(719h) Statistics Pattern Analysis Based k-Nearest Neighbors (SPA-kNN) Fault Detection for Pressure Swing Adsorption Processes

Over the past few decades, process monitoring and fault diagnosis methods have played an important role in the improvement of product quality and operation safety in a variety of industries. Most of these approaches are data-driven multivariate statistical process monitoring (MSPM) methods that are developed for batch or continuous processes. In recent years, periodic operations, such as pressure swing adsorption (PSA) and simulated moving bed (SMB), have gained wider applications in industries. For example, there has been widespread development of PSA systems, with their applications expanding from traditional bulk gas separation and drying, to CO2 sequestration, trace contaminant removal, and many others. Compared to continuous processes operating at steady-state, there are several unique characteristics of PSA processes, which pose significant challenges to process monitoring: (1) it is operated in a periodic or cyclic fashion, which lead to unsteady-state, or highly dynamic process behavior; (2) One cycle consists of many different steps with different operation conditions, which lead to highly complex nonlinear behavior; (3) cycle time is frequently adjusted to meet demand fluctuations, which lead to multimodal operations of the process. As a result, applying traditional MSMP methods, such as principal component analysis (PCA), to the PSA process can lead to high false alarms and/or missed faults. To address the challenges, we previously proposed a data-driven feature space monitoring (FSM) approach based on statistics pattern analysis (SPA). In the SPA framework, we extract the statistics/features that explain process characteristics of each step from original process data. Since these features can better capture process characteristics than the raw process variables, features are used for monitoring the process instead of process variables. In addition, SPA can address the challenges from PSA process, such as unequal step and/or cycle time that requires preprocessing steps (e.g., trajectory alignment or synchronization) for the conventional MSMP methods. In this study, we propose to enhance SPA by integrating it with k nearest neighbor (kNN) rule. The proposed SPA-kNN method has three advantages for the periodic processes. First, SPA-kNN method is well suited for multimodal data set as the models are built locally. Second, SPA-kNN method makes no assumption about the linearity of the data set. When data set has strong nonlinearity, the method can perform efficient process monitoring while the linear MSMP methods such as PCA can lead to high false alarm and/or missed faults. Finally, the proposed method is simple and practical. It can be used in periodic processes as well as batch and/or continuous processes. In addition, we propose a new pipeline or workflow for fault diagnosis using SPA-kNN method. In addition, different strategies for splitting PSA cycles into steps and extracting various features from each step are studied, and guidance of how to select the features based on process characteristics is suggested. Finally, different strategies for reducing false alarms are investigated and their performances are compared.