Abstract: To solve the problems associated with coal mine gas and coal dust explosion disaster alarm method, such as less monitoring means, outdated monitoring technology, high false alarm rate, and high no-alarm rate, and improve the accuracy of coal mine gas and coal dust explosion identification, a sound recognition method of coal mine gas and coal dust explosion based on feature fusion is proposed. By installing mining pickups in the critical monitoring area of underground coal mine to collect the working sound of coal mine underground equipment and environmental sound in real time, etc., the sound signal is extracted from the mixed features MGCC composed of the Mel-scale frequency cepstral coefficient and the Gammatone filter cepstral coefficient. The first 9-dimensional eigenvalues are extracted by the method of principal component analysis(PCA) to form the feature vector characterizing the sound signal, which is input to the support vector machine for training to obtain the sound recognition model of coal mine gas and coal dust explosion. The sound samples to be measured are extracted from their feature vectors and input to the trained sound recognition model for coal mine gas and coal dust explosion, and the classification recognition results are obtained and verified by experiments. Firstly, through the feature extraction experiment of the sound signal, the characteristics of the sound signal features extracted by MGCC and principal component analysis are analyzed. The distribution of feature values of the same sound signal of different durations is stable and does not differ significantly. The distribution of feature values of different sound signals of the same duration differs significantly. Therefore, the feature vectors of the sound signal extracted by MGCC and principal component analysis can effectively characterize the sound signal. Secondly, the recognition rate of the proposed recognition method reaches 97%, which is higher than that of the comparison algorithm. Finally, the experimental results of Bayesian hyperparameter optimization show that the optimized sound recognition model of coal mine gas and coal dust explosion is significantly better than the recognition model before optimization, which can meet the requirements of coal mine gas and coal dust explosion perception and alarm.