This paper, titled “Potential, challenges and future directions for deep learning in prognostics and health management applications,” authored by Olga Fink, Qin Wang, Markus Svensén, Pierre Dersin, Wan-Jui Lee, and Melanie Ducoffe, provides a thorough and insightful evaluation of the current developments, drivers, challenges, potential solutions, and future research needs in the burgeoning field of Deep Learning (DL) as applied to Prognostics and Health Management (PHM) applications.
The authors begin by establishing that the overarching goal of PHM is to optimize maintenance policies for industrial assets, ensuring high availability at minimal costs. This holistic approach encompasses fault detection, fault diagnostics (identifying origin and type), and prognostics (predicting Remaining Useful Life, or RUL), extending beyond mere failure prediction to support optimal maintenance and logistics decisions by considering resources, operating context, and economic consequences.
Despite the extensive monitoring of complex industrial assets and the collection of vast amounts of heterogeneous condition monitoring signals (e.g., temperature, pressure, vibration, images, video), the application of DL for detecting, diagnosing, and predicting faults has been limited. A primary challenge lies in the need for manual feature engineering to derive useful representations from raw data, a process that is highly dependent on expert knowledge, lacks generalization, and struggles with scalability as monitored parameters increase.
The paper highlights DL’s significant potential to address these challenges by offering the ability to automatically process massive and heterogeneous condition monitoring data, extract useful features, learn complex functional and temporal relationships, and transfer knowledge across different operating conditions and units.
However, the authors also critically examine the challenges for broader acceptance of DL models in PHM. These include the need to identify and handle outliers, ensure dynamic learning that distinguishes evolving operating conditions from faults, detect novel degradation types, guarantee robustness under varying conditions, achieve generalization, and provide interpretability of results for domain experts. Interpretability, in particular, is a significant concern for users and experts due to the non-linear, distributed computations of deep neural networks.
The article draws inspiration from DL’s successes in other domains, such as speech recognition, machine translation, natural language processing (NLP), image processing, and recommendation systems. It then delves into existing DL PHM applications, detailing how NLP techniques can process maintenance reports and event logs to identify fault causes, and how computer vision methods, particularly Convolutional Neural Networks (CNNs), are used for automated image and video analysis in inspections (e.g., railway infrastructure, marine structures, jet engines). For sensor data, the paper reviews supervised learning algorithms like Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, 1-D CNNs, and combinations thereof, as well as unsupervised and semi-supervised approaches such as autoencoders for signal reconstruction and clustering for novelty detection in the absence of abundant labels. The crucial aspect of uncertainty quantification in prognostics, particularly for RUL predictions, is also discussed, emphasizing the need for confidence intervals and probability distributions, often estimated via Monte Carlo simulation.
Finally, the authors outline five promising DL directions for PHM applications:
- Transfer Learning: Aims to mitigate distribution discrepancies between different machines or operating conditions, allowing models trained on one system to be adapted to others without extensive re-labeling.
- DL for Fleet Approaches: Addresses the high variability in system configurations and operating conditions across a fleet, proposing methods from sub-fleet clustering to domain alignment in feature spaces.
- Deep Generative Algorithms (e.g., GANs, VAEs): Explores their potential for learning probabilistic distributions of data, particularly for anomaly detection and generating synthetic fault data to address data scarcity.
- Deep Reinforcement Learning (DRL): Offers solutions for dynamic and complex maintenance tasks, such as deciding on maintenance actions or planning task orders, by learning optimal strategies through trial-and-error in a dynamic environment.
- Physics-Induced Machine Learning: Focuses on integrating prior domain knowledge (e.g., physical models, constraints) into DL models to enhance interpretability, reduce data requirements, and improve performance, especially for applications outside image processing.
The paper concludes by identifying critical future research needs, including heterogeneous domain adaptation, ensuring the physical plausibility of generated samples, extending DRL to more complex real-world PHM problems, consolidating physics-induced machine learning approaches, and urgently addressing the lack of large, representative datasets through data augmentation, generation, and cross-company data sharing. This comprehensive review serves as a crucial roadmap for researchers and practitioners in advancing the application of deep learning for effective and efficient Prognostics and Health Management.
Reference:Fink, O., Wang, Q., Svensén, M., Dersin, P., Lee, W.-J., & Ducoffe, M. (2020). Potential, challenges and future directions for deep learning in prognostics and health management applications. Engineering Applications of Artificial Intelligence, 92, 103678. https://doi.org/10.1016/j.engappai.2020.103678
