Abstract: The sensor monitoring data combined with neural network prediction model is the mainstream method of mine belt conveyor operation state prediction. However, using contact sensor to monitor the belt conveyor running state has some problems, such as inconvenient installation and large data error, resulting in low prediction precision of belt conveyor operation state. In order to solve this problem, a prediction method of mine belt conveyor operation state based on audio signal is proposed. Firstly, the high-pass filter and Boll spectral subtraction are used to filter and reduce the noise of the original audio signal during belt conveyor operation. Secondly, the first dimension component (MFCC₀) of Mel-frequency cepstral coefficients (MFCC) of audio signal is extracted by pre-emphasis, framing and windowing, Fourier transform, Mel filter energy calculation, discrete cosine transform, and input to the residual block optimized convolutional neural network combined with long and short term memory network (Res-CNN-LSTM) prediction model to reduce the amount of input data of the prediction model. Finally, the MFCC₀ spatial characteristics of the belt conveyor audio signal are extracted adaptively by CNN with residual blocks, and the dimension of the data is reduced. Moreover, the temporal characteristics of the dimension-reduced data are extracted based on LSTM, so as to improve the prediction precision of the belt conveyor operation state. The experimental results show that MFCC₀ can effectively characterize the audio signal characteristics of belt conveyor in different operation states. Compared with CNN, LSTM, and CNN-LSTM models, the Res-CNN-LSTM model is more accurate in predicting the operation state of the belt conveyor.