Vision-language models exhibit remarkable zero-shot capabilities but suffer significant performance degradation under distribution shifts. While test-time adaptation (TTA) via Low-Rank Adaptation offers a parameter-efficient solution, we identify a fundamental bottleneck in current methods: the reliance on static rank configurations. Because visual inputs inherently possess varying information densities, a fixed rank forces an inevitable optimization compromise, leading to underfitting on complex scenes and overfitting on simple ones. To bridge this gap, we propose Multi-Rank Adaptation (MuRA), a novel framework that dynamically selects and fuses adaptation modules of varying capacities based on token-level visual complexity. MuRA synergizes Multi-Rank Orthogonal Decomposition to provide a superior, knowledge-preserving initialization, and Unified Component Fusion with Continuous Router Updating to sustainably learn semantic-to-rank mappings. Furthermore, we provide rigorous theoretical justifications mathematically proving the necessity and gradient stability of this adaptive mechanism. Crucially, MuRA's dynamic design uniquely thrives at the deepest visual layer, capitalizing on the shortest gradient backpropagation path. Extensive experiments demonstrate that MuRA achieves state-of-the-art accuracy across extensive domain generalization and cross-dataset benchmarks while significantly reducing both computational and memory overhead.
Gengyuan Liu, Nan Wang, Chang Liu et al.· 0 citations
Deep learning achieves widespread success in fault diagnosis. However, its effectiveness is hindered in practical industrial environments due to complex operating conditions and data sparsity. This article proposes a novel model-agnostic meta-learning framework based on a selective state space model (MAML-S3M) to address the challenge of cross-condition few-shot bearing fault diagnosis. The framework introduces three core innovations. First, the continuously stacked selective state space module dynamically adjusts its receptive field, enabling precise feature extraction under varying conditions. Second, the channel attention mechanism is combined with the selective state space model to capture multi-scale features, thereby enhancing the feature extraction capability of the model. Third, the introduction of an explicit information discarding strategy during meta-task optimization refines the meta-learning process, thereby yielding optimal parameters. Extensive experiments on bearing datasets across different operating conditions demonstrate that the proposed MAML-S3M achieves superior diagnostic accuracy, with an average accuracy of 99.18% across six cross-condition tasks, outperforming state-of-the-art methods such as generalized model-agnostic meta-learning (GMAML) by at least 1.1%. The improvements are particularly helpful in scenarios with complex operating conditions and scarce samples, maintaining over 94% accuracy even in the challenging “10-way 1-shot” setting. We have made the paper’s results publicly available on GitHub. The link is as follows:
https://github.com/12138250/MAML-S3M
.
Siyu Liu, Nan Wang, Xueyi Li et al.· Structural Health Monitoring· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
