A degradation prediction method for component based on distribution divergence awareness and multi-granularity temporal modeling

The degradation behavior of rotating components during operation exhibits complex temporal evolution. To overcome the limitations of existing prediction methods, which frequently fail to simultaneously capture the overall degradation trajectory and local fine-grained fluctuations, a multi-temporal-granularity degradation trend prediction approach is proposed. Initially, a degradation indicator is formulated by combining gamma distribution mapping with Kullback–Leibler (KL) divergence computation. Subsequently, a global trend prediction model is constructed, integrating one-dimensional convolution with a Transformer encoder to characterize the long-term degradation trend at a coarse temporal granularity. The output of this global model serves as a constraint for the local fine-grained prediction stage. A residual-based multi-scale convolutional network is then developed as the local prediction model, optimized through a joint loss function incorporating a trend-consistency constraint. This design enables accurate modeling of fine-grained fluctuations in the degradation process and facilitates high-resolution trend prediction. Experimental evaluation on the FEMTO-ST dataset verifies the effectiveness of our method.