Stacked supervised auto-encoder with graph regularization for feature extraction and fault classification in chemical processes

As an unsupervised deep learning method, the stacked auto-encoder (SAE) has been incorporated into many recent successful applications due to its potential to extract deep features from input data. However, the SAE suffers from bad performance in feature extraction and fault classification tasks as it cannot accurately extract task-related features. To solve this problem, a supervised SAE with graph regularization is proposed in this paper. A different supervised learning approach is adopted, in which feature extraction and classification are performed based on graph representation and constructed from raw data and label information. An improved Laplacian regularization term then causes the sample points with the same label to congregate into a cluster and sample points with different labels to repel each other. A fully connected graph with sparsification is then developed to avoid over-fitting and an improved greedy layer-wise pre-training method is employed to avoid the network falling into the local optimum. As an industrial simulation process, the TE process is used to test the performance of the proposed method. Experimental results show that our proposed algorithm is not only feasible, but displays improved results in comparison with other feature extraction and classification algorithms for chemical process.