Abstract: Bearing remaining useful life (RUL) prediction based on vibration signals is of great significance in industry safety production, there are still some problems such as high difficulty in model construction and low prediction accuracy in the field. In order to realize adaptive feature mode extraction and denoising, simplify the model construction process and improve the prediction effect, a variational mode extraction (VME) algorithm based on improved pigeon-inspired optimization (IPIO) algorithm (IPIOVME) and a bearing remaining useful life prediction method based on ConvNeXt-Encoder-gate recurrent unit (GRU)were proposed. Firstly, the pigeon-inspired optimization algorithm was efficient and accurate, which was suitable for the parameter selection of VME, but it was easy to fall into local optimum. The IPIO algorithm was mainly improved by adaptive inertia weight, shrinking encircling mechanism, Levy flight and other methods to improve convergence speed and global convergence ability. Secondly, in order to realize adaptive mode extraction, the objective function of IPIO-VME algorithm was designed, which could effectively extract bearing vibration features according to the characteristics of VME algorithm and bearing vibration signals. Finally, aiming at the problem of cumbersome model construction and low accuracy, the ConvNeXt-Encoder-GRU model was proposed. The data set construction method of interval and continuous sampling was used, and the remaining useful life prediction model was constructed by the method of joint vibration data and characteristic curve. The vibration features were extracted by the ConvNeXt module, the trend features were extracted by the Encoder module of Transformer, and fused by the GRU.The algorithm and prediction model were also experimentally compared and validated. The research results show that IPIO has faster convergence speed and better global convergence ability. Under the test function, after 1 000 iterations, the accuracy can reach up to 1.23×10-9. The ConvNeXt-Encoder-GRU model has high prediction accuracy. The LogCosh index of ConvNeXt-Encoder-GRU on Xi‘’an Jiaotong University-ChangxingSumyoung Technology Co., Ltd. (XJTU-SY)bearing data set could reach 0.001 3, which is better than the single model. The research results have certain guiding significance for the fault feature extraction and remaining useful life prediction of bearings.
Key words:  rolling bearing; remaining useful life(RUL) prediction; improved pigeon-inspired optimization(IPIO); variational mode extraction(VME); ConvNeXt; gate recurrent unit(GRU)