Abstract: To address the problems of weak fault features, scarce high-quality samples, and cross-condition distribution shifts in mining rolling bearings, which lead to insufficient generalization performance of traditional deep learning models, a Mine Rolling Bearing Fault Diagnosis Model Based on Low-Rank Multimodal Fusion and Adversarial Metrics (MTSFCL) is proposed. The superlet transform was used to construct dual-modal input data composed of time-series signals and time-frequency images, which enhanced the multidimensional representation of rolling bearing faults. A lightweight dual-branch feature extraction layer was designed. The temporal branch adopted a Bidirectional Gated Recurrent Unit (BiGRU) enhanced by the Efficient Channel Attention (ECA) mechanism, which captured long-term dependencies in time-series signals while effectively suppressing interference from redundant information. The spatial branch was built on an improved StarNet architecture. Multi-scale convolution and a selective kernel fusion mechanism were used to extract multi-scale fault features from time–frequency images. Element-wise multiplication was used to achieve high-dimensional spatial feature mapping without increasing network depth. A Low-Rank Multimodal Fusion (LMF) module was designed, in which low-rank factors projected temporal and spatial features into a common subspace, and nonlinear fusion was performed through element-wise multiplication, enabling deep interaction between dual-modal features with low computational cost. To improve model generalization performance, a domain adaptation module based on an adversarial metric was constructed by combining the Conditional Domain Adversarial Network (CDAN) with Local Maximum Mean Discrepancy (LMMD) as a metric constraint, thereby reducing marginal and conditional distribution differences between the source domain and the target domain. Experimental results showed that: ① the number of parameters of MTSFCL was only 0.322 1 × 10⁶, and the inference time for a single sample was 2.76 ms. ② The average diagnostic accuracy under a single operating condition reached 99.94%. Under the small-sample condition with only five fault samples for each class, the average diagnostic accuracy reached 94.12%, which was significantly higher than that of high-parameter models such as ViT and VGG16. ③ Under cross-condition scenarios, the average diagnostic accuracy reached 99.28%. Compared with the CDAN domain adaptation method without the LMMD metric constraint, the accuracy increased by 4.27%. High accuracy was also maintained under strong noise interference, demonstrating strong generalization performance and robustness.