Research On Remaining Useful Life Prediction Of Rolling Bearing Based On Multi-Feature Fusion

Rolling bearings are essential components of rotating machinery and their operating conditions is very important to the equipment’s performance and reliability.If the fault occurs,it will affect the normal operation of the equipment and cause heavy losses.Therefore,predicting the Remaining Useful Life of rolling bearings is of utmost importance.The thesis focuses on vibration signals collected throughout the entire life cycle of rolling bearings,and investigates the process of feature extraction,screening,and fusion during the degradation process of these bearings.Based on this,health indicators are constructed to conduct in-depth research on predicting the remaining life of rolling bearings.The main objectives of this study include:(1)Aiming at the problem that it is difficult to fully depict the bearing degradation process by integrating single level features,considering the time sequence characteristics of health indicators,a residual life prediction model based on N-Beats and multi-feature fusion was proposed.Firstly,the low-level features are extracted from the time domain and frequency domain,and the feature comprehensive evaluation value consisting of time correlation,predictability and robustness is used to screen the high quality features.Then,the CNN feature extractor is established to excavate the deep features of the original signal,and the feature set of two levels is combined with PCA to construct HI.Finally,three trend stacks and two general stacks were used to construct the N-BEATS model,which made use of the characteristics of the model embedded sequential decomposition architecture to make the predicted results interpretable.The experimental results show that the correlation,predictability and robustness of the health factors constructed by combining the high and low levels of characteristics are better than that of the single level of health factors.The remaining lifetime model based on N-Beats and multi-feature fusion has higher accuracy than other temporal prediction networks.(2)Aiming at the problem that the remaining life prediction model of rolling bearings is easy to overfit under the condition of small samples and the model parameter training needs to be reset in the face of new tasks,the MAML-N-Beats model was proposed.This model introduces MAML strategy on the basis of N-Beats and multi-feature fusion remaining life prediction model,optimizes the network through multi-task training,uses historical data to learn meta-knowledge,adaptively adjusts initialization parameters,and can quickly apply new tasks,thus alleviating the overfitting problem under the condition of small samples.The experimental results show that compared with other methods,the proposed model can fully extract the features of small samples and learn the degradation law of rolling bearings,and the prediction accuracy and prediction speed are improved.(3)Aiming at the issue of the singularity of feature extractors and input signals in the prediction model for rolling bearing remaining useful life,a prediction model based on Gramian Angular Field(GAF)and Dual-channel ConvNeXt is designed.Firstly,the one-dimensional signal is encoded into a two-dimensional image using GAF,and then the short-term features are extracted using the standard ConvNeXt network.Then,the ConvNeXt network is improved by replacing the two-dimensional convolution with one-dimensional convolution and using the high-quality feature parameter set selected by feature screening as input to extract deep features.Finally,the two networks are fused to form a dual-channel network to achieve deep mining of multidimensional time-series data.Using leave-one-out cross-validation on public datasets,the experimental results show that the proposed method has improved prediction accuracy compared to the single-input ConvNeXt model.