2023 Editorial: Special issue on Physics-informed Machine learning Enabling Fault Feature Extraction and Robust Failure Prognosis
본문
- Journal
- Mechanical Systems and Signal Processing
- Date
- 2023-06
- Citation Index
- SCIE (IF: 7.9, Rank: 2.5%)
- Vol./ Page
- Vol. 192, pp. 110219
- Year
- 2023
- Link
- http://doi.org/10.1016/j.ymssp.2023.110219 105회 연결
The use of machine learning (ML) techniques has resulted in fairly high accuracy in diagnosing faults and predicting failures for a wide range of engineered systems. Nonetheless, production-scale adoption of ML-based predictive maintenance has in part been impeded by challenges resulting from poor generalization of purely data-driven ML models. While these models can work well on often small quantities of fault/failure scenarios that a training dataset has captured, they may fail to perform on so-called out-of-distribution cases that are not well-represented in the training dataset but are encountered in real-world system operations. Such behavior makes ML models unusable in high-value systems where bad predictions have serious consequences on costs or safety. Moreover, data-driven ML models may fail to provide any physical interpretation of identified fault features or final predictions. Only limited research has been carried out so far to adapt ML models to out-of-distribution cases and to offer physical interpretability. More fundamental and applied research is needed to address these problems. Advances in the understanding of how to incorporate physical knowledge into data-driven ML models have the promise of making these models generalize better to unseen, out-of-distribution cases while at the same time explaining underlying physics.
This special issue consists of 17 papers creating new theoretical foundations and models/algorithms of physics-informed ML for fault diagnosis and failure prognosis. According to the applications of interest, these papers can be broadly categorized into four groups: (1) physics-informed ML for machinery health monitoring, encompassing fault diagnosis and failure prognosis; (2) physics-informed ML specifically for the battery health monitoring domain, The scope encompasses (a) state-of-health (SOH) estimation, (b) degradation diagnosis, and (c) degradation prognosis; (3) physics-informed ML for prognosis as applied to various engineered systems that represent rotating components and machinery; and (4) physics-informed ML for other engineering applications. A summary of these papers is provided in 1 Physics-informed ML for machine health monitoring, 4 Physics-informed ML for other engineering applications. This editorial is concluded in Section 5, where we analyze the general themes of these 17 papers and discuss potential topics for future research surrounding physics-informed ML for diagnosis and prognosis.