Lightweight EEG Classification Model Based on EEG-sensor with Few Channels

In the traditional work, based on EEG Sensors monitoring the physiological and emotional state of humans, the time-frequency analysis method is usually used to extract the characteristics of the signal, and then analyzed by SVM, KNN and other methods. However, the recognition accuracy of such methods is not ideal. Deep learning methods such as CNN and RNN have become the research hotspots in the field of EEG analysis. However, due to the high structural risk of the deep learning model, the model is prone to poor generalization ability, long training time, over-fitting, and poor real-time performance. In addition, current EEG data acquisition devices generally require dozens of signal channels to provide a data foundation for accurate analysis of subsequent brain signals, which not only brings high hardware costs, but also is not portable enough to be popularized in daily life. However, due to the weak and susceptible interference of the EEG signal, reducing the number of signal channels under the premise of using the above method may further lead to a worse analysis effect. In order to solve these problems, we propose a new lightweight EEG classification model based on a small number of channel EEG sensors, called Armaiti. Armaiti first uses the 5-channel EEG signal collected by the portable EEG acquisition device to perform blind source signal separation to obtain data from multiple sources, including noise signals such as EOG. Instead of identifying and screening these signals, we use a combination of EEG, EOG and other signals. Then Armaiti performs wavelet packet transform on each signal source, decomposes the signal into different frequency bands, and then inputs it into the lightweight convolutional neural network model designed in this paper to obtain five classifiers. Finally, Armaiti uses the ensemble learning to get the final classification results. In order to verify the performance and practicability of the model, from September 2018 to June 2019, we used 5 channel portable EEG acquisition equipment to collect data on EEG in different scenarios, and built fatigue data sets and applications. The prototype system analyzes the accuracy, training time, loss, precision, recall, ROC (Receiver operating characteristic) and other indicators. The experimental results show that Armaiti exhibits better performance: according to the characteristics of manual extraction, K-means is used for clustering, and the final accuracy rate is 65.6% according to the centroid. The final accuracy rate using SVM to classify is 74.7%. Using the convolutional neural network to analyze EEG signals, the accuracy rate was 80.1%, while the accuracy of Armaiti was 96.4%. In addition, the traditional CNN model has 5.8 times the average training time per round compared to Armaiti.