Fisher Discriminative Sparse Representation Based on DBN for Fault Diagnosis of Complex System

Abstract

Fault detection and diagnosis in the chemical industry is a challenging task due to the large number of measured variables and complex interactions among them. To solve this problem, a new fault diagnosis method named Fisher discriminative sparse representation (FDSR), based on deep belief network (DBN), is proposed in this paper. We used DBN to extract the features of all faulty and normal modes. Features extracted by the DBN were used to learn subdictionaries, then the overcomplete dictionary was constructed by cascading all subdictionaries in order, and each dictionary atom corresponded to class labels. The Fisher discrimination criterion (FDC) was applied to the dictionary learning to ensure smaller within-class scatter but greater between-class scatter. The quadratic programming method was applied to estimate the sparse coefficients simultaneously class by class. Therefore, both the reconstruction error and sparse coefficients were discriminative, so that the reconstruction error after sparse coding can be used for pattern classification. An experiment performed on the Tennessee Eastman (TE) process indicated that compared with the traditional monitoring methods, the FDSR based on DBN produced more interpretable results and achieved superior performance for feature extraction and classification in the field of complex system fault diagnosis. T-distributed stochastic neighbor embedding (t-SNE) appropriately visualized the performance of the proposed method and produced reliable fault diagnosis results.

An Intelligent Fault Diagnosis Method Using GRU Neural Network towards Sequential Data in Dynamic Processes

“…In this case, the results of the proposed method are compared with two deep learning methods: DBN and CNN. In accordance with References [18,19], the neural numbers of DBN were set to 23 20 16 9    , and the CNN consisted of a pair of convolutional layer and pooling layer with a convolution kernel size of 2. The diagnosis results of DBN are shown in Figure 17a, and the mean accuracy rate was 92.09%.…”

“…Deep learning can learn the abstract representation features of the raw data automatically, which could avoid the requirement of prior knowledge. Deep learning is a branch of machine learning algorithms that attempt to model complexity and internal correlation in a dataset by using multiple processing layers, or with complex structures, to mine the information hidden in the dataset for classification or other goals [18]. In recent years, deep learning has developed rapidly in academic and industrial fields.…”

“…In recent years, deep learning has developed rapidly in academic and industrial fields. Tang et al applied deep belief networks (DBNs) to fault feature extraction and diagnosis of the chemical industry and introduced the quadratic programming method to estimate the sparse coefficients simultaneously class by class [18]. Wen et al convert fault signals into two-dimensional (2-D) images and adopt convolutional neural networks (CNNs) to extract the features of the converted 2-D images [19].…”

Diagnosis of Intermittently Faulty Units at System Level

“…It is worth noting that diagnosis of system units at a system level considers a fault in the system as a failure of a system unit. We do not consider the details of what was wrong with the failed unit as it was done, for example, in the fault diagnosis described in [9][10][11]. In view of this, we do not classify the faults as it was done, for example, in [12,13].…”

DBN Structure Design Algorithm for Different Datasets Based on Information Entropy and Reconstruction Error

“…In 2006 and later, Hinton proposed the DBN [3] and CD-K [4] algorithms, which has enabled ANNs to develop from a shallow to deep structure, achieving significant performance improvements. As a typical type of deep network [5], DBNs are widely used in image processing [6][7][8][9][10], speech recognition [11][12][13] and nonlinear function prediction [14], yielding excellent performance. However, DBNs still have many problems worth studying, such as the network structure design [15][16][17][18][19], selection and improvement of training algorithms [20,21], introduction of automatic encoders, and implementation of GPU parallel acceleration [22,23].…”