Research On Deep Learning Mechanism Of Human-computer Interaction Information And Platform Implementation

The "mind control" capability has always been in mankind’s imagination,with the continuous improvement of the computer’s performance,EEG techniques have also achieved great development in recent years and have made great progress in the fields of human-computer interaction,medical assistance,etc.Furthermore,human computer interaction researchers have developed brain interaction system prototypes that allow people to use his or her solely brain to perform various tasks.Because of recurrent neural network method and LSTM neural network,there are some limitations when processing time series data,that is,the gradie nt disappearance or the gradient explosion are easy to occur and insufficient memory.In view of EEG data processing and fine classification for low-cost devices,the mainly contributions of this thesis are as follows:(1)Based on the Stacked LSTM network,this thesis studies a classification model of EEG data for low-cost devices.First,we collect data using low-cost EEG acquisition equipment and construct multi-category EEG datasets.Then,this thesis improves the Stacked LSTM network.By adding a fully connected layer before and after the Stacked LSTM hidden layer,the model could extract multiple features.In addition,models can also get more accurate predictions by the impact of data before the moment.Finally,the thesis proves that the model can achieve nearly ninety percent of the effect in EEG data classification under low-cost equipment.(2)This thesis proposes a category prediction model based on Attention Stacked LSTM about EEG data.And we verify the model performance by experiments.After we describe the advantages of attention mechanism in dealing with EEG data from low-cost devices and its own solution to the problem of insufficient memory,we put the output value and status value of the LSTM hidden layer into the attention module separately,it could not only make up for the inferiority of LSTM network memory deficiency,but also obtain a more detailed EEG data classification effect.The experimental results also verify that the model has an advantage in comparison(such as ram and auc).And the model could achieve nearly ninety eight percent of the effect in EEG data classification.(3)Then this thesis designs an end-to-end system based on the human-computer interaction information under the attention Stacked LSTM.In our system,we describe an overarching software architecture,a hardware(robot based on TX2 depth learning and stm32 machine control),a low-cost off-the-shelf EEG headset,and a state-of-the-art attention Stacked LSTM algorithm to interpret the signals.We have thought an end-to-end wheeled robotic solution based on TX2 deep learning and stm32 machine control.The solution reads the user’s EEG signals and performs tasks at home.And it designs user interaction with a participatory design approach.We named the solution ‘embedded AI’.