Causal Graph Spatial-Temporal Autoencoder for Reliable and Interpretable Process Monitoring

TL;DR

CGSTAE addresses the reliability and interpretability gap in data-driven MSPM by integrating a SSAM-based correlation-graph learner with a GCLSTM-based spatial-temporal autoencoder. A reverse-causal-invariance-based three-step training pipeline learns an invariant causal graph from changing correlations and then uses it for reconstruction. Fault detection and diagnosis are performed with Hotelling’s $T^2$ in the feature space and $SPE$ in the residual space, aided by a discrete causal graph subgraph for root-cause localization. The approach is validated on the Tennessee Eastman process and a real-world air separation process, showing improved detection performance and interpretability relative to correlation-based baselines, with robustness to hyperparameters and process knowledge constraints.

Abstract

To improve the reliability and interpretability of industrial process monitoring, this article proposes a Causal Graph Spatial-Temporal Autoencoder (CGSTAE). The network architecture of CGSTAE combines two components: a correlation graph structure learning module based on spatial self-attention mechanism (SSAM) and a spatial-temporal encoder-decoder module utilizing graph convolutional long-short term memory (GCLSTM). The SSAM learns correlation graphs by capturing dynamic relationships between variables, while a novel three-step causal graph structure learning algorithm is introduced to derive a causal graph from these correlation graphs. The algorithm leverages a reverse perspective of causal invariance principle to uncover the invariant causal graph from varying correlations. The spatial-temporal encoder-decoder, built with GCLSTM units, reconstructs time-series process data within a sequence-to-sequence framework. The proposed CGSTAE enables effective process monitoring and fault detection through two statistics in the feature space and residual space. Finally, we validate the effectiveness of CGSTAE in process monitoring through the Tennessee Eastman process and a real-world air separation process.

Figures (14)

• Figure 1: Network architecture of CGSTAE.
• Figure 2: GCLSTM unit of the spatial-temporal encoder.
• Figure 3: Flowchart of the CGSTAE-based process monitoring.
• Figure 4: Process flow diagram of the TEP.
• Figure 5: ${{\text{T}}^{2}}$ statistic of different methods for fault 11 of TEP. (a) AE (b) LSTM-AE (c) GAE-I (d) GAE-II (e) DGSTAE (f) CGSTAE. The residual-based fault detection methods KDGCN and KG-GCBiGCN do not have ${{\text{T}}^{2}}$ statistic.