Answers are grounded in this paper and cite their sources.
In industrial applications, the extreme imbalance in the distribution between healthy and faulty classes severely limit the generalization capabilities of diagnostic models. Generative Adversarial Networks (GANs) are susceptible to mode collapse, and while Stable Diffusion models can generate high-fidelity images, they lack the domain-specific prior knowledge required for bearing fault diagnosis. Consequently, this research introduces a novel data generation framework, termed the Large Language Model-guided Physics-Constrained Diffusion model (LLM-PCDiff). The proposed framework consists of three core components. First, features are extracted from vibration signals, and a Large Language Model (LLM) is fine-tuned using a lightweight Low-Rank Adaptation (LoRA) strategy. This enables the LLM to generate dual guidance information, comprising both semantic fault descriptions and soft labels. Second, a composite loss function is formulated by integrating physical constraints with a conditional Wasserstein distance, embedding the physical prior knowledge of bearing fault characteristic frequencies into the generative process. Finally, the Variational Autoencoder (VAE) and text encoder of the Stable Diffusion model are frozen, while only the noice prediction network undergoes parameter fine-tuning, leveraging time-frequency diagrams, prompt descriptions, and soft labels to optimize the generation parameters. Experimental results demonstrate that the sample quality of the LLM-PCDiff framework is significantly superior to that of both GAN-based and DDPM-based models. By synergizing the semantic understanding capabilities of LLMs with the high-quality generation capacity of an enhanced diffusion model, this study provides a novel data augmentation strategy for imbalanced bearing fault diagnosis.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: