KTT: Using a Kolmogorov-Arnold Network-Optimized TCN-Transformer Parallel Architecture for Hand Muscle Force Prediction

Hand muscle strength serves as an important indicator for assessing human motor function, providing crucial evidence for the diagnosis and rehabilitation of muscle injury patients. This study employs surface electromyography (sEMG) signals to predict hand muscle strength. However, these signals are complex non-stationary time series, and existing methods exhibit limitations in simultaneously analyzing the short-term dynamic fluctuations, long-term temporal dependencies, and the complex nonlinear relationships between sEMG signals and muscle strength. To address this issue, this paper proposes a Kolmogorov-Arnold Network(KAN)-optimized TCN-Transformer parallel architecture (KTT) for hand muscle strength prediction. The KTT model adopts a parallel design, utilizing the TCN and Transformer modules to respectively extract local temporal features and long-range dependencies. Moreover, it employs a cross-attention mechanism to dynamically calculate attention weights, thereby achieving deep feature fusion. Furthermore, this study introduces the KAN in the field of muscle strength prediction for the first time, replacing traditional feedforward neural network (FFN) layers to adaptively map the nonlinear coupling relationships of multi-channel signals, thereby achieving more efficient nonlinear modeling of sEMG signals. Experimental results indicate that the RMSE of the KTT model reaches 0.1934, while the R² improves to 0.9638 compared to existing methods. Ablation studies quantitatively validate the effectiveness of each module, confirming the architectural innovation of the proposed method.