In order to improve the accuracy of coal mining machine rocker gear fault diagnosis, reduce the model size and facilitate the deployment to more mobile and edge devices, a lightweight densely connected convolutional network (LW-DenseNet) based coal mining machine rocker gear fault diagnosis model is built. Separable convolution is used instead of traditional convolution to reduce model parameters and improve diagnosis efficiency; feature propagation is enhanced by dense connection mechanism to strengthen feature extraction capability. The rocker arm gear vibration signals collected from the coal mining machine rocker arm loading test bench are used to train and verify the effectiveness of the model. The experimental results show that the proposed method can achieve 99.276% classification accuracy with only 0.05MB model size compared with various diagnostic models, and the good generalization of the model is verified using the Case Western Reserve University bearing dataset. Finally, the visual representation of the key layer using t-SNE clearly shows the good feature extraction performance of the model.