A Data-Driven Framework for Planetary Winch Reducer Noise Prediction via Feature Selection and Residual Compensation

Accurate reducer noise prediction is essential for condition monitoring and predictive maintenance of mechanical transmission systems. However, the strong nonlinear coupling between operating conditions and noise responses, together with measurement uncertainties, remains a major challenge for data-driven prediction methods. This study proposes a hybrid prediction framework integrating feature selection, adaptive neural modeling, and residual compensation to improve the accuracy of planetary winch reducer noise prediction. A random forest (RF)-based feature selection strategy is first employed to identify the most informative vibration characteristics associated with reducer noise. Subsequently, a generalized regression neural network (GRNN) optimized by a hybrid whale optimization and bat algorithm (WOA-BAT) is developed to adaptively determine the smoothing factor and enhance nonlinear prediction capability. Furthermore, a residual Kalman compensation (RKC) mechanism is introduced to suppress prediction fluctuations caused by stochastic disturbances and modeling uncertainties. Experimental results demonstrate that the proposed WOA-BAT-GRNN-RKC framework achieves highly accurate noise prediction, with an RMSE of 0.05631 dB and an MAE of 0.027972 dB. The corresponding MAPE is 0.037897%. The proposed approach provides an effective pathway toward intelligent reducer condition monitoring and predictive maintenance.