Dynamic heterogeneous graph attention network for spatiotemporal correlation modeling of mine microseismic events and intelligent rockburst risk level reasoning

Deep mining releases accumulated strain energy violently, and microseismic arrays record the fracturing that precedes rockbursts. Graph neural networks have already been applied to seismic and microseismic tasks such as phase association, source characterisation and event forecasting; what those graphs share is a topology anchored to a fixed sensor array and a single edge semantics, so they cannot follow an event field that turns over as the working face advances. This paper proposes a dynamic heterogeneous graph attention network (DHGAT). Every event enters as a node that retains four feature channels — waveform, source parameter, spatial and temporal — and three edge types (spatial proximity, temporal sequence, feature similarity) are rebuilt inside a 48 h sliding window. Relation-specific multi-head attention aggregates neighbours separately for each edge type, a semantic-level layer learns how much each relation contributes, and a gated recurrent encoder tracks the evolving graph embedding before four risk grades are assigned: no risk, low, moderate and high. On a 24-month catalogue of 7956 events from one deep metal mine, DHGAT reached 91.9 ± 0.4% accuracy, 0.853 ± 0.008 macro-F1 and 0.887 ± 0.006 quadratic weighted Kappa over ten runs, ahead of the strongest static heterogeneous baseline (HAN, 89.2% and 0.788) and of every other baseline at p < 0.01; recall on high-risk events reached 80.0%. Relation weights migrate from spatially dominated during gradual stress build-up to temporally dominated in the final hours before failure, which mirrors the shift from distributed microcracking to localised coalescence, and a perturbation test confirms that the attention carries the discriminative signal. Inference costs 46 ms per window on a single NVIDIA RTX 3090 GPU. All data come from one mine, so the reported thresholds require recalibration before transfer.