Cross roller bearings are vital in precision mechanical systems, where early-stage faults can lead to severe operational failures. While traditional deep learning method, such as CNNs and autoencoders, have shown promise in vibration-based fault diagnosis, their limited semantic modeling capacity restricts robustness under noisy or data-scarce conditions. To address this challenge, a novel diagnostic framework, named DiagLLM, is proposed, in which low-dimensional features extracted by a 1D CNN are embedded into a pre-trained large language model (LLaMA), enabling semantic-level classification through parameter-efficient LoRA tuning. Unlike prior approaches that require handcrafted prompts or domain-specific adaptations, DiagLLM enables direct and lightweight integration of sensor signals into transformer architectures. A comprehensive experimental study is performed on a cross roller bearing test rig, involving diverse feature representations and backbone architectures to systematically assess the diagnostic performance. The results show that DiagLLM consistently outperforms conventional methods in both full-data and few-shot scenarios, demonstrating its effectiveness and adaptability for intelligent fault diagnosis in industrial settings. This work not only enables lightweight integration of sensor signals with LLMs, but also establishes a novel and generalizable paradigm for semantic-level diagnostics in data-scarce and noisy environments.

Fault Diagnosis,Cross Roller Bearings,Large Language Models,CNN,LoRA