Research On Fault Diagnosis Method Of Rolling Bearing Based On One-dimensional Neural Network

In the modern industrial system,fault detection and diagnosis of mechanical equipment plays an important role in maintaining normal production and the order of life,and the bearing is one of the most easily damaged parts of mechanical equipment.Therefore,this paper concentrates on rolling bearings.With the progress of computer science and technology,convolutional neural network,which can process a large amount of data,is widely used in various fields.Therefore,this paper proposes a fault diagnosis method based on one-dimensional convolutional neural network,which can make good use of historical data to automatically learn fault features and perform the task of fault classification.The main contents of the article are as follows:(1)Through experiments on different channel structures,a lightweight onedimensional convolution neural network is proposed.The network has 8 layers,including three convolution layers,three pooling layers and two fully connection layers,which can obtain good diagnosis results for inner race,outer race and ball faults.Aiming at the problem of too many parameters in the fully connection layer,the global average pooling layer is used to replace the fully connection layer to optimize the model,which reduces the number of parameters by more than half and achieves higher diagnostic accuracy.(2)Since there may be noise in the real environment,Gaussian additive white noise is added in test sets to simulate the noise in the real environment,and the model is improved for the case with noise.Using ELU activation function and He initialization method,adding dropout in the input layer to improve the noise robustness of the model,and further improving by using an appropriate number of batch size and gradient decay learning rate.The test shows that the improved model has better diagnosis result.When the signal-to-noise ratio is greater than or equal to-4d B,more than 94% diagnostic accuracy can be achieved.(3)In the real environment,the load may change during the operation of the motor,and the data distribution will change too.It is difficult to diagnose the data with different distribution by using convolutional neural network.In view of this problem,the model is improved.The traditional methods to improve network performance are tested,including increasing network width and depth.They both improve the adaptive performance of the network to a certain extent and increasing network depth is better than increasing network width.The attention mechanism is introduced into onedimensional convolutional neural network,and the channel attention mechanism and channel spatial attention mechanism are tested.Among all the test methods,the channel spatial attention mechanism has best diagnostic effect and greatly improves the adaptive ability of the model.In order to better verify the generalization ability of the model,the data of Western Reserve University is used as the source domain data and the data of Paderborn university is used as the target domain data for model-based transfer learning.In experiment,the migrated model can achieve more than 91% diagnostic accuracy when the signal-to-noise ratio is-6d B,and more than 98% diagnostic accuracy when the signal-to-noise ratio is greater than or equal to-4dB.