Abstract: To acquire the effective feature vectors and classification model for the coal-rock identification based on the working face of fully mechanized caving mining, the acoustic pressure signals of coal and rock were collected by the existing equipment and the designed collection program. By analyzing the acquired acoustic pressure signals in time domain, the feature vectors were acquired which were used as the input vectors of neural network. The principal component analysis method was used to reduce the correlation between time-domain characteristics and the dimensions of the input feature vectors of neural network. Then the BP neural network model was designed. By comparing the gradient descent method and Levenberg-Marquard method, the time consuming of LM training method is significantly less than the gradient descent method. The neural network based on the LM method combing with principal component analysis (PCA) obtains higher identification accuracy and consumes much less time than that not combing with PCA.