Sleep detection method and system based on brain muscle and electroencephalogram deep learning fusion model

Abstract

The application discloses a sleep detection method and system based on a brain-muscle-electro-oculogram deep learning fusion model, the method comprising the following steps: synchronously collecting electroencephalogram (EEG), electromyogram (EMG) and electro-oculogram (EOG) data, and respectively pre-processing the collected EEG, EMG and EOG data; performing data fusion on the characteristic data of the pre-processed EEG, EMG and EOG to obtain fused sleep characteristic signals; and inputting the sleep characteristic signals into a deep learning fusion model, and outputting a prediction result of sleep classification after performing feature extraction and feature fusion on the sleep characteristic signals by using a deep learning network. According to the application, independent signal extraction is performed on different signals, and the signals are fused, so that the problem that a large amount of CPU resources are occupied by the above-mentioned multi-layer network and the calculation method cannot be migrated to an embedded chip for separate use is solved, and the classification result accuracy is improved on the basis.

Description

Technical Field

[0001] This invention relates to the field of sleep detection technology, and in particular to a sleep detection method and system based on a deep learning fusion model of brain-muscle-eye electrocoagulation. Background Technology

[0002] Sleep plays a vital role in everyone's daily life and is closely related to the performance of daily activities. Generally, there are six sleep stages in a sleep cycle. The entire sleep process can be divided into three parts: wakefulness (W), non-rapid eye movement (NREM) sleep, and rapid eye movement (REM) sleep. NREM sleep is further divided into four stages: stage 1 is very light sleep (N1), stage 2 is light sleep (N2), stage 3 is moderate sleep (N3), and stage 4 is deep sleep (N4). In 2007, the American Academy of Sleep Medicine developed new standards that merged stages N3 and N4, which are characterized by slow waves, into a single slow-wave sleep stage.

[0003] Sleep is crucial for the body and brain to recover from fatigue during the sleep phase. However, with increasing life stress, more people are suffering from sleep-related disorders. Sleep disorders have been recognized as a public health hazard and are receiving increasing attention. Staging sleep states using various physiological signals is an effective method for objectively assessing sleep quality.

[0004] However, existing sleep detection methods often employ multi-layered networks to classify and identify electroencephalogram (EEG) signals. For example, in patent application number 202210225368.2, entitled "Multimodal Sleep Staging Method Based on Dynamic Fusion," the obtained sleep physiological electrical signals are processed using networks such as SMDC, S1DCNN, ResNets, AUX, and fusion networks for feature extraction, signal fusion, and sleep level classification. The numerous network layers require significant CPU resources, making them unsuitable for independent use on embedded microcontrollers. Furthermore, feature extraction from multi-sleep physiological electrical signal data, including EEG, EOS, and EMG data, directly relies on S1DCNN and ResNets networks for time and frequency domain feature extraction, neglecting the inherent differences in signal sources. Signals with unclear time or frequency domain features cannot be accurately identified, thus making sleep level classification based on this method prone to inaccurate results.

[0005] Therefore, in order to solve the problems of excessive CPU resource consumption by the aforementioned multi-layer networks, the inability to migrate the calculation method to embedded chips for independent use, and the inaccurate classification results, this application provides a sleep detection method and system based on a brain-muscle-electro-optic deep learning fusion model. Summary of the Invention

[0006] Therefore, the purpose of this invention is to provide a sleep detection method and system based on a deep learning fusion model of brain-muscle-eye electroencephalography. This method extracts and fuses signals independently based on different signals, solving the problem that the above-mentioned multi-layer network occupies too much CPU resources and cannot migrate the calculation method to an embedded chip for independent use. On this basis, it improves the accuracy of classification results.

[0007] To achieve the above objectives, the present invention provides a sleep detection method based on a brain-muscle-electroometry deep learning fusion model, comprising the following steps:

[0008] S1. Simultaneously acquire EEG, EMG, and EEO data, and preprocess the acquired EEG, EMG, and EEO data respectively. The preprocessing process includes:

[0009] 1. After performing time-frequency conversion on the EEG data, calculate the differential entropy of the obtained frequency band data respectively, and use the obtained differential entropy results as the feature data of the EEG signal;

[0010] 2. Decompose the electromyographic signal and synthesize electromyographic spectrum data from the decomposed results. Calculate the sample entropy of the synthesized electromyographic spectrum data as the feature data of the electromyographic signal.

[0011] 3. Decompose the electrooculogram (EOG) signal and synthesize EOG spectral data from the decomposed results. Calculate the sample entropy of the synthesized EOG spectral data as the feature data of the EOG signal.

[0012] S2. The preprocessed EEG, EMG and EOG feature data are fused to obtain the fused sleep feature signal;

[0013] S3. Input the sleep feature signals into the deep learning fusion model, and after feature extraction and feature fusion using the deep learning network, output the predicted results of sleep classification.

[0014] More preferably, in S1, after performing time-frequency conversion on the EEG data, the differential entropy of the obtained frequency band data is calculated, including the following steps:

[0015] After time-frequency conversion of the EEG data, the original EEG signal is converted into five frequency bands: delta, theta, alpha, beta, and gamma.

[0016] Calculate the power spectral density of the EEG signals for each of the five frequency bands;

[0017] The differential entropy of the obtained power spectral density is calculated using the following formula:

[0018] ;

[0019] Where P is the power spectral density of the data in each frequency band, and N represents the number of signal segments.

[0020] More preferably, in S1, the process of decomposing the electromyographic signal and synthesizing electromyographic spectrum data from the decomposed results includes:

[0021] Wavelet basis functions are used to decompose electromyographic signals into multiple scales, and specific layers are selected based on low-frequency coefficients to synthesize electromyographic spectrum data from the decomposed results.

[0022] More preferably, in S2, the fused sleep feature signal is a two-dimensional vector synthesized from preprocessed EEG, EMG, and EEO feature data; the row vectors of the two-dimensional vector are: 8-channel EEG signal, 2-channel EMG signal, and 2-channel EEO signal; wherein, the 8-channel EEG signal, with 5 signal frequency bands in each channel, totaling 40 EEG feature data, serves as the EEG row vector.

[0023] More preferably, in S3, the deep learning network includes a feature extraction layer, a feature fusion layer, and a sleep classification layer;

[0024] The feature extraction layer comprises three parallel convolutional networks, with kernel sizes set to (1, Fs), (1, Fs), (1, Fs), and (1, Fs) respectively. ), (1,16), where Fs is the sampling frequency,

[0025] The padding is set to (0, ...). ), (0, (0,0)

[0026] The three parallel convolutional networks are used to extract different features from the input sleep feature signal, and padding is used to ensure that the output of each network is the same size.

[0027] More preferably, the feature fusion layer includes four CNN convolutional blocks and an SE-Net network, wherein the CNN convolutional blocks are used for feature fusion, and the SE-Net network is used for feature calibration.

[0028] The feature calibration process of the SE-Net network includes:

[0029] The data after feature fusion is then globally pooled; after pooling, two Conv2d convolutional networks are used for classification and activation to complete feature calibration.

[0030] More preferably, the sleep classification layer consists of a Linear fully connected layer and a Softmax normalized exponential function, used to predict N sleep categories of the input data.

[0031] This invention provides a sleep detection system based on a brain-muscle-electroometry deep learning fusion model, used to implement the aforementioned sleep detection method based on the brain-muscle-electroometry deep learning fusion model, comprising:

[0032] The integrated brain-muscle-eye electroacupuncture data acquisition device is used to simultaneously acquire the wearer's electroencephalogram (EEG), electromyogram (EMG), and electrooculogram (EOG) data;

[0033] The main control processing module is used to preprocess the collected EEG, EMG and EEO data; and to fuse the processed EEG, EMG and EEO feature data to obtain the fused sleep feature signal.

[0034] The data is then input into a deep learning fusion model, where feature extraction and fusion are performed using a deep learning network, and the predicted sleep classification results are output.

[0035] In a further preferred embodiment, the data acquisition module is used to filter and amplify the acquired EEG, EMG, and EEG data before inputting them in parallel to the main control processing module via the SPI data interface.

[0036] More preferably, the main control processing module is connected to a data server, which is used to comprehensively evaluate the system's performance based on the sleep classification prediction results output by the main control processing module and the raw data of user bioelectrical signals collected.

[0037] When the data server collects a certain amount of raw bioelectrical signal data from users, it will retrain the deep learning network based on the collected bioelectrical signals of the users themselves, generate network model parameters that are more targeted to the users, and send them to the main control processing module.

[0038] The sleep detection method and system based on the deep learning fusion model of EEG, EMG, and EEG disclosed in this application have at least the following advantages compared with the prior art: In the preprocessing process, this application performs independent feature extraction and fusion based on different signals, which solves the problem of traditional sleep prediction models relying on multi-layer networks and consuming too much CPU resources. According to the method of this application, the method can be transferred to an embedded main control chip to perform sleep classification prediction, which solves the problem that sleep instruments cannot be used independently without a computer. For different signal types of EEG, EMG, and EEG, different convolution kernels and padding parameters are set on the basis of processing, which improves the accuracy of sleep classification and can achieve perfect data fusion. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the sleep detection method based on the deep learning fusion model of brain muscle electrooculography of the present invention.

[0040] Figure 2This is a schematic diagram of the data processing of the sleep detection method and system based on the deep learning fusion model of brain muscle electrooculography of the present invention.

[0041] Figure 3 This is a schematic diagram of the sleep detection system based on the deep learning fusion model of brain muscle electrooculography of the present invention.

[0042] Figure 4 This is a network structure diagram of the deep learning fusion model based on brain-muscle-electro-optical signals of the present invention.

[0043] Figure 5 This is a screenshot of the running program of the deep learning fusion model based on brain-muscle-electro-optical signals in the embodiment. Detailed Implementation

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 1 As shown, one embodiment of the present invention provides a sleep detection method based on a brain-muscle-electroometry deep learning fusion model, characterized by comprising the following steps:

[0046] S1. Simultaneously acquire EEG, EMG, and EEO data, and preprocess the acquired EEG, EMG, and EEO data respectively. The preprocessing process includes:

[0047] 1. After performing time-frequency conversion on the EEG data, calculate the differential entropy of the obtained frequency band data respectively, and use the obtained differential entropy results as the feature data of the EEG signal;

[0048] The EEG signals are concentrated in the 0.5–50 Hz range, with the 0.5–30 Hz range being the primary focus during sleep. Therefore, FFT transformation was used for time-frequency conversion, transforming the raw EEG signals into data in five frequency bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–50 Hz).

[0049] Then, the power spectral density (P) of the five frequency bands of the EEG signals is calculated. The calculation formula is: ,in It is the discrete Fourier transform value of segment n. ( )yes The conjugate function of .

[0050] Finally, the differential entropy (DE) of the EEG signal needs to be calculated to extract its feature data. The formula for calculating the differential entropy of the EEG signal is:

[0051] ;

[0052] Where P is the power spectral density and N represents the number of signal segments.

[0053] 2. Decompose the electromyographic signal and synthesize electromyographic spectrum data from the decomposed results. Calculate the sample entropy of the synthesized electromyographic spectrum data as the feature data of the electromyographic signal.

[0054] Electromyography (EMG) signals range from 0.5 Hz to 500 Hz, with the effective signal frequency representing muscle typically between 30 Hz and 125 Hz. Therefore, a three-layer decomposition using the wavelet basis function sym3 was employed, and layers D1 and D2 were selected to synthesize the EMG spectral data.

[0055] 3. Decompose the electrooculogram (EOG) signal and synthesize EOG spectral data from the decomposed results. Calculate the sample entropy of the synthesized EOG spectral data as the feature data of the EOG signal.

[0056] The frequency of the electrooculogram (EOG) signal is concentrated between 0Hz and 40Hz, with the main frequency between 0Hz and 10Hz. Therefore, a Butterworth low-pass filter is first used for filtering, and then a 7-level decomposition is performed using the wavelet basis function db4. The EOG spectrum data is synthesized by selecting D3 to D7.

[0057] S2. The preprocessed EEG, EMG and EOG feature data are fused to obtain the fused sleep feature signal;

[0058] The data fusion process involves synthesizing the characteristic signals of EEG, EMG, and EOG into a two-dimensional vector. The data of 40 EEG features (8-channel EEG signals, each EEG signal including 5 signal frequency bands), 2 EMG features, and 2 EOG features are combined into a 44-row, 7500-column two-dimensional vector, denoted by S. Then, the data S is input into a neural network for sleep staging.

[0059] S3. Input the sleep feature signals into the deep learning fusion model, and after feature extraction and feature fusion using the deep learning network, output the predicted results of sleep classification.

[0060] like Figure 4 As shown, the neural network for sleep staging consists of three parts: a feature extraction layer, a feature fusion layer, and a sleep classification layer. 1) Feature extraction layer:

[0061] The three network convolutional kernels are set to (1, ..., ...) respectively. ), (1, 、(1, The fill parameters are set to (0, ), ), (0, ), (0, );in , , Three convolutional kernels of different sizes, K1>K2>K3, are used to extract time-domain and frequency-domain features from the data. , , The padding value is the size of the convolution kernel, and the settings of the convolution kernel and the padding value satisfy the following relationship:

[0062] , where i is 1, 2, or 3;

[0063] The three parallel convolutional networks are used to extract different features from the input sleep feature signal, and padding is used to ensure that the output of each network is the same size.

[0064] Specifically, the feature extraction layer comprises three parallel Conv2D convolutional networks, with kernel sizes set to (1, Fs), (1, Fs), (1, Fs), and (1, Fs) respectively. ), (1,16), where Fs is the frequency of use, and the padding parameter is set to (0, ), ), (0, The convolution kernels are set to (0,0) and (1,16) to extract different features from the input data S. Padding is used to ensure that the output of each network is the same size. Larger convolution kernels are used to extract time-domain features, and smaller convolution kernels are used to extract frequency-domain features.

[0065] The feature data extracted by different convolutional networks are merged, and the merged data S1 is input into the feature fusion layer. The calculation formula for feature merging is: S1=Cat[Conv2D(X,Kernel1),Conv2D(X,Kernel2),Conv2D(X,Kernel3)]; Cat[*] is the connection function, Conv2D(*) is the convolutional network, and Kernel1, Kernel2, and Kernel3 represent different convolutional kernel configuration parameters;

[0066] 2) Feature fusion layer:

[0067] Step 1: The first part of the feature fusion layer consists of 4 CNN convolutional blocks for feature fusion. The fused data is S2, where each convolutional block contains four parts: Conv2D convolutional network, BatchNorm2d batch normalization layer, ReLU activation function, and MaxPool2d max pooling layer.

[0068] ;

[0069] in, b( ) represents the weight parameters of a two-dimensional CNN network. The result of the feature extraction layer, This is the output result after fusion.

[0070] The specific parameter configurations for the CNN convolution module are as follows:

[0071] The first convolutional block has 64 output channels, a kernel size of 11, a stride of 1, and padding of 5.

[0072] The second convolutional block has 32 output channels, a kernel size of 7, a stride of 1, and padding of 3.

[0073] The third convolutional block has 64 output channels, a kernel size of 5, a stride of 1, and padding of 2.

[0074] The fourth convolutional block has 128 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0075] The kernel size of the pooling layer MaxPool2d is set to (2,2), and the padding is set to (1,1).

[0076] The second part of the feature fusion layer is the SE-Net (Squeeze and Excitation Networks) network module. The features are calibrated through SE-Net, and the calibrated data is S3. Then, the S2 and S3 data are merged and input into the sleep classification layer.

[0077] The SE-Net module is implemented as follows:

[0078] Step A: The feature-fused data S2 undergoes the following process: Then output A, , This is a global pooling operation;

[0079] Step B, Data A passes through Output B: ,in and Here are two Conv2d convolutional networks, where σ is the sigmoid function. This is the ReLU activation function.

[0080] Step C, Data B after Output S3: .

[0081] 3) Sleep classification layers:

[0082] The sleep classification network consists of a Linear fully connected layer and a Softmax normalized exponential function. It predicts N sleep categories of the input data (N is the number of categories) and sends the final classification results to the data transmission and processing module. The parameters of the Linear fully connected layer are (128*4*31, N).

[0083] This invention also provides a sleep detection system based on a brain-muscle-eye electroreception (BME) deep learning fusion model, used to implement the aforementioned sleep detection method based on the BME deep learning fusion model, comprising:

[0084] The integrated brain-muscle-eye electroacupuncture data acquisition device is used to simultaneously acquire the wearer's electroencephalogram (EEG), electromyogram (EMG), and electrooculogram (EOG) data;

[0085] The integrated brain-muscle-eye electroencephalogram (EEG), electroencephalogram (EEG), and electrooculogram (EOG) acquisition device includes surface acquisition electrodes and lead wires. The surface acquisition electrodes can be worn on the user's head in the form of a headband or eye mask to acquire EEG, EMG, and EOG signals from the wearer's head.

[0086] The bioelectrical signal data collected by the surface acquisition electrode is connected to the integrated data acquisition and analysis system via lead wires, and the collected data is transmitted to the integrated data acquisition and analysis system.

[0087] It also includes three main parts: a data acquisition module, a main control processing module, and a data transmission module.

[0088] The data acquisition module includes three parts: ESD circuit, anti-aliasing filter circuit, and signal amplification and ADC circuit.

[0089] The data acquisition module and the main control processing module are connected via an SPI interface. The main control processing module controls the data acquisition module to acquire bioelectrical data through the SPI interface. The acquired bioelectrical data is used by the main control processing module to complete the sleep stage, and the sleep stage results and bioelectrical data are transmitted to the data server system through the data transmission module.

[0090] The main control processing module is used to preprocess the collected EEG, EMG and EEO data; and to fuse the processed EEG, EMG and EEO feature data to obtain the fused sleep feature signal.

[0091] The data is then input into a deep learning fusion model, where feature extraction and fusion are performed using a deep learning network, and the predicted sleep classification results are output.

[0092] More preferably, the main control processing module is connected to a data server, which is used to comprehensively evaluate the system's performance based on the sleep classification prediction results output by the main control processing module and the raw data of user bioelectrical signals collected.

[0093] When the data server collects a certain amount of raw bioelectrical signal data from users, it will retrain the deep learning network based on the collected bioelectrical signals of the users themselves, generate network model parameters that are more targeted to the users, and send them to the main control processing module.

[0094] Data server systems can operate remotely and are typically built using high-performance servers for data storage and real-time analysis.

[0095] In one specific embodiment, such as Figure 5 The figure shows the parameter list of the sleep staging network for the sleep detection method based on the deep learning fusion model of brain muscle and electrooculography of the present invention:

[0096] like Figure 5 The AI ​​network model for sleep staging based on the brain-muscle-electro-optic deep learning fusion model shown is 4.16MB (param size parameter). Therefore, this model is more suitable for deployment in embedded systems to achieve real-time sleep staging.

[0097] Combination Figure 5 As shown, the changes in feature data after feature extraction and fusion through different networks are as follows:

[0098] Step 1: Input data in parallel. Input three convolutional networks: Conv2d:1-1, Conv2d:1-2, and Conv2d:1-3. The output data size of each network is 32*44*468.

[0099] Step 2: Input the data sequentially into 4 CNN convolutional blocks. After inputting BasicConv2d: 1-4, the output data size is: 64*23*235; after inputting BasicConv2d: 1-6, the output data size is: 32*12*118; after inputting BasicConv2d: 1-8, the output data size is: 64*7*60; after inputting BasicConv2d: 1-10, the output data size is: 128*4*31.

[0100] Step 3: After inputting data into the SE-Net (SEBlock: 1-12) network, the output data size is 128*4*31;

[0101] Step 4: Input the data into the fully connected layer for classification, and the classification network outputs the probabilities of 5 sleep stages.

[0102] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A sleep detection method based on a brain muscle eye depth learning fusion model, characterized in that, Comprise the following steps: S1, synchronously collect electroencephalogram, electromyogram and electrooculogram data, and preprocess the collected electroencephalogram, electromyogram and electrooculogram data respectively, the preprocessing process comprising: S101, after time-frequency conversion of the electroencephalogram data, the obtained frequency band data is calculated for differential entropy, and the obtained differential entropy result is taken as the characteristic data of the electroencephalogram signal; the calculation of differential entropy comprises the following steps: After time-frequency conversion of the electroencephalogram data, the original electroencephalogram signal is converted into delta, theta, alpha, beta and gamma five frequency band data; The power spectral density of the electroencephalogram signal of the five frequency band data is calculated respectively; The obtained power spectral density is calculated for differential entropy by formula respectively: DE= + ; Wherein, P is the power spectral density of each frequency band data, and N represents the number of signal segments; S102. decompose the electromyogram signal, synthesize electromyogram spectrum data from the decomposed result, calculate sample entropy of the synthesized electromyogram spectrum data, and take the sample entropy as the characteristic data of the electromyogram signal; S103. decompose the electrooculogram signal, synthesize electrooculogram spectrum data from the decomposed result, calculate sample entropy of the synthesized electrooculogram spectrum data, and take the sample entropy as the characteristic data of the electrooculogram signal; S2, data fusion is performed on the characteristic data of the preprocessed electroencephalogram, electromyogram and electrooculogram to obtain fused sleep characteristic signals; S3, the sleep characteristic signal is input into a deep learning fusion model, and after feature extraction and feature fusion by using a deep learning network, the prediction result of sleep classification is output; The deep learning network comprises a feature extraction layer, a feature fusion layer and a sleep classification layer; The feature extraction layer includes three parallel convolution networks, and convolution kernels of the three networks are respectively set as (1, ), (1, ), (1, ), and padding parameters are respectively set as (0, ), (0, ), (0, ). , , The three convolution kernels are set as convolution kernels with different sizes, and > > are used for extracting time domain features and frequency domain features in data, 、 、 is a size of a padding value, and the convolution kernel and the padding value are set to meet the following relationship: = , where i is 1, 2, or 3; The three parallel convolution networks are used to extract different features in the input sleep characteristic signal, and the size of the output of each network is ensured to be the same by padding.

2. The sleep detection method based on the brain muscle eye depth learning fusion model according to claim 1, characterized in that, In S1, when the electromyogram signal is decomposed and the decomposed result is synthesized into electromyogram spectrum data, it comprises: The electromyogram signal is decomposed by using wavelet basis function, and the decomposed result is synthesized into electromyogram spectrum data according to the low-frequency coefficient of a specific layer.

3. The sleep detection method based on the brain muscle eye depth learning fusion model according to claim 1, characterized in that, In S2, the fused sleep characteristic signal is a two-dimensional vector synthesized by the characteristic data of the preprocessed electroencephalogram, electromyogram and electrooculogram; the row vectors of the two-dimensional vector are respectively: 8-channel electroencephalogram signal, 2-channel electromyogram signal and 2-channel electrooculogram signal; wherein, the 8-channel electroencephalogram signal, each channel has 5 signal frequency bands, totaling 40 electroencephalogram characteristic data as the electroencephalogram row vector.

4. The sleep detection method based on the brain muscle eye depth learning fusion model according to claim 3, characterized in that, The feature fusion layer comprises 4 CNN convolution blocks and an SE-Net network, the CNN convolution block is used for feature fusion, and the SE-Net network is used for feature calibration; The process of feature calibration by the SE-Net network comprises: The data after feature fusion is globally pooled; After classification and activation by using two Conv2d convolution networks after pooling, the feature calibration is completed.

5. The sleep detection method based on the brain muscle eye depth learning fusion model according to claim 4, characterized in that, The sleep classification layer is composed of Linear full connection layer and Softmax normalization exponential function, which is used for completing the prediction of N sleep classifications of the input data.

6. A sleep detection system based on a brain muscle electroencephalogram and electrooculogram deep learning fusion model, used to implement the sleep detection method based on the brain muscle electroencephalogram and electrooculogram deep learning fusion model in any one of claims 1-5, characterized in that, Comprise: A brain-muscle-eye integrated acquisition device for synchronously collecting electroencephalogram, electromyogram and electrooculogram data of a wearer; The main control processing module is used for pre-processing the collected electroencephalogram, electromyogram and electrooculogram data; performing data fusion on the feature data of the processed electroencephalogram, electromyogram and electrooculogram to obtain fused sleep feature signals; And input into the deep learning fusion model, use the deep learning network to extract and fuse features, and output the prediction result of sleep classification.

7. The sleep detection system based on the brain muscle eye depth learning fusion model according to claim 6, characterized in that, The data acquisition module is used for filtering and amplifying the collected electroencephalogram, electromyogram and electrooculogram data, and then inputting the data into the main control processing module through the SPI data interface in parallel.

8. The sleep detection system based on the brain muscle eye depth learning fusion model according to claim 6, characterized in that, The main control processing module is connected with a data server, and the data server is used for comprehensively evaluating the effect of system operation according to the prediction result of sleep classification output by the main control processing module and the collected original biological electric signal data of the user; When the data server collects a certain amount of original biological electric signal data of the user, the deep learning network is retrained according to the collected biological electric signal of the user, the network model parameters more targeted to the user are generated, and the network model parameters are sent to the main control processing module.

Images