Conditional discriminative autoencoder and condition-driven immediate representation of soft transition for monitoring complex nonstationary processes. (May 2022)

• Record Type: Journal Article
• Authors: Chen, Xu; Zhao, Chunhui
• Is Part Of: Control engineering practice. Volume 122(2022)
• Publication Date: 2022-05
• Subjects: Soft transition immediate representation -- Conditional discriminative autoencoder (CDAE) network -- Condition-driven -- Attention-based evaluator -- Nonstationary process monitoring; Automatic control -- Periodicals; 629.89
• DOI: 10.1016/j.conengprac.2022.105090
• Languages: English
• ISSNs: 0967-0661

Abstract

Abstract: Real industrial processes show typical nonstationary characteristics due to the operating condition switchings and production product changes. However, not all the switchings are instantaneous. The transition characteristics may lead to false alarms or missing alarms in process monitoring. In this work, a condition-driven soft transition representation and monitoring method is proposed to handle this problem. Based on different condition modes initially partitioned by an automatic sequential condition-mode division algorithm, a fine-grained mode recognition strategy is developed to further explore the underlying changes in each condition mode and separate the condition mode into steady and transition sub-modes. Then, different models are designed to reveal different characteristics of steady and transition sub-modes. A conditional discriminative autoencoder (CDAE) network is designed to establish the steady sub-mode model, which can closely describe each steady sub-mode and highlight the difference across steady sub-modes. The transition sub-mode evaluator is developed to immediately describe the transition characteristics by combining the different steady sub-mode features. Finally, an online monitoring strategy is designed to tell the current operation sub-mode and capture the changes in nonstationary processes. A real industrial case illustrates the effectiveness and superiority of the proposed method. Highlights: We propose transition representation concept from condition-driven perspective. A fine-grained mode recognition strategy is proposed. We propose CDAE network to describe each steady sub-mode closely. An attention-based evaluator is designed to describe transition characteristics. BID monitoring indices are designed to distinguish normal operations from faults.