Accidents often occur in the electric power operation sites due to workers not wearing safety equipment. Existing risk detection algorithms primarily focus on improving detection accuracy, however, it ignore the memory and computational costs, making them unsuitable for resource-constrained devices. Therefore, this paper proposes a lightweight network LRDN to detect whether personnel at power operation sites are wearing safety equipment correctly. It can be integrated into the system, suitable for portable intelligent terminal devices, and effectively reduce risks. To reduce model parameters and computational complexity while enhancing feature extraction and fusion, lightweight convolution and feature enhancement modules are introduced into the backbone and neck of the network. To alleviate the gradient vanishing problem and balance computational efficiency with model performance, inverted residual is introduced. Subsequently, in order to further improve the model’s ability to express input features while ensuring the lightweight of the model, and enhancing its performance and generalization ability, the Efficient Channel and Spatial Attention (ECSA) module is introduced. Experimental results show that our method achieves a mean Average Precision (mAP) of 94.91%, F1-score of 0.93. Without compromising model performance, the method reduces Floating Point Operations Per Second (FLOPs) by 29.3%, parameters by 40.9%, and model size by 39.8%. The detection efficiency of LRDN is superior to existing methods.