Decoupled Feature-Temporal CNN: Explaining Deep Learning-Based Machine Health Monitoring

Decoupled Feature-Temporal CNN: Explaining Deep Learning-Based Machine Health Monitoring

IEEE Transactions on Instrumentation and Measurement

10.1109/TIM.2021.3084310

https://doi.org/10.1109/tim.2021.3084310

Chaoyi Zhu, Zhenghua Chen, Rui Zhao, Jinjiang Wang, Ruqiang Yan

Electrical and Electronic Engineering, Instrumentation

27 May 2021

Zhu, C., Chen, Z., Zhao, R., Wang, J., & Yan, R. (2021). Decoupled Feature-Temporal CNN: Explaining Deep Learning-Based Machine Health Monitoring. IEEE Transactions on Instrumentation and Measurement, 70, 1–13. doi:10.1109/tim.2021.3084310

Machine learning, especially deep learning, has been extensively applied and studied in the area of machine health monitoring. For machine health monitoring systems (MHMS), the major efforts have been put in designing and deploying more and more complex machine learning models. Those black-box models are non-transparent towards their working mechanism. However, this research trend brings huge potential risk in real life. Since machine health monitoring itself belongs to high stake decision applications, the outputs of the autonomous monitoring systems should be trustworthy and reliable, which refers to obtain explainability. Then, it comes to the following key question: \textit{Why the deployed MHMS predicts what they predict}. In this paper, we shed some lights on this meaningful research direction: explainable machine health monitoring systems (EMHMS). In EMHMS, the machine doctor could act like a real doctor who can not only make diagnosis but also describe the patient's symptoms. First, we propose a specific convolutional neural network (CNN) structure, named as \textbf{D}ecoupl\textbf{E}d \textbf{F}eature-\textbf{T}emporal CNN (DEFT-CNN), to balance precision-explainability trade-off. Specifically, feature and temporal information have been encoded in different stages of our model. The spatial attention module is added to boost the performance of the model. Then, to explain the decision of the model, we adopt the gradient-based methods to generate features and temporal saliency maps highlighting which kind of features and time steps are key for the model's predictions. At last, we conduct the experimental studies on two real datasets to verify the effectiveness of our proposed framework.

Publisher Copyright

This research / project is supported by the Agency for Science, Technology and Research - AME Programmatic Funds - Learning with Less Data
Grant Reference no. : A20H6b0151

This research / project is supported by the Agency for Science, Technology and Research - Career Development Award - Contrastive Learning for Time Series Domain Adaptation
Grant Reference no. : C210112046

© 2021 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.

https://oar.a-star.edu.sg/communities-collections/articles/17910

0018-9456
1557-9662

Institute for Infocomm Research