In real industrial processes, data are continuously collected from various sensors and actuators, such as flow meters, pressure sensors, pumps, and valves. Although the data collected from each sensor may appear to be independent, they are structurall...
In real industrial processes, data are continuously collected from various sensors and actuators, such as flow meters, pressure sensors, pumps, and valves. Although the data collected from each sensor may appear to be independent, they are structurally related within the process. When an abnormal situation occurs, sensor values do not change independently; instead, they form interrelated change patterns according to the process stage to which each sensor belongs and the physical flow between process units. Therefore, effective anomaly detection in industrial processes requires not only the temporal pattern of each individual sensor but also the underlying process structure.
In this study, a process structural knowledge-based hierarchical hypergraph architecture is proposed to mimic the structure of real industrial processes. The proposed model performs multivariate time-series anomaly detection by hierarchically modeling both sensor-level and process-level representations. First, multi-scale one-dimensional convolution is applied to each sensor time-series window to extract temporal change patterns, and process topology information is injected into sensor embeddings to reflect the positional information of each sensor within the process flow. Then, a process-knowledge is used to aggregate information from sensors that belong to the same process or are structurally related, thereby generating sensor-level representations. Next, the sensor-to-process membership relationship is defined in the form of a hypergraph, and sensor representations belonging to the same process are aggregated into process-level representations. In addition, a directed process flow graph is used to generate process-level representations that reflect upstream and downstream process contexts. Finally, the process-level context is redistributed back to the sensor level, and anomaly scores are computed based on the reconstruction error of sensor representations that incorporate process context.
Furthermore, this study uses the process structural knowledge graph to trace abnormal causes. The attack point information provided by the dataset is defined as a proxy label for the root cause sensor, and root cause candidates are traced using the process knowledge graph, upstream tracing based on process flow, and sensor state changes.
The performance of the proposed model was evaluated using industrial control system datasets. Experimental results show that the proposed model achieved competitive anomaly detection performance compared with baselines, indicating that hierarchical modeling with process structural knowledge can contribute to improved multivariate time-series anomaly detection. In addition, the root cause candidate tracing results show that process structural knowledge can be used to trace not only the sensors where anomalies are observed but also possible causes of abnormal events. These results suggest that incorporating process structural knowledge into industrial multivariate time-series anomaly detection can improve both detection performance and traceability