Abstract: Idler rollers are key load-bearing components of coal mine belt conveyors, and detecting their abnormal states is of great importance. However, in typical underground operating conditions, harsh environments, scarcity of abnormal samples, and strong background noise often lead to poor generalization and high false-alarm rates for conventional monitoring models. To address these issues, this paper proposes an unsupervised anomaly detection method based on a multi-granularity spectrogram and an attention autoencoder (MG-AAE), which is trained solely on acoustic signals from idlers under normal operating conditions without requiring fault labels. First, a multi-granularity composite spectral feature is constructed by combining Mel spectrograms and Mel-frequency cepstral coefficients (MFCCs), jointly capturing both global energy contours and fine-grained acoustic textures. Then, a Gaussian Difference Pyramid (GDP) and Multi-Head Attention (MHA) are incorporated into the encoder, where GDP performs multi-scale modeling and hierarchical decomposition of the composite spectrogram, while MHA adaptively fuses multi-scale features to suppress steady-state background noise and highlight key fault-related frequency bands. Finally, a multi-dimensional reconstruction error is used as the anomaly criterion to realize automatic identification of abnormal idler states. Experimental results show that, when trained using only normal samples, MG-AAE achieves average AUC and pAUC values of 84.2% and 70.4%, respectively, under 0 dB high-noise conditions for four typical industrial machines in the MIMII dataset, outperforming baseline models by 7.3%and 5.6%. On a real idler dataset, MG-AAE attains an AUC of 91.63%, with the mean reconstruction error of abnormal samples being approximately 4.3 times that of normal samples. These results demonstrate that the proposed method exhibits strong cross-machine generalization and low false-alarm characteristics, and can provide effective technical support for abnormal state detection of idler rollers in coal mine belt conveyor systems.