An EEG Signal Recognition Technology Research Based On Deep Learning

From the perspective of signal processing,the brain is an organ whose structure is complex and physiological electrical signals are difficult to extract.It is also the actual“controller”of the entire central nervous system.The nervous in our brain will produce a certain amount of nerve impulses while they're subjected to different stimulates.And what the essence of the nerve impulses is some electrical signals,and the signals reflect the information in many different states in our brain.After the signals were analyzed,we could recognize the intentions of our brain roughly.Therefore,developing a research on EEG signals is not only an important part of exploring Brain Science,but also a major challenge to the Natural Science in the 21^stt century.Brain-Computer Interface?BCI?is a communication system which not rely on normal output pathways that composed of peripheral nerves and muscles^[1].It analyzes these signals through digital processing techniques by extracting the scalp EEG signals,thereby“reading”the purpose of the human brain,while also providing a communication and control channel between the user and the external facilities.It's a new type which involvessuchasNeuroscience,Computerscience,andSignalprocessing technologies,that has been widely concerned recently.Now it is mainly used in military and rehabilitation medicine and other fields.Among them,rehabilitation robots have developed rapidly which help the patients with ALS,myasthenia gravis and many other nervous system disease recover some certain control and communication skills.The BCI system will be divided into different systems which based on different stimulates produce different EEG signals,such as P300,steady-state visual evoked,and motor imagery.Among them,motion imagery BCI system takes the brain motor cortical rhythms that stimulated by motor imaging as the input signals,then through the signal processing part,which is included into the core part of the system,to determining the types of motor imaging,and then encode the motor imagery types as a control command to achieve communication and control functions of the human brain and external devices eventually.However,as the characteristics of EEG signals,such as easy disturbance,dynamic,transient,low signal-to-noise ratio,and non-stationary,the development and application of the system are facing more challenges severely.Therefore,how to effectively extract the features of EEG signals and match the best classifier is the focus of the signal processing part of the system.This dissertation based on the EEG signals of motor imagery BCI system,which develops a research mainly from the aspects of preprocessing,feature extraction and classification.What's more,the signals are analyzed and processed from time and frequency domain,deep learning network,and then be matched with the best classifier.This dissertation proposes preprocessing based on S algorithm,feature extraction by convolution neural network,feature classification by support vector machine to verify the feasibility of the algorithms by using the EEG data of international standard database and the self-acquired EEG signals.This dissertation includes the research work as follows:Firstly,the introduction is the first chapter of this dissertation.It mainly introduces the background,significance,purpose of the research,and also analyzes the development status of brain-computer interface at home and abroad in detail.At the same time,the experiences and methods that should be learned in the research,then makes a brief introduction to the work flow of brain-computer interface.Secondly,a block diagram analysis is made on the overall system module of the subject.The entire BCI system which includes four modules,the signal acquisition module,the signal processing module,an external device module and feedback module.The signal processing module is the core of the system and the focus of this dissertation.This module includes:preprocessing,feature extraction and classification.Thirdly,feature extraction of EEG signals in time-frequency domain is presented.A feature extraction algorithm based on CNN is proposed.Fourthly,the classification algorithm.This paper analyzes the best matching classifiers from support vector machine?SVM?,Bayesian classifier?BLDA?,gradient Boosting?GB?,so that achieve a high classification accuracy.Designing a 5-layer feature extraction and learning of EEG signals,and the network is adjusted via using BP feedback algorithm to reduce errors.Fifthly,the feasibility of the algorithm is verified by preprocessing,feature extraction and feature classification of EEG data provided by the international standard database.At the same time,we designed an experiment on imagining the bending of the thumb of the left hand or the right hand,utilized the method proposed in this dissertation to classify and identify,and further verified the effectiveness of the algorithm.Finally,showing a summary of this research during the graduate student's career,and putting forward deficiencies aspects for the work.