Abstract: Under complex underground operating conditions, mechanical noise generated by belt friction and coal flow impacts, airflow-induced disturbance noise, and coupled noise from multiple devices are superimposed, causing fault-related acoustic signatures of idlers to be easily masked by environmental noise. Meanwhile, the acquisition of abnormal idler samples is difficult and annotation costs are high, making traditional supervised learning-based idler abnormal condition detection methods hard to generalize effectively. To address these issues, an unsupervised idler abnormal condition detection method based on Multi-Granularity Attention Autoencoder (MG-AAE) was proposed, which used only normal-condition idler sounds for model training and required no fault labels. A multi-granularity composite acoustic feature composed of Mel spectrograms and Mel-Frequency Cepstral Coefficients (MFCCs) was constructed to jointly capture energy contours and fine-grained acoustic signatures. A Gaussian Difference Pyramid (GDP) and a Multi-Head Attention (MHA) mechanism were introduced into the encoder to perform multi-scale modeling and adaptive weighted fusion, thereby suppressing steady background noise and highlighting key fault-related frequency bands. A multi-dimensional reconstruction mean-square error was used as the anomaly criterion to achieve automatic identification of idler abnormal conditions. Experimental results showed that, when trained using only normal samples, the MG-AAE model demonstrated excellent performance in cross-device and real-world operating conditions. Evaluation on four typical device categories in the MIMII dataset showed that, under a strong noise condition of 0 dB, the average area under curve (AUC) and local AUC (pAUC) of the MG-AAE model reached 84.2% and 70.4%, respectively, representing improvements of 7.3% and 5.6% over the Autoencoder model. On real idler data, the AUC reached 95.47%, and the reconstruction error of abnormal samples was approximately 1.40 times that of normal samples. These results indicate that the proposed method has good cross-device generalization and a low false alarm rate, and provides effective technical support for abnormal condition detection of idlers in coal mine belt conveyor systems.