Research on Lightweight Unsafe State Recognition Algorithm Based on Dual-teacher Distillation Network

This research is dedicated to the development of lightweight unsafe state recognition algorithms for new energy stations to optimize safety monitoring and real-time risk assessment. Given that traditional deep learning models are computationally redundant, have large parameters, and are not suitable for resource-constrained environments, we have constructed a lightweight CNN model by innovatively integrating integrating multi-scale modules and knowledge distillation techniques. The model accurately locates anomalies through multi-scale feature modules, and adopts a dual-teacher distillation strategy to reduce the number of parameters and storage requirements, ensuring recognition accuracy while realizing a lightweight model. This algorithm not only meets the needs of edge computing, but also follows the principle of data confidentiality, providing an efficient solution for the safety monitoring of new energy stations.