categories.
- Supervised learning: Supervised learning estimates a function from paired features and labels for inference or prediction, primarily through regression and classification.Regression uses quantitative responses, whereas classification uses qualitative responses and can involve binary or multiclass outputs.
- Semi-supervised learning: Semi-supervised learning estimates a function when only some observations have labels, bridging supervised and unsupervised approaches.It applies when some feature vectors have associated responses while others do not.
- Reinforcement learning: Reinforcement learning explores inputs and searches for an optimal input under a desired output using a reward function rather than a fixed dataset.Its distinctive setup differs from the preceding categories because it does not operate with a fixed dataset.
- Models and algorithms: Available models range from linear regression and decision trees to support-vector, random-forest, Bayesian, Gaussian-process, and deep-neural-network methods.Gaussian Process Regression and Bayesian networks provide uncertainty quantification in regression, while DNNs are increasingly used for high-dimensional nonlinear problems.
- Unsupervised learning: Unsupervised learning examines unlabeled inputs to discover patterns, especially through clustering and anomaly detection.Clustering divides observations into similar groups, while anomaly detection identifies observations that differ from expected behavior.
3. Machine learning applications in reliability engineering and safety applications: overview and current status
The paper reviews current ML use in reliability engineering and safety, synthesizing major publications, introducing algorithms, and emphasizing deep learning and future opportunities.
- Current status: The review examines current ML use in reliability engineering and safety applications.It presents this review as part of its synthesis of the growing literature.
- Review scope: The authors sample major publications across the ML categories and subcategories discussed in the paper.They explicitly make no claim of exhaustiveness because the literature is growing.
- Deep learning: The review separately discusses deep learning to highlight its growing popularity and distinctive advantages despite overlap with the other ML categories.Deep learning receives a dedicated subsection within the reliability and safety review.
- Motivation and opportunities: ML applications generally seek more novel and accurate results from datasets for reliability- and safety-informed decisions and accident prevention.The paper frames these objectives in relation to system design and operation.
- Application areas: Applications discussed include remaining-useful-life estimation, anomaly and fault detection, health monitoring, maintenance planning, and degradation assessment.These recurring application themes organize part of the reviewed literature.
3.1 Supervised learning applications
Supervised ML applications in reliability and safety span regression and classification, including RUL estimation, degradation prediction, fault diagnosis, and reliability assessment.
- Regression: Supervised regression is widely used for RUL estimation and degradation prediction in reliability and safety applications.Applications include aircraft engines, railway tracks, turbine components, rolling bearings, and batteries.
- Regression: RUL prediction supports maintenance planning by reducing downtime and adverse consequences associated with run-to-failure strategies.The paper identifies accurate RUL and degradation estimates as important for safety-critical systems and condition-based maintenance.
- Regression: Data-driven degradation models use historical sensor data without requiring prior system knowledge, but their accuracy and computational intensity depend on data quality and quantity.The paper contrasts these models with physics-based approaches that require extensive prior knowledge and expertise.
- Classification: Classification ML tools support fault detection and identification across aircraft engines, transformers, pipes, bearings, wind turbines, software, and forest fires.The reviewed literature includes both binary and multiclass fault-classification applications.
- Classification: A reliability surrogate model achieved high prediction accuracy with an order of magnitude less computational effort than traditional Monte Carlo simulation.Other reviewed classifiers improved screening accuracy and precision while reducing output variability between human appraisers.
3.2 Unsupervised learning applications
Unsupervised learning applications emphasize clustering and anomaly detection for reliability and safety, especially where operational labels are limited or unavailable.
- Clustering: Clustering has been applied to network reliability, railway degradation, wind-turbine failures, nuclear fault detection, structural damage, bearing faults, and mining operations.The reviewed studies use clustering for both application-specific diagnosis and broader reliability modeling.
- Clustering: Clustering extracted structural network properties and supported hierarchical modeling for network reliability and criticality analysis.Artificial Ant Clustering identified operational modes and two induction-motor fault types despite limited operating-mode information.
- Clustering: Artificial Ant Clustering produced significant improvements in clustering error rates compared with other methods for its particular application.The reported improvement held even when information about the motors’ operating modes was limited.
- Anomaly detection: Anomaly detection is suited to early fault or damage detection because unlabeled data are widely available while labeled data are expensive or rare.Reliability and safety datasets may contain labels for nominal operating conditions only, motivating semi-supervised use of anomaly-detection tools.
- Anomaly detection: LSTM demonstrated near-real-time anomaly detection in spacecraft telemetry spanning more than 700 channels.The reviewed anomaly-detection literature also includes aircraft recorders, gas turbines, spacecraft, induction motors, and structural damage.
- Anomaly detection: A CNN merged feature extraction and unsupervised classification, detecting anomalies directly from raw sensor data without preprocessing or input transformation.This design addresses the computational cost of feature extraction and the dependence of detection accuracy on feature selection.
- Anomaly detection: An SDAE+ELM model improved gas-turbine anomaly-detection sensitivity and specificity while automatically extracting features from exhaust-gas-temperature measurements.The model inferred combustor operating condition and detected anomalies in industrial turbines.
3.3 Semi-supervised learning applications
Semi-supervised learning combines limited labeled data with abundant unlabeled data for fault detection, classification, and RUL prediction, while the review cautions against assuming universal gains.
- Applications: Semi-supervised learning is used in reliability and safety mainly for fault detection and identification, with additional applications in RUL prediction.Reviewed systems include cooling fans, centrifugal pumps, turbofan engines, induction motors, and solar arrays.
- Opportunities: Semi-supervised methods have significant, barely tapped potential for reliability and safety applications.The review identifies this potential across fault detection, PHM, and RUL prediction.
- RUL prediction: Semi-supervised RUL prediction improved accuracy and reduced result scatter compared with supervised learning, but one study used labels for 50% of the data.That labeled proportion was much larger than typically found in semi-supervised applications.
- RUL prediction: A variational auto-encoder approach achieved good RUL prediction accuracy using labeled data comprising as little as 1% of the dataset.The method trained on labeled and unlabeled data simultaneously.
- Methods: Co-training repeatedly pseudo-labels the most confident predictions on unlabeled data and retrains the supervised model on the expanded dataset.The process continues until a termination criterion is reached and further improvement is unavailable.
- Fault detection: A graph-based method achieved 99% fault-detection and classification accuracy for solar PV arrays with less than 2% labeled data under real working conditions.It used existing voltage, current, and operating-temperature measurements without additional hardware and leveraged self-learning.
- Caution: Semi-supervised learning can degrade performance when its underlying assumptions are violated, despite publication bias toward positive results.Safe semi-supervised learning studies when unlabeled data can be integrated without degrading performance.
3.4 Reinforcement learning applications
Reinforcement learning addresses reliability and safety problems involving large state and action spaces, including maintenance, sampling, and network operations. Safe reinforcement learning adds safety constraints to avoid hazardous states during learning and deployment.
- Applications: A Deep Reinforcement Learning solution for cellular handover outperformed alternatives on user QoS while requiring fewer parameters to tune.The objectives included reducing dropped calls and unsatisfied customers, increasing throughput, and balancing cell load.
- Applications: Reinforcement learning supports inspection, maintenance, structural reliability assessment, and cellular-network handover applications.Reported applications include large-scale infrastructure maintenance, reliability sampling, and wireless user handover.
- Applications: Deep reinforcement learning identifies efficient inspection and maintenance policies for large-scale infrastructure systems and near-optimal policies for traditionally intractable problems.The review presents DRL as a response to the computational challenges of conventional decision-making tools.
- Applications: A DRL sampling method achieved higher failure-probability prediction accuracy than Monte Carlo Simulation and Latin hypercube sampling.The sampling space is treated as the state, sample selection as the action, and a reward function guides point selection.
- Safe reinforcement learning: Safe reinforcement learning incorporates safety considerations into learning to avoid hazardous or unsafe states during learning and deployment.Safety can be encoded through the optimization criterion, exploration process, risk metrics, or external knowledge about undesirable states.
4 There are other definitions of safe reinforcement learning, see for example Munos et al. [149]
Deep learning uses layered feature extraction to handle complex and high-dimensional data. The reviewed applications report strong performance in diagnostics, anomaly detection, and remaining-useful-life prediction, while the authors qualify that its advantage need not be universal.
- Deep learning foundations: Deep learning organizes connected computational neurons in multiple layers, with each layer extracting features from lower-level representations.This nested hierarchy enables high-level latent features to be learned from simpler features and raw data.
- Distinctive advantages: Deep learning is suited to high-dimensional reliability and safety problems because its hierarchical representations can counter the curse of dimensionality.The review connects this advantage to applications such as prognostics and health management.
- Applications: Deep learning applications are growing across anomaly detection, fault classification, remaining-useful-life estimation, maintenance planning, and PHM.The review presents these areas as examples of the expanding use of DL in reliability and safety.
- Applications: A deep belief network achieved superior classification accuracy to shallow ML methods such as SVM for system health diagnosis.The model was used to classify different system health states and benchmarked against shallow methods.
- Applications: Deep autoencoders support unsupervised anomaly detection for high-dimensional time-sequential data, including flight time series.A multimodal deep autoencoder was validated on NASA’s DASHlink open database with excellent reported anomaly-detection performance.
- Applications: A CNN-based RUL model using multiple turbofan-engine sensor signals reported higher accuracy and lower RMSE than shallow ML methods.Validation used NASA’s turbofan engine degradation dataset.
- Overall assessment: The review reports that DL significantly outperformed shallow ML algorithms in applications to date, while cautioning that this need not be universal.The authors nonetheless judge DL’s return on investment more favorably for reliability and safety applications than shallow methods.
4. Future opportunities: promising reliability and safety applications of ML
The review proposes a short, non-exhaustive set of future opportunities for ML in reliability and safety. It emphasizes uncertainty-aware deployment, higher-level PHM, richer safety data integration, wearable sensing, and advanced models for complex engineering systems.
- Accident databases and safety-management systems: Linking accident and non-accident databases could support exposure measurement, cross-context comparisons, and identification of predictive features associated with accidents and injuries.These linkages are discussed in the context of occupational injury prevention.
- Uncertainty quantification and deployment: Future ML deployments should quantify and reduce uncertainty rather than provide only point predictions, especially for RUL estimation.Confidence intervals and identified uncertainty sources are presented as more useful for informed decision-making than a single RUL value.
- Fleet-level and system-of-systems PHM: The review identifies fleet-level and system-of-systems PHM as promising directions beyond component- or subsystem-level applications.Examples include satellite mega-constellations, truck fleets, and wind-turbine farms, with challenges involving human factors, systems engineering, and varying operating conditions.
- Accident databases and safety-management systems: ML could extract more insight from accident databases and safety-management systems by integrating heterogeneous safety data across time scales and collection methods.The proposed applications include accident-precursor identification, safety interventions, near-miss analysis, and exposure measurement.
- Wearable computing and predictive safety analytics: Wearable computing, sensors, and body cameras could enable predictive safety analytics, injury prevention, and technician support during maintenance and inspection.The review also describes possible integration of body-camera recognition with permit-to-work systems, subject to proper design, testing, and rollout.
- Advanced ML models: The authors highlight DGP and GAN models as promising candidates for fault detection, classification, uncertainty analysis, and advanced PHM in complex engineering systems.They cite their accuracy and efficiency with high-dimensional data as reasons for anticipating broader adoption.
- Scope and research agenda: The proposed future agenda is explicitly short and non-exhaustive, inviting researchers to complement the authors’ selection.The review also calls for better integration of ML with risk-analysis tools and concepts and greater transparency in comparative model selection.
5. Conclusion
The paper surveys ML applications in reliability engineering and safety, organizing fragmented literature into a coherent overview and roadmap. It also identifies future opportunities while noting that the reliability and safety of ML/AI systems themselves remain outside its scope.
- Motivation: The paper responds to a large, fragmented literature that can be difficult to navigate and integrate coherently.It frames the review as a way to provide structure for an expanding analytical landscape.
- Review approach: It first maps ML categories, sub-categories, tasks, and corresponding models and algorithms.
- Review approach: It reviews selected publications across ML categories and discusses Deep Learning's growing popularity and distinctive advantages.
- Future opportunities: The paper looks ahead by outlining promising opportunities for leveraging ML to advance reliability and safety considerations.
- Scope boundary: The review does not address the reverse relationship: the reliability and safety of ML systems and, more broadly, AI systems.The authors identify this as an important topic requiring careful attention and further treatment.
- Scope boundary: Future work is needed on risk-analysis methodologies, accident-risk reduction, and fundamental failure mechanisms for ML/AI systems.The paper points to an introduction to this topic and proposes examining it more carefully in future work.