Temporal supervised generative adversarial functional causal model for root cause diagnosis

Abstract

In this paper, a novel causal inference model, called temporal supervised generative adversarial functional causal model (TSGA-FCM), is established for root cause diagnosis of industrial processes. First, a causal generation module (CGM) for multivariate time-series data is developed to infer causal relationships through a functional causal mechanism loss. Moreover, a temporal supervised generative adversarial network is established for joint training of the CGM. The parameters of the CGM are optimized via a combination of functional causal mechanism loss, reconstruction loss, and temporal supervised loss. The capability in temporal feature extraction is enhanced by reducing temporal distribution feature differences between generated data and original data. Using both the extracted static and temporal features, a directed acyclic causality graph is derived to pinpoint the root cause. Finally, a benchmark process and a real industrial process are utilized to validate the effectiveness of the proposed TSGA-FCM. Using the temporal supervised generative adversarial network, the proposed TSGA-FCM effectively extracts temporal feature-based causal inference to avoid unnecessary symmetry assumption of the traditional autoregressive-based root cause diagnosis (RCD) methods. The proposed method makes novel contributions to data-driven causal inference and demonstrates practical application value in an important heavy-plate rolling process.