IterPrompt: An Iterative Prompt-Based Finetuning Alignment Method for Fault Diagnosis of Wind Turbine Gearbox

Large language model (LLM) has developed rapidly in recent years, but the high training cost of domain-specific LLM targeted at real engineering problems has led to its limited application in real industrial environments. To address the above problems, an iterative prompt-based finetuning alignment method is presented in the paper for fault diagnosis of wind turbine gearbox. In this method, the base model is first finetuned and compressed to adapt the specific-domain task as well as improve computational efficiency without degrading model performance. And then, an optimal prompt template is constructed based on self-supervised prompt optimization (SPO) to improve the model generalization performance in fault diagnosis of wind turbine gearbox. Furthermore, the structured input with textual form is designed based on signal features extracted from original signals to meet the requirement of textual data input of LLM, and the pretrained knowledge of LLM is aligned further by the iterative prompt strategy based on post-feedback. By comparing with other prompt learning methods, the model finetuned by the method proposed in the paper has better performance and higher test accuracy on the task of wind turbine fault diagnosis, which verifies the effectiveness of the method proposed in this paper.