Research On Feature Selection Method Of Sleep Monitoring Signal Based On Reinforcement Learning

With the development of society,the quality of sleep is decreasing year by year due to the pressure from various aspects,more and more patients are suffering from sleep disorders.Improving sleep quality and treating sleep disorders have become hot issues in current society.Sleep staging is the basis for studying the sleep structure and treating sleep disorders.And obtaining a set of efficient features is the key to sleep staging.Effective sleep signal features can not only improve the performance of the algorithm,but also provide a basis for subsequent decisions.Therefore,the feature selection of sleep signals is an important research topic.The existing end-to-end deep learning models have good performance compared with traditional machine learning methods.However,they often only focus on the results of sleep stage,and lack reasonable explanation for the features.The iterative feedback mechanism of reinforcement learning algorithm enables the model to record the action of each step while training the model,providing a basis for the explicability of the results.Therefore,this dissertation proposes a feature selection model for sleep monitoring signals based on the deep reinforcement learning strategy to effectively locate the key features in the signals while improving the sleep staging performance.Firstly,this dissertation performs data pre-processing on the sleep monitoring signals.The EEG signals from the Fpz-Cz poles and the horizontal EEG signals from the Sleep-EDF-39 dataset collected by Bob Kemp are used as study data in this dissertation.The impact of large changes in the feature variables in the sleep monitoring signals on the model is reduced by data normalization operations.And the overlapping sliding windows are used to divide the monitored signals and construct the minimum units of the features while maintaining the time dependence.After that,the EEG and EYE signals are fused by vector-based and label-based units respectively,and divided into four datasets.The experimental results show that the fused signals by vector-based units are more effective.Secondly,this dissertation constructs a deep reinforcement learning-based sleep monitoring signals feature selection model,the Sleep Data Distilled(SD-Distilled)model,addresses the common problem of ignoring key connectivity features in feature selection.The model is divided into three parts: Policy Network(PNet),Subset Optimization Model(SOM)and Classification Network(CNet).PNet takes a random strategy to extract the actions corresponding to each environment to generate an action sequence for the current sleep signal.SOM converts action sequences into structured representations.CNet performs classification based on the structured representation generated by SOM and provides reward calculation for PNet.The SOM selects only the important signal features that are relevant to the task.The model transforms the feature selection problem into a sequential decision problem,thus solving the feature selection problem using the Policy Gradient method in reinforcement learning.Finally,this dissertation organizes and analyzes the action selection sequences in the policy network,extracts the action operations,establishes the correspondence between the actions and the original signal data and performs feature selection.After that,the selected features are integrated according to the time dependence,short data feature sequences are eliminated,and adjacent features are merged.Then,the final obtained data features are analyzed and visually displayed.The experimental results show that the sleep staging accuracy of the method in this dissertation is 86.43%,which is better than the sleep staging performance of other models.It is clear from the final feature display that the EEG signal is more effective in distinguishing the N1,N2,and N3 phases than the EEG signal,and the EEG signal is more effective in distinguishing the NREM and REM phases.The model proposed in this dissertation can identify close to artificially labeled sleep signal features,which can be a good aid for subsequent research on the treatment of sleep disorders.