Stacked Auto-encoder With Improved Graph Regularization And Its Application In Fault Diagnosis

With the development and progress of sensor technology,communication technology,the Internet of Things,data storage,and the widespread use of distributed control systems in the industrial sector,a large amount of data has been accumulated over time,which has driven the development of industrial big data.The emergence of neural networks and large-scale parallel computing has further intensified the fusion of deep learning and traditional industries.As an unsupervised deep learning method,stacked auto-encoder(SAE)has been successfully applied in many fields due to their potential to extract deep features from input data.However,SAEs do not perform well in feature extraction and fault classification tasks because they cannot accurately extract task-relevant features.To address this issue,this paper proposes a supervised SAE algorithm with graph regularization and validates the proposed fault feature extraction and classification method using the chemical process Tennessee Eastman(TE)simulation platform and real wind turbine data.The specific work of this paper includes:1)A stacked auto-encoder with Laplacian graph regularization is proposedThe proposed model introduces the regularization expression based on Laplacian graph embedding and combines it with deep autoencoder to develop a new model called stacked auto-encoder with Laplacian graph regularization.This approach incorporates manifold structure features into the feature extraction model,improving the effectiveness of feature extraction and enhancing the capability of the original SAE model in identifying and extracting geometric features.Experiments are conducted to compare and analyze the impacts of different graph structures and similarity matrix construction methods on feature extraction and classification results.2)A stacked auto-encoder with improved Laplacian graph regularization is proposedDelving into the study of graph regularization methods,an improved Laplacian graph regularization method combined with a deep autoencoder model based on discriminant analysis is proposed.Based on the analysis of the problems in graph regularization and deep auto-encoder,an improved layer-by-layer greedy supervised pre-training method is also proposed.In terms of graph construction,a fully connected graph with sparsification is proposed to address the overfitting problem in the combination of stacked auto-encoder and graph regularization.The experimental results of two different application scenarios,TE process and wind turbine bearing,show that the proposed algorithm is feasible and robust,and has stronger feature extraction ability and higher classification accuracy compared to other feature extraction and classification algorithms.