| CPC G05B 23/0281 (2013.01) [G05B 23/024 (2013.01)] | 9 Claims |
|
1. An unsupervised fault diagnosis method for mechanical equipment based on an adversarial flow model, comprising steps of:
(1) data preprocessing: converting a mechanical vibration signal into a frequency domain signal, and normalizing an amplitude value of the frequency domain signal into a range of [0, 1];
(2) prior distribution designing: designing a mixture of Gaussian distribution with K subdistributions, wherein K is determined by the number of mechanical equipment status;
(3) model construction: constructing an unsupervised fault diagnosis model by combining an autoencoder, a flow model, and a classifier;
(4) model training: training the unsupervised fault diagnosis model by using various classes of status data, along with the designed prior distribution, preset training steps, loss functions, and an optimization algorithm; and
(5) fault diagnosis: inputting status data of mechanical equipment into the trained unsupervised fault diagnosis model to obtain a data clustering result and a fault diagnosis result,
wherein in step (3), the autoencoder is constructed by an encoder and a decoder, low-dimensional features of inputted data are learned by using the encoder, dimensionality of the low-dimensional features is the same as the dimensionality of the designed mixture of Gaussian distribution, and the inputted data is reconstructed by inputting the low-dimensional features into the decoder; the flow model maps the low-dimensional features of the inputted data into a Gaussian distribution with the same dimensionality as the low-dimensional features, and obtains mapped features; a combination of the autoencoder and the flow model is called a feature extractor; and
the classifier classifies the mapped features and prior features separately.
|