Anomaly detection for industrial time series in process industry using informed machine learning with graph attention networks

In Industry 4.0, detecting anomalies in multivariate time series for industrial device monitoring is a significant challenge. Inherent data biases in the training dataset may cause traditional models to learn spurious correlations, resulting in outcomes that do not align with expert knowledge. Consequently, the integration of knowledge-based representations with sequential data is essential to enhance the capacity to capture complex patterns of high-level semantics and provide meaningful explanations. This paper presents Composite Knowledge Fusion Data with Graph Attention Networks (CKDGAT), an unsupervised anomaly detection method for process industry production monitoring. CKDGAT utilizes a two-layer graph attention network architecture to capture variable interactions and temporal dependencies, fusing these elements to generate new features. A multi-head stochastic attention mechanism is employed to model knowledge-based information. A reconstruction module leverages these features to reconstruct input multivariate time series and generate anomaly scores. Experiments demonstrate that CKDGAT outperforms state-of-the-art baseline models on the vertical roller mill and secure water treatment testbed datasets. Additionally, further analysis indicates that CKDGAT provides interpretable explanations for detected anomalies. • Proposes a detection model bridging temporal data patterns with domain knowledge. • Designs a dual-layer graph attention network for data-knowledge fusion. • Introduces stochastic multi-head attention to enable an interpretable architecture. • The proposed model outperforms baselines on two process industry datasets.