SLAM: Self-supervised learning for predictive maintenance
- Reference number
- ID20-0019
- Project leader
- Xiong, Ning
- Start and end dates
- 210301-260301
- Amount granted
- 2 500 000 SEK
- Administrative organization
- Mälardalen University, Västerås
- Research area
- Information, Communication and Systems Technology
Summary
Machine learning has been widely used in predictive maintenance to learn to predict potential failures of machinery equipment or systems using previous data records. Currently various supervised learning techniques are being exploited in this area. However, they all require labelled training data, which are highly expensive to acquire. Moreover, the batch-mode of supervised learning does not account for dynamic properties and therefore cannot adapt to drifting conditions of the equipment or systems of interest. This project will develop self-supervised and continual learning methods to promote wider accessibility to data-driven predictive maintenance in power networks. First, we will attempt to circumvent the costly work of data annotation by investigating new self-supervised learning schemes that use attributes derived from input data as supervisory signals. Second, we will manage the learning to precede in a continual manner in terms of data streams (current and voltage signals) that are continuously generated from power networks. The feature of continual (and life-long) learning is of high merit to support more informed and accurate maintenance decisions by handling evolving conditions of power networks such as aging effects of electrical components. Case studies with data collected from power stations will be performed to evaluate the efficacy of the proposed method.
Popular science description
Predictive maintenance is increasingly important for industrial companies to achieve cost effective and highly reliable performance. Machine learning has been widely used in predictive maintenance to build a prediction model for fault detection or prognosis. But the current methods of learning in this area have two limitations: (1) Labels of large amount of data need to be annotated prior to learning, which is very costly; (2) Learning is usually performed offline and hence the acquired models cannot adapt to variations and drifting conditions. SLAM aims to overcome the above limitations. It will develop a self-supervised and continual learning framework in support of predictive maintenance in power network systems. Self-supervised learning will be performed in the framework that rely on input data as supervisory signals such that data annotation is no longer required. Learning will also be implemented in a continual manner to enable life-long knowledge augmentation as well as to keep track of system evolution. The framework developed in SLAM will consist of three learning modules: (1) online learning of normal system behaviour for anomaly detection; (2) joint deep learning and clustering for aging trend identification; (3) continual learning of the classification model for deterioration state prediction. Combining the results from the three learning modules will offer an integrated approach for more reliable assessment of the deterioration status of power networks. Case studies with sensor data collected from power stations will be carried out to evaluate the feasibility and performance of the proposed methods. .