Research On Cross-Domain SEMG Gesture Recognition Based On Transfer Learning

Surface electromyography(s EMG)has been widely used in the field of gesture recognition,prosthetic control,rehabilitation and medical treatment due to its non-invasive acquisition,rich human intention information and other characteristics.However,because of the existence of user differences and other factors,current researches have poor robustness and universality,which cannot be truly applied to the market.In addition,the differences in the privacy protection and acquisition process of EMG data make different datasets exist as islands and lead to the huge resistance in data integration.Therefore,based on the idea of transfer learning and federated learning,this thesis aims to reduce the performance degradation of the motion intention perception recognition model caused by cross-domain differences between cross-users and cross-datasets.The main research contents and innovations are as follows:(1)A cross-subject gesture recognition method based on transfer learning is proposed.By extracting the Mean Absolute Value,Waveform Length,Zero crossing,Signal Slope Change and 4-order Auto-regression Coefficient and selecting the optimal feature combination,the multi-EMG feature image dataset is constructed based on the designed conversion method from one-dimensional signals to two-dimensional images using features.In six and thirteen gestures of cross-subject recognition experiments,the proposed method improves the recognition accuracy by 13.25% and7.4%,respectively,compared with optimal results obtained by five traditional classifiers.The experimental results not only confirm that this method can effectively solve the problem of performance degradation caused by cross-user differences in s EMG pattern recognition,but also verify that there is "knowledge" that can be transferred and learned from each other between the field of computer vision and s EMG pattern recognition.(2)The feasibility of two-dimensional multi-feature EMG images in cross-dataset problems is explored.Comparing the performance of one-dimensional EMG signals in cross-datasets,it is found that the results of interaction between datasets are not all positive under the one-dimensional dimension,and the data integration is difficult due to the differences of acquisition processes such as acquisition equipment and gesture types.In the two-dimensional dimension,simply through common methods in the image field can effectively solve the integration problems caused by different numbers of acquisition electrodes and the large difference of signal amplitude that cannot be handled by traditional methods,and can also improve the model recognition performance.In this thesis,ISRMyoI,Ninapro-DB4 and Ninapro-DB5 datasets are used.The experimental results between two datasets show that different datasets can be utilized to improve the recognition accuracy,which provides a new idea for s EMG data integration.(3)A gesture recognition framework based on federation-transfer learning is proposed.At first,the idea of federated learning is introduced.The server/client framework is simulated on a single machine.Then,the global model trained by federated learning is used as the benchmark model for transfer learning,which is fine-tuned through secondary training.Finally,the fine-tuned model is tested in thirteen gestures classification.Combined with the experimental results of different numbers of training epoch,it is found that taking a network model trained by one or more s EMG datasets as the benchmark model for transfer can effectively reduce the number of iteration required for model convergence.In the cross-subject experiment of thirteen gestures recognition,the classification accuracy is further improved by 1.75% compared with the experimental results in Section 3.The experimental results verify that the gesture recognition scheme can effectively solve the problem of resource waste caused by inconsistent datasets,alleviate the pressure of data integration,and improve the accuracy of gesture recognition.