Abstract: To address the time-consuming, labor-intensive, and experience-dependent issues of traditional process action analysis methods in the field of industrial engineering, this paper proposes an intelligent detection method for process actions based on skeleton sequences using action recognition technology to replace traditional manual decomposition methods. A human body posture estimation model is built using a 2D camera and the MediaPipe framework to obtain skeleton sequences, and relevant evaluation metrics are introduced for action quantitative analysis. Also, a convolutional gated recurrent unit (CNN-GRU) based action classification model is trained using skeleton data. Experiments are conducted on a self-built process action dataset, demonstrating that the proposed CNN-GRU model achieves higher accuracy with fewer parameters compared with LSTM and GRU models. Furthermore, by comparing the inference results with standard operating procedures, abnormal actions are identified, providing an effective solution for process action recognition and analysis, which helps to standardize production operations and improve production efficiency.

Key words: human posture estimation, skeleton sequences, convolutional neural networks-gated recurrent unit (CNN-GRU), process action recognition