A local–global transformer for distributed monitoring of multi-unit nonlinear processes

Abstract

Nonlinear modeling of modern industrial processes with multi-unit, large-scale characteristics is very challenging. Centralized modeling involving all process variables at a time may neglect local behaviors. And most local–global modeling methods tend to ignore the correlation between units. To preserve the intra-unit information and inter-unit correlation, this paper proposes a local–global transformer (LGT) for distributed process monitoring. First, the local representation of each unit is extracted based on feedforward neural networks (FNN). Considering that the units have a fixed order in the process, the designed orthogonal positional encoding (OPE) is added to the local representation to obtain the token of each unit, which also enhances the local behaviors. Then the attention mechanism in the transformer can adaptively adjust the attention to different units and learn the inter-unit correlation from the tokens to extract global features. Finally, the distributed monitoring framework and the variable contribution rate are combined to achieve fault detection and location. The proposed LGT demonstrates the feasibility through a numerical simulation. Extensive experimental results on Tennessee Eastman (TE) process and three-phase flow (TPF) process show the superiority of LGT. The source code of LGT can be found in https://github.com/YiQian-137/Local--global-transformer.