Myoelectric Pattern Recognition Research Articles

Myoelectric prosthesis hand grasp control following targeted muscle reinnervation in individuals with transradial amputation.

Despite the growing availability of multifunctional prosthetic hands, users' control and overall functional abilities with these hands remain limited. The combination of pattern recognition control and targeted muscle reinnervation (TMR) surgery, an innovative technique where amputated nerves are transferred to reinnervate new muscle targets in the residual limb, has been used to improve prosthesis control of individuals with more proximal upper limb amputations (i.e., shoulder disarticulation and transhumeral amputation). The goal of this study was to determine if prosthesis hand grasp control improves following transradial TMR surgery. Eight participants were trained to use a multi-articulating hand prosthesis under myoelectric pattern recognition control. All participated in home usage trials pre- and post-TMR surgery. Upper limb outcome measures were collected following each home trial. Three outcome measures (Southampton Hand Assessment Procedure, Jebsen-Taylor Hand Function Test, and Box and Blocks Test) improved 9-12 months post-TMR surgery compared with pre-surgery measures. The Assessment of Capacity for Myoelectric Control and Activities Measure for Upper Limb Amputees outcome measures had no difference pre- and post-surgery. An offline electromyography analysis showed a decrease in grip classification error post-TMR surgery compared to pre-TMR surgery. Additionally, a majority of subjects noted qualitative improvements in their residual limb and phantom limb sensations post-TMR. The potential for TMR surgery to result in more repeatable muscle contractions, possibly due to the reduction in pain levels and/or changes to phantom limb sensations, may increase functional use of many of the clinically available dexterous prosthetic hands.

Similarity function for one-shot learning to enhance the flexibility of myoelectric interfaces.

This study aims to develop a flexible myoelectric pattern recognition (MPR) method based on one-shot learning, which enables convenient switching across different usage scenarios, thereby reducing the re-training burden. First, a one-shot learning model based on a Siamese neural network was constructed to assess the similarity for any given sample pair. In a new scenario involving a new set of gestural categories and/or a new user, just one sample of each category was required to constitute a support set. This enabled the quick deployment of the classifier suitable for the new scenario, which decided for any unknown query sample by selecting the category whose sample in the support set was quantified to be the most like the query sample. The effectiveness of the proposed method was evaluated by experiments conducting MPR across diverse scenarios. The proposed method achieved high recognition accuracy of over 89% under the cross-scenario conditions, and it significantly outperformed other common one-shot learning methods and conventional MPR methods (p < 0.01). This study demonstrates the feasibility of applying one-shot learning to rapidly deploy myoelectric pattern classifiers in response to scenario change. It provides a valuable way of improving the flexibility of myoelectric interfaces toward intelligent gestural control with extensive applications in medical, industrial, and consumer electronics.

A Transformer-based Multi-task Learning Framework for Myoelectric Pattern Recognition Supporting Muscle Force Estimation.

Simultaneous implementation of myoelectric pattern recognition and muscle force estimation is highly demanded in building natural gestural interfaces but a challenging task due to the gesture classification accuracy degradation under varying muscle strengths. To address this problem, a novel method using transformer-based multi-task learning (MTL-Transformer) for the prediction of both myoelectric patterns and corresponding muscle strengths was proposed to describe the inherent characteristics of an individual gesture pattern under different force conditions, thereby improving the accuracy of myoelectric pattern recognition. In addition, the transformer model enabled the characterization of long-term temporal correlations to ensure precise and smooth estimation of the muscle force. The performance of the proposed MTL-Transformer framework was evaluated via experiments of classifying eleven hand gestures and estimating the corresponding muscle force simultaneously, using high-density surface electromyogram (HD-sEMG) recordings from forearm flexor muscles of eleven intact-limbed subjects. The MTL-Transformer framework yielded high classification accuracy (98.70±1.21%) and low root mean square deviation (12.59±2.76%), and outperformed other two common temporally modelling methods significantly (p < 0.05) in terms of both improved gesture recognition accuracies and reduced muscle force estimation errors. The MTL-Transformer framework is demonstrated as an effective solution for simultaneous implementation of myoelectric pattern recognition and muscle force estimation. This study promotes the development of robust and smooth myoelectric control systems, with wide applications in gestural interfaces, prosthetic and orthotic control.

Reduce the User Burden of Multiuser Myoelectric Interface via Few-Shot Domain Adaptation.

Due to physiological and anatomical variations across users, myoelectric interfaces trained by multiple users cannot be adapted to the unique hand movement patterns of the new user. Most current work requires the new user to provide one or more trials per gesture (dozens to hundreds of samples), applying domain adaptation methods to calibrate the model and achieve promising movement recognition performance. However, the user burden associated with time-consuming electromyography signal acquisition and annotation is a key factor hindering the practical application of myoelectric control. As shown in this work, once the number of calibration samples is reduced, the performance of previous cross-user myoelectric interfaces will degrade due to the lack of enough statistics to characterize the distributions. In this paper, a few-shot supervised domain adaptation (FSSDA) framework is proposed to address this issue. It aligns the distributions of different domains by calculating the distribution distances of point-wise surrogates. Specifically, we introduce a positive-negative pair distance loss to find a shared embedding subspace where each scarce sample from the new user will be closer to the positive samples and away from the negative samples of multiple users. Thus, FSSDA allows every target domain sample to be paired with all source domain samples and optimizes the feature distance between each target domain sample and the source domain samples within the same batch, instead of direct estimation of the data distribution of the target domain. The proposed method is validated on two high-density EMG datasets, which achieves the averaged recognition accuracies of 97.59% and 82.78% with only 5 samples per gesture. In addition, FSSDA is also effective even when only one sample per gesture is provided. The experimental results show that FSSDA greatly reduces the user burden and further facilitates the development of myoelectric pattern recognition techniques.

Neurostimulation artifact removal for implantable sensors improves signal clarity and decoding of motor volition.

As the demand for prosthetic limbs with reliable and multi-functional control increases, recent advances in myoelectric pattern recognition and implanted sensors have proven considerably advantageous. Additionally, sensory feedback from the prosthesis can be achieved via stimulation of the residual nerves, enabling closed-loop control over the prosthesis. However, this stimulation can cause interfering artifacts in the electromyographic (EMG) signals which deteriorate the reliability and function of the prosthesis. Here, we implement two real-time stimulation artifact removal algorithms, Template Subtraction (TS) and ε-Normalized Least Mean Squares (ε-NLMS), and investigate their performance in offline and real-time myoelectric pattern recognition in two transhumeral amputees implanted with nerve cuff and EMG electrodes. We show that both algorithms are capable of significantly improving signal-to-noise ratio (SNR) and offline pattern recognition accuracy of artifact-corrupted EMG signals. Furthermore, both algorithms improved real-time decoding of motor intention during active neurostimulation. Although these outcomes are dependent on the user-specific sensor locations and neurostimulation settings, they nonetheless represent progress toward bi-directional neuromusculoskeletal prostheses capable of multifunction control and simultaneous sensory feedback.

Multi-user myoelectric pattern recognition for shoulder motion intent detection

Myoelectric interfaces have a large range of applications in rehabilitation and prosthetic fields. The electromyographic signal (EMG) generates different patterns of activations according to a specific motion activity allowing the extraction of task-discriminant features, and leading to an efficient external device control [1]. The application of such methodologies in the motion intention detection (MID) is challenging, since this requires to predict the subjects’ intentions while attempting to generate a certain movement [2]. Despite good results have been obtained for the intra-subject scenario [3], the inter- subject generalization still remains an issue. The aim of this work was to investigate myoelectric interfaces for multi-user application on data relative to the shoulder MID. The dataset was composed by eight subjects performing six shoulder movements [4]. The signals were segmented into 150 ms windows with an overlap of 75 ms and four time-domain features were extracted. The feature matrix was split into a calibration and testing subsets. The spectral regression (SR) was used to reduce the dimension of the feature space and with the canonical correlation analysis (CCA) the reduced feature set was projected into a new space, where data related to different subjects are maximally correlated. The obtained projection coefficients were then used on the testing subset. The SVM classifier was trained on the projected calibration subset of a user A and the projected testing subset of a user B was used for testing. In this way, the classifier was trained on data belonging to a certain subject, and then tested on unseen data of a different person. Outcomes demonstrated that the presented approach boosted the classification performance, even if only one trial was used for the training session, with respect to the case in which neither the SR nor the CCA were applied. Indeed, the F1-score increased from 28% up to 53%. It can be also appreciated that increasing the number of repetitions for the calibration subset, the F1-score was further improved, up to 70% already with two repetitions. The accuracy presented a similar behaviour (Fig. 1). The coupling between SR and CCA proved to be reliable for the multi-user oriented shoulder MID. In MID problems, the movement recognition is performed relying on features extracted from a very short time period and on a transient epoch of the EMG signal, before the movement was fully achieved. For this challenging task, the present architecture proved to be reliable in an inter-subject scenario. Indeed, it provided high classification performances training on one subject and then testing on an unseen one, hence supporting the use of CCA for the development of subject-independent myoelectric interfaces.

Detecting the universal adversarial perturbations on high-density sEMG signals

Myoelectric pattern recognition is a promising approach for upper limb neuroprosthetic control. Convolutional neural networks (CNN) are increasingly used in dealing with the electromyography (EMG) signal collected by high-density electrodes due to its capacity to take full advantage of spatial information about muscle activity. However, it has been found that CNN models are very vulnerable to well-designed and tiny perturbations, such like universal adversarial perturbation (UAP). As shown in this work, the CNN-based myoelectric pattern recognition method can achieve a classification accuracy of more than 90%, but can only achieve a classification accuracy of less than 20% after the attack. This type of attack poses a big security concern to prosthetic control. To the best of our knowledge, there is no study on the detection of adversarial attacks to the myoelectric control system. In this paper, a correlation feature based on Chebyshev distance between the adjacent channels is proposed to detect the attack for EMG signals, which will serve as early warning and defense against the adversarial attacks. The performance of the detection framework is assessed with two high-density EMG datasets. The results show that our method has a detection rate of 91.39% and 93.87% for the attacks on both datasets with a latency of no more than 2 ms, which will facilitate the security of muscle–computer interfaces.

Incremental Learning and Fault-Tolerant Classifier for Myoelectric Pattern Recognition Against Multiple Bursting Interferences

Bursting interference that causes a sudden and significant change in surface electromyography (sEMG) characteristics, can reduce the stability and security of myoelectric assistive robots. Current adaptation strategies for progressive-generated interference are incapable of dealing with bursting interference. To address this problem, an incremental learning and fault-tolerant classifier (ILFTC) was proposed by combining a Gaussian mixture model (GMM) ensemble and linear discriminant analysis (LDA), in conjunction with online update and marginalization schemes. Subsequently, an ILFTC-based myoelectric pattern recognition (MPR) strategy was developed to improve the robustness of MPR against multiple interferences, including outlier motion and missing/fault data owing to electrode loosening. Experiments on hand/wrist motions were conducted to validate the anti-interference performance of the ILFTC. Experimental results showed that the ILFTC could effectively resist the two types of bursting interference and produce a significant improvement in the recognition performance over traditional classifiers, as well as the methods presented in previous studies. The results show that the proposed method has the potential to enhance the robustness of myoelectric assistive robots.

Incremental Adaptive Gesture Classifier for Upper Limb Prostheses

Myoelectric pattern recognition is widely used to control upper limb prostheses. However, the non-stationary characteristics of electromyography (EMG) signals, caused by physiological changes (e.g. muscle fatigue) or non-physiological changes (e.g. the electrode- skin impedance), hinder the use of prostheses and deteriorate the performance of the myoelectric control system. In this paper, a set of robust features is proposed to be integrated with adaptive learning techniques in order to improve the myoelectric performance. Four types of features are proposed, namely the H, HI, AI, and AIH features. The H features correspond to the histogram-oriented gradient (HOG) algorithm of the High-Density surface Electromyography map (HD-sEMG map). On the other hand, the HI features are generated by combining the H features and the intensity feature that is evaluated from the HD-sEMG map. AI is the intensity features calculated from the segmented HD-sEMG maps constructed in the individual channel. Finally, AIH features are obtained by combining the H features and AI features. Offline and online adaptive tests are conducted to evaluate the proposed features. The results show that employing the proposed AI features with an adaptive classifier improves the classification accuracy from 91.58% to 97.2% in online classification setups. Results also show that the AI features are more robust against noise than TD features. The average classification accuracy is reduced by 0.7% when additive White Gaussian noise is applied, in comparison to 5.8% reduction in accuracy when TD features is used. The results confirm the robustness of the proposed features extracted from the HD-sEMG map.

Improved swarm-wavelet based extreme learning machine for myoelectric pattern recognition

Myoelectric signal generated by muscles is one of the bio-signals which is used by humans to control equipments. To achieve this purpose, a good myoelectric pattern recognition (M-PR) is required. The applied classifier and extracted feature sets greatly affect the success of M-PR. This paper proposes a hybrid and fast classifier, extreme learning machine (ELM) which is enhanced by improved hybrid particle swarm optimization with wavelet mutation (improved swarm-wavelet). ELM is, in essence, a single-hidden layer feed-forward neural network that keeps off iterative learning to save the training time. In addition to improving the actual performance of M-PR, we also evaluate the optimization of ELM using improved swarm-wavelet in this paper. The swarm-wavelet is improved by using the particle refresh and applying velocity improvement to avoid trapped in local minima. In ELM, the improved swarm-wavelet is used to find the most suitable parameters to increase the classification accuracy. Furthermore, this paper provides comparisons of improved swarm-wavelet-ELM, swarm-wavelet-ELM and standard swarm-ELM. The experimental results show that the improved swarm-wavelet-ELM, our proposed method, is the most accurate classifier with mean accuracy of 99.6%.

Myoelectric Pattern Recognition Performance Enhancement Using Nonlinear Features.

The multichannel electrode array used for electromyogram (EMG) pattern recognition provides good performance, but it has a high cost, is computationally expensive, and is inconvenient to wear. Therefore, researchers try to use as few channels as possible while maintaining improved pattern recognition performance. However, minimizing the number of channels affects the performance due to the least separable margin among the movements possessing weak signal strengths. To meet these challenges, two time-domain features based on nonlinear scaling, the log of the mean absolute value (LMAV) and the nonlinear scaled value (NSV), are proposed. In this study, we validate the proposed features on two datasets, the existing four feature extraction methods, variable window size, and various signal-to-noise ratios (SNR). In addition, we also propose a feature extraction method where the LMAV and NSV are grouped with the existing 11 time-domain features. The proposed feature extraction method enhances accuracy, sensitivity, specificity, precision, and F1 score by 1.00%, 5.01%, 0.55%, 4.71%, and 5.06% for dataset 1, and 1.18%, 5.90%, 0.66%, 5.63%, and 6.04% for dataset 2, respectively. Therefore, the experimental results strongly suggest the proposed feature extraction method, for taking a step forward with regard to improved myoelectric pattern recognition performance.

Rejecting Novel Motions in High-Density Myoelectric Pattern Recognition Using Hybrid Neural Networks.

The objective of this study is to develop a method for alleviating a novel pattern interference toward achieving a robust myoelectric pattern-recognition control system. To this end, a framework was presented for surface electromyogram (sEMG) pattern classification and novelty detection using hybrid neural networks, i.e., a convolutional neural network (CNN) and autoencoder networks. In the framework, the CNN was first used to extract spatio-temporal information conveyed in the sEMG data recorded via high-density (HD) 2-dimensional electrode arrays. Given the target motion patterns well-characterized by the CNN, autoencoder networks were applied to learn variable correlation in the spatio-temporal information, where samples from any novel pattern appeared to be significantly different from those from target patterns. Therefore, it was straightforward to discriminate and then reject the novel motion interferences identified as untargeted and unlearned patterns. The performance of the proposed method was evaluated with HD-sEMG data recorded by two 8 × 6 electrode arrays placed over the forearm extensors and flexors of 9 subjects performing seven target motion tasks and six novel motion tasks. The proposed method achieved high accuracies over 95% for identifying and rejecting novel motion tasks, and it outperformed conventional methods with statistical significance (p < 0.05). The proposed method is demonstrated to be a promising solution for rejecting novel motion interferences, which are ubiquitous in myoelectric control. This study will enhance the robustness of the myoelectric control system against novelty interference.

Unknown Motion Rejection in Myoelectric Pattern Recognition Using Convolutional Prototype Network

Traditional myoelectric pattern recognition (MPR) systems usually perform well under laboratory conditions. However, in practice, unknown motions that do not appear in the training phase will inevitably disturb MPR systems, leading to dramatic performance degradation in motion classification. Generative and discriminative models are two representative ways to solve this problem in the literature. They only learn the patterns of known motions and reject unknown motion samples endowed with different patterns. However, the generative models often perform poorly in categorizing known motions compared with the discriminative models. While in contrast, the discriminative models always perform relatively inferiorly in distinguishing unknown motions. In this paper, a novel myoelectric control method based on the convolutional prototype network (CPN) is proposed. CPN combines the advantage of the generative model and the discriminative model in an end-to-end manner. It constructs multiple prototypes of known motion classes and identifies samples of known/unknown classes using the template (i.e., prototype) matching. The proposed method is evaluated using the high-density surface electromyography (HD-sEMG) data recorded while performing seven known and six unknown hand motions. Compared to the latest deep learning-based study, our proposed method can yield a significantly higher accuracy of 97.87% in identifying unknown motions and reduce training and testing time by more than 80%, indicating that it is a promising solution to realize robust myoelectric control in practice.

Design of an Effective Prosthetic Hand System for Adaptive Grasping with the Control of Myoelectric Pattern Recognition Approach.

In this paper, we develop a prosthetic bionic hand system to realize adaptive gripping with two closed-loop control loops by using a linear discriminant analysis algorithm (LDA). The prosthetic hand contains five fingers and each finger is driven by a linear servo motor. When grasping objects, four fingers except the thumb would adjust automatically and bend with an appropriate gesture, while the thumb is stretched and bent by the linear servo motor. Since the change of the surface electromechanical signal (sEMG) occurs before human movement, the recognition of sEMG signal with LDA algorithm can help to obtain people’s action intention in advance, and then timely send control instructions to assist people to grasp. For activity intention recognition, we extract three features, Variance (VAR), Root Mean Square (RMS) and Minimum (MIN) for recognition. As the results show, it can achieve an average accuracy of 96.59%. This helps our system perform well for disabilities to grasp objects of different sizes and shapes adaptively. Finally, a test of the people with disabilities grasping 15 objects of different sizes and shapes was carried out and achieved good experimental results.

Domain Adaptation With Self-Guided Adaptive Sampling Strategy: Feature Alignment for Cross-User Myoelectric Pattern Recognition

Gestural interfaces based on surface electromyographic (sEMG) signal have been widely explored. Nevertheless, due to the individual differences in the sEMG signals, it is very challenging for a myoelectric pattern recognition control system to adapt cross-user variability. Unsupervised domain adaptation (UDA) has achieved unprecedented success in improving the cross-domain robustness, and it is a promising approach to solve the cross-user challenge. Existing UDA methods largely ignore the instantaneous data distribution during model updating, thus deteriorating the feature representation given a large domain shift. To address this issue, a novel method is proposed based on a UDA model incorporated with a self-guided adaptive sampling (SGAS) strategy. This strategy is designed to utilize the domain distance in a kernel space as an indicator to screen out reliable instantaneous samples for updating the classifier. Thus, it enables improved alignment of feature representations of myoelectric patterns across users. To evaluate the performance of the proposed method, sEMG data were recorded from forearm muscles of nine subjects performing six finger and wrist gestures. Experiment results show that the UDA method with the SGAS strategy achieved a mean accuracy of 90.41% ± 14.44% in a cross-user classification manner, outperformed the state-of-the-art methods with statistical significance ( ). This study demonstrates the effectiveness of the proposed UDA framework and offers a novel tool for implementing cross-user myoelectric pattern recognition towards a multi-user and user-independent control.

Learning Non-Euclidean Representations With SPD Manifold for Myoelectric Pattern Recognition.

How to learn informative representations from Electromyography (EMG) signals is of vital importance for myoelectric control systems. Traditionally, hand-crafted features are extracted from individual EMG channels and combined together for pattern recognition. The spatial topological information between different channels can also be informative, which is seldom considered. This paper presents a radically novel approach to extract spatial structural information within diverse EMG channels based on the symmetric positive definite (SPD) manifold. The object is to learn non-Euclidean representations inside EMG signals for myoelectric pattern recognition. The performance is compared with two classical feature sets using accuracy and F1-score. The algorithm is tested on eleven gestures collected from ten subjects, and the best accuracy reaches 84.85%±5.15% with an improvement of 4.04%~20.25%, which outperforms the contrast method, and reaches a significant improvement with the Wilcoxon signed-rank test. Eleven gestures from three public databases involving Ninapro DB2, DB4, and DB5 are also evaluated, and better performance is observed. Furthermore, the computational cost is less than the contrast method, making it more suitable for low-cost systems. It shows the effectiveness of the presented approach and contributes a new way for myoelectric pattern recognition.

HD-sEMG Signal Denoising Method for Improved Classification Performance in Transhumeral Amputees Pros thesis Control.

Surface myoelectric pattern recognition (sMPR) based control strategy is a popularly adopted scheme for multifunctional upper limb prostheses. Meanwhile, above-elbow amputees (transhumeral: TH) usually have limited residual arm muscles, that mostly hinder the provision of requisite signals necessary for physiologically appropriate sMPR control. Hence, the need to maximally explore the limited signals to realize adequate sMPR control scheme in practical settings. This study proposes an effective signal denoising method driven by Multi-scale Local Polynomial Transform (MLPT) concept that can improve the signal quality, thus allowing adequate decoding of TH amputees' motion intent from high-density electromyogram (HD-sEMG) signals. The proposed method's performance was systematically investigated with HD-sEMG signals obtained from TH amputees that performed multiple classes of targeted upper limb movement tasks, and compared with two common signal denoising methods based on wavelet transform. The obtained results show that the proposed MLPT method outperformed both existing methods for motion tasks decoding with over 13.0% increment in accuracy across subjects. The possibility of generating distinct and repeatable myoelectric contraction patterns using the MLPT based denoised HDs-EMG recordings was investigated. The obtained results proved that the MLPT method can better denoise and aid the reconstruction of myoelectric signal patterns of the amputees. Therefore, this suggest the potential of the MLPT method in characterizing high-level upper limb amputees' muscle activation patterns in the context of sMPR prostheses control scheme.

Real-Time and Offline Evaluation of Myoelectric Pattern Recognition for the Decoding of Hand Movements

Pattern recognition algorithms have been widely used to map surface electromyographic signals to target movements as a source for prosthetic control. However, most investigations have been conducted offline by performing the analysis on pre-recorded datasets. While real-time data analysis (i.e., classification when new data becomes available, with limits on latency under 200–300 milliseconds) plays an important role in the control of prosthetics, less knowledge has been gained with respect to real-time performance. Recent literature has underscored the differences between offline classification accuracy, the most common performance metric, and the usability of upper limb prostheses. Therefore, a comparative offline and real-time performance analysis between common algorithms had yet to be performed. In this study, we investigated the offline and real-time performance of nine different classification algorithms, decoding ten individual hand and wrist movements. Surface myoelectric signals were recorded from fifteen able-bodied subjects while performing the ten movements. The offline decoding demonstrated that linear discriminant analysis (LDA) and maximum likelihood estimation (MLE) significantly (p < 0.05) outperformed other classifiers, with an average classification accuracy of above 97%. On the other hand, the real-time investigation revealed that, in addition to the LDA and MLE, multilayer perceptron also outperformed the other algorithms and achieved a classification accuracy and completion rate of above 68% and 69%, respectively.

Shoulder Motion Intention Detection Through Myoelectric Pattern Recognition

In this letter, a motion intention detection (MID) problem from surface electromyographic (sEMG) signals, involving upper limbs, was faced through a pattern recognition approach. Linear discriminant analysis (LDA) and multinomial logistic regression (MLR) were used to tackle a multiclass classification for eight healthy subjects. The sEMG signals were segmented with a window centered on movement onset. Different feature sets, i.e., engaging time domain (TD) and frequency domain (FD), were used to fit the models. Moreover, principal component analysis was employed to reduce the whole TD+FD space. In this case, both models performed satisfactorily, reaching mean accuracy of 88.8 (LDA) and 91.8% (MLR). Finally, a heuristic method is proposed to evaluate feature importance. The results presented here support the use of pattern recognition control to solve MID problems, highlighting the possibility to integrate FD features to the commonly used TD ones, as in other myoelectric pattern recognition problems, e.g., hand gesture recognition.

Metric learning for novel motion rejection in high-density myoelectric pattern recognition

Most traditional myoelectric pattern recognition systems can only identify limited patterns and are prone to be disturbed by unknown motion tasks. This paper is aimed to develop a robust myoelectric control method towards rejecting novel/unknown patterns. In the proposed method, we first convert high-density surface electromyogram (HD-sEMG) signals into a series of feature images. Next, a metric-learning guided convolutional neural network (CNN) is utilized to extract discriminative representations of the images. Compared to separable representations from common CNN, discriminative property characterizes representations in both the compact intra-class variations and separable inter-class differences. Subsequently, we train multiple autoencoders (AEs) to reject representations from any novel pattern that appeared significantly different from target patterns. The performance of the proposed method was evaluated using HD-sEMG signals recorded by two pieces of flexible 6×8 high-density electrode array placed over forearm extensors and flexors of nine subjects during performing seven target motion tasks and six complicated novel motion tasks. The proposed method can identify and reject novel patterns with high accuracy of 94.28%, which is significantly better than a widely adopted traditional method of 77.49% (p < 0.05). This work demonstrated the validity of applying metric learning in alleviating novel motion interference, which is inevitable in myoelectric control. This work will enhance the robustness of myoelectric control systems.