A Fault Diagnosis Method Based on Autoencoder for Large Shafting Speed ​​Fluctuation

Abstract

The invention discloses a fault diagnosis method based on the shafting rotation speed large fluctuation of an automatic encoder. The method comprises the steps of 1, carrying out fast Fourier transformation and amplitude normalization on samples which belong to a rotation speed 1 and a rotation speed 2 in a source domain and a target domain respectively; 2, training the automatic encoder by meansof the signal of the rotation speed 1; 3, adding the MMD penalty term automatic coding migration learning algorithm to the automatic coding algorithm, and carrying out training by utilizing training samples with the two rotation speeds; 4, training a softmax characteristic classifier added with the MMD penalty term to classify extracted features. In this way, the fault diagnosis can be achieved onfault signals which are not known to the rotation speed 2. According to the invention, the intelligent diagnosis is carried out on fault signals in the large-rotating-speed fluctuation state throughthe migration learning algorithm. By introducing the MMD penalty term, the accuracy of a model after the source domain sample training is improved for the target domain sample is improved, so that intelligent fault diagnosis under the large fluctuation condition of the rotation speed is achieved.

Description

technical field [0001] The invention belongs to the technical field of vibration signal intelligent fault diagnosis, and relates to a fault diagnosis method based on an automatic encoder for large fluctuations in the rotating speed of a shaft system. Background technique [0002] With the advent of the era of big data, the fault diagnosis method based on equipment vibration signals has changed from traditional signal processing methods to deep learning methods. The steps of applying deep learning for intelligent fault diagnosis are generally to train a weight matrix capable of extracting sample features through an unsupervised learning algorithm, and then classify the sample features extracted by the weight matrix through a supervised learning algorithm. At present, the unsupervised learning algorithms commonly used mainly include sparse Boltzmann machine, autoencoder, sparse autoencoder, sparse coding, independent component analysis, etc., all of which improve the diagnosis...

AI Technical Summary

Problems solved by technology

This will lead to repeated training of the model, which will affect the effect and accuracy of the final fault diagnosis in practical applications

Images