Abstract: The operating environment of coal mine belt conveyors is complex, and multi-source interference causes severe noise contamination in vibration signals, making fault features indistinct. Fault diagnosis based solely on mechanism analysis using signal decomposition and frequency-domain feature reconstruction makes it difficult to achieve accurate and stable identification. Fault diagnosis of coal mine belt conveyors highly depends on the stability of large-scale data transmission between underground and surface systems, which makes it difficult to realize real-time and high-precision fault diagnosis at the edge side. To address these problems, an edge-side fault diagnosis and lightweight modelling approach for coal mine belt conveyors was proposed. A signal preprocessing method combining Ensemble Empirical Mode Decomposition (EEMD) and Denoising Autoencoder (DAE) was adopted to achieve deep denoising of vibration signals, and feature reconstruction of the denoised signals was conducted based on correlation coefficients. A joint fault diagnosis method combining mechanism analysis and a data-driven model was established, and a convolutional neural network was used to perform local feature extraction and deep feature mining, which effectively compensated for the insufficient adaptability of a single fault diagnosis method. To meet the deployment requirements of edge-side fault diagnosis models in underground environments, convolutional channels and fully connected layer neurons with less importance were pruned to effectively remove redundant structures in the model. The experimental results showed that the joint fault diagnosis method combining mechanism analysis and a data-driven model achieved an average accuracy of 98.54%. After lightweighting, the number of parameters, computational cost, model size, and memory usage were reduced by 24.5%, 22.7%, 22.8%, and 24.4%, respectively, compared with those before lightweighting, while the decreases in average accuracy, recall, and F1 score were all less than 1%, achieving a balance between diagnostic accuracy and computational resource consumption and improving the feasibility of deploying the fault diagnosis model in underground edge environments.