Research On The Extension Of Extreme Learning Machine And Its Application In EEG Classification

The brain-computer interface(BCI)is a system that allows its users to control external devices which are independent of peripheral nerves and muscles with brain activities.The BCI gathers the user's intentions by analyzing the signals generated from brain activity,and electroencephalography(EEG)has been widely used in BCI systems.Without any physical movement,EEG-based BCI systems can control a computer program.BCIs promise to revolutionize many application areas,such as text input systems,rehabilitation devices for stroke patients,new gaming input devices and systems that can react to the user's mental states and so on.The development of BCI technology has brought great convenience to the daily activities of patients with severe disabilities.The classification algorithm directly affects the practicability of BCIs,which is one of the most critical problems in BCIs.This paper focuses on the exploration of EEG classification algorithm.Compared with traditional classification algorithms,extreme learning machine(ELM)has the advantages of strong generalization ability and fast learning speed.This paper focuses on the extensions of ELM for improving its classification performance by taking ELM as the basic classifier and introducing sparse learning and active learning,and applies the proposed methods to BCIs.The main work of this paper is given as follows:(1)In order to fully learn EEG signals which has low signal-noise ratio,the FDDL-ELM method has been proposed,which owns a layer-wise framework and takes full advantages of ELM and sparse learning.Firstly,the common spatial pattern(CSP)algorithm was adopted to perform spatial filtering on raw EEG data.Secondly,the Fisher discrimination criterion was employed to learn a structured dictionary and obtain sparse coding coefficients.Then new high level features were obtained by reconstructing,and these new features are identified.The proposed method was evaluated on UCI datasets and BCI Competition Datasets ?a,? and BCI Competition ? Datasets ?a.Experiment results show that our method achieved superior performance.(2)It is expensive to annotate these raw EEG signals.In real situations,unlabeled EEG signals are widespread,whereas traditional supervised learning methods cannot obtain good performance with a small number of labeled data.To solve this problem,a novel double-criteria active learning method has been proposed,which integrates active learning to ELM and makes full use of unlabeled data.Firstly,a relatively large batch of unlabeled examples were selected with an uncertainty strategy.Then,the improved diversity strategy was used to evaluate the diversity of the unlabeled candidate data.Finally,a tradeoff parameter was introduced to select the most informative and representative data for labeling.Extensive experiments were conducted using benchmark datasets and BCI Competition ? Dataset ?a to evaluate the efficacy of the proposed method.Experimental results show that the performance of the new algorithm exceeds or matches those of several state-of-the-art active learning algorithms.