Information Entropy-Based Fault Detection for Quality-Related Processes

Canonical Correlation Analysis (CCA) is currently very popular in the field of data-driven fault detection. The CCA method usually assumes that the data obeys a Gaussian distribution, but the sample data collected in the actual complex industrial production process often has non-Gaussian characteristics, and the CCA method is not very suitable for modeling this kind of data. In order to extend the applicability of the CCA method from Gaussian data space to non-Gaussian data space, this paper proposes an information entropy-based quality-related process fault detection method (ECCA) based on CCA. First, the ECCA approach takes as its objective function to maximize the mutual information of typical pairs of variables in the space of process and quality variables. The objective function is also optimized using Particle Swarm Optimization (PSO) algorithm to solve for the optimal pairs of typical variables; Second, when considering the extraction of the next pair of independent typical variable pairs, the extraction of independent typical variable pairs can be achieved by resorting to the framework of the Partial Least Squares (PLS) algorithm in combination with the minimization of the information entropy of the residuals in the space of the process and quality variables; Third, a truncation criterion for the number of typical variables is identified; Fourth, T
²
and Te
²
statistics are constructed to monitor the system; Finally, the ECCA method is applied to synthetic data and TE process data for comparison with the traditional CCA method. Experimental validation shows that the ECCA method has better monitoring performance than the traditional CCA method in monitoring quality-related process faults.