Edge-Deployable Lightweight Fault Pattern Recognition and Early Warning Model for Key Components of Wind Power Equipment

To address the problems of difficult fusion of multi-source heterogeneous data, weak fault features, difficult edge deployment and delayed warning for key components of wind power equipment under complex operating conditions, this paper proposes an edge-deployable lightweight fault pattern recognition and hierarchical early warning model, named LW-FDWP. The model integrates SCADA data, vibration signals, acoustic signals and operation and maintenance logs. Through key feature selection, lightweight multi-branch encoding and reliability-guided dynamic fusion, it realizes efficient representation and collaborative modeling of multi-source information. On this basis, fault classification, residual-driven risk assessment and dynamic-threshold warning mechanisms are introduced to achieve integrated outputs of fault recognition and early warning. Meanwhile, structural pruning, parameter quantization and knowledge distillation are adopted to achieve lightweight model design and meet the deployment requirements of edge devices. Experiments are conducted on the constructed WPF-Multi dataset. The results show that the proposed method achieves 94.28% accuracy, 93.53% F1-score and 0.976 AUC in the fault recognition task, improving over the Transformer model by approximately 2.83%, 2.94% and 2.3%, respectively. In the warning task, the average advance warning time reaches 26.8 h, while the false alarm rate and missed alarm rate are reduced to 3.92% and 4.16%, respectively. Edge deployment experiments show that the model has only 4.1M parameters and 2.9G FLOPs, and its inference latency on Jetson Orin NX is only 8.6 ms. The results verify the effectiveness and engineering application value of the proposed method in intelligent operation and maintenance of wind power equipment.