Research On Robust Myoelectric Control Method Based On Pattern Recognition

With the integration of human-computer interaction devices into every aspect of human life,the interaction between humans and machines has become closer.With the assistance of computing devices,it makes the rehabilitation of amputees,the application of AR/VR,and the deployment of industrial automation possible.As an important implementation of human-machine interaction,pattern recognition based on surface myoelectricity is highly favored for its strong anti-interference capability and good adaptability to noisy environments.In recent years,myoelectric pattern recognition has energized the field of human-computer interaction,making myoelectric control not only used in prosthetic control,but also in consumer electronics.However,consumer electronics often require higher robustness,i.e.,they are usually used across scenarios and across multiple users.Specifically,a set of control instructions is difficult to achieve natural interaction,so facing different usage scenarios requires myoelectric devices to switch instruction sets quickly;similarly,myoelectric control products often need to alternate between multiple users,and the myoelectric signal varies greatly between users due to its non-smooth random characteristics,so it needs to be switched easily and quickly between users.A robust EMG pattern recognition method should be robust to new actions and new users,but traditional EMG pattern recognition methods often face these challenges with poor user experience and severe loss of recognition rate.This paper proposes a series of solutions around robustness to new actions and new users.(1)For robust to new action pattern recognition,this study aims to develop a flexible myoelectric pattern recognition(MPR)method based on One-shot learning,which can easily switch between different usage scenarios and thus reduce the retraining burden.First,a One-shot learning model based on twin neural networks is constructed to evaluate the similarity of any given sample pair.In new scenarios involving a new set of gesture categories and/or a new user,only one sample per category is needed to form a support set.This enables the rapid deployment of a classifier suitable for the new scenario that determines any unknown query sample by selecting the category in the support set whose samples are quantified as most like the query sample.The effectiveness of the proposed method is evaluated by conducting experiments on MPR in different scenarios.The method achieves a high recognition accuracy of over 89% in cross-scene conditions and significantly outperforms other common One-shot learning methods and traditional MPR methods(p<0.01).This study demonstrates the feasibility of applying One-shot learning to rapidly deploy myoelectric pattern classifiers in response to scene changes.It provides a valuable approach to improve the flexibility of myoelectric interactions to achieve intelligent gesture control for a wide range of applications in medical,industrial,and consumer electronics.(2)The aim of this study is to develop a cross-user myoelectric pattern recognition method that is robust to new users,based on a Classification and Contrastive Semantic Alignment Loss(CCSA)function.This method improves the recognition rate of the classifier on new users by calculating the distance between data distributions in different domains,shrinking the differences between data of the same class,and expanding the distances between data of different classes.The proposed method is a model that can switch between domain generalization and domain adaptation without changing its structure.On the one hand,when there is no labeled data for the new user,a calibration dataset is constructed using the source domain data.By calculating the distribution distance between the data in the calibration dataset,a general feature space is generated to improve the generalization ability of the model on new users.On the other hand,when labeled data for the new user is available,the distribution distance between the original user and the calibration dataset is calculated.The CCSA is used to shrink the distance between the two distributions to improve the adaptability of the model to the new user.Finally,the effectiveness of the proposed method was evaluated by conducting cross-user myoelectric pattern recognition experiments in different scenarios.The proposed method achieved recognition accuracy of over 73% in various extreme cross-user conditions,significantly outperforming other transfer learning-based cross-user myoelectric pattern recognition methods(p < 0.05).This study demonstrates that applying CCSA loss can improve the generalization ability and adaptability of cross-user myoelectric pattern recognition.It provides an advanced deployment method for myoelectric interaction in the consumer electronics field.