Abstract: Aiming at the problems of complex sound source and insignificant characteristics in the audio recognition of roller fault of belt conveyor in coal mine, an audio recognition method of roller fault based on convolution neural network and linear regression is proposed. Firstly, the audio signal along the roller is collected by the MEMS pickup carried by the inspection robot of the belt conveyor. Based on the wavelet autocorrelation denoising technology, the sound is preprocessed to suppress the background noise signal in the audio signal and optimize the data quality. Secondly, using the voiceprint spectrum separation technology, the HPSS (Harmonic Percussive Source Separation) method is used to separate the harmonic and shock wave components to enhance the sound signal characteristics of the roller fault. Based on MFCC (Mel Frequency Cepstrum Coefficient) voiceprint feature extraction method, the voiceprint feature information of the roller in the harmonic-shock wave is analyzed, the sound spectrum is generated, and the voiceprint representation ability of the roller fault is improved. Then, a harmonic-shock wave weak classifier is constructed based on multi-scale residual convolutional neural network, and a sound quality weak classifier is constructed based on multiple linear regression. Finally, based on two weak classifiers, using the spectrogram and sound quality features as data sources, fusion of multimodal faulty features and enrich data dimensions, based on the spectrogram and sound quality features, residual convolutional neural network computing image features, fast fitting of audio basic features using multiple linear regression, a roller fault voiceprint representation model combining convolutional neural network and linear regression is generated for joint training. The sample weight of the model training is optimized by the Focal Loss loss function to improve the accuracy of the model for roller fault recognition.The method in this paper is used to analyze and verify the audio information of the fault roller of the belt conveyor actually collected in Guojiawan Coal Mine of Yulin. The results show that the detection rate of roller fault reaches 95.79 %, and the detection accuracy reaches 95.60 %.