PowerXGraph++: Regulation-Constrained Spatiotemporal Evidence Graphs with Counterfactual Explanations for Power-Worksite Safety Monitoring

Automated power-worksite monitoring requires accurate event recognition and auditable explanations. This paper presents PowerXGraph++, a regulation-constrained spatiotemporal evidence-graph framework for missing PPE, boundary crossing, unsafe proximity, falls, and fire/smoke. Typed evidence tokens are encoded by a transformer; full-context prediction is separated from explanation selection, and explanations are tested for sufficiency, necessity, sparsity, and regulation consistency. Distillation, event-boundary counterfactuals, background-invariance training, and non-inferiority checkpoint selection preserve predictive performance. On annotation-derived construction-PPE graphs, three-seed Macro-F1, violation recall, and rationale F1 were 0.989, 0.965, and 0.901. Under the same frozen predictor, the supervised, grammar-aware selector attained rationale F1 of 0.983 and regulation consistency of 1.000, although the post hoc baselines lacked equivalent supervision. A single YOLO reproduction run using seed 17 on 197 matched person graphs yielded Macro-F1 of 0.844, missing-PPE recall of 0.591, and rationale F1 of 0.660. Expert evaluation produced Fleiss’ kappa of 0.740 and method-to-majority rationale F1 of 0.870. Held-out-camera reasoning and a two-camera fence-crossing pilot provide additional evidence under supplied-evidence conditions; the results do not establish automatic generation of every semantic predicate or unrestricted raw video performance.