A Lightweight Method Integrating Dynamic Frequency-Aware Convolution and SoftPool for Abnormal Sound Detection in Wind Turbines

Abstract

Aiming at the problems of insufficient feature expression capability and high model computation complexity in traditional wind turbine group abnormal sound detection methods, this paper proposes a lightweight detection method based on improved MobileNetV3 network. First, the SincNet bandpass filter and Mel spectrum are integrated to construct multi-dimensional acoustic features, taking into account the original signal time-domain features and frequency-domain features. Second, a dynamic frequency-aware convolution (DFC) module is introduced into the MobileNetV3 network architecture to adaptively adjust the parameters of the convolution kernel through the frequency-domain attention mechanism to strengthen the frequency feature capture of abnormal sounds; the feature downsampling process is optimized by combining with SoftPool to reduce the loss of high-frequency information. On the dataset of Danish University of Science and Technology, the AUC reaches 94.71%, and the number of parameters is only 2.38M, which is 62.3% lower than the mainstream model ResNet-18, providing a high-precision edge-end solution for wind turbine status monitoring.