Remaining Useful Life Prediction For Rolling Bearing Based On Parallel Feature Extraction And Deep Learning

As the core component of rotating machinery,rolling bearing plays an indispensable role in industrial manufacture.However,under the influence of complex environment and intensive operation,rolling bearing is prone to failure.Therefore,remaining useful life prediction of rolling bearing is helpful to the maintenance of equipment and improves reliability and safety,which has great realistic value and research significance.In intelligent manufacturing,abundant monitoring means are applied to inspect the health condition of equipment and obtain massive condition monitoring data.With powerful data processing ability,deep learning can effectively extract hidden information from data.Therefore,this thesis takes rolling bearing as the subject of research and studies the remaining useful life prediction method based on deep learning.According to the characteristics of bearing vibration signal,existing data preprocessing method is optimized,and a deep learning model with parallel structure is constructed to predict remaining useful life of bearing more accurately.The specific research contents and innovations are summarized as follows:(1)Due to the fact that bearing vibration signal contains a large amount of information unrelated to degradation process,this thesis identifies the key frequency band through frequency domain analysis,which eliminates the influence of irrelevant information and reduces data volume.Aiming at the problem that existing methods are difficult to identify the initial weak fault quickly and accurately,a new method for first prediction time based on frequency energy ratio is proposed.To retain the global variation trend of time-frequency features in the normalization process,a synthetic time-frequency graph is constructed,which makes features reflect the degradation process of bearing better.The synthetic time-frequency graph and the normal time-frequency graph are used as the inputs of the life prediction model respectively,and the comparison of the prediction results shows its advantage.(2)To improve the feature extraction ability of the model,this thesis builds a life prediction model based on multi-scale convolutional neural network and multi-head attention mechanism.More comprehensive features are extracted by multi-scale convolutional neural network from the synthetic time-frequency graph and allocated attention by multi-head attention mechanism through their correlation and importance,which represent more effective feature information.The effectiveness of the model is verified by comparative experiments.(3)Directly extracting time information from feature information leads to large network complexity.A prediction model with parallel structure is constructed in this thesis.The model can separately extract time information of the bearing degradation process and combine it with feature information for life prediction.Through comparative experiments,this parallel structure can utilize time information and feature information and reduce the interference of the previous deviation to the current prediction,which improves the accuracy of life prediction.