A VR dizziness state detection and classification method based on a multi-dimensional Taylor net

Abstract

The application provides a VR vertigo state detection classification method based on a multidimensional Taylor network, including the following steps: S1. Taking typical electrodes representing different brain areas as classification features, obtaining corresponding data sets and dividing them into a training set and a test set; S2. Establishing a multidimensional Taylor network model, the multidimensional Taylor network model including an input layer, an intermediate layer, a fully connected layer, a softmax layer and an output layer; S3. Inputting data in the training set into the network to perform training, and using a cross-entropy loss function to evaluate the data processing effect of the multidimensional Taylor network model; S4. Testing data in the test set by using the multidimensional Taylor network model trained in step S3, obtaining a predicted classification label of the VR vertigo state, so as to realize classification and identification of the VR vertigo state. The application can effectively reduce the calculation complexity and the training time, and improve the detection speed of the VR vertigo state without reducing the classification accuracy.

Description

Technical Field

[0001] This invention relates to the field of VR motion sickness state detection and classification, and in particular to a VR motion sickness state detection and classification method based on a multidimensional Taylor network. Background Technology

[0002] Virtual Reality (VR) technology is a novel and practical technology that emerged in the 20th century. It is widely used in industries such as medicine, education, military, and film. With social progress and continuous development of science and technology, the demand for VR technology is increasing across various sectors. However, users often experience dizziness during VR experiences, and subjective assessments are easily affected by individual differences and environmental factors, failing to accurately reflect the user's level of dizziness.

[0003] The ability to detect and classify VR-induced dizziness levels can significantly promote the development of the VR industry. Electroencephalogram (EEG) signal analysis is an effective method for analyzing the electrical activity signals of the cerebral cortex, objectively reflecting physiological symptoms such as fatigue and stress. Combining EEG signal analysis with machine learning to classify dizziness states has significant research and application value for detecting dizziness levels. Commonly used machine learning classifiers include K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and Random Forests. However, each classification method has its limitations. For example, KNN has a high computational cost, especially when the number of features is very large. SVM is difficult to implement with large-scale training samples and struggles with multi-classification problems. Random Forests, due to their inherent complexity, require more training time than other similar algorithms and may overfit in noisy classification or regression problems. Summary of the Invention

[0004] Purpose of the invention

[0005] The purpose of this invention is to propose a VR dizziness state detection and classification method with low computational complexity, short training time, and the ability to improve detection speed without reducing classification accuracy.

[0006] Technical solution

[0007] A VR motion sickness state detection and classification method based on multidimensional Taylor networks includes the following steps:

[0008] S1. Take typical electrodes representing different brain regions as classification features. Based on the selection of classification features, obtain the corresponding dataset. Determine the label of the vertigo state corresponding to each group of data in the dataset according to the actual measurement table, and divide the dataset into training set and test set.

[0009] S2. Establish a multidimensional Taylor network model, which includes an input layer, intermediate layers, fully connected layers, a softmax layer, and an output layer;

[0010] The input layer receives external input variables, which are represented as follows:

[0011] X(k) = {x1(k), x2(k), ... x i (k)…,x n (k)} T (k = 1, 2, 3... l);

[0012] Where, x i (k) represents all data containing typical electrode signals of the i-th brain region, l represents the total number of people containing data containing typical electrode signals of all brain regions, and n represents the number of different types of typical electrode signals of different brain regions contained in the input variable.

[0013] The input variables obtained from the input layer are normalized and then passed to the intermediate layer, where they undergo m-th term expansion. Each input variable in the intermediate layer is weighted and summed for each power-order product term. Values ​​are then assigned to the intermediate layer nodes, and the corresponding weight vector matrix W is represented as follows:

[0014]

[0015] Where w represents the elements of the weight vector, and N(n,m) represents the total number of product terms in the approximation expansion;

[0016] intermediate layer node y i (i = 1, 2, 3... n) is represented as

[0017]

[0018] Where, λ j,i This represents the variable x in the j-th product term of the approximation expansion. i powers of;

[0019] The intermediate layer nodes are processed using an activation function to obtain the intermediate layer vector α, which is represented as follows:

[0020]

[0021] The number of nodes in the fully connected layer is the same as the number of data types. Each node in the fully connected layer obtains the intermediate layer vector α, performs a weighted summation, and assigns a value to each node. The corresponding weight vector σ is represented as follows:

[0022]

[0023] Where s represents the number of data categories;

[0024] Fully connected layer node t i (i = 1, 2, ..., s) is represented as

[0025]

[0026] The nodes of the fully connected layer are processed using an activation function to obtain the input vector β of the softmax layer. The input vector of the softmax layer is represented as follows:

[0027]

[0028] The softmax layer calculates the class probability value P using the input vector β. i (i = 1, 2, ..., s), the probability value of the category P i The calculation formula is

[0029]

[0030] The output layer obtains the probability value P of the class. i And output it.

[0031] S3. Input the data from the training set into the network, perform training, obtain the class probability values ​​as output, and evaluate the data processing performance of the multidimensional Taylor network model using the cross-entropy loss function. The expression for the cross-entropy loss function is:

[0032]

[0033] Among them, Y i This is a tag matrix with an l·s structure based on one-hot encoding technology. Each row of the label matrix has exactly one element that is 1, and all other elements are 0. The smaller the calculated result of the cross-entropy loss function, the better the training effect.

[0034] Training ends after all data in the training set has been processed.

[0035] The multidimensional Taylor network model proposed in this invention has a simple network structure, which reduces the difficulty of network training and effectively reduces computational complexity.

[0036] S4. Use the multidimensional Taylor network model trained in step S3 to test the data in the test set to obtain the predicted classification label of VR dizziness state, thereby realizing the classification and recognition of VR dizziness state.

[0037] Furthermore, the actual measurement form examines three conditions: gastrointestinal discomfort (N), eye discomfort (O), and orientation disorder (D). Gastrointestinal discomfort (N) includes general discomfort, increased salivation, sweating, nausea, poor concentration, stomach upset, and hiccups. Eye discomfort (O) includes general discomfort, fatigue, headache, eye strain, difficulty focusing, poor concentration, and blurred vision. Orientation disorder (D) includes difficulty focusing, nausea, headache, blurred vision, dizziness (with eyes open), dizziness (with eyes closed), and vertigo. Each condition is represented by a value of 0 or 1, with 1 indicating the presence of a specific discomfort and 0 indicating the absence of a specific discomfort. The values ​​for N, O, and D are the sum of the values ​​for their respective conditions. The total score (T) is calculated using the following formula:

[0038] T = N + O + D × 0.34;

[0039] When a sample's total score is higher than the average score, it is considered to have a state of dizziness and is labeled 1; when a sample's total score is lower than the average score, it is considered not to have a state of dizziness and is labeled 0.

[0040] The practical measurement table proposed in this invention contains a systematic and rich set of physiological characteristics, providing a reliable basis for data analysis.

[0041] Furthermore, the activation function is the ReLU function, with the following form:

[0042] This invention uses the ReLU function, which makes it easier for the model to represent complex nonlinear functions using polynomial functions, further simplifying the network structure and reducing the difficulty of training.

[0043] Beneficial effects

[0044] This invention proposes a VR vertigo state detection and classification method based on a multidimensional Taylor network. It is based on a systematic and rich actual measurement table of physiological characteristics. The network model itself has a scientific and concise structure, low computational complexity, short training time, and can improve the detection speed without reducing the classification accuracy. Attached Figure Description

[0045] Figure 1 This is a logic flowchart of the present invention;

[0046] Figure 2 This is a diagram of the multidimensional Taylor network structure proposed in this invention;

[0047] Figure 3 This is a diagram illustrating the data detection and classification results of an embodiment of the present invention. Detailed Implementation

[0048] To enable those skilled in the art to better understand the present invention, the invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0049] A VR motion sickness state detection and classification method based on multidimensional Taylor networks includes the following steps:

[0050] S1. Typical electrodes representing 30 brain regions are selected as classification features. Based on the selection of classification features, a corresponding dataset is obtained. The label for each group of data in the dataset corresponding to the dizziness state is determined according to the actual measurement table. The dataset is then divided into a training set and a test set. The actual measurement table examines three conditions: gastrointestinal discomfort (N), eye discomfort (O), and orientation disorder (D). Gastrointestinal discomfort (N) includes general discomfort, increased salivation, sweating, nausea, poor concentration, stomach upset, and hiccups. Eye discomfort (O) includes general discomfort, fatigue, headache, eye fatigue, difficulty focusing, poor concentration, and blurred vision. Orientation disorder (D) includes difficulty focusing, nausea, headache, blurred vision, dizziness (with eyes open), dizziness (with eyes closed), and vertigo. The values ​​for each condition are either 0 or 1. A value of 1 indicates the presence of a specific discomfort state, while a value of 0 indicates the absence of a specific discomfort state. The values ​​of N, O, and D are the sum of the values ​​for each specific condition. The formula for calculating the total score T is...

[0051] T = B + O + D × 0.34;

[0052] When a sample's total score is higher than the average score, it is considered to have a state of dizziness and is labeled 1; when a sample's total score is lower than the average score, it is considered not to have a state of dizziness and is labeled 0.

[0053] S2. Establish a multidimensional Taylor network model, which includes an input layer, intermediate layers, fully connected layers, a softmax layer, and an output layer;

[0054] The input layer receives external input variables, which are represented as follows:

[0055] X(k) = {x1(k), x2(k), ... x i (k)…,x n (k)} T (k = 1, 2, 3... l);

[0056] Where, x i (k) represents all data containing typical electrode signals of the i-th brain region, l represents the total number of people containing data containing typical electrode signals of all brain regions, and n represents the number of different types of typical electrode signals of different brain regions contained in the input variable.

[0057] The input variables obtained from the input layer are normalized and then passed to the intermediate layer for quadratic expansion. Each input variable undergoes a weighted summation of its product terms at each power level in the intermediate layer. Values ​​are then assigned to the intermediate layer nodes, and the corresponding weight vector matrix W is represented as follows:

[0058]

[0059] Where w represents the elements of the weight vector, and N(n,2) represents the total number of product terms in the approximation expansion;

[0060] intermediate layer node y i (i = 1, 2, 3... n) is represented as

[0061]

[0062] Where, λ j,i This represents the variable x in the j-th product term of the approximation expansion. i powers of;

[0063] The intermediate layer nodes are processed using an activation function, specifically the ReLU function, which has the following form: The intermediate layer nodes are processed to obtain the intermediate layer vector α, which is represented as follows:

[0064]

[0065] The number of nodes in the fully connected layer is the same as the number of data types. Each node in the fully connected layer obtains the intermediate layer vector α, performs a weighted summation, and assigns a value to each node. The corresponding weight vector σ is represented as follows:

[0066]

[0067] Where s represents the number of data categories;

[0068] Fully connected layer node t i (i = 1, 2, ..., s) is represented as

[0069]

[0070] The nodes of the fully connected layer are processed using the activation function described above to obtain the input vector β of the softmax layer. The input vector of the softmax layer is represented as follows:

[0071]

[0072] The softmax layer calculates the class probability value P using the input vector β. i (i = 1, 2, ..., s), the probability value of the category P i The calculation formula is

[0073]

[0074] The output layer obtains the class probability value P. i And output it;

[0075] S3. Input the data from the training set into the network, perform training, obtain the class probability values ​​as output, and evaluate the data processing performance of the multidimensional Taylor network model using the cross-entropy loss function. The expression for the cross-entropy loss function is:

[0076]

[0077] Among them, Y i This is a tag matrix with an l·s structure based on one-hot encoding technology. Each row of the label matrix has exactly one element that is 1, and all other elements are 0. The smaller the calculated result of the cross-entropy loss function, the better the training effect.

[0078] Training ends after all data in the training set has been processed.

[0079] S4. The multidimensional Taylor network model trained in step S3 is used to test the data in the test set to obtain the predicted classification label of VR dizziness state, thereby realizing the classification and recognition of VR dizziness state. The 50 actual test data points input in this embodiment are shown in the following six tables. Each table contains the test signals from different brain regions, where all data points with the same number constitute a complete test data point. The final prediction result is as follows: Figure 3 As shown.

[0080] Table 1

[0081]

[0082]

[0083]

[0084] Table 2

[0085]

[0086]

[0087]

[0088] Table 3

[0089]

[0090]

[0091]

[0092] Table 4

[0093]

[0094]

[0095] Table 5

[0096]

[0097]

[0098] Table 6

[0099]

[0100]

[0101]

[0102] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A VR motion sickness state detection and classification method based on multidimensional Taylor networks, characterized in that, Includes the following steps: S1. Take typical electrodes representing different brain regions as classification features. Based on the selection of classification features, obtain the corresponding dataset. Determine the label for the dizziness state corresponding to each group of data in the dataset according to the actual measurement table, and divide the dataset into training set and test set. S2. Establish a multidimensional Taylor network model, which includes an input layer, intermediate layers, fully connected layers, a softmax layer, and an output layer; The input layer receives external input variables, which are represented as follows: X(k)={x1(k),x2(k),…x i (k)…,x n (k)} T (k=1,2,3...l); Where, x i (k) represents all data containing typical electrode signals of the i-th brain region, l represents the total number of people containing data containing typical electrode signals of all brain regions, and n represents the number of different types of typical electrode signals of different brain regions contained in the input variable. The input variables obtained from the input layer are normalized and then passed to the intermediate layer, where they undergo m-th term expansion. Each input variable in the intermediate layer is weighted and summed for each power-order product term. Values ​​are then assigned to the intermediate layer nodes, and the corresponding weight vector matrix W is represented as follows: Where w represents the elements of the weight vector, and N(n,m) represents the total number of product terms in the approximation expansion; intermediate layer node y i (i = 1, 2, 3... n) is represented as Where, λ j,i The variable x in the h-th product term of the approximation expansion i powers of; The intermediate layer nodes are processed using an activation function to obtain the intermediate layer vector α, which is represented as follows: The number of nodes in the fully connected layer is the same as the number of data types. Each node in the fully connected layer obtains the intermediate layer vector α, performs a weighted summation, and assigns a value to each node. The corresponding weight vector σ is represented as follows: Where s represents the number of data categories; Fully connected layer node t i (i = 1, 2, ..., s) is represented as The nodes of the fully connected layer are processed using the activation function described above to obtain the input vector β of the softmax layer. The input vector of the softmax layer is represented as follows: The softmax layer calculates the class probability value P using the input vector β. i (i = 1, 2, ..., s), the probability value of the category P i The calculation formula is The output layer obtains the probability value P of the class. i And output it; S3. Input the data from the training set into the network, perform training, obtain the class probability values ​​as output, and evaluate the data processing performance of the multidimensional Taylor network model using the cross-entropy loss function. The expression for the cross-entropy loss function is: Among them, Y i This is a tag matrix with an l·s structure based on one-hot encoding technology. Each row of the label matrix has exactly one element that is 1, and all other elements are 0. The smaller the calculated result of the cross-entropy loss function, the better the training effect. Training ends after all data in the training set has been processed. S4. Use the multidimensional Taylor network model trained in step S3 to test the data in the test set to obtain the predicted classification label of VR dizziness state, thereby realizing the classification and recognition of VR dizziness state.

2. The VR motion sickness state detection and classification method based on multidimensional Taylor networks according to claim 1, characterized in that: The actual measurement table examines three conditions: gastrointestinal discomfort (N), eye discomfort (O), and orientation disorder (D). Gastrointestinal discomfort (N) includes general discomfort, increased salivation, sweating, nausea, poor concentration, stomach upset, and hiccups. Eye discomfort (O) includes general discomfort, fatigue, headache, eye strain, difficulty focusing, poor concentration, and blurred vision. Orientation disorder (D) includes difficulty focusing, nausea, headache, blurred vision, dizziness (with eyes open), dizziness (with eyes closed), and vertigo. Each condition is represented by a value of 0 or 1, where 1 indicates the presence of a specific discomfort and 0 indicates the absence of a specific discomfort. The values ​​for N, O, and D are the sum of the values ​​for each specific condition. The total score (T) is calculated using the following formula: T = N + O + D × 0.34; When a sample's total score is higher than the average score, it is considered to have a state of dizziness and is labeled 1; when a sample's total score is lower than the average score, it is considered not to have a state of dizziness and is labeled 0.

3. The VR motion sickness state detection and classification method based on multidimensional Taylor networks according to claim 1, characterized in that: The activation function is the ReLU function, with the function form as follows:

Images