Audio Classification of Water Supply Pipeline Leakage Based on MobileNetV3 Convolutional Neural Network

To accurately identify the leakage sound of the urban water supply network, this paper proposes an audio classification and recognition method for water supply pipeline leakage based on MobileNetV3. First, the audio files in the ROPP dataset were enhanced offline, the leakage signal was converted into a logarithmic Mel spectrogram and the spectral subtraction was used to achieve data noise reduction. The Attention Mechanism module and MobileNetV3 neural network training were used to extract image features. Finally, the Softmax function was used to classify the leaked audio. The experimental results show that this method can make the classification precision of water leakage category reach 99.40%, and the recall rate reaches 99.20%.