| Gesture recognition based on surface electromyography(sEMG)is a crucial part of human-machine interfaces.Myoelectric control has been developed by leaps and bounds since it was put forward in the 1960s.In early years,a simple "on-off’ control strategy was used,which mainly relied on the amplitude of raw sEMG signals.Although it can guarantee relatively stable control performance,it is hard to increase the number of control degrees-of-freedom(DoFs),resulting in limited functions.Recently,the emergence of myoelectric pattern recognition(MPR)has brought new vigor and vitality to the field of myoelectric control.The MPR technology aims to automatically discover regularities in sEMG signals through machine learning algorithms and with these regularities to enable dexterous control of multiple DoFs.Many studies have reported satisfactory gesture recognition performances under laboratory conditions.However,some impediments face when transitioning the MPR technology to clinical situations.Among them,electrode shift and novel motion interference seriously restrict the practicality of the MPR.The electrode shift is inevitable during re-wearing the sEMG electrodes,and the induced variations in sEMG data may not comply with the rules learned by the previous MPR system,leading to misclassifications of the movement intentions.Besides,it is impossible to force the users to conduct only the target motion tasks,and the variability in real life makes the novel motion tasks unavoidable.As a result,frequent erroneous outputs by the traditional MPR method may provide negative feedback to users.A robust MPR method is expected to be robust to the electrode shift and has the ability to reject novel motions.Nevertheless,it isn’t easy to achieve these abilities with common MPR methods.On this basis,a series of solutions for addressing issues brought from both the electrode shift and the novelty interference are presented in this dissertation:(1)In order to reduce the influence of electrode shift,a preprocessing approach was first performed to convert high-density surface electromyogram(HD-sEMG)signals into a series of images,and the electrode shift appeared as pixel shift in these images.Then,two solutions were presented,namely,robust myoelectric control against electrode shift using data augmentation and dilated convolutional neural network,and an object-detection-based myoelectric control method against electrode shift.In the first method,a data augmentation approach was applied to the training data from just one position(no shift)to simulate HD-sEMG images derived from fictitious shift positions.The dilated convolutional neural network(DCNN)was subsequently adopted for classification.Compared to the standard convolutional neural network,DCNN always contained a larger receptive field that was supposed to be adept at mining wider spatial contextual information in HD-sEMG images.This property was further confirmed to facilitate the classification of myoelectric patterns using HD-sEMG.Furthermore,the other method was designed to identify and match partially overlapped regions between the training images of muscular activation and any given testing image derived from any unknown position,in an unsupervised manner.This approach was regarded to calibrate the electrode shift.Thus,the data within the overlapped region were used for pattern recognition,thereby improving the performance of gesture classification.The proposed methods’ performances were evaluated with HD-sEMG data recorded by a 10×10 electrode array placed over forearm extensors of ten subjects during their performance of six wrist and finger extension tasks.Under various actual electrode shift conditions,the proposed methods achieved a mean classification accuracy of more than 95%,which was significantly better than other common methods.Therefore,the proposed methods are practical solutions for robust myoelectric control against electrode array shifts.(2)To alleviate the influence brought by novel motions,two novel motion rejection methods were proposed using hybrid neural networks and the idea of metric learning,respectively.The rejection of novel motions relies on accurate descriptions of different target motions,making novel motions in significantly different appearances.The novel motion rejection method was designed within a structure of hybrid neural networks.First,a convolutional neural network(CNN)was used to extract spatio-temporal information conveyed in the HD-sEMG data.Given the target motion patterns well-characterized by the CNN,then a series of autoencoder networks were applied to learn correlation in the spatio-temporal information within samples of every target pattern respectively,where samples from any novel pattern were expected to be significantly different from those from target patterns.Subsequently,a metric-learning guided CNN was proposed to extract discriminative representations of the images.Compared to separable representations from standard CNN,discriminative property characterized representations in a trend of reduced intra-class variations and enlarged inter-class differences.Enlarging inter-class differences made the features to be sensitive to differences in HD-sEMG patterns,implying a larger distance between novel and target motions.The performances of the proposed methods were 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 motions and six complicated novel target motions.The proposed method can identify and reject novel patterns with high accuracy of more than 90%,which was significantly better than a widely adopted traditional method.This work helps to enhance the robustness of myoelectric control systems against the interference of novel motions.In this dissertation,a series of solutions for improving MPR systems’ robustness and practicability are presented and offered,by the means of applying deep learning technologies as well as specifically considering physiological characteristics of the muscle activation. |
PDF Full Text Request