A Metric Learning-based Health Indicator for Early Degradation Point Detection of Machines

The construction of a reasonable health indicator (HI) to describe the degradation process of equipment and identify the early degradation point (EDP) is the focus of machines condition monitoring. However, HIs based on manual feature extraction rely on prior knowledge. Moreover, most of the methods with good trendability are not sensitive to early fault of machines due to the presence of strong noise. To address above problems, a HI construction method based on metric learning is proposed for identifying EDPs of machines. Firstly, a triplet network stacked with residual-attention (RA) block is constructed and optimized by genetic algorithm (GA) for feature extraction. Then, triplet loss is used to promote intra-class aggregation and inter-class separation for EDP detection. After that, convex optimization (CO) is applied to assign corresponding weights to the features for optimizing the trendability of the degradation process. Finally, the weighted relative similarity between two samples is calculated as the GA-RA-CO-HI of machines. The effectiveness of the proposed method is verified on two benchmarks. The results show that the proposed GA-RA-CO-HI can effectively detect EDP while promising desirable trendability and monotonicity.