Condition-Driven Soft Transition Modeling and Monitoring Strategy for Complex Nonstationary Process

Abstract

Abstract Due to the frequent switching of operation condition, the real industrial processes show typical nonstationary characteristics. In some cases, the switching is frequent and may not be instantaneous, revealing typical transition characteristics different from steady operation. In this work, a condition-driven soft transition modeling and monitoring method is proposed to deal with this problem. The condition modes are obtained by an automatic sequential condition-mode division algorithm, then a fine-grained mode recognition strategy is developed to further separate the condition mode into steady and transition submodes. The steady submode model is established by designing a conditional autoencoder network which can more closely describe each steady submode and facilitate evaluation of the relationship between transition submode and each steady submode. Finally, an online monitoring strategy is designed which can capture the nonstationary process changes. A real industrial case illustrates the effectiveness and superiority of the proposed method, which establishes a more accurate model for nonstationary processes by revealing the transition feature.