Abstract
Abstract
The vibration data collected by the fault experiment are analyzed in the time domain and frequency domain respectively, and the conclusion shows that the frequency characteristics of some vibration data cannot be fully characterized due to noise interference. In order to solve the problem of noise interference, this paper uses the VMD algorithm based on the kurtosis peak to denoise the collected data, constructs wavelet time-frequency map samples through the wavelet transform method to provide input sources for the subsequent convolutional neural network, and uses the image samples for fault diagnosis and identification. [1] Through the convolutional neural network built to diagnose and identify the different states of the multistage centrifugal pump, it can be seen that the recognition accuracy, training time, and recognition time of a single sample of various image samples in the three cases of no noise reduction, noise reduction and a small number of samples can be found through the method of comparative analysis, and compared with the traditional SVM method, the superiority of the convolutional neural network using image recognition can be found.