Research On Myoelectric Pattern Recognition Based On Spiking Neural Network

Due to its safety,portability and low cost,surface electromyography(sEMG)is often used to control peripheral devices after pattern recognition and is widely used in the field of myoelectric control such as prosthetic control and human-computer interaction.The decisive factor in the performance of myoelectric control systems is the myoelectric pattern recognition algorithm.With the development of artificial intelligence,myoelectric pattern recognition can be implemented by traditional machine learning,Deep neural network(DNN)and Spiking Neural Network(SNN)respectively,however,most of the current myoelectric pattern recognition algorithms suffer from poor algorithm robustness,heavy user training burden and high power consumption.To promote the application of myoelectric control technology,combining the advantages of SNN in processing temporal physiological signals,this thesis constructs various SNN-based myoelectric pattern recognition schemes starting from spike encodings,spike neurons and synaptic structures,and focuses on exploring the feasibility of SNN in reducing user training burden,improving algorithm robustness and reducing power consumption.Its main work and findings can be summarised as follows:(1)Framework for SNN-based myoelectric pattern recognition is constructed based on spike encodings,spike neurons and synaptic structures.First,based on the types of synaptic structures,a Spike Recurrent Neural Network(SRNN),a Spiking Recurrent Convolutional Neural Network(SRCNN)and a Spiking Long Short-term Memory Network(SLSTM)are designed.Secondly.Second,based on Leaky Integrate and Fire(LIF)neurons,two improved models,LIF based on voltage and current effect and Exponential LIF(ELIF),are introduced,and ELIF is identified as a suitable model in terms of spike release rate,gesture recognition accuracy and training speed.Finally,the performance of poisson encoding,delayed encoding and adaptive time-contrast encoding is compared in terms of spike release rate,gesture recognition accuracy and encoding time window,and adaptive time-contrast encoding is identified as the optimal encoding.(2)Performance testing experiments are carried out for both SRCNN and SLSTM network architectures.Specifically,gesture recognition experiments are carried out in user-independent mode,electrode-shift mode and different training test ratios using the ES-GE dataset containing 9 types of gestures and the 30-GE dataset containing 30 types of gestures as target sets respectively.The performance of SRCNN,SLSTM,CNN,LSTM and Linear Discriminant Analysis(LDA)classifiers are compared and analysed in terms of gesture recognition accuracy,number of gesture repetitions included in the training set and network computational power consumption,and we found that SNN has absolute advantages in improving model classification accuracy and reducing user training burden,and basically meets or even exceeds the robustness of existing DNN in the electrode-shift and user-independent experiments,meanwhile has obvious advantages of low power consumption and low storage resources.The research results of this thesis are valuable in promoting the practical application of myoelectric control systems.The proposed SNN-based myoelectric pattern recognition framework has various advantages such as reducing user training burden,enhancing robustness,reducing latency,reducing power consumption and reducing storage resources,and is more suitable for the implementation of real-time myoelectric control systems.