基于音频数据的风电机组异常检测与分类算法研究OA

A semi-supervised anomaly detection and classification framework are proposed to address limited accuracy in anomaly detection of wind turbines caused by highly heterogeneous audio data,scarce anomalous samples,and the difficulty in identifying faults of unknown categories.The core innovation is a two-stage strategy:(i)an improved GANomaly model for efficient anomaly detection,and(ii)a dynamic feature memory bank that performs fine-grained anomaly classification via feature-similarity matching to facilitate identification of categories beyond the training scope.Experiments on a real-world wind-turbine audio dataset show that the proposed approach achieves an F1-score of 99.39%for anomaly detection and an overall accuracy of 96.75%for multi-class anomaly classification,offering an effective solution for class imbalance and challenges in identifying anomalies in unknown categories in industrial scenarios and providing a reliable tool for intelligent wind-turbine operation and maintenance.