Health indicator adaptive construction method of rotating machinery under variable working conditions based on spatiotemporal fusion autoencoder

Health indicators (HI) can effectively reveal potential faults and express the degradation process of rotating machinery in engineering, which are significant for health state assessment, prognostic and decision-making. Nevertheless, most classical HI construction methods have some problems of inadequate spatiotemporal feature extraction, neglect of working conditions and individual discrepancy, and difficulty in adapting to complex degradation, leading to poor model feature expression and adaptability. To overcome these challenges, this paper proposes a new HI adaptive construction method of rotating machinery (HCPTSCAE). A spatiotemporal fusion autoencoder neural network integrating pyramid convolution and Transformer is proposed to extract the signal’s deep spatiotemporal degradation features. Then, condition domain alignment and individual degradation alignment are introduced as homogeneity constraints to reduce the discrepancy of conditions and individuals. On this basis, an autoencoder structure with adaptive weight is used to adjust the model and automatically construct HI based on the quadratic function degradation rule. The effectiveness and applicability of the HCPTSCAE network are validated by the Xi’an Jiaotong University (XJTU) bearing degradation dataset and our lab’s reducer dataset. The mean comprehensive score for different bearings is 0.7283, showing an average increase of 0.2026 compared with other methods. The mean comprehensive score for different reducers is 0.6680, with an average increase of 0.1664 compared with other methods. Moreover, the results indicate that HCPTSCAE has advantages in finding the early state degradation point and predicting remaining useful life, which promotes the trend consistency of the samples’ same features.