Electroencephalogram signal monitoring method for driver

Abstract

According to the electroencephalogram signal monitoring method for the driver, the electroencephalogram signals are processed under the non-denoising condition, the problem that useful information of data of the electroencephalogram signals is filtered out together during denoising is solved, it is ensured that the detection precision is higher, and meanwhile compared with a common analysis method with the electroencephalogram signals subjected to denoising processing as objects, the electroencephalogram signal monitoring method has the advantages that the accuracy is higher; according to the method, the steps are simplified, the execution efficiency is higher, and the method is more suitable for scenes with high real-time requirements.

Description

Technical Field

[0001] The present application relates to the field of intelligent vehicle technology, and in particular to a method for monitoring driver's electroencephalogram signals. Background Art

[0002] How to promptly remind the driver of changes in EEG signals during driving, or to perform intelligent braking of the vehicle in a timely manner based on changes in the driver's EEG signals, is a technical direction worth exploring for intelligent driving systems.

[0003] Hans Berger first proposed the electroencephalography (EEG). By recording the EEG displayed during the epileptic stage of the EEG signal, the data of the EEG signal is collected, and then the extracted data is subjected to simple calculation, denoising and filtering, providing a data basis for subsequent epilepsy research. Epileptic seizures produce a variety of brain waves: irregular waveforms such as short waves and slow waves, as well as many relevant values ​​of the EEG active areas in the brain that are of reference value. By observing the waveform of the EEG signal, scholars found that there are significant differences between the waveforms of EEG signals during and without epileptic seizures. The signal complexity during epileptic seizures is reduced, and the entropy value used to study the complexity of the time series can also be used as a feature standard for epilepsy detection. Bai Dongmei and others used the sample entropy of nonlinear dynamics as the EEG feature signal and found that the sample entropy is very effective as a feature for epilepsy detection. However, there are also defects when using entropy for research: in epilepsy features, the sample entropy calculation is complex and time-consuming. At present, other nonlinear feature research parameters include approximate entropy, Hurst exponent and other related dimensions.

[0004] Of course, there are devices in the prior art that monitor human EEG signals based on the theories and methods mentioned above, but most of them are devices for use in medical systems, while the equipment used in smart vehicles must be a real-time detection system, which requires timeliness in the calculation of EEG characteristics. However, most of the EEG signal detection devices in the prior art are based on nonlinear characteristics for calculation, which is highly complex and time-consuming, and cannot meet the real-time requirements of intelligent driving systems. Summary of the invention

[0005] In order to solve the problem that the EEG detection method in the prior art does not meet the real-time requirements of the intelligent driving system, the present application provides a driver's EEG signal monitoring method, which can accurately monitor the occurrence of EEG signals with epilepsy characteristics in the driver's brain waves in real time.

[0006] The technical solution of the present application is as follows: a method for monitoring driver's electroencephalogram signals, comprising the following steps: S1: Prepare EEG acquisition equipment; The brain wave acquisition device acquires the driver's brain wave signals in real time based on the cerebral cortex electrode patch, and a data communication link exists between the brain wave acquisition device and the vehicle intelligent control system; S2: Construct the brain network structure; the construction process specifically includes the following steps: a1: define network nodes; Selecting the leads at the electrode patch positions set in the EEG acquisition device as network nodes; the number of EEG channels is the same as the number of electrode patches; a2: Perform variational mode decomposition on the EEG signal, select and calculate the separated IMF components corresponding to each brain channel, perform coupling operations on the brain network edges through phase-locked values, and then obtain the weighted network matrix; a3: Use the density method to binarize the values ​​in the network so that the number of nodes connected is the same; set the sparsity and construct the brain network according to the sparsity; The sparsity includes: density range and step size; S3: Setting the classification feature parameters for classification; The characteristic parameters for classification include: brain network characteristic parameters and sample entropy characteristics; The brain network characteristic parameters include: global efficiency, local efficiency, clustering coefficient and characteristic path length; S4: Construct an epilepsy detection classification model based on the classification algorithm model; S5: constructing a training data set, and using the training data set to train the epilepsy detection classification model to obtain a trained epilepsy detection classification model; S6: embedding the epilepsy detection classification model into the intelligent vehicle control system; extracting the brain network characteristic parameters based on the brain wave signals collected by the brain wave acquisition device, constructing the characteristic parameters for classification together with the sample entropy characteristics, sending the characteristic parameters for classification into the trained epilepsy detection classification model for classification, and outputting the classification results.

[0007] It is further characterized by: The EEG acquisition device includes 18 electrode patches; In step a2, for the continuous EEG signal, the epileptic and non-epileptic time periods are segmented and processed respectively; the EEG signal is decomposed into a specified number of n IMF components by using the variational mode decomposition method, and the nodes of the brain network are formed through the EEG channels. The brain channel signals are phase-coupled through the phase-locked value, and the coupling relationship established between the channels is used to establish the edges of the brain network, so as to construct n different brain networks; In step a2, n is set to 6; The EEG signal was decomposed into 6 IMF components and 6 different brain networks were constructed; In the sparsity, the density range is set to 15% to 45%, and every 5% is set as a step length; In step S4, an epilepsy detection classification model is constructed based on a support vector machine algorithm model; In step S5, the training data set is constructed after extracting the non-denoised EEG information.

[0008] The present application provides a method for monitoring EEG signals for drivers, which processes EEG signals under non-denoising conditions, thereby eliminating the problem of filtering out useful information of EEG signal data during denoising, ensuring higher detection accuracy. At the same time, compared with general analysis methods using EEG signals that have been denoised as the object, the method simplifies the steps, has higher execution efficiency, and is more suitable for scenarios with high real-time requirements. The method can specify the desired number of modes by selecting variational mode decomposition, and can also effectively avoid mode aliasing, thereby further improving classification accuracy. The method constructs a brain network, comprehensively compares and selects features such as characteristic path length, clustering coefficient, global efficiency, local efficiency, and sample entropy to extract features, divides the feature values ​​into a test set and a training set to train a support vector machine, and provides a classification model for subsequent epilepsy detection, ensuring that EEG signals can be effectively and accurately classified and identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 Building processes for brain networks; Figure 2 This is a brain network flowchart based on epilepsy detection; Figure 3 This is a comparison chart of detection performance under different classifiers; Figure 4 It is the AUC curve diagram under support vector machine; Figure 5 Example 1 for different network detection results; Figure 6 Example 2 for different network detection results; Figure 7 Example 3 shows different network detection results. DETAILED DESCRIPTION

[0010] In order to ensure the timeliness of monitoring, the method realizes epilepsy detection based on brain network in a noisy environment. The application includes a method for monitoring driver's electroencephalogram signals, which includes the following steps.

[0011] S1: Prepare brain wave collection equipment for the driver; The brain wave acquisition device collects the driver's brain wave signals in real time based on the cerebral cortex electrode patch, and there is a data communication link between the brain wave acquisition device and the vehicle intelligent control system. The specific implementation can be based on the vehicle health monitoring device in the prior art.

[0012] S2: Construct brain network structure.

[0013] Structural networks, functional networks, and causal networks are generally considered to be the main building blocks of brain networks. The purpose of building brain networks is to visualize the signal transmission information within the brain.

[0014] The construction process is as follows Figure 1 As shown in the figure, there are three steps to construct a brain network: (1) define the network nodes. (2) measure the edges and calculate the coupling relationship between the nodes. (3) select the size of the threshold. Most network studies focus on binary networks, while the binary nature of weighted networks is determined by appropriate threshold selection. The detailed steps are shown below.

[0015] a1: define network nodes; The most important thing in building a brain network is how to choose appropriate nodes and edges, which will affect the subsequent calculation of network properties, analysis of topological characteristics and neurobiological interpretation. When selecting nodes, information elements that can represent each area of ​​the brain should be considered as an important factor to ensure that the information contained in the nodes is the same and complete, and the inherent differences between nodes should also be maintained. Generally speaking, the constructed nodes can fully represent the brain areas with important research value in the electrodes inside the brain. In the study of brain networks, the number of nodes is used to represent the distribution range of brain network areas.

[0016] Because the research object is the EEG signal collected by the cerebral cortex electrode patch, this method selects the lead at the electrode patch position set in the EEG acquisition device as the network node; the number of EEG channels is the same as the number of electrode patches; the EEG acquisition device includes 18 electrode patches, that is, a node composed of 18 selected EEG channels.

[0017] a2: Edge measurement, calculating the coupling relationship between nodes.

[0018] The length from one node to another in the network is represented as an edge. There are many methods for calculating the strength of coupling relationships. Methods for calculating undirected network edges include classical methods in the time domain, such as correlation, cross-correlation, and frequency domain methods, such as coherence, phase-locking value, phase lag index, and weighted phase lag index. In this method, the EEG signal is subjected to variational mode decomposition, and the phase-locking value has strong noise resistance, so the IMF component corresponding to the separation of each brain channel is calculated, and the coupling operation of the brain network edge is performed through the phase-locking value, thereby obtaining a weighted network matrix. Ensure that this method is suitable for the analysis of EEG waves in a noisy environment.

[0019] In step a2, the continuous EEG signal is segmented and processed for the epileptic and non-epileptic periods respectively; the variational mode decomposition (VMD) method is used to decompose the EEG signal into a specified number of n IMF components, and the nodes of the brain network are formed through the EEG channels. The brain channel signals are phase-coupled through the phase-locked value, and the coupling relationship between the channels is used to establish the edges of the brain network, thereby constructing n different brain networks.

[0020] In this embodiment, the value of n is 6; the EEG signal is decomposed into 6 IMF components, and 6 different brain networks are constructed.

[0021] a3: Select threshold.

[0022] When analyzing brain networks, the construction of weighted networks or binary networks is selected according to the specific research content. To convert a weighted network into a 0 and 1 value network, an appropriate threshold should be selected. Binarization is to compare a pre-set threshold with the weight of the edge established in the weighted network. If the weight value of the network edge is greater than the threshold, a connection relationship is established for the nodes at both ends of the edge, and the weight of the edge is set to 1. On the contrary, it is considered that the nodes at both ends have no connection relationship, and the weight of the edge is set to 0. The sparsity of binary network construction and the study of subsequent network characteristics are inseparable from the choice of thresholds. Smit DJA et al. studied the characteristics of resting brain networks under multiple thresholds of brain networks, and analyzed the characteristics of brain networks under thresholds with an average node degree range of 3 to 7.

[0023] In this method, after calculating the relationship matrix between channels, the threshold is determined to convert it into a binary matrix so as to further construct the brain network. It is worth noting that when the threshold is set too large, the connection edges in the brain network will be reduced and some important connections will be ignored. When the threshold is set too small, there will be too many connection edges in the brain network, thus introducing some wrong connections. Therefore, the setting of the threshold is very important for the establishment of the network.

[0024] In this method, the density method is used to binarize the values ​​in the network so that the number of nodes connected is the same; the sparsity is set, and the brain network is constructed according to the sparsity; the sparsity includes: density range and step size. In this embodiment, in the sparsity, the density range is set to 15% to 45%, and every 5% is set as a step size.

[0025] S3: Setting the classification brain network feature parameters for classification.

[0026] In this method, complex EEG signals are classified by brain network characteristic parameters to show the difference between epileptic seizures and non-epilepsy. Complex brain networks can be studied through many attribute characteristics. In order to study the impact of brain activity during epileptic seizures, the classification brain network characteristic parameters in this method include: global efficiency, local efficiency, clustering coefficient, characteristic path length and sample entropy. Among them, the four brain network characteristic parameters of global efficiency, local efficiency, clustering coefficient and characteristic path length have small computational complexity and can also better represent the functional characteristics of the brain network, which is especially suitable for application scenarios with timeliness requirements. At the same time, representative sample entropy features are selected in nonlinear dynamic analysis for feature extraction. In this method, clustering coefficient, global efficiency, characteristic path length, local efficiency and sample entropy are selected to express the significant differences between EEG signals during epileptic seizures and non-epilepsy. These features are used as parameters for epilepsy detection and classification, which can accurately realize the recognition and classification of EEG signals.

[0027] Among them, Characteristic Path Length is a global characteristic attribute that describes the network. It refers to the average value of the path length between two nodes in a network, from one node through zero or several nodes to the target node, expressed as: ; In the formula, the number of network nodes is represented by N, and the minimum number of edges connecting two nodes in all paths from node i to node j is represented by d ij .

[0028] Global efficiency refers to the possibility of each node in the network being connected to another node. The higher the value of global efficiency, the greater the possibility that there is a path from node to node in the network and that there is a path between nodes to transmit information to each other. It is used to judge the efficiency of network transmission and information processing. This advantage is more obvious when calculating in non-connected networks such as busy networks. Global efficiency E global It is expressed as: ; In the formula, the number of network nodes is represented by N, and the shortest path length from network node j to node k is represented by d j,k .

[0029] In graph theory, the clustering coefficient is used to measure the degree of connection between a node and other surrounding nodes. The smaller the weight, the closer the node is to the node, and the larger the weight, the farther the node is. In this application, the local clustering coefficient is used to study the local characteristics of the network, reflecting the local connectivity of the nodes in the network. The goal of the clustering coefficient is to compare the degree of aggregation of the group with the degree of aggregation that can be achieved. The clustering coefficient c of node i is i It is expressed as: ; In the formula, the k edges of node i are represented as k i , there are at most (k(k-1)) / 2 edges between nodes, and E is the number of connected nodes.

[0030] Regarding local efficiency, different partitions within the brain complete the functional area distribution of each part. In this method, by comparing the information processing rate and transmission efficiency in the partitions, observing the local efficiency index represented by them, we can observe the network characteristics under the quantitative details of the brain network. The local efficiency e i It is expressed as: ; In the formula, the number of network nodes is represented by N, and the shortest path length between node i and node j is represented by d ij .

[0031] S4: constructing an epilepsy detection classification model based on the classification algorithm model. In this embodiment, an epilepsy detection classification model is constructed based on a support vector machine algorithm model.

[0032] In this method, the changes in EEG signal characteristics during epileptic seizures are studied by constructing a brain network. The specific brain network structure process based on epilepsy detection is as follows: Figure 2 As shown, in this embodiment, the brain network is constructed by each component of the EEG signal in 18 channels to complete the classification detection of epilepsy. The detection time of EEG signal is long, and the continuous EEG signal is segmented and processed for epilepsy and non-epileptic periods according to a frequency of 256 and a period of 2 seconds. The variational mode decomposition method is used to decompose the EEG signal into 6 IMF components, and the nodes of the brain network are formed by 18 EEG channels. The brain channel signals are phase-coupled through the phase-locked value, and the coupling relationship established between the channels is used to establish the edge of the brain network, and 6 different brain networks are constructed.

[0033] The classification feature parameters used for epileptic EEG signal detection set in this method include: brain network features and sample entropy features.

[0034] The properties of the brain network studied include clustering coefficient, global efficiency, modularity, node betweenness, etc., and the nonlinear dynamics analysis includes fractal dimension, Lyapunov, sample entropy and other features. In this method, the characteristic path length, global efficiency, clustering coefficient, and local efficiency feature values ​​are extracted, and representative sample entropy features are selected for feature extraction in nonlinear dynamics analysis. After feature extraction, these features are screened by one-way analysis of variance, and it is found that the clustering coefficient, global efficiency, characteristic path length, local efficiency, and sample entropy parameters have significant differences when epilepsy occurs and when it does not occur. Therefore, these features are used as parameters for epilepsy detection and classification.

[0035] Commonly used classification models include support vector machine, decision tree and random forest. This method selects these three classification models to complete the construction of epilepsy detection classification model. The EEG signal data set is divided into a training set and a test set. The extracted features are first trained with the training set to create a model. The features that can complete epilepsy detection are put into the three classification models of support vector machine, decision tree and random forest for classification. Experiments show that SVM is the best and completes the goal of epilepsy detection classification based on brain network.

[0036] S5: Construct a training data set, and use the training data set to classify the epilepsy detection model to obtain a trained epilepsy detection classification model. In this embodiment, the training data set is constructed after extracting non-denoised EEG information.

[0037] S6: Embed the epilepsy detection classification model into the intelligent vehicle control system; extract brain network feature parameters based on the EEG signals collected by the EEG acquisition device, construct classification feature parameters together with the sample entropy features, send the classification feature parameters to the trained epilepsy detection classification model for classification, and output the classification results.

[0038] During the specific implementation, a module formed by the brain network structure based on epilepsy detection will be installed on the vehicle computer, which can detect whether epilepsy is triggered for the cured driver. When triggered, the vehicle system takes over the driving of the vehicle, enters automatic driving to a safe position and calls 120 for emergency treatment in time.

[0039] In order to confirm the effectiveness of this method, the following experiments and analyses were performed. Accuracy, sensitivity, receiver operating characteristic curve and specificity were selected as performance metrics. The definitions of accuracy (ACC), sensitivity (SEN) and specificity (SPE) are as follows.

[0040] ; ; ; In epilepsy detection, epileptic seizure samples are divided into four combined categories based on the actual results and predicted results: True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). ACC is the overall prediction standard. The higher the value, the better the classification. SEN is the prediction standard for positive examples, and SPE is the prediction standard for negative classes.

[0041] The Receiver Operating Characteristic (ROC) curve measures classification performance and generalization, and draws a numerical curve to represent the classification result. Sensitivity is also called the true positive rate, and specificity is called the false positive rate. The sum of sensitivity and specificity is 1. In the ROC curve, the vertical axis is the true positive rate, the horizontal axis is the false positive rate, and the area under the ROC curve (Area Under ROC, AUC) determines the classification result.

[0042] Based on this method, a brain network was constructed and feature parameters for classification were extracted. Then, three classification models, support vector machine, decision tree and random forest, were selected to implement epilepsy classification detection. The results are shown in Figure 2. Figure 3 shown.

[0043] Depend on Figure 3 It is known that the values ​​of sensitivity, specificity and accuracy of support vector machine (SVM), decision tree and random forest are all over 80%, and the difference between the three values ​​is small. Considering the individual differences, SVM support vector machine, which is suitable for small samples to avoid dimensionality disaster, is selected for classification detection. Support vector machine shows the best classification results, with an accuracy of 93.5%, a specificity of 88.9% and a sensitivity of 85%.

[0044] After classification by support vector machine, the ROC curve and AUC values ​​are as follows: Figure 4 As shown. Figure 4 It is known that the AUC value is 0.88, the accuracy is 93.5%, the specificity is 88.9%, and the sensitivity is 85%, achieving the expected effect of more than 90% accuracy. In the non-denoising environment, the classification of epilepsy detection is performed, and the AUC value is close to 0.9. The classification effect is obvious and can be used to detect epileptic seizures in epilepsy patients.

[0045] In order to highlight the advantages of the brain network in this application, different networks were used and the same classification method was selected to detect epileptic seizures in epilepsy patients. The results are as follows Figure 5 , Figure 6 , Figure 7As shown in the figure, the horizontal axis is the name of the network model, and the vertical axis is the detection result. This experiment selects the time network, space-time network and the brain network model in this application.

[0046] based on Figure 5 From the data, it can be seen that in the support vector machine classification method, the sensitivity, specificity and accuracy of the brain network detection results designed by the present application are better than those of the spatial network and the temporal network, with a sensitivity of 85%, a specificity of 89% and an accuracy of 94%, while the sensitivity, specificity and accuracy of the space-time network and the temporal network are between 70% and 80%, and the detection accuracy is lower.

[0047] based on Figure 6 From the data, it can be seen that in the decision tree classification method, the sensitivity, specificity and accuracy of the brain network detection results designed by the present application are better than those of the spatial network and the temporal network, with a sensitivity of 90%, a specificity of 92% and an accuracy of 95%. The sensitivity, specificity and accuracy of the spatiotemporal network are between 80% and 90%, and the sensitivity, specificity and accuracy of the temporal network are the lowest.

[0048] based on Figure 7 From the data, it can be seen that in the random forest classification method, the sensitivity, specificity and accuracy of the brain network detection results designed by the present application are better than those of the spatial network and the temporal network, among which the sensitivity is 96%, the specificity is 82%, and the accuracy is 93%. The sensitivity and accuracy are much higher than those of the space-time network and the temporal network, and the specificity is slightly better than that of the space-time network and the temporal network.

[0049] In summary, since the brain network constructed in this application can process non-denoised EEG signals and extract more epilepsy feature information, the denoised EEG signals will delete some useful epilepsy feature information; and the variational mode decomposition is used to classify and extract different types of epilepsy feature information. Compared with the collective extraction of feature information of space-time networks and time networks, the brain network constructed in this application is better.

Claims

1. A method for monitoring driver's electroencephalogram signals, comprising the following steps: S1: Prepare EEG acquisition equipment; The brain wave acquisition device acquires the driver's brain wave signals in real time based on the cerebral cortex electrode patch, and a data communication link exists between the brain wave acquisition device and the vehicle intelligent control system; S2: Construct the brain network structure; the construction process specifically includes the following steps: a1: define network nodes; Selecting the leads at the electrode patch positions set in the EEG acquisition device as network nodes; the number of EEG channels is the same as the number of electrode patches; a2: Perform variational mode decomposition on the EEG signal, select and calculate the separated IMF components corresponding to each brain channel, perform coupling operations on the brain network edges through phase-locked values, and then obtain the weighted network matrix; a3: Use the density method to binarize the values ​​in the network so that the number of nodes connected is the same; set the sparsity and construct the brain network according to the sparsity; The sparsity includes: density range and step size; S3: Setting the classification feature parameters for classification; The characteristic parameters for classification include: brain network characteristic parameters and sample entropy characteristics; The brain network characteristic parameters include: global efficiency, local efficiency, clustering coefficient and characteristic path length; S4: Construct an epilepsy detection classification model based on the classification algorithm model; S5: constructing a training data set, and using the training data set to train the epilepsy detection classification model to obtain a trained epilepsy detection classification model; S6: embedding the epilepsy detection classification model into the intelligent vehicle control system; extracting the brain network characteristic parameters based on the brain wave signals collected by the brain wave acquisition device, constructing the characteristic parameters for classification together with the sample entropy characteristics, sending the characteristic parameters for classification into the trained epilepsy detection classification model for classification, and outputting the classification results.

2. The method for monitoring driver's electroencephalogram signals according to claim 1, characterized in that: The brain wave acquisition device includes 18 electrode patches.

3. The method for monitoring driver's electroencephalogram signals according to claim 1, characterized in that: In step a2, for the continuous EEG signal, the epileptic and non-epileptic time periods are processed in segments respectively; The variational mode decomposition method is used to decompose the EEG signal into a specified number of n IMF components. The nodes of the brain network are formed through the EEG channels. The brain channel signals are phase-coupled through the phase-locked value. The coupling relationship between the channels is used to establish the edges of the brain network, and n different brain networks are constructed.

4. The method for monitoring driver's electroencephalogram signals according to claim 1, characterized in that: In step a2, n is set to 6; The EEG signals are decomposed into 6 IMF components and 6 different brain networks are constructed.

5. The method for monitoring driver's electroencephalogram signals according to claim 1, characterized in that: In the sparsity, the density range is set to 15% to 45%, and each 5% is set as a step length.

6. The method for monitoring driver's electroencephalogram signals according to claim 1, characterized in that: In step S4, an epilepsy detection classification model is constructed based on the support vector machine algorithm model.

7. The method for monitoring driver's electroencephalogram signals according to claim 1, characterized in that: In step S5, the training data set is constructed after extracting the non-denoised EEG information.

Images