Remaining useful life prediction: from statistical modeling to federated approach

• Main Author: Gan, Xin
• Other Authors: Wen Yonggang
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/156410
• Institution: Nanyang Technological University

Smart manufacturing is getting increasingly popular in different areas, including aircraft systems, uninterruptible power supply systems, hydraulic systems, etc. Meanwhile, the capability of intelligent machines doing edge computing is growing rapidly. Predictive maintenance is applied to different areas to use edge data to reduce risks of operating machines since any unplanned downtime of the machine could result in unmeasurable loss. Remaining Useful Life Prediction, determines whether or when machine fails based on current and past data representing system conditions, is an essential component of predictive maintenance. Statistical models and machine learning techniques are able to do Remaining Useful Life (RUL) Prediction. We studied both methods and present a demo to show how can we implement a Remaining Useful Life Prediction system and do risk analysis. However, during the development of the Remaining Useful Life Prediction system, we found that with the increasingly stringent privacy restrictions, the centralization of data from edge devices are limited in reality. The situation is further aggravated by components isolation. Different machine components are not able to share data with each other since they belong to different manufacturers or have component-specific sensitivity. To tackle these privacy matters, we present Federated Remaining Useful Life Prediction(FedRUL) for isolated system components, which implements federated learning, a distributed training method, to Remaining Useful Life Prediction. FedRUL uses machine learning method to do RUL prediction. FedRUL preserves data privacy by aggregating partial model parameters instead of data, from clients to a central server. To make it feasible and with good performance: (1) We add local dimension transfer layer to transfer all input to the same dimension; (2) We only aggregate the layers except local dimension transfer layer. (3) We apply momentum update on model parameter aggregation to address the importance of the local models to optimize the performance. In this work, we use intuitive methods to implement a Remaining Useful Life Prediction system, construct a basic analysis for Remaining Useful Life Prediction addressing component isolation, apply FedRUL to alleviate the performance drop introduced by component isolation.