Prognostics and Systems Health Management: A Review

Enrico Zio’s 2022 paper in Reliability Engineering and System Safety, “Prognostics and Health Management (PHM): Where are we and where do we (need to) go in theory and practice”, is not a general survey. It focuses on the main obstacles to the full practical deployment of PHM in condition-based and predictive maintenance, and on the methods that may overcome them. The topic matters for the digital transition of industry and Industry 4.0, where digital, physical and human elements interact as socio-cyber-physical systems, and for any sector, including healthcare technology, that relies on safe and reliable equipment.

What PHM does

PHM is a computation-based paradigm that uses physical knowledge, information and data from structures, systems and components to detect anomalies in equipment and processes, diagnose degradation states and faults, and predict how degradation will evolve towards failure by estimating remaining useful life (RUL). These outputs support just-in-time maintenance that avoids accidents, maximises availability and minimises costs.

Challenges to practical deployment

Physics of the problem. Degradation processes are often complex, poorly known, dynamic and highly non-linear.

Data availability. Real data contain missing or erroneous values, labelled degradation data are scarce, big data are difficult to manage, and operating and environmental conditions change over time.

Requirements of the solutions. Models must be accurate and precise, but also transparent, explainable and interpretable, especially in safety-critical and regulated settings. The supporting technology network must also be secure in terms of availability, data integrity, confidentiality and authentication.

Uncertainty. Incomplete knowledge, randomness in future use, model inaccuracy and measurement error all affect PHM results, so robust uncertainty quantification is essential.

Promising methods

Data-driven methods are increasingly preferred to model-based approaches. Zio highlights several directions:

Generative adversarial networks for fault detection when fault data are missing, by learning the distribution of normal-condition data and scoring anomalies with auto-encoders and reconstruction error.

Optimal transport theory for fault detection directly on raw signals, without reconstruction or feature extraction and without distributional assumptions, for example by using the Wasserstein distance between healthy and monitored data.

Reservoir computing and conceptors for fault diagnosis when degradation labels are missing and long-term dependencies matter: conceptors extracted from run-to-failure data are clustered into degradation levels, which then train a convolutional neural network for real-time diagnosis.

Echo-state networks for prognostics, predicting RUL directly from incomplete, irregular, event-based measurements thanks to their memory of input history.

Interpretability methods, including LIME for individual predictions, Shapley values for feature importance, sensitivity analysis, integration of prior physical knowledge, and visualisation tools such as partial dependence, individual conditional expectation and accumulated local effect plots.

Outlook

PHM promises better industrial performance, reliability and safety, but substantial theoretical and practical work is still needed before these benefits can be fully realised in practice.

APA Reference:

Zio, E. (2022). Prognostics and Health Management (PHM): Where are we and where do we (need to) go in theory and practice. Reliability Engineering & System Safety, 218, 108119. https://doi.org/10.1016/j.ress.2021.108119