Abstract: Power personnel behavior recognition is a critical component for the safe operation and maintenance of the power system. However, current personnel behavior recognition algorithms, which primarily rely on support vector machines and multi-layer perceptrons for behavior classification, have a number of shortcomings, including low recognition accuracy, insufficient consideration of the interactions between human skeletons, and poor mobility and universality. To address these challenges, we proposes a novel behavior classification decoder based on a self-attention and cross-attention mechanism, which fully considers the associations between human skeletons. Compared to the traditional classification methods, the proposed approach improves the classification accuracy by approximately 10%~20%, and outperforms the deep learning MLP classification methods by more than 2%. To implement behavior recognition, we use a two-stage encoder-decoder architecture method, which has good extensibility while making the decoder suitable for the back end of any pose estimation network. Additionally, we use a grouped decoding method to overcome the quadratic complexity induced by the attention mechanism, which enables the decoder to extend to include more behavior categories, thus being more universal. The proposed behavior recognition algorithm achieves the optimal recognition effect in the personnel image data set based on the substation working scenarios. The comprehensive recognition rate reaches 91.1%, which verifies the efficacy and practicality of the proposed power personnel behavior classification method.