Process monitoring using principal component analysis and stacked autoencoder for linear and nonlinear coexisting industrial processes
Abstract With the relationships between industrial process variables becoming more complex, linear and nonlinear relationships coexist in most processes, both of which should be considered simultaneously to improve monitoring effect. Focusing on this issue, the paper proposes a novel principal component analysis-stacked autoencoder (PCA-SAE) model for fault detection. In this model, PCA and SAE respectively deals with linear and nonlinear components. Besides, PCA plays a role in separating the two components. As a linear mapping method, PCA is supposed to extract only linear features and leave the nonlinear part. And this is accomplished by adjusting its cumulative percent variance (CPV) of features. After that, the remaining nonlinear part is modeled by SAE. Comprehensive statistics are established to monitor the two parts of processes. The proposed method achieves 86.5% average fault detection rate in Tennessee Eastman (TE) process, higher than pure PCA, pure SAE, and many other conventional methods; and it successfully detects the fault that neither PCA nor SAE is able to achieve in a wind power generation process.
Paper
Full text
Process monitoring using principal component analysis and stacked autoencoder for linear and nonlinear coexisting industrial processes
Semantic Scholar · Engineering · 2020
Abstract
Abstract With the relationships between industrial process variables becoming more complex, linear and nonlinear relationships coexist in most processes, both of which should be considered simultaneously to improve monitoring effect. Focusing on this issue, the paper proposes a novel principal component analysis-stacked autoencoder (PCA-SAE) model for fault detection. In this model, PCA and SAE respectively deals with linear and nonlinear components. Besides, PCA plays a role in separating the two components. As a linear mapping method, PCA is supposed to extract only linear features and leave the nonlinear part. And this is accomplished by adjusting its cumulative percent variance (CPV) of features. After that, the remaining nonlinear part is modeled by SAE. Comprehensive statistics are established to monitor the two parts of processes. The proposed method achieves 86.5% average fault detection rate in Tennessee Eastman (TE) process, higher than pure PCA, pure SAE, and many other conventional methods; and it successfully detects the fault that neither PCA nor SAE is able to achieve in a wind power generation process.