Low-Density sEMG-Based Pattern Recognition of Unrelated Movements Rejection for Wrist Joint Rehabilitation

Abstract

sEMG-based pattern recognition commonly assumes a limited number of target categories, and the classifiers often predict each target category depending on probability. In wrist rehabilitation training, the patients may make movements that do not belong to the target category unconsciously. However, most pattern recognition methods can only identify limited patterns and are prone to be disturbed by abnormal movement, especially for wrist joint movements. To address the above the problem, a sEMG-based rejection method for unrelated movements is proposed to identify wrist joint unrelated movements using center loss. In this paper, the sEMG signal collected by the Myo armband is used as the input of the sEMG control method. First, the sEMG signal is processed by sliding signal window and image coding. Then, the CNN with center loss and softmax loss is used to describe the spatial information from the sEMG image to extract discriminative features and target movement recognition. Finally, the deep spatial information is used to train the AE to reject unrelated movements based on the reconstruction loss. The results show that the proposed method can realize the target movements recognition and reject unrelated movements with an F-score of 93.4% and a rejection accuracy of 95% when the recall is 0.9, which reveals the effectiveness of the proposed method.

Keywords: autoencoder (AE); convolutional neural network (CNN); surface electromyography (sEMG); unrelated movements rejection; wrist joint rehabilitation training.

Figure 1

The overview of the system: (a) Bilateral rehabilitation based on joint pattern recognition from sEMG signals [4]; (b) The overview of the proposed method in this paper.

Figure 2

Experimental devices and data acquisition scheme: (a) Thalmic Myo armband and schematic diagram of the experimental data acquisition; (b) The seven target movements; (c) The five unrelated movements.

Figure 3

The sEMG images correspond to the target movements of 8 channels’ sEMG signals.

Figure 4

The architecture of the CNN model developed in this paper.

Figure 5

The architecture of the discriminant network using multiple autoencoders.

Figure 6

Visualization of deep features extracted from all subjects and the scattered points represent samples with colors to indicate different motions: (a–i) Visualization of deep features when $\mathsf{\lambda }$ is set as $0$, $5\times {10}^{-6}$, $1\times {10}^{-5}$, $1.5\times {10}^{-5}$, $2\times {10}^{-5}$, $1\times {10}^{-4}$, $2.5\times {10}^{-4}$, $5.0\times {10}^{-4}$, and $1\times {10}^{-3}$ respectively.

Figure 6

Visualization of deep features extracted from all subjects and the scattered points represent samples with colors to indicate different motions: (a–i) Visualization of deep features when $\mathsf{\lambda }$ is set as $0$, $5\times {10}^{-6}$, $1\times {10}^{-5}$, $1.5\times {10}^{-5}$, $2\times {10}^{-5}$, $1\times {10}^{-4}$, $2.5\times {10}^{-4}$, $5.0\times {10}^{-4}$, and $1\times {10}^{-3}$ respectively.

Figure 7

The mean F-score (a) and mean accuracy (b) of all subjects when the recall is defined as $0.85$ and $0.9$, respectively, achieved by the trained models with different $\mathsf{\lambda }$.

Figure 8

Reconstruction loss distribution achieved by the trained model: (a) Reconstruction loss distribution on the validation dataset; (b) Reconstruction loss distribution of all unrelated movements; (c) Pearson correlation(PC) distribution for each movement on the test dataset and all unrelated samples when the recall is 0.85; (d) Pearson correlation(PC) distribution for each movement of test dataset and all unrelated samples when the recall is 0.9.

Figure 8

Reconstruction loss distribution achieved by the trained model: (a) Reconstruction loss distribution on the validation dataset; (b) Reconstruction loss distribution of all unrelated movements; (c) Pearson correlation(PC) distribution for each movement on the test dataset and all unrelated samples when the recall is 0.85; (d) Pearson correlation(PC) distribution for each movement of test dataset and all unrelated samples when the recall is 0.9.

Figure 9

ROC curves are described by the CNN-AE with softmax loss and center loss and CNN-AE with softmax loss when the recall is 0.85 and 0.9.

Figure 10

Confusion matrices averaged over all subjects: (a,b) Confusion matrices for the proposed method on the test dataset and all unrelated samples when the recall is defined as 0.85 and 0.9 (b); (c,d) Confusion matrices for CNN-AE-S method on the test dataset and all unrelated samples when the recall is defined as 0.85 and 0.9; Neutral and T1–T6 denote seven target movements, whereas R1–R5 denote rejection movements.

Figure 10

Confusion matrices averaged over all subjects: (a,b) Confusion matrices for the proposed method on the test dataset and all unrelated samples when the recall is defined as 0.85 and 0.9 (b); (c,d) Confusion matrices for CNN-AE-S method on the test dataset and all unrelated samples when the recall is defined as 0.85 and 0.9; Neutral and T1–T6 denote seven target movements, whereas R1–R5 denote rejection movements.

Figure 11

Reconstruction loss distribution achieved by the CNN-AE-S. (a) Reconstruction loss distribution in the validation dataset; (b) Reconstruction loss distribution of all unrelated movements.