A Spatiotemporal Fusion Autoencoder-Based Health Indicator Automatic Construction Method for Rotating Machinery Considering Vibration Signal Expression

Rotating machinery is widely applied in various industries, and its health indicator (HI) construction is significant in the data-driven status assessment and remaining useful life (RUL) prediction; however, most existing HI construction methods adopt manual features and simple fusion models, which are hard to detect early fault points and quantify degradation trends due to insufficient feature completeness and poor nonlinear characterization. To overcome the mentioned issues, this article proposes a novel integrated HI automatic construction method by coupling multimode samples of vibration signals. To construct the unsupervised HI automatically, a deep spatiotemporal fusion autoencoder network (MSCLACAE) is developed by integrating multiscale convolution (MSCNN), convolutional long short-term memory network (ConvLSTM), and attention mechanism (AM). On this basis, a quadratic function-based shape constraint is introduced to improve the performance of HI constructed by the MSCLACAE network. The effectiveness of the proposed method is verified by the standard bearing dataset from Xi’an Jiaotong University, the average comprehensive score under different bearings is 0.7327, which is 0.1835 higher than other methods on average; moreover, the proposed method is also tested by the reducer platform, and the comprehensive score is 0.9144, which is increased by 0.2712 averagely compared with different methods; furthermore, the experimental results verify that MSCLACAE not only can find early degradation points or state degradation points earlier but can also predict the RUL with lower error.