Abstract: To address the issues of insufficient generalization performance in traditional deep learning models caused by weak fault features, scarcity of high-quality samples, and cross-condition distribution shifts in mining rolling bearings, a lightweight fault diagnosis model based on Low-rank Multimodal Fusion and Adversarial Metric (MTSFCL) is proposed. The Superlet Transform (SLT) is utilized to convert vibration signals, constructing dual-modal input data of time-series and time-frequency diagrams to enhance the fault representation of mining rolling bearings from multiple dimensions. A lightweight dual-branch feature extraction layer is designed, where the temporal branch employs a Bidirectional Gated Recurrent Unit (BiGRU) enhanced by the Efficient Channel Attention (ECA) mechanism, effectively suppressing redundant information interference while capturing long-term dependencies in time-series signals. The spatial branch is based on an improved StarNet architecture, using multi-scale convolution and a Selective Kernel (SK) fusion mechanism to extract multi-scale fault features from time-frequency diagrams, and utilizing element-wise multiplication to achieve the mapping of high-dimensional spatial features without increasing network depth. A Low-rank Multimodal Fusion (LMF) module is designed, utilizing low-rank factors to project temporal and spatial features into a common subspace and performing non-linear fusion via element-wise multiplication, achieving deep interaction of dual-modal features at a low computational cost. Combining Conditional Domain Adversarial Network (CDAN) with Local Maximum Mean Discrepancy (LMMD) as a metric constraint, an adversarial metric domain adaptation module is constructed to jointly reduce the marginal and conditional distribution discrepancies between the source and target domains, improving the model's generalization performance. The Case Western Reserve University (CWRU) bearing dataset is selected for fault diagnosis experiments, and the experimental results show that: the parameter size of MTSFCL is only 0.3221 M; the average diagnostic accuracy of MTSFCL is 99.91% under single operating conditions; the average diagnostic accuracy is 94.12% under small-sample conditions with only 5 fault samples per class; the average diagnostic accuracy is 99.28% under cross-condition scenarios; and high accuracy is maintained under noise interference.