Sleep staging method based on multi-view fusion and improved dense connection network

Abstract

The invention relates to a sleep staging method based on multi-view fusion and an improved dense connection network. The sleep staging method comprises the following steps: acquiring single-channel sleep EEG data as an initial data set; preprocessing the initial data set to form a three-channel two-dimensional time-frequency spectrogram as a target data set; constructing an improved dense connection network as a sleep staging network model by taking the dense connection network as a reference; and sending the target data set into the constructed network model to complete sleep staging discrimination. In order to solve the problems that sleep electroencephalogram frequency components are complex, weak dynamic information is difficult to capture and individual differences of subjects are large, a two-dimensional time-frequency spectrogram generated by single-channel EEG signals serves as a main feature information channel, corresponding spectrograms are generated with the help of FNS and LNDP methods, and local dynamic information representation of the EEG signals is enhanced. And the improved dense connection network can perform more comprehensive spectrogram information perception, so that the sleep staging performance is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram signal processing, and in particular to a sleep staging method based on multi-view fusion and improved densely connected network. Background Art

[0002] Sleep is a vital physiological process in daily life and has a profound impact on personal physical and mental health. The monitoring and tracking of sleep quality has attracted widespread attention, and sleep staging is an important means of evaluating sleep quality.

[0003] The manual sleep scoring process by experts is time-consuming, labor-intensive and subjective. Currently, there are many studies on automatic sleep staging based on single-channel EEG signals. Currently, it is mainly divided into two categories: one is to use specific algorithms to extract manually designed features from EEG signals. Although it can capture relevant information, it requires a deep understanding of sleep signals, and the extracted features may lead to poor generalization performance of the staging model due to individual differences; the other is to automatically learn EEG features through deep learning models. This method has good expressive power for features and avoids complex feature engineering. However, sleep staging technology still faces challenges, such as large individual differences among subjects and limited ability of the model to represent local hidden dynamics of sleep EEG. Summary of the invention

[0004] Based on the above-mentioned problems existing in the prior art, the purpose of an embodiment of the present invention is to provide a sleep staging method based on multi-view fusion and improved densely connected network, with the two-dimensional time-frequency spectrum generated by a single-channel EEG signal as the main feature information pathway, and the spectrogram generated based on the FNS and LNDP methods to assist in enhancing the local dynamic representation of the EEG signal, so that the improved densely connected network can perform more comprehensive spectrogram information perception, thereby improving the sleep staging performance.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is: a sleep staging method based on multi-view fusion and improved dense connection network, including:

[0006] S1, obtain single-channel sleep EEG data as the initial data set;

[0007] S2, preprocessing the initial data set to form a three-channel two-dimensional time-frequency spectrum as a target data set;

[0008] S3, based on the densely connected network, an improved densely connected network is constructed as a sleep staging network model;

[0009] S4, sending the target data set into the constructed network model to complete the sleep stage discrimination.

[0010] Further, in S2, the initial data set is preprocessed to form a three-channel two-dimensional time-frequency spectrum as a target data set, including:

[0011] Step S21, using a 0.5-45 Hz bandpass Butterworth filter to filter the original EEG signal to remove background noise and power line interference;

[0012] Step S22, transforming the filtered single-channel EEG signal into a two-dimensional time-frequency spectrum as the main feature information pathway of the sleep staging network model;

[0013] Step S23, using the FNS method to process the filtered EEG sequence into an FNS sequence, and using the two-dimensional time-frequency spectrum of the FNS sequence as the second input path of the sleep staging network model to assist the network model in capturing hidden dynamic information in the sleep EEG signal;

[0014] Step S24, using the LNDP method to process the filtered EEG sequence into an LNDP sequence, and using the two-dimensional time-frequency spectrum of the LNDP sequence as the input path three of the sleep staging network model to assist the network model in capturing local changes in the EEG signal.

[0015] Furthermore, the FNS method constructs a structure function and calculates its cumulative average within a sliding window to generate a dynamic time series for characterizing the hidden dynamic information in the EEG signal.

[0016] Furthermore, the LNDP method constructs a new time series by comparing the values ​​of adjacent signal points and performing binarization, which is used to characterize the local changes of the time series and express the key information in the weak signal.

[0017] Furthermore, in S3, the sleep staging network model includes a shallow feature extraction module, a deep feature extraction module and a classifier module. The shallow feature extraction module includes a convolution layer and a maximum pooling layer, the deep feature extraction module includes 4 dense attention blocks and 3 improved transition layers, and the classifier module includes a fully connected layer and a softmax layer. The fully connected layer processes the features obtained by the feature extraction module into one-dimensional features, and then the softmax layer maps the one-dimensional features into the probabilities of each sleep stage to obtain the final sleep staging recognition result.

[0018] Furthermore, the dense attention block of the deep feature extraction module is composed of several dense attention layers, and the dense attention layer includes a dense layer and a channel attention CAM module.

[0019] Furthermore, the CAM module consists of an average pooling layer, a convolutional layer and an activation function layer, which is used to extract the correlation between different channels in the feature map so that the network can focus on more effective information.

[0020] Furthermore, the sleep staging network model uses a cross entropy function as a loss function during training, and adopts a stochastic gradient descent method to optimize the parameters of the network model to reduce random errors, and the learning rate is set to 1e-4.

[0021] The beneficial effects of the present invention are as follows: the sleep staging method based on multi-view fusion and improved densely connected network of the present invention uses the two-dimensional time-frequency spectrum generated by a single-channel EEG signal as the main feature information pathway, and uses the FNS and LNDP methods to enhance the local dynamic information representation of the EEG signal, so that the sleep staging network model can capture the complex patterns in the EEG and the correlation information between pathways; the channel attention mechanism CAM is integrated into the densely connected network, and the dense connection method unique to the densely connected network is used to realize the transmission and reuse of features, and at the same time, the features of different channels are weighted to enhance the attention to more useful channels, thereby improving the performance of the sleep staging model. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0023] In the figure:

[0024] Figure 1 A flowchart of a sleep staging method based on multi-view fusion and improved densely connected network provided in Embodiment 1 of the present invention;

[0025] Figure 2 An algorithm flow chart of the FNS method provided in the first embodiment of the present invention;

[0026] Figure 3 An algorithm flow chart of the LNDP method provided in Embodiment 1 of the present invention;

[0027] Figure 4 A schematic diagram of a dense connection mechanism of a densely connected network provided in the first embodiment of the present invention;

[0028] Figure 5 A schematic diagram of the structure of a transition layer provided in Embodiment 1 of the present invention;

[0029] Figure 6 A schematic diagram of the structure of an improved densely connected network constructed according to the first embodiment of the present invention;

[0030] Figure 7 A schematic diagram of the structure of a dense attention layer provided in Embodiment 1 of the present invention;

[0031] Figure 8 A schematic diagram of the structure of the channel attention module CAM provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] The embodiment of the present invention provides a sleep staging method based on multi-view fusion and improved dense connection network, including: obtaining single-channel sleep EEG data as an initial data set; preprocessing the initial data set to form a three-channel two-dimensional time-frequency spectrum as a target data set; constructing an improved dense connection network as a sleep staging network model based on a dense connection network; sending the target data set into the constructed network model to complete sleep staging discrimination. The sleep staging method based on multi-view fusion and improved dense connection network of the present invention targets the problems of complex frequency components of sleep EEG, difficulty in capturing weak dynamic information, and large individual differences among subjects. The two-dimensional time-frequency spectrum generated by the single-channel EEG signal is used as the main feature information pathway, and the corresponding spectrogram is generated by means of FNS and LNDP methods to enhance the local dynamic information representation of the EEG signal, so that the dense connection network can perform more comprehensive spectral information perception. In addition, the channel attention mechanism is integrated into the dense connection network to further enhance the information interaction between network channels, thereby improving the sleep staging performance.

[0034] The following is a detailed description of the implementation details of the sleep staging method based on multi-view fusion and improved densely connected network in this embodiment. The following content is only for the convenience of understanding the implementation details, which is not necessary for the implementation of this solution. The specific process of this embodiment is as follows: Figure 1 shown.

[0035] Step S1, obtaining single-channel sleep EEG data as an initial data set.

[0036] Specifically, sleep EEG data refers to the electroencephalogram data recorded during an individual's sleep. The Fpz-Cz channel EEG data of 20 subjects in the public Sleep-EDF dataset were used, with a sampling frequency of 100 Hz and a signal length of 30 seconds per segment. The sleep staging experiment of the proposed method was conducted on the sleep EEG data of WAKE, N1, N2, N3 and REM stages.

[0037] Step S2, preprocessing the initial data set to form a three-channel two-dimensional time-frequency spectrum as a target data set.

[0038] Specifically, the specific steps of preprocessing the initial data set to form a three-channel two-dimensional time-frequency spectrum as the target data set are as follows:

[0039] Step S21, using a 0.5-45 Hz bandpass Butterworth filter to filter the original EEG signal to remove background noise and power line interference.

[0040] Step S22, transforming the filtered single-channel EEG signal into a two-dimensional time-frequency spectrum diagram as the main feature information pathway of the sleep staging network model.

[0041] Step S23, using the FNS method to process the filtered EEG sequence into an FNS sequence, and using the two-dimensional time-frequency spectrum of the FNS sequence as the second input path of the sleep staging network model to assist the network model in capturing the hidden dynamic information in the sleep EEG signal.

[0042] Specifically, if Figure 2 As shown, the FNS method is used to process the filtered EEG sequence into an FNS sequence, and the two-dimensional time-frequency spectrum of the FNS sequence is used as the second path of the sleep staging network model input, which assists the network model in capturing the hidden dynamic information in the sleep EEG signal. The specific steps include:

[0043] Step S231, using the second-order difference moment to construct the structure function Ω(τ), the formula is:

[0044]

[0045] Among them, Ω(τ) is the structure function, x(t) is the filtered EEG signal, T is the window length, T = 2s, τ is the time lag, τ∈(0,T / 2].

[0046] The autocorrelation function formula of the signal x(t) is:

[0047]

[0048] Among them, ψ(τ) is the autocorrelation function, x(t) is the filtered EEG signal, T is the window length, T = 2s, τ is the time lag, τ∈(0,T / 2].

[0049] The structure function is further transformed into the formula:

[0050] Ω(τ)=2[ψ(0)-ψ(τ)]

[0051] Among them, Ω(τ) is the structure function, ψ(τ) is the autocorrelation function, and ψ(0) is the autocorrelation function of the signal x(t) when there is no time lag. That is, the calculation of the structure function is converted into the calculation of the autocorrelation function.

[0052] Step S232, calculate the cumulative average of the structure function, the calculation formula is:

[0053]

[0054] Among them, Q k is the cumulative average of the structure function corresponding to the kth time window, α is the sliding window factor and α∈(0,0.5], here α=0.13, k∈[1,L], L is the number of segments after the signal sliding window.

[0055] Step S233, generate a dynamic time series. In order to describe the dynamic changes hidden in the EEG signal during the sleep staging task, the data of this dynamic change is reconstructed based on the structure function within a limited time window. Let ΔT be the sliding step of the time window, where ΔT is preset to 0.05s, then the dynamic time series C = {C k}(k=1,…,L-1), the specific calculation formula is:

[0056]

[0057] Among them, C k is a dynamic time series, Q k+1 is the cumulative average of the structure function corresponding to the k+1th time window, Q k is the cumulative average of the structure function corresponding to the kth time window, ΔT is the sliding step of the time window, and T is the window length.

[0058] Step S24, as Figure 3 As shown in the figure, the filtered EEG sequence is processed into an LNDP sequence using the LNDP method, and the two-dimensional time-frequency spectrum of the LNDP sequence is used as the input path of the sleep staging network model to assist the network model in capturing local changes in the EEG signal. The LNDP method constructs a new time series by comparing the values ​​of adjacent points of the signal and performing binarization to characterize the local changes of the time series, so that the key information in the weak signal can be expressed. The specific steps are as follows:

[0059] Step S241, set the number of adjacent points m, take m = 8; then for signal point S Cn (n=1,…,N and N=3000), select m / 2 adjacent points in the forward and reverse directions, denoted as p={p m ,p m-1 ,…,p1,p0}, the missing points of boundary points in the forward and reverse directions are padded with zeros.

[0060] Step S242, calculate the difference between the selected m+1 consecutive points, i.e., d i =p i -p i+1 ,i=0,1,…,m-1.

[0061] Step S243, calculate signal point S Cn The LNDP value is denoted as Z n , then we get the LNDP sequence Z = {Z n}(n=1,…,N and N=3000), specifically:

[0062]

[0063] Among them, Z n To calculate the signal point S Cn LNDP value, d i is the difference between i and i+1, and m is the number of adjacent points.

[0064] The above-mentioned conversion of the filtered EEG sequence, FNS sequence, and LNDP sequence into a two-dimensional time-spectrogram adopts the wavelet transform method, and then the size of the formed three-channel time-spectrogram is adjusted to 64×64 to facilitate the convolution operation. The wavelet transform adopts the Morse wavelet basis.

[0065] Step S3, building an improved densely connected network as a sleep staging network model based on the densely connected network.

[0066] Specifically, the existing densely connected network model is composed of dense blocks and transition layers stacked alternately. Its dense connection mechanism is as follows: Figure 4 As shown in Figure 1. Each dense block consists of multiple dense layers, each of which contains a batch normalization layer, an activation function layer, and a convolutional layer. Figure 5 As shown in the figure, the transition layer consists of batch normalization, ReLU activation function, 1×1 2D convolution layer with a step size of 1, and an average pooling layer. The dense connection mechanism of the densely connected network is reflected in the fact that each dense layer in the dense block is spliced ​​with all the previous layers in the channel dimension. This dense connection enables each layer in the dense block to obtain information from all the previous layers, avoiding information loss, promoting the transmission and reuse of features, and enabling the network to extract richer and more representative features. The setting of the transition layer helps to reduce the size of the feature map and the number of feature maps, thereby controlling the growth rate of the network while maintaining the depth of the network and the diversity of features.

[0067] like Figure 6 As shown, in this embodiment, the channel attention mechanism is integrated into the densely connected network model to construct a sleep staging network model based on the improved densely connected network. The sleep staging network model includes a shallow feature extraction module, a deep feature extraction module and a classifier module.

[0068] Specifically, the shallow feature extraction module consists of a 2D convolution layer and a maximum pooling layer, which is mainly used to extract shallow features of the network. The convolution kernel size of the 2D convolution layer is 7×7, the step size is 2, and the boundary padding value is 3; the pooling window size of the maximum pooling layer is 5×5, and the step size is 2.

[0069] Specifically, the deep feature extraction module is composed of 4 dense attention blocks and 3 improved transition layers stacked alternately. Each dense attention block is composed of several dense attention layers. The structure of the dense attention layer is as follows: Figure 7 The specific structure of the deep feature extraction module is as follows: the first dense attention block contains 6 dense attention layers, followed by the first transition layer; followed by the second dense attention block consisting of 12 dense attention layers, followed by the second transition layer; the third dense attention block contains 32 dense attention layers, followed by the third transition layer; finally, the fourth dense attention block also consists of 32 dense attention layers.

[0070] like Figure 7 As shown in Figure 1, the dense attention layer includes a dense layer and a channel attention module. The dense layer includes a batch normalization layer, an activation function layer, and a convolution layer. A CAM module is added between the activation function layer and the convolution layer to form a dense attention layer. The CAM module consists of an average pooling layer, a convolution layer, and an activation function layer. It is mainly used to extract the correlation between different channels in the feature map so that the network can focus on more effective information. Its structure is as follows: Figure 8 As shown in Figure 2, the size of the pooling layer and convolution layer in the CAM module is 1×1, and the stride is 1. Let the input of the CAM module be the feature F∈R H×W×C , where H is the length of each channel feature, W is the width of each channel feature, and C is the number of channels. The CAM module calculation formula is:

[0071] A C (F) = σ(K*GAP(F)+b)

[0072] Among them, A C (F)∈R 1×1×C is the attention vector after operation on the channel dimension, K is the weight matrix of the one-dimensional convolution kernel, GAP(F) is the feature map aggregation operation after global averaging, b is the bias vector, and σ is the sigmoid activation function.

[0073] The output calculation formula of CAM is:

[0074]

[0075] Among them, F'∈R H×W×C is the output of the CAM module, A C (F)∈R1×1×C is the attention vector after operation on the channel dimension, is element-wise multiplication.

[0076] The CAM module obtains the weighted feature representation by multiplying the attention vector with the input feature. The CAM method learns the global information of the feature map and assigns different weights to each channel, thereby enhancing effective features and suppressing invalid features.

[0077] In addition to introducing CAM in the densely connected network, this embodiment also makes targeted adjustments to the network to meet the needs of the sleep staging task. Specifically, most of the pooling layers in the network are replaced with maximum pooling to enhance the network's sensitivity to small differences and fluctuations in the frequency spectrum. In addition, the size of some transition layers is enlarged to increase the receptive field of the model so that global features in the EEG frequency spectrum can be more effectively extracted.

[0078] The classifier module includes a fully connected layer and a softmax layer. The fully connected layer processes the features obtained by the feature extraction module into one-dimensional features, and then the softmax layer maps the one-dimensional features into the probabilities of each sleep stage to obtain the final sleep stage recognition results.

[0079] Step S4, sending the target data set to the constructed network model to complete the sleep stage discrimination.

[0080] Specifically, the specific steps of sending the target data set into the sleep staging network model to complete the sleep staging discrimination include:

[0081] Step S41, dividing the target data set into a training set, a validation set and a test set.

[0082] Specifically, the target dataset is divided into training set, validation set and test set in a ratio of 7:2:1.

[0083] Step S42: input the training set into the sleep staging network model for model training.

[0084] Specifically, the batch value set during model training is 128, the training cycle is set to 100, and the sleep staging network model uses the cross entropy function as the loss function during training. The stochastic gradient descent method is used to optimize the parameters of the network model to reduce random errors, and the learning rate is set to 1e-4. The calculation formula of the cross entropy loss function is:

[0085]

[0086] Among them, a is the output of the deep feature extraction network after passing through the softmax layer; y is the sleep stage label, and M is the number of categories, where M=5.

[0087] Step S43, inputting the validation set into the trained sleep staging network model to determine the optimal model parameters.

[0088] Specifically, the validation set is sent to the trained sleep staging network model for model evaluation, and the optimal model weight parameters are determined based on the staging performance of the trained model on the validation set.

[0089] Step S44, inputting the test set into the sleep stage network model with determined parameters to obtain the sleep stage recognition result.

[0090] Specifically, the specific steps of inputting the test set into the sleep stage network model with determined weight parameters to obtain the sleep stage recognition result include:

[0091] Step S441: convert the three-channel spectrogram X 64×64×3 As the input of the network model, the shallow feature d0 is obtained after passing through the shallow feature extraction module. 14×14×64 , where 14×14 is the size of the feature map and 64 is the number of channels of the feature map.

[0092] Step S442, input d0 into the deep feature extraction module for deep feature extraction. Take the first dense attention block as an example. The first dense attention block consists of 6 dense attention layers, and the structure of each dense attention layer is: batch normalization, Relu activation function, CAM module, 2D convolution layer. The convolution kernel size of the 2D convolution layer is 3×3, the step size is 1, the boundary filling value is 1, and the number of convolution kernels is 32. Then the output D of the first dense attention block is:

[0093] D = concatenate(d0,d1,...,d6)

[0094] Where d is the output of each 2D convolutional layer in the dense attention layer, and the size is 14×14×32. Therefore, the input d0 14×14×64 After the first dense attention block, the output D∈R is obtained 14×14×256 .

[0095] Step S443, D passes through the first transition layer, the convolution kernel size of the convolution layer is 1×1, the step size is 1, the convolution kernel data volume is 128, the maximum pooling layer window size is 5×5, and the step size is 2. Then the output after the first transition layer is:

[0096] T=H(D)

[0097] Among them, H is a nonlinear transformation function, including batch normalization, Relu activation function, 2D convolution and maximum pooling operation. The input D is passed through the first transition layer to obtain the output T∈R 5×5×128 .

[0098] Step S444, T passes through the second dense attention block, the second transition layer, the third dense attention block, the third transition layer, and the fourth dense attention block in sequence to obtain the final feature output z∈R 1×1×1664 Among them, the convolution kernel size of the 2D convolution layer of each dense attention layer in the second, third, and fourth dense attention blocks is 3×3, the stride is 1, and the number of convolution kernels is 32; the convolution kernel size in the second transition layer is 1×1, the stride is 1, the number of convolution kernels is 256, the maximum pooling layer window size is 3×3, and the stride is 2; the convolution kernel size in the third transition layer is 1×1, the stride is 1, the number of convolution kernels is 640, and the maximum pooling layer window size is 2×2, and the stride is 2. Then the output T∈R of the first transition layer 5 ×5×128 The output size after the second dense attention block is 5×5×512, then the output size after the second transition layer is 2×2×256, then the output size after the third dense attention block is 2×2×1280, then the output size after the third transition layer is 1×1×640, and finally the output size after the fourth dense attention block is 1×1×1664.

[0099] In step S445, the fully connected layer in the classifier module processes the feature z obtained by the deep feature extraction module into a one-dimensional feature, and maps it into the probabilities of the five stages after passing through the softmax layer to obtain the final sleep stage recognition result.

[0100] In order to demonstrate the superiority of this algorithm, the sleep staging network model using the same data set is compared with other algorithms. The results are shown in Table 1.

[0101] Table 1 Comparison of sleep staging performance of the present invention with other algorithms

[0102]

[0103] The first embodiment of the present invention provides a sleep staging method based on multi-view fusion and improved dense connection network, including obtaining single-channel sleep EEG data as an initial data set; preprocessing the initial data set to form a three-channel two-dimensional time-frequency spectrum as a target data set; building an improved dense connection network as a sleep staging network model based on the dense connection network; and sending the target data set into the constructed network model to complete sleep staging discrimination. The sleep staging method based on multi-view fusion and improved dense connection network of the present invention uses the two-dimensional time-frequency spectrum generated by the single-channel EEG signal as the main feature information pathway, and uses FNS and LNDP methods to enhance the local dynamic information representation of the EEG signal, so that the sleep staging network model can capture the complex patterns in the EEG and the correlation information between pathways; the channel attention mechanism CAM is integrated into the dense connection network, and the dense connection method unique to the dense connection network is used to realize the transmission and reuse of features, and at the same time, the features of different channels are weighted to enhance the focus on more useful channels, thereby improving the performance of the sleep staging model.

[0104] The step division of the above methods is only for the purpose of clear description. When implemented, they can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this patent; adding insignificant modifications to the algorithm or process or introducing insignificant designs without changing the core design of the algorithm and process are all within the scope of protection of this patent.

[0105] The above is only an embodiment of the present invention. The common sense such as the known specific structure and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field know all the common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A sleep staging method based on multi-view fusion and improved densely connected network, characterized in that: include: S1, obtain single-channel sleep EEG data as the initial data set; S2, preprocessing the initial data set to form a three-channel two-dimensional time-frequency spectrum as a target data set; S3, based on the densely connected network, an improved densely connected network is constructed as a sleep staging network model; S4, sending the target data set into the constructed network model to complete the sleep stage discrimination.

2. The sleep staging method based on multi-view fusion and improved densely connected network according to claim 1, characterized in that: In S2, the initial data set is preprocessed to form a three-channel two-dimensional time-frequency spectrum as a target data set, including: Step S21, using a 0.5-45 Hz bandpass Butterworth filter to filter the original EEG signal to remove background noise and power line interference; Step S22, transforming the filtered single-channel EEG signal into a two-dimensional time-frequency spectrum as the main feature information pathway of the sleep staging network model; Step S23, using the FNS method to process the filtered EEG sequence into an FNS sequence, and using the two-dimensional time-frequency spectrum of the FNS sequence as the second input path of the sleep staging network model to assist the network model in capturing hidden dynamic information in the sleep EEG signal; Step S24, using the LNDP method to process the filtered EEG sequence into an LNDP sequence, and using the two-dimensional time-frequency spectrum of the LNDP sequence as the input path three of the sleep staging network model to assist the network model in capturing local changes in the EEG signal.

3. The sleep staging method based on multi-view fusion and improved densely connected network according to claim 2, characterized in that: The FNS method constructs a structure function and calculates its cumulative average within a sliding window to generate a dynamic time series for characterizing the hidden dynamic information in the EEG signal.

4. The sleep staging method based on multi-view fusion and improved densely connected network according to claim 2, characterized in that: The LNDP method constructs a new time series by comparing the values ​​of adjacent signal points and performing binarization, which is used to characterize the local changes of the time series and express the key information in the weak signal.

5. The sleep staging method based on multi-view fusion and improved densely connected network according to claim 1, characterized in that: The sleep staging network model includes a shallow feature extraction module, a deep feature extraction module and a classifier module. The shallow feature extraction module includes a convolution layer and a maximum pooling layer, the deep feature extraction module includes 4 dense attention blocks and 3 improved transition layers, and the classifier module includes a fully connected layer and a softmax layer. The fully connected layer processes the features obtained by the feature extraction module into one-dimensional features, and then the softmax layer maps the one-dimensional features into the probabilities of each sleep stage to obtain the final sleep staging recognition result.

6. The sleep staging method based on multi-view fusion and improved densely connected network according to claim 5, characterized in that: The dense attention block consists of several dense attention layers, which include a dense layer and a channel attention CAM module.

7. The sleep staging method based on multi-view fusion and improved densely connected network according to claim 6, characterized in that: The CAM module consists of an average pooling layer, a convolutional layer, and an activation function layer, which is used to extract the correlation between different channels in the feature map so that the network can focus on more effective information.

8. The sleep staging method based on multi-view fusion and improved densely connected network according to claim 1, characterized in that: The sleep staging network model uses a cross entropy function as a loss function during training, and adopts a stochastic gradient descent method to optimize the parameters of the network model to reduce random errors, and the learning rate is set to 1e-4.

Images