A deep learning based health index construction method with contrastive learning

Health index (HI) can help equipment maintenance personnel better understand the health status of equipment. However, how to construct a HI generation model with robust predictive performance and strong interference-resistant ability is still a pressing problem to be solved. This paper presents a new HI construction method that combines unsupervised learning with contrastive learning. In the proposed method, a multi-granularity contrastive learning module is designed to extract in-depth feature from the data. This module operates at both the instance and subsequence levels, ensuring comprehensive feature extraction, and its introduction enhances the interference resistance of the HI generation model. Furthermore, this approach exclusively utilizes the monotonicity of the HI to design the target loss function, ensuring that the model maintains excellent predictive performance across various scenarios. To address the issues that may arise when constructing an unsupervised HI generation model solely based on monotonicity, such as unclear trends and periodic monotonicity, this paper innovatively introduces a localization loss function to tackle these problems. The effectiveness of the proposed HI generation method are evaluated by assessing the performance of the generated HI in remaining useful life (RUL) prediction. The experimental results indicate that this method exhibits robust predictive performance across various scenarios.