Real-time sensor data collected from Time Sensitive Networks (TSN) systems can serve as the primary source of data for identifying user ac­tion intent, providing critical support for applications such as human-computer interaction and exercise health. Therefore, this paper proposes a deep spatio­temporal neural network-based table tennis action recognition method to ad­dress the difficulty of recognizing table tennis actions in TSN systems, where the action amplitude is small, the action frequency is fast, and the action recog­nition difficulty is high. The method includes a video dataset of nine table ten­nis action techniques, and uses a frame sequence of 16 frames before and after the key action as the input to the network model, which performs rapid infer­ence calculations and identifies the action category. Experimental results demonstrate that the proposed method can recognize athlete actions in real­time, achieving an accuracy of 93.89% for the nine categories on the self-built dataset. Compared with other intelligent action recognition algorithms, the pro­posed algorithm increases classification accuracy by 2.92% to 6.37%.