Abstract: Intelligent health state identification of belt conveyor idler bearings faces challenges including insufficient degradation characterization by a single modality, poor complementarity and consistency between modalities, and difficulty extracting features that link health state classes to continuous degradation. To address these challenges, a bearing health state identification method based on acoustic–vibration fusion for Health Indicator (HI) construction and Multi-Granularity Ordinal Contrastive Learning (MGOCL) was proposed. Time- and frequency-domain features were extracted from preprocessed vibration and acoustic signals, and degradation-sensitive features were selected based on correlation, monotonicity, and robustness. An acoustic–vibration fusion method for HI construction based on Mamba-Bidirectional Cross Attention (Mamba-BiCA) was developed. Dual-branch Mamba encoders captured long-term temporal dependencies in acoustic and vibration features throughout bearing degradation. Bidirectional cross attention enhanced interactions between the two modalities, and physical constraints were incorporated to construct HIs for five health states: healthy, good, fair, degraded, and failed. MGOCL combined discrete state supervision with continuous degradation constraints to preserve interclass separability, the ordering of health states, and intraclass degradation continuity in the learned representations. Health states were identified using k-nearest neighbor classification. Experimental results showed that the method achieved identification accuracies of 0.972, 0.918, and 0.965 on subsets G1, G2, and G4 of the XJTU public dataset, respectively, with an inference latency of only 6.183 ms. On a self-built dataset collected using a belt conveyor simulation test rig, the identification accuracy reached 0.947, outperforming comparison models including XGBoost, MLP, and MSCCNN.