通过采用交叉验证的方法，权衡网络的拟合不足和过拟合的情况。并最终，将得

By adopting the method of cross-validation, the under-fitting and over-fitting of the network are weighed. Finally, the obtained network parameters are fixed, 4 test sets are tested, and the classification of the final test set is used as an important indicator to measure the performance of the neural network classifier. Use LF, HF, LF/HF, and body movement value respectively as network input, carry on training, the training result is shown as in Fig. 5. <br> <br>(a1) BP Net (LF+Act) calculated value <br>(a2) BP Net (LF+Act) real value <br> <br>(b1) BP Net (HF+Act) calculated value <br>(b2) BP Net (HF+Act) real value <br> <br>(c1 ) BP Net (LF/HF+Act) calculated value <br>(c2) BP Net (LF/HF+Act) actual value <br> <br>(d1) GA-BP Net (LF+Act) calculated value <br>(d2) GA-BP Net (LF+ Act) actual value <br> <br>(e1) GA-BP Net (HF+Act) calculated value <br>(e2) GA-BP Net (HF+Act) actual value <br> <br>(f1) GA-BP Net (LF/HF+Act) calculated value <br>(f2) ) GA-BP Net (LF/HF+Act) true value <br>Figure 5 After the calculation results of different algorithms are <br>passed through the previous training and verification, it can be seen that the training effect of using LF/HF and body movement value as the network input is the best. Select LF/HF and body motion values ​​as the best set of result parameters in the network input, and the accuracy of the test after the model is fixed. <br>Table 2 accuracy of different algorithms <br><br>algorithm category <br>processed signal <br>accuracy of <br>this algorithm <br>BCG signal and body motion signal<br><br>Hidden Markov Model[7] <br>BCG signal <br>79.43% <br>Convolutional Neural Networks[23] <br>BCG signal <br>74% <br>Classification based on observations[11] <br> <br>Body motion signal <br>69% <br>Classifier Algorithm[24] <br>Body motion signal <br>78% <br>through Table 2, compared with other articles The analyzed signal characteristics and the algorithm used, this article is based on the BCG signal and body motion signal for feature extraction, the comprehensive recognition accuracy rate is 83.14%, and a good recognition effect is achieved. <br><br><br>Discussion and conclusion The <br>quality of sleep directly affects people's normal life and physical and mental health. In this paper, the BP neural network algorithm based on BCG signal and body motion signal has a high consistency between the discrimination results of sleep and wakefulness and the calibration, especially the correct recognition of sleep. <br>In order to improve the recognition speed, heart rate variability and body movement values ​​are selected as the characteristic values, the calculation amount is greatly reduced, and the recognition speed is improved. At the same time, the time-frequency domain characteristics of the heart rate variability are used to reduce the environmental interference caused by the direct use of the cardiac shock signal and improve the robustness of the algorithm. The use of heart rate variability and body movement values ​​can better feed back the characteristics of different stages of sleep, and the process of collecting heart shock signals and body movement signals is more convenient. Of course, when it is applied to thousands of people, the recognition error may become larger. It is necessary to continue to study the extraction of more characteristic parameters from the heart shock signal and optimize the neural network model to further improve the recognition rate of sleep staging.<br>This article uses cardiac shock signals and body motion signals to provide a simple and effective method for the identification of sleep stages. The method combined with wearable devices can be used for home sleep monitoring and can also be used as a preliminary screening for sleep diseases. It can be used as a preliminary screening diagnosis of sleep diseases in clinical medicine, and it can also facilitate users to monitor sleep quality at home; moreover, it can better solve the problem of polysomnography affecting normal sleep during the monitoring process. <br><br><br><br>references

By using a cross-validation approach, you weigh the underfit and overfitt of the network. Finally, the resulting network parameters will be fixed, 4 test sets will be tested, with the final test set classification as an important indicator to measure the performance of neural network classifiers. Using LF, HF, LF/HF, the training is carried out with the dynamic value as the network input, and the training results are shown in Figure 5.<br> <br>(a1) BP Net (LF-Act) calculation values. <br> (a2) The true value of BP Net (LF-Act).<br> <br>(b1) BP Net (HF-Act) calculation values. <br> (b2) The true value of BP Net (HF-Act).<br> <br>(c1) BP Net (LF/HF-Act) calculation values. <br> (c2) THE true value of BP Net (LF/HF-Act).<br> <br>(d1) GA-BP Net (LF-Act) calculation value. <br> (d2) The true value of GA-BP Net (LF-Act).<br> <br>(e1) GA-BP Net (HF-Act) calculation value. <br> (e2) The true value of GA-BP Net (HF-Act).<br> <br>(f1) GA-BP Net (LF/HF-Act) calculation value. <br> (f2) GA-BP Net (LF/HF-Act) true value.<br>Figure 5 The results calculated by different algorithms.<br>After the previous training and verification, it can be seen that the LF/HF and body value as the network input training effect is the best. Select the LF/HF and motion values as the best set of result parameters in the network input, and pin the model to the test accuracy.<br>Table 2 The accuracy of different algorithms.<br><br>The algorithm category. <br>Process the signal. <br>Accuracy.<br>This article algorithm. <br>BCG signals and motion signals. <br><br>Hidden Markov Model <br>BCG signal. <br>79.43%<br>Convolutional Neural Networks<br> BCG signal. <br>74%<br>Classification base on observations<br> <br>Physical signal. <br>69%<br>Classifier Algorithms <br>Physical signal. <br>78%<br>Through Table 2, the characteristic extraction of BCG signal and physical signal is based on the signal characteristics and the algorithm used in other articles, and the comprehensive recognition accuracy is 83.14%, and the recognition effect is better.<br><br>Discussions and conclusions.<br>The quality of sleep directly affects people's normal life and physical and mental health. In this paper, the bp neural network algorithm based on BCG signal and motion signal is used to determine the consistency of sleep and awakening, especially the correct recognition of sleep.<br>In order to improve the recognition speed, the heart rate variability and physical activity value were selected as the characteristic value, the calculation amount was greatly reduced, and the recognition speed was improved. At the same time, the time frequency domain characteristics of heart rate variability are used to reduce the environmental interference caused by the direct use of heart impact signals and improve the robustness of the algorithm. The heart rate variability and motion value can be used to feedback the characteristics of different stages of sleep, while the process of heart shock signal and movement signal is more convenient. Of course, when applied to thousands of individuals, the recognition error may become erred, and further research is needed to extract more feature parameters from heart shock signals and optimize neural network models to further improve the recognition rate of sleep phases.<br>In this paper, the use of heart shock signal and physical signal for the recognition of sleep stage to provide a relatively simple and effective method. This method, combined with wearable devices, can be used for home sleep monitoring and can also be used as a primary screening for sleep disorders. It can be diagnosed as a primary screening of clinical sleep diseases, and it can also be convenient for users to monitor sleep quality at home, and it can better solve the problem of multi-conductor sleep meter affecting normal sleep during monitoring.<br><br>References.

By using cross validation method, the situation of insufficient fitting and over fitting of network is balanced. Finally, the network parameters are fixed, and four test sets are tested. The classification of the final test set is used as an important index to measure the performance of neural network classifier. LF, HF and LF / HF are used as network inputs respectively with body motion value to train. The training results are shown in Fig. 5.<br>(A1) BP net (LF + act) calculated value<br>(A2) BP net (LF + act) true value<br>(B1) BP net (HF + act) calculated value<br>(B2) BP net (HF + act) true value<br>(C1) BP net (LF / HF + act) calculated value<br>(C2) BP net (LF / HF + act) true value<br>(D1) GA BP net (LF + act) calculated value<br>(D2) GA BP net (LF + act) true value<br>(E1) GA BP net (HF + act) calculated value<br>(E2) GA-BP net (HF + act) true value<br>(F1) calculated value of GA-BP net (LF / HF + act)<br>(F2) GA-BP net (LF / HF + act) true value<br>Fig. 5 calculation results of different algorithms<br>After the previous training and verification, it can be seen that the training effect of LF / HF and body dynamic value as network input is the best. LF / HF and body dynamic value are selected as the best result parameters in the network input, and the test accuracy after the model is fixed.<br>Table 2 accuracy of different algorithms<br>Algorithm category<br>Processing signals<br>Accuracy<br>The algorithm in this paper<br>BCG signal and body movement signal<br>Hidden Markov Model[7]<br>BCG signal<br>79.43%<br>Convolutional Neural Networks[23]<br>BCG signal<br>74%<br>Classification based on observations[11]<br>Body movement signal<br>69%<br>Classifier Algorithm[24]<br>Body movement signal<br>78%<br>According to table 2, compared with the signal features and algorithms used in other articles, this paper extracts features based on BCG signal and body motion signal, and the comprehensive recognition accuracy rate is 83.14%, and good recognition effect is achieved.<br>Discussion and conclusion<br>The quality of sleep directly affects people's normal life and physical and mental health. In this paper, BP neural network algorithm based on BCG signal and body motion signal is used to distinguish sleep and wakefulness, which is consistent with the calibration results, especially the correct recognition of sleep.<br>In order to improve the recognition speed, heart rate variability and body movement value are selected as the characteristic values, which greatly reduces the amount of calculation and improves the recognition speed. At the same time, the time-frequency characteristics of HRV are used to reduce the environmental interference caused by direct sampling impulse signal and improve the robustness of the algorithm. Heart rate variability (HRV) and body movement value (SV) can be used to feed back the characteristics of different sleep stages. At the same time, the acquisition process of cardiac impulse signal and body movement signal is more convenient. Of course, when it is applied to thousands of people, the recognition error may become larger. It is necessary to continue to study the extraction of more feature parameters from cardiac shock signal and optimize the neural network model to further improve the recognition rate of sleep staging.<br>This paper provides a simple and effective method for the recognition of sleep stages by heart impulse signal and body movement signal. This method, combined with wearable devices, can be used for home sleep monitoring and screening of sleep diseases. It can be used as the preliminary screening and diagnosis of sleep diseases in clinical medicine, and it can also facilitate users to monitor sleep quality at home; moreover, it can better solve the problem that polysomnography affects normal sleep in the monitoring process.<br>reference<br>