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

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.

💡 Research Summary

The paper introduces a novel framework called Causal Graph Spatial‑Temporal Autoencoder (CGSTAE) to address two persistent challenges in multivariate statistical process monitoring (MSPM): reliability under changing operating conditions and interpretability of black‑box models. CGSTAE consists of two tightly coupled components.

First, a Correlation Graph Structure Learning module based on a Spatial Self‑Attention Mechanism (SSAM) processes sliding‑window slices of the multivariate time series. For each window X(t) the model computes query and key matrices (Q = XW_Q, K = XW_K) and forms an attention matrix A(t) = σ(QKᵀ/√w). The resulting attention weights are interpreted as dynamic pairwise similarity scores, i.e., the adjacency matrix of a time‑varying correlation graph. This allows the network to capture non‑stationary relationships among process variables without any prior knowledge of the underlying physical connections.

Second, the authors recognize that correlation graphs do not convey causality and therefore propose a three‑step Causal Graph Structure Learning algorithm that exploits a “reverse perspective” of the causal invariance principle. The principle states that true causal mechanisms remain stable across different operating regimes, even when observed correlations fluctuate. By aggregating correlation graphs obtained under diverse conditions, the algorithm extracts the invariant portion of the edges, yielding a directed causal adjacency matrix. The three steps are: (1) pre‑training the autoencoder with the SSAM‑derived correlation graphs, (2) learning the causal graph by averaging, sparsifying, and optionally incorporating domain knowledge to enforce directionality, and (3) fine‑tuning the entire model while keeping the causal graph fixed. The method assumes sufficient interventions in the data so that the invariant causal effect can be identified.

With the causal graph in hand, the Correlation Graph module is discarded and the second component—a Spatial‑Temporal Encoder‑Decoder built from Graph Convolutional LSTM (GCLSTM) units—takes over. Each GCLSTM cell performs a graph convolution (GC) on the node features using the causal adjacency, then feeds the result into the standard LSTM gates (forget, input, output). The encoder runs forward through the window, producing a final cell state c(t) and hidden state h(t). The decoder runs backward, reconstructing the original sequence via a fully‑connected output layer.

Two monitoring statistics are derived from the trained model: (i) a T² statistic computed on the latent hidden representations (feature space) and (ii) a SPE (Squared Prediction Error) statistic computed on the reconstruction residuals (residual space). Because both statistics are grounded in the learned causal graph, they are more robust to spurious correlations and provide clearer diagnostic insight than traditional PCA‑based T²/SPE.

Experimental validation is performed on the benchmark Tennessee Eastman (TE) process and on a real‑world air‑separation plant. On TE, CGSTAE achieves an average detection delay of 2–3 samples across 21 fault scenarios and reduces false alarm rates to below 1 %, outperforming recent GNN‑based MSPM methods by 12–18 % in F1‑score. On the air‑separation data, which exhibits strong non‑stationarity and sensor noise, CGSTAE improves F1‑score by more than 15 % and cuts detection delay by roughly 20 % relative to baseline models.

Key contributions are: (1) a unified pipeline that learns dynamic correlation graphs via self‑attention and converts them into stable causal graphs using the reverse causal invariance principle, (2) the integration of a causal graph into a GCLSTM encoder‑decoder for joint spatial‑temporal feature extraction and reconstruction, (3) dual monitoring statistics that leverage both latent features and residuals for reliable fault detection, and (4) thorough empirical evidence of superior performance on both simulated and real industrial processes.

Limitations include the need for sufficiently diverse operating conditions (interventions) to satisfy the invariance assumption, increased computational cost of GCLSTM on large‑scale graphs, and the reliance on expert validation to confirm the directionality of discovered causal edges. Future work may explore Bayesian causal graph learning for sample efficiency, lightweight graph‑Transformer alternatives to reduce complexity, and online adaptation mechanisms that update the causal graph in real time as new data arrive.