Causal inference of multivariate time series in complex industrial systems

Abstract

In complex industrial systems, causal inference plays a crucial role in improving production and tracing faults. The causal inference of industrial systems encompasses two main steps. First, it aims to discover causal relationships between variables. Second, it also attempts to determine time delays between variables that are causally linked, caused by factors such as industrial responses and pipeline transport. However, existing causal inference methods have certain limitations when applied to complex process industrial systems with highly dynamic and noisy environments. Therefore we propose the Causal-Transformer, a causal model based on the deep structure of the transformer. In our model, a multi-head causal attention mechanism is designed to discovery causal relationships, remove indirect coupling effects in causality, and eliminate the effects of confounding factors in causality through a causal verification step. To the best of our knowledge, we propose for the first time a novel delay discovery method to address the dynamic delay problem caused by the environment. In summary, our approach addresses the problem of causal inference in complex industrial systems. We conducted extensive numerical experiments to verify the effectiveness of our model. Moreover, we applied the model to the actual polymerization process of polyester fiber, further validating its performance in a real industrial system.