The figure is the confusion matrix corresponding to the method in this paper, through which the identification of each type of fault in the sample can be understood.
Each row represents the real health status of the, and each column represents the predicted results of the model. It can be seen from the figure that each state contains 100 test samples. Because the signals of bearing inner ring and outer ring are similar in a single fault, the model mistakenly identifies one sample of bearing inner ring fault as bearing outer ring fault. At the same time, the effectiveness of Sanc is proved by confusion matrix: Sanc filters out the periodic signal (mainly gear signal) in the original signal, resulting in the composite fault signal similar to the single fault signal, so the model mistakenly identifies the bearing inner ring fault as gear broken and bearing inner ring fault.
By using Sanc to preprocess the signal, the gearbox vibration signal is separated into periodic signal component and random signal component, and then 1d-cnn is used to extract and identify the random signal components containing bearing fault features. The results show that after the original signal is separated into periodic signal and random signal, the strong gear vibration interference is effectively suppressed, and the recognition accuracy of 1d-cnn is improved. Compared with the traditional mechanical fault diagnosis method, it reduces the dependence on human factors. At the same time, compared with other machine learning methods such as SVM, RNN and 1d-cnn, the advantages and effectiveness of the text method are verified.