Principal Polynomial Analysis for Fault Detection and Diagnosis of Industrial Processes

Abstract

Real-time process monitoring is crucial to improve the productivity, process safety, and product quality. In this paper, a novel fault detection and diagnosis technique based on a principal polynomial analysis (PPA) is proposed. PPA is a nonlinear modeling technique, which describes the data using a set of flexible principal polynomial components. Compared with the PCA-based methods, PPA is more effective in capturing the intrinsic nonlinear geometry structure of the process data. Moreover, compared with other nonlinear methods, such as kernel-based and neural-networks-based methods, PPA has the appealing features of straightforward out-of-sample extension, volume-preservation, and invertibility. In addition, two new types of fault detection and diagnosis statistics are derived. The effectiveness of the proposed PPA-based monitoring method was verified through its applications to a nonlinear numerical example and an industrial benchmark process. The application results have demonstrated that the proposed method has superior fault detection and diagnosis performance than the conventional PCA-based and kernel PCA-based methods.

Keywords

• Fault detection and diagnosis
• nonlinear processes
• process monitoring
• principal polynomial analysis