Improving the Precision of Hidden Danger Recognition in Power Dispatch Duty Logs through RLHF Multi-round Human Feedback Mechanism

Abstract

The problem of hidden danger identification in power dispatch duty logs is that professional text semantics are complex and expert annotations are scarce, resulting in insufficient recognition accuracy. This paper proposes an optimization method based on multi-round RLHF (Reinforcement Learning from Human Feedback). The reward model is trained through interactive expert feedback to drive the fine-tuning of the BERT model, and active learning is combined to screen high-value samples to achieve continuous improvement in the accuracy of hidden danger identification. A multi-dimensional reward function based on semantic similarity and hidden danger severity is designed. The reward model is trained using real-time expert scoring of the model output to quantify the recognition accuracy. With the reward model as the optimization target, the PPO (Proximal Policy Optimization) algorithm is used to fine-tune the pre-trained BERT model for multiple rounds. Active learning combines uncertainty sampling and diversity sampling strategies to give priority to log texts with low model prediction confidence and large semantic differences. Expert annotation data, reward model output, and active learning samples are jointly included in the training cycle to gradually improve model performance. Experiments show that the multi-round RLHF optimization framework significantly improves the precision and recall of hidden danger identification, can effectively deal with the scarcity of expert annotations, and shows a high coverage rate in long-tail hidden danger identification, demonstrating strong professional text semantic understanding capabilities and practical value.