Cross-Domain Fault Diagnosis via Meta-Learning-Based Domain Generalization

In recent years, Intelligent Fault Diagnosis (IFD) technology, as a promising method, has been a hot research topic in the field of condition monitoring and diagnosis systems, which is the focus of ensuring industrial production safety and productivity. However, the utilities of many existing IFD methods are limited by the poor-quality monitoring data, most of which are unlabeled, non-stationary, and collected from various working conditions. In addition, the unavailability of the testing data in the IFD model training phase makes the problem more challenging but more practical. In the paper, a simple-structured one-dimensional convolutional neural network(1-D CNN) with a feature extractor, a classifier, and a meta-optimizer is constructed to tackle the tricky cross-domain issues. A scalable meta-learning-based domain generalization strategy is proposed to reduce the gap among the multi-source domains. As a result, the network can learn common fault knowledge from multiple related but different source domains and then be used to analyze new target domains. Two case studies verify the effectiveness, real-time performance, and application prospects of the proposed training strategy in cross-domain fault diagnosis tasks.