LLM-VSFD: cross-modal learning for few-shot vibrating screen fault diagnosis

Abstract As a critical piece of equipment in mineral processing production lines, the accuracy of fault diagnosis for vibrating screens directly impacts production continuity and safety. However, the scarcity of fault samples in industrial settings severely restricts the practical application of existing intelligent diagnostic methods. To address this issue, this paper introduces large language models (LLMs) for vibrating screen fault diagnosis (VSFD) and proposes the LLM-VSFD framework to tackle the diagnostic challenges under few-shot conditions. The framework employs a two-stage strategy: (1) the cross-modal semantic alignment stage, which combines contrastive learning and classification tasks to optimally map time-series vibration signals to the LLM semantic space, activating the LLM’s cross-modal understanding capability and mitigating feature aliasing issues; (2) the domain-adaptive fine-tuning stage, where a hybrid prompt module, incorporating signal features, adaptive semantics, and learnable prompts, is designed, coupled with the low-rank adaptation parameter-efficient fine-tuning strategy, to fully unlock the diagnostic potential of the LLM. Experimental results show that under the 5-shot setting on a real vibrating screen dataset, LLM-VSFD achieved an accuracy of 84.74%, an improvement of about 30% points over the next best benchmark method. Additionally, it achieved accuracies of 80.59% and 86.45% on the Case Western Reserve University bearing dataset and Jiangnan University bearing dataset, respectively, showing an average improvement of about 10% points over the benchmark method, thus validating the framework’s strong generalization ability.