基于骨架序列的发电厂作业人员异常行为识别研究OA

To address the challenges of topological mismatch in skeleton recognition,data degradation,and insufficient recall of rare anomalies,caused by the coupling of strong electromagnetic interference,dust,and vibration in power plants,this paper proposes an Electromagnetic-Robust Spatio-Temporal Heterogeneous Synergy Network(STHS-Net).First,an electromagnetic-quantified spatiotemporal graph(ERSTG)is constructed by embedding measured electromagnetic intensity into dynamic topology and pruning with inter-frame memory updating to suppress spurious connections.Then,a coupled noise separation and quality regeneration strategy(CNSQR)is proposed,employing tensor robust principal com-ponent analysis combined with bone-length preservation regularization to achieve low-rank completion.Subsequently,a dual-branch heterogeneous feature fusion(DBHFF)module encodes potential-energy mutation and anomaly-activated gated deep features,forming a tripartite collaborative network comprising physically preserved,deep anomaly,and global context pathways.A gradient-balanced re-weighting mechanism is further adopted to enhance the recall of rare abnormal behaviors.Experimental results show that,compared with methods such as GCN-RNN and STTN,STHS-Net improves the mean average pre-cision(mAP)by 2.1-10.4 percentage points and rare-class recall(RCR)by 3.3-12.7 percentage points,while achieving a per-frame inference time of 8.7 ms,a frame rate of 115 FPS,and a model size of 28.6 MB,providing a deployable solution for strong-disturbance industrial scenarios.