Structure-prior relation graph learning for spatio-temporal anomaly detection in multi-entity collaborative systems

Spatio-temporal anomaly detection in multi-entity collaborative systems aims to identify abnormal states from multivariate time series generated by dynamically coupled variables. Fully data-driven relation learning may recover unstable or spurious dependencies when normal observations are limited, noisy, or weakly informative about rare coordination failures. This paper studies structure-prior relation graph learning (SPRGL), which augments a GDN-style predictive graph backbone with weak structural priors. Variable groups, topology files, or process connectivity are represented as a structure-prior graph and united with the embedding-derived candidate graph before graph-attention prediction. The technical contribution is therefore the candidate-edge fusion mechanism, rather than the established embedding, Top- k graph learning, attention forecasting, or residual-scoring components. We also construct CTown-RB as a semi-synthetic benchmark containing five injected relation-breaking scenarios based on public CTown topology and normal BATADAL operating data. Under the reported comparison with PCA-SPE, Isolation Forest, and LOF, SPRGL obtains the highest reported F1 and AUC-PR in the SMD, BATADAL, and CTown-RB settings shown in Table 1 ; a separate HAI experiment examines sensitivity to the residual-aggregation size k s . On CTown-RB, the combined prior changes PA-F1 from 0.5651 to 0.8734 and detects all five injected scenarios, while also reducing Precision. These results delimit the current evidence to the reported baselines, datasets, and archived evaluation outputs.