Research On Rolling Bearings Fault Detection Method Based On Deep Learning

Rolling bearings are core components in the industrial field,widely used in mechanical production and automation industries.Maintenance and inspection of these core components are major expenses in the industrial field,requiring professional personnel to conduct regular inspections,resulting in waste of time and manpower.The detection of the remaining service life of bearings requires comprehensive judgment based on the current health status and load conditions of the equipment,while the purpose of fault detection is to give corresponding health status based on the vibration signal during equipment operation.Therefore,it is crucial to achieve intelligent rolling bearing fault detection and effective Remaining Useful Life(RUL)detection.With the development of the industrial Internet and the advancement of artificial intelligence technology,a large amount of research and work has applied convolutional neural networks to rolling bearing fault detection and RUL detection,in order to improve the efficiency of rolling bearing fault detection and the accuracy of RUL detection,and reduce the manual intervention and impact on the system.However,most of the existing work analyzes and processes a single input signal,and the extracted features are obtained from unidirectional acquisition.These methods still have problems with insufficient and incomplete use of signals collected from different positions or different types of signals,and do not consider the correlation features between different position information.In terms of rolling bearing fault detection,most high-precision detection models still have the problem of high complexity and large model parameters,while data-driven methods for solving the problem of effective RUL detection of bearings are not suitable for processing excessively long time series,and using comprehensive networks also has the problem of too many module parameters.To address these issues,this thesis proposes two convolutional neural network models based on multi-head attention mechanism.The Multi-headed Attention First layer Wide Convolutional Neural Network(MAFWCNN)model based on the multi-head attention mechanism is proposed to address the problem of insufficient use of correlation information in fault detection methods.It uses the multi-head attention mechanism to process multiple sources of information for autonomous weight allocation,analyzes the correlation characteristics between multiple sources of data and fault detection results from multiple perspectives,and then uses wide kernel convolution to process fusion information.Experimental results on two public datasets show that the MAFWCNN model can accurately identify the fault level and position of rolling bearings through vibration signals,with high detection accuracy and reliability.The Multi-headed Attention First layer Wide Temporal Convolutional Neural Network(MAFWTCN)model based on the multi-head attention mechanism is proposed to address the problem that data-driven methods are not good at processing excessively long time series.It uses the dilated causal convolution module to process wide temporal acquisition signals,fully utilizes the information collected from different positions through the multi-head attention mechanism,amplifies the correlation features with the full life of the bearing,and efficiently utilizes computing resources.Experimental results show that the MAFWTCN model can fit the full life process of the bearing,detect the current operating state of the bearing based on the input signal,and output the effective remaining life ratio of the bearing.