Abstract: With the acceleration of intelligent construction of coal mines, efficient recognition of coal and rock has become a technical problem to be solved urgently in intelligent coal mining. The existing coal and rock recognition methods under complex coal mine geological conditions have problems of low precision, poor universality and are difficult to apply in engineering. In order to solve the above problems, a coal and rock recognition model based on improved 1-dimensional convolutional neural network (1DCNN) is proposed. Based on the 1DCNN, a plurality of continuous convolution layers are used for extracting one-dimensional vibration signal features. The global average pool (GAP) layer is used for replacing the full connection layer. The model training parameters are reduced, and computing resources are saved. At the same time, a cosine annealing attenuation method with a linear hot start is adopted for optimizing the learning rate. Therefore, the model training is prevented from falling into a local minimum region, and the training quality is improved. In order to intuitively describe the feature extraction process and classification capability of the improved 1DCNN model for coal and rock cutting vibration data, the t-distributed stochastic neighbor embedding (t-SNE) manifold learning algorithm is used to visually analyze the feature learning process of the model. The results show that the improved 1DCNN model can realize the recognition of coal and rock cutting states well through feature learning layer by layer. Based on the measured vibration data obtained in the process of coal and rock cutting of the MG 650/1590-WD shearer in a mine in Shaanxi province, the model is trained and the result shows that the accuracy of the improved 1DCNN model is 99.91% on the training set and 99.32% on the test set. The model can be directly used to classify the original vibration signals of the shearer in coal and rock cutting, and can effectively identify the cutting state of coal and rock. Compared with traditional machine learning, ensemble learning and the unmodified 1DCNN model, the improved 1DCNN model has obvious advantages. The average recognition accuracy rate reaches 99.56%. The calculation cost is greatly saved, and the model recognition speed is improved.