A Novel Method for Hand Movement Recognition Based on Wavelet Packet Transform and Principal Component Analysis with Surface Electromyogram

As an input method of signal language, the hand movement classification technology has developed into one of the ways of natural human-computer interaction. The surface electromyogram (sEMG) signal contains abundant human movement information and has significant advantages as the input signal of human-computer interaction. However, how to effectively extract components from sEMG signals to improve the accuracy of hand motion classification is a difficult problem. Therefore, this work proposes a novel method based on wavelet packet transform (WPT) and principal component analysis (PCA) to classify six kinds of hand motions. The method applies WPT to decompose the sEMG signal into multiple sub-band signals. To efficiently extract the intrinsic components of the sEMG signal, the classification performance of different wavelet packet basis functions is evaluated. The PCA algorithm is used to reduce the dimension of the feature space composed of the features reflecting hand motions extracted from each sub-band signal. Besides, to ensure higher classification performance while reducing the dimension of the feature space by the PCA algorithm, the classification performance of different dimensions of the feature space is compared. In addition, the effects of the variability of the sEMG signal and the size of the window on the proposed method are further analyzed. The proposed method was tested on the sEMG for Basic Hand Movements Data Set and achieved an average accuracy of 96.03%. Compared with the existing research, the proposed method has better classification performance, which indicates that the research results can be applied to the fields of exoskeleton robot, rehabilitation training, and intelligent prosthesis.


Introduction
Hand movements are the most meaningful and elementary form of human daily communication and represent the intentions expressed by people [1,2]. As a signal language input method, hand movement classifcation has important theoretical research signifcance and practical application value in the feld of human-computer interaction [3]. As a result, hand movement classifcation technology that allows humans to communicate with computers more efciently, conveniently, and naturally has developed into an important part of the feld of artifcial intelligence [4,5]. Many studies have classifed hand movements in terms of computer vision and wearable sensors [6][7][8]. Compared with computer vision, sEMG signals collected by wearable sensors are an ideal source for hand motion classifcation [9,10]. Although there are many studies using sEMG signals for hand motion classifcation, how to extract the efective components of sEMG signals to achieve accurate hand motion classifcation is still a challenging problem.
According to the method of collecting data, hand movements classifcation can be divided into two categories: hand movement classifcation based on computer vision and hand movement classifcation based on wearable sensors [2]. Based on computer vision hand movement classifcation, hand movement images are captured by the camera and then the feature is obtained through image processing technology to classify the hand movements [6]. Sharma et al. [7] used a convolutional neural network to recognize images of Indian sign language gestures collected with an RGB camera.
Kumar et al. [11] proposed two viewpoint-set-up gesture classifcation methods. Teir experimental results show that compared with a single-camera system, this method has high classifcation accuracy even when simple classifers such as nearest neighbors and decision trees are used. However, the performance of hand movement classifcation based on computer vision is often afected by factors such as light intensity, shooting distance, shooting angle, and occlusion of sight [12]. Hand movement classifcation based on wearable sensors overcomes the above-mentioned problems by detecting signals generated by hand movements. Jiang et al. [13] designed a real-time gesture classifcation wristband based on the sensor fusion of sEMG and inertial measurement unit. Teir initial experimental results show that the classifcation accuracy of air and surface gestures is 92.6% and 88.8%, respectively. Wearable sensors for hand movement classifcation include accelerators, inertial measurement units, gyroscopes, and sEMG signal sensors [14][15][16][17]. Compared with the signals recorded by other wearable sensors, the sEMG signal refects the human body's electrophysiological response to various activities and has inherent advantages in predicting actions and distinguishing between passive and active activities [18]. Te sEMG signal has become the primary research approach in the felds of hand movement classifcation, activity recognition, and gait analysis [19][20][21].
Te sEMG signal is the superposition of action potentials of many motor units in time and space, which refects the body's movement intention [22,23]. Te sEMG signal is an unstable bioelectric signal with diferent frequency components at diferent moments [24]. Te representative timefrequency analysis methods of the sEMG signal include fast Fourier transform (FFT) [25], short-term Fourier transform (STFT) [26,27], Wigner-Ville distribution (WVD) [28,29] and Hilbert-Huang transform (HHT) [30], and wavelet transform (WT) [18,31,32]. Te prerequisite for FFT and STFT to efectively analyze the signal is that the signal is stable [33]. It is obvious that they cannot efectively refect the time-frequency characteristics of nonlinear and nonstationary sEMG signals. Compared with STFT, WVD can interpret the signals better [34]. However, WVD may have cross-terms that carry important information about the relationship between signal components to mask the original signal, making it difcult to interpret the time-frequency information of the signal [35,36]. Te time-frequency analysis of the signal by the HHT method consists of two steps, namely, empirical mode decomposition and the Hilbert transform. Te HHT method is suitable for nonlinear and nonstationary signal analysis, but the computational cost is higher than that of WT and other methods [24]. Te WT analyzes the local characteristics of the signal at diferent time periods and in diferent ranges by calculating the convolution of the signal and the wavelet basis function [37]. Te WPT is an important extension of the WT, which can efectively analyze the frequency components of the signal [38]. Existing research has shown that WPT with many resolution levels can efectively extract the components in the signal to obtain better classifcation results [31,39].
Feature extraction plays a vital role in the classifcation of hand movements based on sEMG signals [40]. Feature extraction is to convert the sEMG signal into a compact and information-rich feature space. Te sEMG signal feature extraction methods are generally divided into time-domain (TD) features and frequency-domain (FD) features [41]. Te TD feature is the TD statistics obtained by directly performing statistical analysis on the signal amplitude [42]. Common TD features include root mean square (RMS), variance (VAR), mean absolute value (MAV), waveform length (WL), and zero crossing (ZC) [43][44][45]. Arief et al. [46] evaluated fve TD features to fnd the best way to minimize the complexity of implementation and reduce the cost of information processing. Te FD features are extracted from the frequency spectrum of the sEMG signal [47]. According to the research of hand movement classifcation based on sEMG signals, the FD features extracted from sEMG signals mainly include median frequency (MDF), mean frequency (MNF), and mean power (MNP) [41]. Phinyomark et al. [48] proposed two modifed FD features for robust feature extraction.
To improve the classifcation performance of hand movements based on sEMG signals, we propose a novel method for hand movement classifcation. In this method, the sEMG signal is decomposed into multiple sub-band signals by WPT; the TD and FD features are extracted from the frequency band signals; the PCA is used to eliminate redundant features; and machine learning is used as a predictive model to achieve the purpose of accurately classifying hand movements. Te method proposed in this work is of great signifcance to the development of humancomputer interaction, clinical medicine, and prosthetic control. Te fow chart of this work is shown in Figure 1.
Based on the above analysis, the motivation of this research is to achieve a high-accuracy hand movement recognition method that can accurately extract efective information from sEMG signals. In this work, we adopt WPT to decompose the sEMG signal into multiple frequency band components to represent the most important information in the sEMG signal. To select a suitable wavelet basis function, we evaluated the classifcation accuracy of diferent wavelet packet functions for hand movements. MAV, RMS, MNF, and MDF were extracted from sub-band signals obtained by the WPT decomposition of sEMG signals. Te extracted features are projected into a low-dimensional space by PCA to remove unimportant features. Machine learning classifers are used to recognize hand movements and compare the corresponding recognition performance.
Te rest of the work is structured as follows: Section 2 describes the database and methods used in this work in detail. Section 3 shows the results. Section 4 discusses the proposed method. Section 5 presents the conclusion.

Related Works
Te sEMG signal data set used in this work is obtained from the publicly available sites. Te sEMG signal contains abundant information on body movements and has great potential for hand motion classifcation [49]. However, the 2 Computational Intelligence and Neuroscience analysis of sEMG signals is a challenging problem due to the fact that the acquisition of sEMG signals is afected by electrode displacement, muscle structure diferences, and muscle contraction strength [50]. Based on this challenging problem, many scholars have conducted in-depth research on the classifcation of hand movements based on sEMG signals [14,32]. Nishad et al. [8] apply tunable-Q wavelet transformbased flter-bank (TQWT-FB) for decomposition of cross covariance of sEMG signals. Kraskov entropy features are extracted from each sub-band signal. RELIEFF is used for feature ranking. Features with signifcant diferences are selected and input into the k-NN classifer for hand motion classifcation. Sapsanis et al. [10] utilized empirical mode decomposition to decompose sEMG signals into intrinsic mode functions (IMFs). Eight features (IEMG, ZC, VAR, SSC, WL, WAMP, kurtosis, and skewness) are extracted from IMFs and raw sEMG signals. A simple linear classifer was used to achieve the classifcation of the six hand movements. Ruangpaisarn et al. [51] proposed a method based on singular value decomposition and SMO to classify six basic hand movements. Te authors propose the V2M-SVD method for feature extraction of sEMG signals, and the SMO classifer is used for the classifcation of six hand movements. Yavuz et al. [52] proposed a method based on cepstral analysis to classify basic hand movements from sEMG signals. By calculating the mel-frequency cepstral coefcients (MFCCs), the cepstral analysis technique is used to extract the TD features of the sEMG signals. Te extracted feature vectors are composed of MFCCs, and then, a generalized regression neural network (GRNN) is used to classify basic hand movements. Fatimah et al. [53] proposed an automatic recognition algorithm for hand movements based on the Fourier decomposition method (FDM). Te method adopts FDM to decompose the sEMG signal into Fourier intrinsic band functions (FIBFs). Te kurtosis, entropy, and L1 norm of each FIBF are extracted as features and fed into a machine learning classifer to classify hand motion.
According to the research mentioned above, it is a challenging problem to efectively extract intrinsic components from sEMG signals to achieve accurate hand motion classifcation. Terefore, this work proposes a novel approach to improve the classifcation performance of hand motions. Te contributions of this work are presented as follows: (i) Te hand motion classifcation performance of WPT based on diferent wavelet basis functions is evaluated (ii) Te classifcation performance of the proposed method based on diferent feature space dimensions is compared (iii) Te robustness of the proposed method is analyzed (iv) Te classifcation performance of the proposed method with diferent window sizes is tested

Data Set.
Te sEMG signal data set used in this work is obtained from the following publicly available accessible URL: https://archive.ics.uci.edu/ml/datasets/ sEMG+for+Basic+Hand+movements [10]. Te signals were taken from two diferential sEMG sensors, and the signals were transmitted to a 2-channel sEMG system by Delsys Bagnoliâ handheld sEMG systems. Tere were two forearm sEMG electrodes (fexor capri ulnaris and extensor capri radialis, longus and brevis) held in place by elastic bands and the reference electrode in the middle, in order to gather information about the muscle activation. Te sEMG signal data set was collected from 5 healthy subjects (2 males and 3 females; age: 20-22 years). Te sampling frequency of the sEMG signal is 500 Hz. Te signals were band-pass fltered using a Butterworth Band Pass flter with low and high cutofs at 15 Hz and 500 Hz, respectively, and a notch flter at 50 Hz to eliminate line interference artifacts. Te subjects were asked to perform repeatedly the following six diferent hand movements: cylindrical (CY), tip (TI), hook (HO), palmar (PA), spherical (SP), and lateral (LA). Te force and speed of each hand movement are determined by the subject's willingness, and the recording time of the sEMG signal is 6 seconds. Figure 1 shows the six hand movements used in this work. Each subject repeats these hand movements 30 times. More details of the data set are introduced in the research [53][54][55]. Figure 2 shows the sEMG signal samples analyzed in this work.

Methods.
To improve the classifcation performance of hand movements based on sEMG signals, we propose a novel method for hand movement classifcation. In this method, the sEMG signal is decomposed into multiple subband signals by WPT; the TD and FD features are extracted from the frequency band signals; the PCA is used to eliminate redundant features; and machine learning is used as a predictive model to achieve the purpose of accurately classifying hand movements. Te method proposed in this work is of great signifcance to the development of humancomputer interaction, clinical medicine, and prosthetic control. Te fow chart of this work is shown in Figure 3.

Wavelet Packet Transform.
Te WPT is a sophisticated decomposition algorithm that can subdivide the high-frequency and low-frequency components of a signal [39,56]. Te defnition of WPT is as follows.
Assuming that the orthogonal scaling function φ(t) and the wavelet function ψ(t) have a two-scale relationship, where h(k) and g(k) represent the flter coefcient in multiresolution analysis.
Defne the recursive function sequence: (2) Each sub-band signal obtained by WPT will be decomposed into two sub-band signals of high and low frequency by two flters of high frequency and low frequency. Terefore, when the number of decomposition layers is the nth layer, the number of sub-band signals is 2n. Each layer of sub-band signal contains the entire frequency range of the original signal, which also refects that WPT is a sophisticated signal analysis method.

Feature Extraction.
Te feature extraction of sEMG signals is a key step in the classifcation of hand movements based on sEMG signals [40]. In this work, four features are extracted from the sub-band signals obtained by decomposing the sEMG signal by WPT to classify hand movements [41,42]. Tese features were selected on the basis of previous studies that showed their usefulness in distinguishing hand movements based on sEMG signals [2,43]. Te details of the four features are as follows.
Te MAV feature is the average value of the absolute value of the sEMG signal amplitude of the segment, which represents the energy of the sEMG signal [23]. Te expression of the MAV is as follows: where N is the window length of the sEMG signal and i is the ith sample point. Te RMS is a measure of the amplitude of the sEMG signal [23]. RMS is defned as Te MNF is the sum of the product of the sEMG power spectrum and the frequency divided by the total sum of the spectrum intensity [23]. Te expression of MNF is where f j is the frequency of the spectrum at frequency bin j, P j is the sEMG power spectrum at frequency bin j, and M is the length of the frequency bin. Te MDF is a frequency that divides the frequency spectrum into two regions with equal amplitude [23]. Te expression of MDF is

Feature Dimension Reduction.
Te PCA is a multivariate statistical method that can map high-dimensional space data to low-dimensional space and reduce the redundancy of high-dimensional space data [57]. Te core idea of PCA is to analyze the input data and project it in the direction with the least information loss and the greatest variance [49]. Te process of PCA dimensionality reduction is as follows: calculate the covariance matrix of the decentralized sample data X � x 1 , x 2 , . . . , x n , where n is the dimension of the sample data: Calculate the eigenvalues of C and the corresponding eigenvectors. Arrange the eigenvectors according to the size of the corresponding eigenvalues, and take the eigenvectors corresponding to the frst k larger eigenvalues to form a matrix P.
Reconstruct the reduced dimensionality data space: Te reduced dimensionality data space P contains most of the information of the original sample data X, which efectively simplifes the modal classifcation problem.

Classifcation.
In this work, three classifers, namely, K-nearest neighbor (KNN), support vector machine (SVM), and bagging, are used to classify hand movements. Te details of these classifers are as follows.
(1) KNN. Te principle of the KNN classifer is to fnd the K data points closest to a specifc sample point in the training set based on a certain distance measurement for a given data set, and then, the label with the most categories in the K samples is used as the label of the fnal prediction sample [58,59]. Te KNN classifer is a supervised learning algorithm and has excellent performance in various biomedical signal processing applications. In this work, the parameters of the KNN classifer include that the distance metric is Euler distance and the method to determine the label of the sample to be tested is the majority voting method.

Computational Intelligence and Neuroscience
(2) SVM. Te SVM is a supervised learning algorithm based on interval maximization and has the advantages of high computational efciency and strong generalization ability [48,60,49]. Te purpose of the SVM is to fnd the optimal hyperplane to maximize the sample interval of diferent classes of the hyperplane. For linear inseparable data, the kernel function technology is needed to map the linear inseparable feature vector to the high-dimensional linear separable feature space. In this work, the Gaussian kernel function was selected as the SVM kernel function to classify hand movements.
(3) Bagging. Bagging is implemented based on the bootstrap sampling method; that is, random sampling with replacement is performed on a given training set and the obtained m subsets of the same size are used as the new training set [53,61]. Train the basic classifcation algorithm on these m training sets to get m models. Te classifcation results of the models are voted on, and the category with the most votes is used as the classifcation result. Bagging can reduce the variance of the basic classifer to obtain a more stable and accurate classifcation performance. In this work, the decision tree is selected as the basic classifer of bagging to classify hand movements.

Performance Evaluation.
In this work, accuracy, recall, precision, and F1-score (F1) are selected to evaluate the classifcation performance of diferent classifcation models [55]. Te equations for these four indicators are given as follows: where true positive (TP) represents that the classifcation category of the model and the actual category of the sample are both positive, true negative (TN) represents that the classifcation category of the model and the actual category of the sample are both negative; false negative (FN) represents that the classifcation category of the model is negative but the actual category of the sample is positive; and false positive (FP) represents that the classifcation category of the model is positive but the actual category of the sample is negative.

Decomposition Performance of WPT.
Te amplitude of the sEMG signal generated by diferent hand movements of the same muscle is diferent. In order to accurately extract the features of the sEMG signal from diferent hand movements, WPT is used to decompose the sEMG signal. As shown in Figure 4, the sEMG signal is decomposed by the three-layer WPT to obtain the frequency band signal. Figure 4 shows that the waveforms of signals in diferent frequency bands have unique properties and efectively refect the intention of hand movements. Terefore, each frequency band signal contains rich features required for hand movement classifcation. Table 1, the efect of fve diferent wavelet packet basis functions on the accuracy of hand action classifcation is evaluated.

Wavelet Packet Basis Function Selection. As shown in
Combining the classifcation accuracy of the three classifers, the classifcation accuracy of hand movements based on the wavelet packet basis function of dmey is the highest and the classifcation accuracy of hand movements based on the wavelet packet basis function of sym3 is the lowest. Te comprehensive classifcation accuracy of fve diferent wavelet packet basis functions from high to low is dmey, fk8, coif2, db4, and sym3 (dmey > fk8 > coif2 > db4 > sym3). Te combination of dmey wavelet packet basis function and KNN classifer achieves the highest classifcation accuracy of 97.01%, and the combination of db4 wavelet packet basis function and SVM classifer achieves the lowest classifcation accuracy of 90.91%. Table 1 shows that the classifcation accuracy of hand movements based on the dmey wavelet packet basis function is obviously better than other wavelet packet basis functions, and the highest classifcation accuracy is 97.01%.

Evaluation of sEMG Signal Classifcation Performance.
In order to ensure that PCA reduces the dimensionality of the feature space while ensuring high classifcation accuracy of hand movements, the classifcation accuracy of hand movements in fve diferent low-dimensional feature spaces (the extracted feature space is reduced from 64 features to 10, 20, 30, 40, and 50 features) is evaluated. Table 2 shows the KNN classifer, the SVM classifer, and the bagging classifer achieved 96.03%, 94.50%, and 90.08% classifcation accuracy in the 30-dimensional feature space, respectively, which is better than the accuracy of the corresponding other-dimensional feature spaces. A comprehensive comparison of the classifcation accuracy of hand movements in fve diferent low-dimensional feature spaces shows that the 30-dimensional feature space achieves the best classifcation performance and, combined with the KNN classifer, has the highest classifcation accuracy of 96.03%. We also perform statistical analysis of the Kruskal-Wallis test on feature space with dimension 30 (P < 0.05 indicates the statistical diference between features) [53]. As shown in Table 3, there is a statistical diference between the 30-dimensional features obtained by PCA dimensionality reduction.
Te accuracy, recall, precision, and F1 of each hand movement are calculated to evaluate the classifcation performance of the proposed method. Table 4 shows the accuracy, recall, precision, and F1 of each hand movement based on the KNN classifer. Te hand movements of TI are classifed with the highest accuracy, recall, precision, and F1, the hand movements of CY are classifed with the lowest accuracy, recall, and F1, and the hand movements of PA are classifed with the lowest precision. Table 5 presents the classifcation performance of each hand movement based on the SVM classifer. Te hand movements of TI are detected with the highest accuracy, precision, and F1, and the hand movements of HO are detected with the highest recall. On the contrary, the hand movements of HO are detected with the lowest accuracy, precision, and F1, the hand movements of LA are detected with the lowest recall. As shown in Table 6, the hand movements of TI are classifed by the bagging classifer with the highest accuracy, recall, precision, and F1, the hand movements of CY are classifed by the bagging classifer with the lowest accuracy, precision, and F1, and the hand movements of PA classifed by the bagging classifer have the lowest recall. Tables 3-5 show that the classifer has the best classifcation performance for hand movements of TI and the worst classifcation performance for hand movements of CY.

Discussion
In this work, we use WPT to decompose the sEMG signal into multiple sub-band signals to further analyze the intention of hand movements. However, the performance of WPT to decompose the sEMG signal is afected by the wavelet basis function. Terefore, it is necessary to select a suitable wavelet basis function to provide the best classifcation performance of hand movements. As shown in Table 1, we evaluated the classifcation accuracy of hand movements of fve diferent wavelet packet basis functions (dmey, fk8, coif2, db4, and sym3). Compared with the wavelet packet basis functions of fk8, coif2, db4, and sym3, the wavelet packet basis functions of dmey have advantages in the classifcation of hand movement intentions. Tis can be explained by the research of Shi et al. [56]; that is, the waveform of the dmey wavelet is similar to the sEMG signal, and it has strong compactness and fast attenuation performance. For this reason, the dmey wavelet is capable of analyzing the small change information in the sEMG signal, which is benefcial in improving the classifcation performance of hand movements. Te MAV, RMS, MNF, and MDF were extracted from sub-band signals obtained from the three-level WPT decomposition sEMG signal. Terefore, the feature space extracted in this work contains 64 features. In order to reduce the dimension of the feature space, the PCA is applied. As shown in Table 2, the accuracy of the feature space of fve diferent dimensions is evaluated. When the feature dimension is 10, the classifcation accuracy of hand movements is the lowest, which may be caused by the lowdimensional feature space failing to efectively refect the intention of hand movements [43]. As the dimension of feature space increases, the information of hand movements contained in feature space also increases, which leads to the improvement of the classifcation accuracy of hand movements [49]. When the dimension of feature space is 30, the classifcation accuracy of hand movements reaches its highest. However, when the dimension of the feature space exceeds 30, the classifcation accuracy of hand movements decreases, which can be explained by the reduced classifcation accuracy caused by the redundancy among features of the feature space with higher dimensions [43]. When the    dimension of feature space exceeds 50, the classifcation accuracy of hand movements may be improved, but the computational complexity also increases. Terefore, a feature space with a dimension of 30 was selected to classify hand movements in this work. In addition, the statistical analysis results in Table 3 also fully prove that the feature space selected with dimension 30 can reduce redundancy among features and retain most information of hand movement intention.
To evaluate the efect of the variability of the sEMG signal on the performance of the proposed model, a robustness analysis is performed on the proposed model. Specifcally, diferent levels of noise are added to the original sEMG signal, and then, the data with noise is processed according to the proposed model in this research. Next, the perturbed data is input into the trained classifer, and the recognition accuracy of hand movements is obtained. Given that the magnitude level of the raw input data is 10 − 2 , the corresponding noise level is set to 1 × 10 − 2 , 1 × 10 − 1 , 3 × 10 − 1 , 5 × 10 − 1 , and 8 × 10 − 1 . Table 7 shows the robustness results for diferent noise levels. In Table 7, the proposed method achieves the best hand motion classifcation performance when the noise level is 1 × 10 − 2 and the hand motion classifcation performance decreases slightly with the increase of the noise level, which demonstrates that the method proposed in this research has strong robustness. Te strong robustness of the proposed method may beneft from the ability of WPT to analyze the intrinsic components of sEMG signals. In addition, Table 7 also shows that the KNN classifer has the best robustness, followed by the SVM classifer and the bagging classifer. Te KNN classifer has the strongest robustness, mainly because of the advantage of the KNN algorithm being insensitive to outliers. According to the above robustness analysis, it is concluded that the method proposed in this research has good generalization ability, can accept a wider range of data, and is suitable for real-life applications in the case of unavoidable data disturbances such as electrode displacement and muscle fatigue.
To test the classifcation accuracy of the proposed method with diferent window sizes, fve tests were performed: 200, 250, 300, 350, and 400 samples. Te average classifcation accuracy of hand movements for diferent window sizes is shown in Table 8. It can be seen that larger window sizes have better average classifcation accuracy. Tis may be attributed to the fact that the larger window size contains more hand motion information. Table 8 shows that the classifer achieves the highest classifcation accuracy in the case of 400 samples. Although a larger window size may achieve better classifcation performance, the computational cost of the proposed method is also higher. Considering the classifcation performance and computational cost, this research analyzes the proposed method based on the window length of 400 samples.
As shown in Table 9, the classifcation performance of the methods proposed in this work is compared with that of studies performed on the same data set. Akben et al. [14] applied fltering and histogram calculation to the energy values of sEMG signals, and then, the correlation between histogram values was calculated by the consistent correlation method as the features. Teir experimental results show that the cascaded-structure classifer achieves the best average classifcation accuracy of 94.72%. Iqbal et al. [17] proposed a method to classify hand movements from sEMG signals based on singular value decomposition and PCA. Tey applied singular value decomposition to sEMG signals to extract singular values and the mean and variance of the frst fve principal components to classify hand movements with an accuracy of 86.71%. Too et al. [32] evaluated the hand movement classifcation accuracy of sixteen features that were extracted from sEMG signals via discrete wavelet transform. Teir results showed that the combination of WL, MAV, enhanced WL, and enhanced MAV achieved an average accuracy of 94.22%. Bergil et al. [57] used the fourlevel symmetric WT to decompose sEMG signals and calculated the energy, mean value, standard deviation, and entropy of wavelet components as features. Te PCA was applied to feature space dimensionality reduction, and the KNN classifer achieved 94.96% average accuracy of hand movement classifcation. Compared with the existing studies, the proposed method achieves an average classifcation accuracy of 96.03%, which proves the superiority of the proposed method in the feld of hand movement classifcation.
Inevitably, there are several limitations to this work. First, this work only considers WPT as a sEMG signal decomposition method. In future work, we will apply other popular methods such as empirical mode decomposition and variational mode decomposition to the processing of sEMG signals. Second, this work only classifes six commonly used hand movements, and future work can extend this method to more hand movements. Finally, in the future, we will work on improving the accuracy by improving the proposed method, focusing on tuning and testing for applications in upper limb amputees.

. Conclusions
In this work, we propose a novel method for hand movement classifcation based on WPT. We also evaluate the efect of WPT based on diferent wavelet basis functions on hand movement classifcation, and the experimental results show that the dmey wavelet basis function has the highest classifcation accuracy. In addition, PCA is used to reduce the dimension of the feature space composed of MAV, RMS, MNF, and MDF to 30 dimensions to achieve high classifcation accuracy. Te KNN classifer was used to classify six kinds of hand movements and achieved an average classifcation accuracy of 96.03%. Compared with the classifcation performance of existing research, the proposed method has obvious advantages in classifcation accuracy. Te research results can be applied to exoskeleton robots, rehabilitation training, and intelligent prosthetics [62][63][64][65].

Disclosure
Te funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; and in the decision to publish the results.

Data Availability
Te data used to support the fndings of this study is available from the corresponding author upon request.

Conflicts of Interest
Te authors declare no conficts of interest.