Process monitoring using a distance-based adaptive resonance theory

Existing forms of adaptive resonance theory, e.g., ART2 and Fuzzy ART, employ similarity-based vigilance measures and contrast enhancement that is analog in nature. They use the "fast" or "fast-commit-slow-recode" learning rules, which do not guarantee convergence of clustering results. Therefore, they are not suitable for process sensor pattern monitoring which required geometrically based classifications. A modified version of the adaptive resonance theory, DART, was developed. DART uses a distance-based vigilance measure, a contrast enhancement procedure that is around the center of the prototype instead of around the null input, and the Kohonen learning rule to ensure convergence when accepting inputs that are highly correlated dynamically. The necessities of such modifications were demonstrated using a simple mathematical example: the Leonard-Kramer problem. The ability of DART to isolate different faults from operation history and to monitor operation in an adaptive manner for a complex plant is demonstrated using the Tennessee-Eastman problem. Although the process exhibits highly nonlinear dynamic behavior, DART is able to obtain classifications that are geometrically based in the sensor pattern space and are closely associated with various fault origins. Given such classifications, the nonlinear nature of the movements of this complex process in the sensor pattern space can be easily visualized. Therefore, dynamic operation can be closely monitored, and prewarning for imminent shutdown can also be provided.