Rolling Bearing Health Indicator: From Design to Modeling and Evaluation

As a key component of industrial machinery, accurate prediction of the degradation trend of rolling bearings is crucial for equipment safety. However, traditional health indicator (HI) extraction methods often suffer from feature redundancy, and prediction models lack the ability to capture spatial dimension features, leading to significant prediction errors. To address these issues, 16 time-frequency domain features were first extracted, and a new HI was constructed by combining the Gaussian Process latent variable model (GPLVM) for non-linear feature fusion and exponentially weighted moving average (EWMA) for smoothing. Additionally, a spatial-temporal convolutional long short-term memory network (ST-CNet) was proposed, which integrates a 3-layer CLSTM, fully connected layers, and batch normalization to effectively capture local and long-term spatiotemporal dependencies. Case studies on IMS bearing datasets show that the constructed HI accurately describes the degradation process, and ST-CNet achieves superior performance with lower MAE and RMSE compared to existing methods.