A novel VAE assistant prior network for health indicator construction under incomplete full life degradation data

作者：Jiechen Sun, Funa Zhou, Chaoge Wang, Xiong Hu, Tianzhen Wang · 发表于：Measurement Science and Technology · 年份：2025 · DOI：10.1088/1361-6501/ae1857 · 被引用次数：2 · 研究领域：Machine Fault Diagnosis Techniques、Reliability and Maintenance Optimization、Advanced Battery Technologies Research

Abstract The health indicator (HI) reflects the current operational status of the equipment and affects the accuracy and reliability of the remaining useful life (RUL) prediction model. However, the majority of existing methods for constructing HI are developed based on complete full lifecycle data. In fact, it is difficult to collect complete full lifecycle data due to reasons such as differences in equipment service time and communication packet loss. Due to the lack of temporal continuity and incomplete coverage of degradation stages in fragmented data, traditional HI construction methods fail to capture consistent degradation trends and become ineffective under incomplete data conditions. We proposed a novel variational auto-encoders assistant prior (AP-VAE) network to enhance latent degradation representation in the presence of temporally fragmented and incomplete lifecycle data. Unlike traditional VAE models, instead of pre-assuming a fixed prior distribution, the proposed method introduces an assistant neural network designed to learn the optimal prior distribution of the latent space from the data itself. The loss function of the proposed model is elaborately designed by integrating reconstruction error, Kullback–Leibler divergence between the posterior and adaptive prior distributions, and an improved maximum mean discrepancy regularization to ensure the latent variables capture consistent degradation tendencies under incomplete lifecycle data. The AP-VAE network is ...