Industrial process fault monitoring method based on time information enhancement graph convolution auto-encoder

Abstract

The invention discloses an industrial process fault monitoring method based on a time information enhancement graph convolution auto-encoder, and the method designs a variable-guided time warping Gaussian kernel method to measure the correlation degree between variables, and introduces a sparse maximization function to adaptively reserve the correlation of significant variables. Constructing a cross-time spatial perception graph structure, and comprehensively representing a spatial dependency relationship between industrial process variables and a cross-time spatial dependency relationship; according to the method, a multi-layer graph convolution auto-encoder network is constructed to extract local space-time related features, and meanwhile, a convolution long-short-term memory network is connected between a coding layer and a decoding layer to capture long-term time-dependent information of a sequence, so that the space-time related feature extraction capability is enhanced, and data reconstruction is assisted. According to the method, the space-time correlation characteristics of normal data in the industrial process can be fully extracted, the reconstruction capability of the model for the normal data is improved, higher monitoring precision is realized under the unsupervised task of the industrial process with the complex space-time correlation, and the method is superior to the prior art.

Images