A compressor fault probability prediction and early warning method based on OCSSA–VMD–CNN–GRU-attention and chi-square residual mapping

This study proposes a fault probability prediction and early-warning method based on the Osprey–Cauchy–Sparrow Search Algorithm (OCSSA), Variational Mode Decomposition (VMD), CNN–GRU-Attention, and chi-square residual mapping for compressor monitoring signals characterized by strong noise, non-stationarity, multiscale behavior, and limited labeled fault samples. Compressor pressure and motor current are used as input signals to characterize pressure-response dynamics and load-state variation. OCSSA–VMD is employed to adaptively decompose the raw sequences, thereby improving multiscale feature separability and suppressing noise. A one-step CNN-GRU-Attention predictor is then constructed, in which CNN extracts local time-varying patterns, GRU captures temporal dependencies, and temporal attention emphasizes the historical moments that contribute most to the current prediction. Based on prediction residuals in the healthy stage, a statistical baseline is established. The current deviation is quantified by the squared standardized residual and mapped to the instantaneous fault probability through the cumulative distribution function of a chi-square distribution with one degree of freedom. A moving average is further applied to obtain a smoothed fault probability curve, and online warning is achieved by combining a dynamic threshold with a consecutive exceedance criterion. In the final validation run, the proposed method achieves RMSE, MAE, R², and MAPE values of 0.4746, 0.2147, 0.9879, and 6.54%, respectively. All four reported fault events in the April–May and June–August validation periods are successfully detected at the event level, with an average warning lead time of 21.25 h.