Investigation into the prediction of arm joint rotation acceleration utilizing signal fusion and time-series network methodologies

In the present investigation, to enhance the precision of human movement intention detection, a dual-modality approach was proposed, integrating both surface electromyography (sEMG) and mechanomyography (MMG) signals. These signals, representing the nerve potential activity and the vibrational characteristics of muscle contractions, respectively, were utilized to train a predictive model for estimating arm joint rotational acceleration. Participants with intact shoulder joints were enrolled in this study, during which both MMG and sEMG signal were acquired using wireless sensor technology. In this research, The BiTCN-BiGRU-Attention algorithm, an integration of Bidirectional Temporal Convolutional Networks (BiTCN), Bidirectional GRU (BiGRU) architecture and Muti-Attention layer, was proposed for acceleration prediction. What’s more, the BiTCN-BiGRU-Attention algorithm was developed by combining the Black-winged Kite Algorithm (BKA) for the optimization of hyperparameters. The Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm was employed to remove random noise of both MMG and sEMG signal from the acquired data. Various methodologies were employed to substantiate the superior performance of the CEEMDAN and BKA-BiTCN-BiGRU-Attention algorithm. Utilizing comparative analyses with conventional algorithms, including backpropagation neural networks (BP), random forests (RF), and support vector machines (SVM), the BKA-BiTCN-BiGRU-Attention model demonstrated superior predictive performance, yielding a prediction accuracy with mean squared error (MSE) of 0.00153, root mean squared error (RMSE) of 0.0128, mean absolute error (MAE) of 0.0098, and a R ² of 0.990.The comparative analysis with conventional signal decomposition techniques, including Empirical Mode Decomposition (EMD), Variational Mode Decomposition (VMD), and Concurrent Ensemble Empirical Mode Decomposition (CEEMD), has revealed that the MMG and sEMG signal processed via the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm exhibit the minimum envelope entropy. This finding indicates that the resultant sub-signals derived from CEEMDAN decomposition are characterized by the lowest levels of random noise. The amalgamation of the sub-signals residing within the respective frequency band was executed, resulting in the formation of MMG and sEMG signal.