Bearing intelligent fault diagnosis

Bearing vibration signal is a kind of time series data, and its time dimension characteristic plays a key role in classification. Using convolutional neural network (CNN) alone to diagnose bearing fault will cause the loss of time dimension information. This results in the decline of diagnosis accuracy. To solve the above problems, a bearing fault diagnosis model combining one-dimensional CNN, bidirectional gated recurrent unit (Bi GRU) and attention mechanism is proposed. Firstly, CNN is used to adaptively extract the local space characteristic of one-dimensional vibration signals. Secondly, the characteristic information is taken as the input of the Bi GRU. Bi GRU is used to perform time dimension fusion on the extracted characteristic information. The attention mechanism is introduced to weigh the characteristic information of a plurality of moments so as to extract a more critical fault characteristic. Finally, the fault characteristic is input into a full connection layer to obtain a classification result, so as to realize intelligent fault diagnosis of the bearing. The experimental result shows the following points. ① On the confusion matrix of the test set, the classification of the bear running state is basically correct. Only some mark types are not completely classified correctly. But the recall rate is more than 95%, and the total fault recognition accuracy rate is 99.3%. ② The t-SNE technology is used to visualize the data after dimensionality reduction processing. The data of each running state of the bearing are well gathered in their own space. Only a small amount of data are mixed into other areas, which shows that the model has strong characteristic extraction capability. ③ Under the condition of constant load, the average accuracy of fault diagnosis of this model is 0.8%, 0.6% and 0.3% higher than that of one-dimensional CNN, Bi GRU and attention CNN models respectively. ④ Under the condition of variable load, this model has better stability than SVM, one-dimensional CNN, Bi GRU, attention CNN and other models. When the load is 2.25 kW, the accuracy rate is more than 85%. The model has the capability to extract one-dimensional CNN local characteristics and the capability to model Bi GRU time-dependent information. The model can further fuse time dimension information among the characteristics after acquiring the bear signal local complex characteristics. And the attention mechanism can further pay attention to the characteristics more relevant to faults. Therefore, the model has better precision.