A method for identifying a car sickness state based on Dropout-MTN

Abstract

The application relates to a car sickness state recognition method based on Dropout-MTN. A brand-new strategy and network structure design are adopted. The brain surface is preset with each detection position brain electric signal waveform, and the brain electric detection sample is constructed in combination with corresponding real car sickness state categories. The Dropout layer is connected. The input layer, the polynomial layer, the full connection layer, the Softmax layer and the output layer are sequentially connected in series from the input end to the output end. The network training is executed. The car sickness state recognition model is obtained and is used for actual application. The design scheme simplifies the model complexity, improves the model generalization ability, enables the designed Dropout-MTN to quickly realize high-precision classification of the car sickness state, and has simple network structure, does not need to perform a large amount of network training, has low calculation complexity, and can improve the detection speed without reducing the classification accuracy compared with common machine learning classifiers.

Description

TECHNICAL FIELD

[0001] The application relates to a method for identifying a car sickness state based on Dropout-MTN, and belongs to the technical field of electroencephalogram signal analysis and detection. BACKGROUND

[0002] Car sickness is a kind of dizziness, which usually occurs during long-distance travel or riding on a vehicle, a ship or the like. The causes of car sickness are various, including diseases of the motion perception system, abnormalities of the inner ear sensory organs and disorders of the brain in processing sensory information. The symptoms of car sickness include dizziness, nausea, vomiting, sweating and increased heart rate. Although car sickness is usually a temporary discomfort, it can affect the quality of life and daily activities of some people. Therefore, it is crucial to diagnose car sickness, and in order to objectively diagnose the car sickness state, analysis based on electroencephalogram signals has been widely applied. At present, deep neural network models are widely used in the analysis of electroencephalogram signals, but these deep neural network models have complex structures and many parameters, and both the training and testing of the models are very time-consuming. Therefore, it is difficult to meet the demand for rapid diagnosis of the car sickness state. SUMMARY

[0003] The technical problem to be solved by the application is to provide a method for identifying a car sickness state based on Dropout-MTN, which uses a completely new design logic and can improve the detection speed without reducing the classification accuracy.

[0004] The application adopts the following technical solution to solve the above technical problem: the application designs a method for identifying a car sickness state based on Dropout-MTN, and performs the following steps A to C to obtain a car sickness state identification model, and then performs step i to apply the car sickness state identification model to complete the identification of the car sickness state of a target person for a target time period.

[0005] Step A. Based on a sample detection time period equal to the target time period, the electroencephalogram waveforms of the brain surface of each detection position of a preset number of to-be-analyzed persons for the sample detection time period are obtained, and a multi-dimensional electroencephalogram detection signal waveform of the to-be-analyzed persons for the sample detection time period is formed, and based on a preset each car sickness state category including a non-car sickness state category, the true car sickness state category of the to-be-analyzed persons for the sample detection time period is obtained, and the multi-dimensional electroencephalogram detection signal waveform and the true car sickness state category form an electroencephalogram detection sample, that is, the electroencephalogram detection samples corresponding to each to-be-analyzed person are obtained, and then step B is entered.

[0006] Step B. A to-be-trained network is constructed, in which an input layer, a polynomial layer, a full connection layer, a Softmax layer and an output layer are connected in series from the input end to the output end, and then step C is entered.

[0007] Step C. Based on the auxiliary to-be-trained network composed of a Dropout layer in series between a polynomial layer and a full connection layer in the to-be-trained network, according to each electroencephalogram detection sample, taking the multi-dimensional electroencephalogram detection signal waveform in the electroencephalogram detection sample as input and the motion sickness state category as output, the auxiliary to-be-trained network is trained, and then the training of the to-be-trained network is realized, and a motion sickness state recognition model corresponding to the to-be-trained network is obtained.

[0008] Step i. The electroencephalogram signal waveforms of the target person's brain surface at each detection position for a target time length are collected to form a multi-dimensional electroencephalogram detection signal waveform of the target person for the target time length, and the motion sickness state recognition model is applied to obtain the motion sickness state category of the target person for the target time length, and the motion sickness state recognition of the target person for the target time length is completed.

[0009] As a preferred technical solution of the present application: the step A includes the following steps A1 to A3:

[0010] Step A1. Based on a sample detection time length equal to the target time length, for each of the preset number of persons to be analyzed, the electroencephalogram signal waveforms of the brain surface of the person to be analyzed at each detection position corresponding to each time point in the sample detection time length are obtained, and the average value of the square of the electroencephalogram signal waveform corresponding to each detection position is taken as the electroencephalogram signal waveform corresponding to the sample detection time length of each detection position, and the multi-dimensional electroencephalogram detection signal waveform of the person to be analyzed corresponding to the sample detection time length is formed, and then step A2 is entered;

[0011] Step A2. Based on each motion sickness state category including a non-motion sickness state category, according to the SSQ scale corresponding to each person to be analyzed, for each of the preset number of persons to be analyzed, the values of parameters N, O and D corresponding to each time point in the sample detection time length are obtained, and the average value of the total score TS of each time point is calculated according to the following formula:

[0012] TS=N+O+D×3.74

[0013] The average value of the total score TS of each time point is obtained, and the true motion sickness state category of the person to be analyzed corresponding to the sample detection time length is obtained according to the preset average value and the corresponding relationship of each motion sickness state category, and then step A3 is entered;

[0014] Step A3. Taking the multi-dimensional electroencephalogram detection signal waveform and the true motion sickness state category of the person to be analyzed corresponding to the sample detection time length as the electroencephalogram detection sample, the electroencephalogram detection sample corresponding to each person to be analyzed is obtained.

[0015] As a preferred technical solution of the present application: the input layer in step B is used to receive multi-dimensional electroencephalogram detection signal waveforms x1(k),..., x n (k),..., x E (k) composed of brain surface preset respective detection position corresponding sample detection time length segment of electroencephalogram signal waveforms, and forward the polynomial layer, 1≤n≤E, E represents the number of detection positions of the brain surface corresponding to the multi-dimensional electroencephalogram detection signal waveforms, x n (k) represents the electroencephalogram signal corresponding to the kth sample detection time length segment of the nth detection position of the brain surface;

[0016] The polynomial layer receives the multi-dimensional electroencephalogram detection signal waveforms x1(k),..., x n (k),..., x E (k), and according to the preset highest expansion term M corresponding to the network to be trained, the output Y(k) of the polynomial layer is constructed as follows, and is transmitted to the full connection layer, 1≤M≤E;

[0017] Y(k) = [y1(k)... y n (k)... y E (k)] = [W1... W n ... W E ] T ·V(k)

[0018] Wherein, y n (k) represents the output of the polynomial layer corresponding to the nth detection position, W n = [w n，1 ... w n，l ... w n，L(E，M) ], 1≤l≤L(E, M), L(E, M) represents the number of monomials in the polynomial based on E and M, W n represents the weight vector corresponding to the nth detection position in the polynomial layer, w n，l represents the lth weight in W n , w n，l is a training parameter in the network to be trained, V(k) = [1, x1(k)... x E (k), x1(k)x2(k)... (x E (k)) 2 ,..., x1(k)... x M (k)... (x E (k)) M ] T , V(k) represents the expansion of each term of the multi-dimensional electroencephalogram detection signal waveforms with respect to the highest expansion term M.

[0019] As a preferred technical solution of the present invention: Regarding the fully connected layer in step B, based on the output Y(k) of the polynomial layer, the output F(k) of the fully connected layer is constructed as follows, and transmitted to the Softmax layer;

[0020] F(k) = [f1(k) … f c (k) … f C (k)] T =S{[Λ1 … Λ c … Λ C ] T ·Y(k)}

[0021] Among them, f c (k) represents the output of the fully connected layer corresponding to the c-th motion sickness state category, Λ c =[Λ c，1 … Λ c，n … Λ c，E ], 1≤c≤C, where C represents the number of categories of motion sickness, Λ c Λ represents the weighted column vector of the c-th motion sickness category. c，n Λ represents the weight from the nth detection location to the cth motion sickness category. c，n These are the parameters to be trained in the network to be trained, and S represents the ReLU activation function.

[0022] As a preferred embodiment of the present invention: Regarding the Softmax layer in step B, the output F(k) based on the fully connected layer is F(k) = [f1(k) … f c (k) … f C (k)] T ,according to The output of the Softmax layer is P(k) = [p1(k) … p c (k) … p C (k)] T And transmit it to the output layer, where e represents the natural constant, p c (k) represents the probability corresponding to the c-th motion sickness state category; the output layer determines the motion sickness state category corresponding to the multidimensional EEG detection signal waveform received by the input layer based on the probabilities corresponding to each c-th motion sickness state category, and outputs the result; p c (k) represents the probability of the fully connected layer outputting the c-th motion sickness state category.

[0023] As a preferred embodiment of the present invention: in step C, regarding the auxiliary network to be trained, which is composed of a Dropout layer connected in series between the polynomial layer and the fully connected layer in the network to be trained, the output Y(k) of the polynomial layer is based on [y1(k) …y n (k) … yE [k], the relationship follows a Bernoulli distribution with a preset parameter β, and the parameters r1, ..., r1 at each detection position are respectively. n ... r E For each detection location, press r n ·y n The result of (k) for y n (k) is updated, and the updated value is the output Y(k) of the Dropout layer = [y1(k) … y n (k) … y E (k)], and transmit to the fully connected layer, where 1≤n≤E, E represents the number of detection locations on the brain surface corresponding to the multidimensional EEG detection signal waveform, y n (k) represents the output of the polynomial layer with respect to the nth detection position, r n This represents the parameter corresponding to the nth detection position, which follows a Bernoulli distribution with a preset parameter β.

[0024] As a preferred technical solution of the present invention: in step C, regarding the training process of the auxiliary network to be trained, regarding the training parameters W1, W2, ..., W under each iteration... E With the parameters to be trained Λ1, Λ2, ..., Λ C The mini-batch gradient descent method is used to update the data as follows;

[0025]

[0026] W'1, W'2, ..., W' E Represented sequentially as W1, W2, ..., W E The updated results are: Λ'1, Λ'2, ..., Λ' C Represented in sequence as Λ1, Λ2, ..., Λ C In the updated result, η represents the learning rate of the parameters, and H represents the result of the loss function.

[0027] As a preferred technical solution of the present invention: in step C, during the training process of the auxiliary network to be trained, the cross-entropy loss function is used as follows to optimize each training parameter in the network to be trained;

[0028]

[0029] Where Q represents the number of EEG samples in the training set during the training of the auxiliary network to be trained, and C represents the number of categories of motion sickness; y q，c Let y denote the sign function. If the true motion sickness category of the q-th EEG sample in the training set is equal to the motion sickness category of the c-th sample, then y q，c= 1, if the real car sickness state category of the qth electroencephalogram detection sample in the training set is not equal to the cth car sickness state category, then y q，c = 0; P q，c represents the probability of the qth electroencephalogram detection sample in the training set belonging to the cth car sickness state category under the assistance of the training of the to-be-trained network, H q represents the loss function result of the qth electroencephalogram detection sample in the training set under the assistance of the training of the to-be-trained network, and H represents the loss function result of the training set under the assistance of the training of the to-be-trained network.

[0030] The car sickness state recognition method based on Dropout-MTN has the following technical effects compared with the prior art by adopting the above technical scheme.

[0031] (1) The car sickness state recognition method based on Dropout-MTN is designed, a new strategy and network structure are adopted, the electroencephalogram signal waveform of each detection position on the brain surface is preset, the electroencephalogram detection sample is constructed in combination with the corresponding real car sickness state category, the Dropout layer is contacted, the to-be-trained network is sequentially connected in series from the input end to the output end, the input layer, the polynomial layer, the full connection layer, the Softmax layer and the output layer, the network training is performed, the car sickness state recognition model is obtained for actual application, the model complexity is simplified, the model generalization ability is improved, the designed Dropout-MTN can quickly realize high-precision classification of car sickness states, the network structure is simple, a large amount of network training is not needed, the calculation complexity is low, compared with the commonly used machine learning classifier, the detection speed can be improved without reducing the classification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 is a schematic diagram of the brain surface preset each detection position for electroencephalogram signal waveform extraction in the embodiment of the application;

[0033] Figure 2 is a network architecture schematic diagram of the to-be-trained network combined with the Dropout layer designed by the application;

[0034] Figure 3 is a schematic diagram of the convergence speed in the training of the auxiliary to-be-trained network compared with other commonly used networks designed by the application. DETAILED DESCRIPTION

[0035] The specific embodiments of the application will be further described in detail below with reference to the accompanying drawings.

[0036] The car sickness state recognition method based on Dropout-MTN is designed, and in actual application, the following steps A to C are specifically designed and executed to obtain the car sickness state recognition model.

[0037] Step A. Based on a sample detection duration equal to the target duration, for a preset number of individuals to be analyzed, obtain the EEG signal waveforms at each preset detection location on the brain surface of the individual to be analyzed, corresponding to the sample detection duration. This constitutes a multidimensional EEG detection signal waveform for the sample detection duration corresponding to the individual to be analyzed. Simultaneously, based on preset categories of motion sickness, including non-motion sickness, obtain the actual motion sickness category for the sample detection duration corresponding to the individual to be analyzed. The multidimensional EEG detection signal waveform and the actual motion sickness category constitute the EEG detection sample, thus obtaining the EEG detection sample corresponding to each individual to be analyzed. Then proceed to Step B.

[0038] In actual implementation, step A above is specifically designed and executed as follows: steps A1 to A3.

[0039] Step A1. Based on a sample detection duration equal to the target duration, for a preset number of individuals to be analyzed, obtain the EEG signal waveforms at each time point in the sample detection duration corresponding to each preset detection location on the surface of the brain of the individual to be analyzed. The average of the squares of the EEG signal waveforms corresponding to each detection location is used to construct the EEG signal waveforms corresponding to each detection location in the sample detection duration, thus constructing the multidimensional EEG detection signal waveforms for the sample detection duration corresponding to the individual to be analyzed. Then proceed to step A2.

[0040] In practical application, a 32-channel EEG data acquisition device was used to conduct VR dizziness tests on 23 individuals to obtain their electroencephalogram (EEG) waveforms. To improve classification efficiency, such as... Figure 1 As shown, 30 typical electrodes representing different brain regions on the surface of the brain were designed and extracted, including FP1, FP2, F11, F7, F3, FZ, F4, F8, F12, FT11, FC3, FCZ, FC4, FT12, T7, C3, CZ, C4, T8, CP3, CPZ, CP4, P7, P3, PZ, P4, P8, O1, OZ, and O2. The EEG input sample data had dimension E, highest expansion order M, and sample data length l. The dataset was then normalized to obtain the sample data matrix E·l. l refers to the sample data length. In this experiment, E = 30, and l = 3960. The method for obtaining these two datasets is shown below:

[0041] 200 samples (including 31 channels of EEG signals) per second, 3 minutes for each person to be analyzed, and not directly inputting 200 samples per second into the trained network, but using an energy value to represent the 200 samples. The energy value is the square average of the 200 sample signals, that is, each sample is squared first, and then the average value is taken as the energy signal, representing the EEG signal waveform of one second. Therefore, one second is equivalent to one energy value, and 3 minutes have 180 energy values. Because only 22 of the 23 subjects successfully completed the experiment and collected data, the total amount of data is 180*22=3960. The final input dimension of the MTN is E·1=30*3960.

[0042] Step A2. Based on the preset various car sickness state categories including the non-car sickness state category, according to the SSQ scale corresponding to each person to be analyzed, the values of parameters N, O, and D of each person to be analyzed at each time point in the sample detection time period are obtained, and the average value of the total score TS of each time point is obtained according to the following formula:

[0043] TS=N+O+D×3.74

[0044] The total score TS of each time point in the sample detection time period corresponding to each person to be analyzed is calculated, and the average score of the total score TS of each time point is obtained, and the real car sickness state category of the sample detection time period corresponding to each person to be analyzed is obtained according to the preset average score and the corresponding relationship of each car sickness state category, and then step A3 is entered.

[0045] In actual application, for example: the N, O, and D values of the first person to be analyzed in the motion sickness SSQ scale are 6.8, 6.09, and 6.06 respectively, and the TS value is calculated to be 35.5544, similarly, the TS values of all persons to be analyzed are calculated, and the average value is 20.367, and the label of the TS value greater than the average value is marked as 1, and the label of the TS value less than the average value is marked as 0. The label here is not the feeling of the person to be analyzed every second in the 3-minute experiment to obtain the corresponding label. But after the experiment, the person to be analyzed gives a general feeling, and the label corresponding to the feeling is taken as the label of the EEG signal in the whole 3-minute experiment.

[0046] Step A3. The multi-dimensional EEG detection signal waveform of the sample detection time period corresponding to each person to be analyzed and the real car sickness state category are taken as the EEG detection sample, that is, the EEG detection sample corresponding to each person to be analyzed is obtained.

[0047] Step B. Construct a trained network with input layer, polynomial layer, full connection layer, Softmax layer, and output layer in sequence from input end to output end, and then enter step C.

[0048] AsFigure 2 As shown, in the actual design and application of the network to be trained, the input layer is used to receive multidimensional EEG detection signal waveforms x1(k), ..., x2(k) consisting of EEG signal waveforms corresponding to sample detection time segments at each preset detection location on the brain surface. n (k), ..., x E (k), and forward the polynomial layer, 1≤n≤E, where E represents the number of detection locations on the brain surface corresponding to the multidimensional EEG detection signal waveform, x n (k) represents the EEG signal at the nth detection location on the brain surface corresponding to the kth sample detection duration.

[0049] The polynomial layer is applied to the received multidimensional EEG detection signal waveforms x1(k), ..., x n (k), ..., x E (k), based on the preset highest expansion term M corresponding to the network to be trained, construct the output Y(k) of the polynomial layer as follows, and transmit it to the fully connected layer, 1≤M≤E;

[0050] Y(k)=[y1(k) … y n (k) … y E [(k)]=[W1 … W n … W E ] T ·V(k)

[0051] Among them, y n (k) represents the output of the polynomial layer at the nth detection position, W n =[w n，1 … w n，l …w n，L(E，M) ], 1≤l≤L(E,M), where L(E,M) represents the number of monomials in the polynomial based on E and M, W n w represents the weight vector corresponding to the nth detection position in the polynomial layer. n，l W n The l-th weight, w n，l These are the trainable parameters in the network to be trained, V(k) = [1, x1(k)...x1(k)]. E (k), x1(k)x2(k)…(x E (k)) 2 ,…,x1(k)…x M (k)…(x E (k)) M ] T V(k) represents the expansion of the multidimensional EEG detection signal waveform with respect to each term of the highest-order expansion term M.

[0052] like Figure 1As shown, the output of the polynomial layer is connected to the full connection layer, the input dimension of the full connection layer is consistent with the dimension E of the feature, and the output layer is consistent with the type number C, so it can be used in the case of different dimensions of input and output. In addition, in order to further improve the fitting accuracy of Dropout-MTN, the activation function ReLU is added behind each output node of the full connection layer to improve the ability of Dropout-MTN to describe nonlinear characteristics.

[0053] In the application, the full connection layer, based on the output Y(k) of the polynomial layer, constructs the output F(k) of the full connection layer as follows, and transmits it to the Softmax layer;

[0054] F(k) = [f1(k) … f c (k) … f C (k)] T = S{[Λ1 … Λ c … Λ C ] T ·Y(k)}

[0055] Where f c (k) represents the output of the full connection layer corresponding to the cth car sickness state category, Λ c = [Λ c，1 … Λ c，n … Λ c，E ], 1≤c≤C, C represents the number of categories of car sickness state categories, Λ c represents the weight column vector of the cth car sickness state category, Λ c，n represents the weight from the nth detection position to the cth car sickness state category, Λ c，n is a to-be-trained parameter in the to-be-trained network, and S represents the operation of the activation function ReLU.

[0056] The Softmax layer, based on the output F(k) = [f1(k) … f c (k) … f C (k)] T of the full connection layer, constructs the output P(k) = [p1(k) … p c (k) … p C (k)] T of the Softmax layer according to and transmits it to the output layer, e represents the natural constant, and p c (k) represents the probability corresponding to the cth car sickness state category; the output layer determines the car sickness state category corresponding to the multi-dimensional electroencephalogram detection signal waveform received by the input layer according to the probability corresponding to each cth car sickness state category, and outputs; p c(k) represents the probability of the full connection layer corresponding to the cth motion sickness state category output.

[0057] Step C. Based on the auxiliary to-be-trained network composed of the Dropout layer in series between the polynomial layer and the full connection layer in the to-be-trained network, according to each electroencephalogram detection sample, taking the multi-dimensional electroencephalogram detection signal waveform in the electroencephalogram detection sample as input and the motion sickness state category as output, the auxiliary to-be-trained network is trained, and then the training of the to-be-trained network is realized, and a motion sickness state recognition model corresponding to the to-be-trained network is obtained.

[0058] In the training process, in order to alleviate the overfitting problem of the above designed to-be-trained network and improve the generalization ability of the network, the Dropout algorithm is introduced in the training process, that is, a dropout layer is added in the to-be-trained network structure, as shown in Figure 1 . In order to prevent model overfitting, the nth node y n (k) output by the hidden layer is discarded with a probability p, n = 1, 2,..., E. For this purpose, a matrix composed of 0 and 1 is generated as follows:

[0059]

[0060] where r n ~ Bernoulli(p), n = 1,..., E, the matrix is a diagonal matrix, and the elements r N on the diagonal are 0 or 1, which obeys the Bernoulli distribution with parameter p (preset).

[0061] That is, in specific practical applications, regarding the auxiliary to-be-trained network composed of the Dropout layer in series between the polynomial layer and the full connection layer in the to-be-trained network, based on the output Y(k) = [y1(k) … y n (k) … y E (k)] of the polynomial layer, in connection with the Bernoulli distribution with parameter preset β, the parameters r1, …, r n , …, r E corresponding to each detection position, respectively, for each detection position, according to the result of r n · y n (k), y n (k) is updated, and Y(k) = [y1(k) … y n (k) … y E (k)] as the output of the Dropout layer is updated and transmitted to the full connection layer, where 1 ≤ n ≤ E, E represents the number of detection positions of the brain surface corresponding to the multi-dimensional electroencephalogram detection signal waveform, y n (k) represents the output of the polynomial layer about the nth detection position, and r nA Bernoulli distribution with a compliance parameter of a preset β, corresponding to an nth detection position.

[0062] Regarding the training of the auxiliary network to be trained, in view of the mesh multi-layer structure of the multi-dimensional Taylor network, the weight of the MTN classifier is not estimated by using the traditional least square method in the design of the present application, but the Mini-Batch Gradient Descent (MBGD) method is used. The advantage of this method is that a batch of samples is considered in each iteration, so that the global optimum can be achieved. Therefore, in the process of training the Dropout-MTN classifier, i.e. training the auxiliary network to be trained, the MBGD method is used to update the weight of the model, in which the weight is updated once for each batch of samples. Two-thirds of the sample data are used as training data, and the remaining data are used as test data. The length of the training data is defined as T, and in addition, the highest expansion term, the parameter learning rate, the number of training iterations, the number of samples processed per batch, and the batch size are respectively denoted as m, η, l, and T bz . The process of completing one forward propagation and reverse propagation in the network with all data is defined as an epoch. We use n e to define the number of epochs. Then T, T bz and n e satisfy the following relationship:

[0063]

[0064] In the actual training process of the auxiliary network to be trained, regarding the training parameters W1, W2, …, W E and the training parameters Λ1, Λ2, …, Λ C in each iteration, the Mini-Batch Gradient Descent method is used to update them in the following way;

[0065]

[0066] wherein W'1, W'2, …, W' E represent W1, W2, …, W E respectively, Λ'1, Λ'2, …, Λ' C represent Λ1, Λ2, …, Λ C respectively, η represents the parameter learning rate, and H represents the loss function result.

[0067] And in the training process, the cross-entropy loss function is used as follows to achieve the optimization of each training parameter in the network to be trained;

[0068]

[0069] wherein, Q represents the number of electroencephalogram detection samples in the training set in the process of training the auxiliary to-be-trained network, and C represents the number of categories of the car sickness state categories; y q，c represents a symbol function, y q，c = 1 if the real car sickness state category of the qth electroencephalogram detection sample in the training set is equal to the cth car sickness state category, and y q，c = 0 if the real car sickness state category of the qth electroencephalogram detection sample in the training set is not equal to the cth car sickness state category; P q，c represents the probability that the qth electroencephalogram detection sample in the training set belongs to the cth car sickness state category output by the auxiliary to-be-trained network, H q represents the loss function result of the qth electroencephalogram detection sample in the training set with respect to the car sickness state category under the auxiliary to-be-trained network training, and H represents the loss function result of the training set under the auxiliary to-be-trained network training.

[0070] In actual application, as shown in FIG. Figure 3 Compared with other commonly used methods, the convergence speed of the auxiliary to-be-trained network Dropout-MTN designed in the application is the fastest, and the classification accuracy is the highest.

[0071] Based on the obtained car sickness state recognition model, the following step i is further designed: applying the car sickness state recognition model to complete the car sickness state recognition of the target person corresponding to the target time length period.

[0072] Step i. Collecting the electroencephalogram signal waveforms of the target person corresponding to the target time length period at each preset detection position on the surface of the brain, constructing a multi-dimensional electroencephalogram detection signal waveform of the target person corresponding to the target time length period, and applying the car sickness state recognition model to obtain the car sickness state category of the target person corresponding to the target time length period, thereby completing the car sickness state recognition of the target person corresponding to the target time length period.

[0073] The car sickness state recognition method based on Dropout-MTN designed in the above technical solution adopts a new strategy and network structure design, constructs electroencephalogram detection samples based on the electroencephalogram signal waveforms at each preset detection position on the surface of the brain in combination with the corresponding real car sickness state category, contacts the Dropout layer, executes network training on the to-be-trained network sequentially connected in series from the input end to the output end, that is, the input layer, the polynomial layer, the full connection layer, the Softmax layer and the output layer, obtains the car sickness state recognition model for actual application, and simplifies the model complexity and improves the model generalization ability. At the same time, the designed Dropout-MTN can quickly realize high-precision classification of the car sickness state, and the network structure is simple, does not need to perform a large amount of network training, and has low computational complexity. Compared with commonly used machine learning classifiers, the detection speed can be improved without reducing the classification accuracy.

[0074] The embodiments of the present application are described in detail above with reference to the accompanying drawings, but the present application is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present application.

Claims

1. A motion sickness state recognition method based on Dropout-MTN, characterized in that: Perform steps A through C to obtain a motion sickness recognition model, and then proceed with the following steps. The motion sickness recognition model is applied to identify the motion sickness status of the target person for a corresponding target time period. Step A. Based on a sample detection duration equal to the target duration, for a preset number of individuals to be analyzed, obtain the EEG signal waveforms of each preset detection location on the surface of the brain of the individual to be analyzed, corresponding to the sample detection duration. This constitutes a multidimensional EEG detection signal waveform for the sample detection duration of the individual to be analyzed. Simultaneously, based on preset categories of motion sickness, including non-motion sickness, obtain the actual motion sickness category for the sample detection duration of the individual to be analyzed. The multidimensional EEG detection signal waveform and the actual motion sickness category constitute the EEG detection sample, thus obtaining the EEG detection sample corresponding to each individual to be analyzed. Then proceed to Step B. Step B. Construct the training network by sequentially connecting the input layer, multinomial layer, fully connected layer, softmax layer, and output layer from the input end to the output end, and then proceed to step C; Step C. Based on the auxiliary training network formed by the dropout layer connected in series between the multinomial layer and the fully connected layer in the network to be trained, according to each EEG detection sample, the multidimensional EEG detection signal waveform in the EEG detection sample is used as input and the motion sickness state category is used as output to train the auxiliary training network, thereby realizing the training of the network to be trained and obtaining the motion sickness state recognition model corresponding to the network to be trained. In step C, the auxiliary network to be trained, consisting of a Dropout layer cascaded between the multinomial layer and the fully connected layer in the network to be trained, is based on the output of the multinomial layer. The connection parameter is preset. Bernoulli distribution, and parameters corresponding to each detection location. For each detection location, according to The results, for Perform an update, updating the output of the Dropout layer. And transmit to the fully connected layer, where, , This indicates the number of detection locations on the brain surface corresponding to the multidimensional EEG detection signal waveform. Represents the polynomial layer with respect to the th The output of each detection location This indicates that the parameter is a preset value. Bernoulli distribution, corresponding to the first Parameters for each detection location; step The system collects EEG signal waveforms at preset detection locations on the target person's brain surface for each target time period, forming a multidimensional EEG detection signal waveform for the target person for the target time period. Then, it applies a motion sickness recognition model to obtain the motion sickness category of the target person for the target time period, thus completing the motion sickness recognition of the target person for the target time period.

2. The motion sickness recognition method based on Dropout-MTN according to claim 1, characterized in that: Step A includes the following steps A1 to A3: Step A1. Based on a sample detection duration equal to the target duration, for a preset number of individuals to be analyzed, obtain the EEG signal waveforms at each time point in the sample detection duration corresponding to each preset detection location on the surface of the brain of the individual to be analyzed. The average of the squares of the EEG signal waveforms corresponding to each detection location is used to construct the EEG signal waveforms corresponding to each detection location in the sample detection duration, thus constructing the multidimensional EEG detection signal waveforms for the sample detection duration corresponding to the individual to be analyzed. Then proceed to step A2. Step A2. Based on the preset categories of motion sickness, including those without motion sickness, and according to the SSQ scale corresponding to each person to be analyzed, obtain the parameters at each time point in the sample detection period for each person to be analyzed, for a preset number of people to be analyzed. , , The value is calculated according to the following formula: ； The total score of the person to be analyzed at each time point in the sample detection time segment was calculated. This allows us to obtain the total score at each point in time. The average score is calculated, and the correspondence between the preset average score and each motion sickness state category is used to obtain the true motion sickness state category of the sample detection time period for the person to be analyzed, and then proceed to step A3. Step A3. Use the multidimensional EEG detection signal waveform and the actual motion sickness state category of the corresponding sample detection time segment of the person to be analyzed as EEG detection samples to obtain the EEG detection samples corresponding to each person to be analyzed.

3. The motion sickness recognition method based on Dropout-MTN according to claim 1, characterized in that: In step B, the input layer is used to receive multidimensional EEG detection signal waveforms composed of EEG signal waveforms corresponding to sample detection durations at each preset detection location on the brain surface. And forward the polynomial layer, , This indicates the number of detection locations on the brain surface corresponding to the multidimensional EEG detection signal waveform. Indicates the first layer of the brain surface The detection location corresponds to the first EEG signals of each sample over a specified time period; The polynomial layer is used for the waveform of the received multidimensional EEG detection signal. Based on the preset highest expansion term corresponding to the network to be trained Construct the output of the polynomial layer As follows, and transmit it to the fully connected layer. ; ； in, Indicates the polynomial layer corresponding to the first The output of each detection location , , Indicates based on , The number of monomials in a polynomial Represents the corresponding polynomial in the layer. A weight vector for each detection location. express The Middle Each weight, These are the parameters to be trained in the network to be trained. , This indicates that the waveform of the multidimensional electroencephalogram (EEG) signal is about the highest order of expansion. The expansion of each term.

4. The motion sickness recognition method based on Dropout-MTN according to claim 1, characterized in that: Regarding the fully connected layer in step B, the output is based on the polynomial layer. Construct the output of the fully connected layer As shown below, and then transmitted to the Softmax layer; ； in, This indicates that the fully connected layer corresponds to the first... Output of each motion sickness state category. , , This represents the number of categories of motion sickness. Indicates the first A weighted column vector of each motion sickness state category. Indicates the first The detection location to the first The weights of each motion sickness category, These are the parameters to be trained in the network to be trained. This represents the ReLU activation function operation.

5. The motion sickness recognition method based on Dropout-MTN according to claim 1, characterized in that: Regarding the Softmax layer in step B, the output is based on the fully connected layer. ,according to Construct the output of the Softmax layer And transmit it to the output layer. Represents the natural constant. Indicates the corresponding number The probability of each motion sickness state category; the output layer is based on the probability of each corresponding motion sickness state category. The probability of each motion sickness state category is used to determine the motion sickness state category corresponding to the multidimensional EEG detection signal waveform received by the input layer, and then the category is output. This indicates that the fully connected layer corresponds to the first... The probability of outputting each motion sickness state category.

6. The motion sickness recognition method based on Dropout-MTN according to claim 1, characterized in that: In step C, regarding the training process of the auxiliary network to be trained, the parameters to be trained under each iteration are... With the parameters to be trained The mini-batch gradient descent method is used to update the data as follows; ； ； In order to represent The updated results In order to represent The updated result, where η represents the learning rate. This represents the result of the loss function.

7. A motion sickness recognition method based on Dropout-MTN according to claim 1 or 6, characterized in that: In step C, during the training process of the auxiliary network to be trained, the cross-entropy loss function is used as follows to optimize each training parameter in the network to be trained. ； in, This indicates the number of EEG samples collected during the training set of the network to be trained. The number of categories representing motion sickness; The sign function is defined as the function whose value is the first in the training set. The true motion sickness category of the EEG test sample is equal to the first Each motion sickness condition is categorized as follows: If the training set The true motion sickness category of each EEG test sample is not equal to the first Each motion sickness condition is categorized as follows: ; This represents the output of the training set under the auxiliary training of the network to be trained. The EEG sample belongs to the first The probability of each motion sickness state category Indicates the first training set under auxiliary training of the network to be trained. The loss function results for identifying motion sickness categories from a single EEG sample. This represents the loss function result of the training set under the auxiliary training of the network to be trained.

Images