Abstract: To solve the problems of high false alarm rate and leakage rate of coal mine gas and coal dust explosion disaster alarm methods, and improve the accuracy of coal mine gas and coal dust explosion perception, sound identification method of coal mine gas and coal dust explosion based on wavelet scattering transform was proposed: install mining sound pickup equipment in the critical monitoring area of coal mine underground, and collect equipment working sound and environmental sound in real time, the wavelet scattering coefficients were obtained from the collected sound by wavelet scattering transform, the wavelet scattering coefficients of the sound signal were constructed, the collected the 11-dimensional feature parameters consisting of small gradient dominance, large gradient dominance, energy, gray average, gradient average, gray mean square difference, gradient mean square difference, correlation, gray entropy, gradient entropy, mixing entropy were obtained by calculating the image gray gradient co-generation matrix of the wavelet scattering coefficient map, which constituted the feature vector characterizing the sound signal, and were input to the support vector machine for training to obtain the coal mine the 11-dimensional feature vectors were obtained by extracting the gray gradient covariance matrix of the wavelet scattering coefficient map of the sound signal to be measured, and bring it into the trained coal mine gas and coal dust explosion sound recognition model for sound recognition classification, it verified by experiments. The wavelet scattering coefficients of different sounds and their feature parameter distribution characteristics were analyzed, the wavelet scattering coefficients of gas and coal dust explosion sounds and their 11-dimensional feature vectors were significantly different from other sounds in the coal mine, the feasibility of the proposed feature extraction method was demonstrated. Support vector machine hyperparameter optimization experiments completed by Bayesian optimization to select hyperparameters that better fit the training model, and the recognition experimental results show that the recognition rate of the proposed method was 95.77%, which was significantly better than other comparison algorithms. It can meet the needs of coal mine gas and coal dust explosion recognition.