250934Enhance publishing丨Unsupervised Anomaly Detection of Hydro-Turbine Generator Acoustics by Integrating Pre-Trained Audio Large Model and Density Estimation

In recent years, non-contact acoustic measurement has received widespread attention as an effective detection method. However, it is difficult to collect abnormal acoustic signals during the actual operation of hydro-turbine generator units, which restricts the application of traditional anomaly detection methods and supervised learning-based classification strategies in this field. To address these challenges, this paper proposes an unsupervised acoustic anomaly detection method for hydro-turbine generators that fuses pre-trained large audio models with density estimation k-nearest neighbors (k-NN).

First, the validity of general audio features extracted by pre-trained audio models in anomaly detection is verified. Subsequently, a parameter fine-tuning strategy integrating attention statistical pooling and warm-up is designed to achieve model transfer optimization. In the inference stage, a density estimation k-NN is designed to realize a robust distance metric.

Experimental results show that the proposed method achieves a multi-metric harmonic mean of 98.7% in the wind tunnel environment and up to 99.9% in the slip ring chamber, providing a feasible and excellent solution for acoustic anomaly detection in hydropower stations.