SERVICE LIFE PREDICTION METHOD FOR IN-SERVICE BEARINGS BASED ON THE IMPROVED CNN-LSTM MODEL

ObjectiveAiming at the problems of complex parameter adjustment and limited prediction accuracy in traditional convolutional neural network-long short-term memory (CNN-LSTM) models, an improved remaining useful life prediction method was proposed to enhance the accuracy and stability of life prediction for in-service rolling bearings.MethodsFirstly, the golden sine strategy was integrated into the golden sparrow search algorithm (GSSA) to improve its global and local search capabilities, enabling adaptive optimization of key parameters in the CNN-LSTM model. Secondly, a feature screening system based on correlation, monotonicity, and robustness was constructed to select highly sensitive degradation features. Finally, using the PHM2012 bearing dataset, a GSSA-CNN-LSTM prediction model was established, and its effectiveness was validated through comparisons with back propagation (BP) neural network and CNN-LSTM model.ResultsThe results showed that the proposed GSSA-CNN-LSTM model reduced the root mean square error, mean absolute error, and mean square error by 67.61%, 83.71%, 80.89% and 61.18%, 78.78%, 51.02%, respectively, compared with the BP neural network and CNN-LSTM models, while the determination coefficient was closer to 1, demonstrating significant improvements in prediction accuracy and robustness.