Tunnel abnormal sound recognition based on multi-channel convolutional neural network

• Published in: 2022 18th International Conference on Computational Intelligence and Security (CIS) pp. 112 - 116
• Main Authors: Lang, Julin, Zheng, Sheng
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-12-2022
• Subjects: CNN; Computational modeling; Feature extraction; Image recognition; Multi channel; Roads; Speech recognition; Stability analysis; Tensors; Voice recognition
• DOI: 10.1109/CIS58238.2022.00031

In order to improve the accuracy of sound detection for abnormal sound events in road tunnels, a multi-channel convolutional neural network model is designed in this paper. The model stitches the extracted low-level features into a tensor feature map as an input channel of the model, and the Mel spectrogram of the sound sample as another input channe. And the above two channels together constitute the input of the model, so that the model learns as many features as possible. The unique advantages of convolutional neural networks in image recognition are used to extract deeper recessive features to improve the accuracy of model recognition and realize the classification of sound signals. Experimental results show that the multi-channel convolutional neural network model has higher classification accuracy than the traditional sound classification model. Moreover, the two-channel input model has more features than the single-channel input model, which improves the stability of the model to a certain extent.