A Key Departure Scene Recognition Method Based on EEG Data and Fuzzy Neural Networks

Abstract

The application provides a key disengagement scene recognition method based on electroencephalogram data and a fuzzy neural network, and comprises the following steps: S1, a driving scene is designed, and electroencephalogram data of a driver is collected; S2, the electroencephalogram data is preprocessed to clean the data, and then a time-frequency analysis method is used to extract electroencephalogram features; S3, a fuzzy convolutional neural network model is constructed to process the electroencephalogram features; and S4, a disengagement scene involved in the experiment is recognized according to the electroencephalogram feature processing result. The method can realize the following: the driving scene is recognized by means of electroencephalogram signals, the understanding and cognition of the driver to the scene are integrated into automatic driving, the key disengagement scene faced is pre-judged, the defects of the automatic driving strategy are improved, and the method has an important reference role for the design and application of a vehicle control system in automatic driving.

Description

Technical Field

[0001] This invention relates to the fields of cognitive neuroscience and pattern recognition, specifically to a method for identifying key out-of-scene conditions based on electroencephalogram (EEG) data and fuzzy neural networks. Background Technology

[0002] The research and development of autonomous driving technology has entered a stage of rapid development, and road testing is a crucial link in the iterative updates and commercial applications of autonomous driving technology. However, in areas with complex road testing scenarios, especially those with ambiguous right-of-way, critical disengagement phenomena can easily occur, namely, events where vehicle control is switched to a human driver. Driving involves complex operations such as perception, judgment, and decision-making, requiring not only physical coordination but also the brain's cognition, judgment, reaction, and coordination to govern and direct the driver's driving operations.

[0003] Electroencephalography (EEG) is a spontaneous and random bioelectrical signal that rapidly reflects a driver's psychological activities and cognitive behavior with high temporal resolution. It is an important physiological indicator of a driver's mental and psychological state. Therefore, extracting EEG features and constructing neural networks for processing can enable the identification of critical disengagement scenarios, offering valuable insights for the design and application of vehicle control systems in autonomous driving.

[0004] Existing methods for recognizing driving scenarios typically employ traditional neural networks to build models. These networks, primarily trained and reasoned based on mathematical models, often fail to simulate human judgment, decision-making abilities, and fuzzy logic. This makes it difficult to explain their internal decision-making processes, resulting in poor interpretability in driving scenario applications. Summary of the Invention

[0005] The present invention aims to provide a key decoupling scene recognition method based on EEG data and fuzzy neural networks to solve the above problems.

[0006] The technical solution of this invention is: a key scene-out recognition method based on electroencephalogram (EEG) data and fuzzy neural networks, comprising:

[0007] Step S1: Design a driving scenario and collect the driver's EEG data;

[0008] Step S2: The EEG data is first preprocessed to clean the data, and then the EEG features are extracted using time-frequency analysis.

[0009] Step S3: Construct a fuzzy convolutional neural network model to process EEG features;

[0010] Step S4: Based on the EEG feature processing results, identify the detachment scenarios involved in the experiment.

[0011] Preferably, in step S1,

[0012] Driving scenarios are classic scenarios for critical disengagement in autonomous driving and the main causes of critical disengagement, including: lane changing to avoid a disabled vehicle, road construction, intersection maneuvering, and other vehicles entering the road.

[0013] The process of collecting the driver's EEG data includes: collecting 3 seconds of EEG data when the driver makes lane-changing and steering decisions, and processing and analyzing this 3-second EEG data to obtain the driver's EEG data.

[0014] Preferably, step S2 specifically includes:

[0015] Step S21, preprocessing the EEG data; specifically including:

[0016] Step S211: The collected EEG data is filtered in a frequency band of 0.5Hz-70Hz.

[0017] Step S212: Remove electrodes that show large-scale drift and abnormal fluctuations in EEG data, i.e., invalid electrodes;

[0018] Step S213: Segment the EEG data according to the requirements of subsequent analysis;

[0019] Step S214: Use independent component analysis to remove artifacts and obtain independent components that are consistent with the number of electrodes. Then, use a plug-in to automatically identify artifacts and manually remove them.

[0020] Step S22 involves performing time-frequency analysis on the preprocessed EEG data; specifically:

[0021] Step S221: Select wavelet analysis as the method for time-frequency analysis;

[0022] Step S222: Select the Daubechies wavelet basis function for analyzing EEG data;

[0023] Step S223: Perform wavelet analysis on the preprocessed EEG data using the Daubechies wavelet basis function.

[0024] Change

[0025] Wavelet transform decomposes a signal into sub-signals with different frequency ranges, and obtains information on frequency and time scales by calculating the wavelet coefficients of each sub-signal.

[0026] Step S224: Determine the number of layers for wavelet decomposition;

[0027] Step S225: Select wavelet coefficients or wavelet energy as EEG features according to the analysis purpose;

[0028] Frequency localization analysis was performed using wavelet transform to extract EEG features.

[0029] Preferably, the fuzzy convolutional neural network in step S3 comprises five layers: an input layer, a fuzzification layer, a rule layer, an inference layer, and a defuzzification layer; wherein,

[0030] 1) The input layer is used to convert the EEG features extracted by wavelet decomposition into a data format that the neural network can understand as input data; that is, to convert it into a vector or matrix as input to the input layer.

[0031] 2) The fuzzing layer fuzzifies the data from the input layer (converting EEG features into data that the neural network can understand), transforming the input into a set of membership functions; specifically, each feature is converted into a fuzzy set, and each fuzzy set is described by a set of membership functions.

[0032] Suppose the input vector is X = (x1, x2, ..., x...). n ) T , where x i It is the value of the i-th input variable, using the fuzzy set F. j (x) represents the fuzzification result of the j-th input variable, where j = 1, 2, ..., n;

[0033] In fuzzy set theory, fuzzification operations are typically defined using membership functions; for each input variable x i and each membership function μ j (x), fuzzification result F j (x i It can be calculated using the following formula:

[0034] F j (x i )=μ j (x i )

[0035] 3) Rule layer: Defines the fuzzy rules of the fuzzy inference system, which describe the relationship between input features and output;

[0036] In the rule layer, each rule typically contains a condition part and a conclusion part. The condition part uses the membership degree of the input features to match the rule, and the conclusion part gives the corresponding output.

[0037] The rules include the following: Rule 1: If the energy of a certain frequency band in the feature vector exceeds a threshold, it is judged as an active brainwave state; Rule 2: If the energy of all frequency bands in the feature vector is below a threshold, it is judged as a resting brainwave state; Rule 3: If the energy of multiple frequency bands in the feature vector is close to the threshold but does not exceed it, it is judged as a moderately active brainwave state.

[0038] The main function of the rule layer is to perform logical judgments or transformations on the input variables according to predefined rules or conditions in order to obtain the required output, which is a label representing the EEG state, including active state, resting state or moderately active state.

[0039] The specific rule operation process is as follows: Assume there are m rules, each rule consists of a condition and a conclusion; for each rule i, the condition can be expressed as:

[0040] IF x1 is A i1 and x2 is A i2 and ... and x n is A in THEN y is B i

[0041] Among them, A ij The input variable x i The membership value of B under the j-th membership function i This is the corresponding output;

[0042] Based on the intersection operation of fuzzy sets, we can calculate the activation degree of each rule, representing the degree to which the rule matches the input; the activation degree of rule i can be calculated as:

[0043] w i =min{u i1 (x1),μ i2 (x2),...,μ in (x n )}

[0044] Where, μ ij (x i ) is the input variable x i The j-th membership function μ j The membership value under (x) indicates that there are n input variables;

[0045] 4) Inference layer: Using the activation degree and conclusion of the rule, the conclusion is the label of the output EEG state, and the value of each output variable is calculated; for each output variable y, the fuzzy output result G(y) can be calculated as:

[0046]

[0047] Among them, w i B is the activation degree of rule i. i This is the corresponding output, where m represents the total number of rules;

[0048] 5) Defuzzification layer: This layer converts the fuzzy values ​​of the output variables into precise values, i.e., by mapping the fuzzy values ​​of the output variables to a precise value range; the specific calculation process is as follows:

[0049] Suppose there is an output variable y, which is the output of the inference layer. Its fuzzification result is G(y), and the defuzzification result can be expressed as:

[0050]

[0051] Among them, y i μ is the i-th element in the blurred result of y. i It is y i The membership value is k, where k is the size of the fuzzy set.

[0052] Preferably, in step S4,

[0053] Step S41: Evaluate the performance and effectiveness of the trained model using accuracy, recall, and F1 score as evaluation metrics.

[0054] Step S42: The trained model classifies the output data to identify four departure scenarios: lane change to avoid a disabled vehicle, road construction, intersection game, and social vehicles cutting in.

[0055] Step S43: Based on the output of the classifier, it can be determined whether the driver is out of the scene.

[0056] The beneficial effects of this invention are as follows:

[0057] 1. This invention utilizes fuzzy logic to simulate human judgment and decision-making abilities and the fuzzy logic thinking of the human brain, and leverages the advantages of neural networks such as learning ability, parallel processing ability, and fault tolerance to provide a new method for exploring brain states in critical detached scenarios.

[0058] 2. This invention uses EEG signals to identify driving scenarios, integrates the driver's understanding and cognition of the scenario into autonomous driving, predicts key departure scenarios, improves the defects of autonomous driving strategies, and thus has important reference value for the design and application of vehicle control systems in autonomous driving. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the key decoupling scene recognition method based on EEG data and fuzzy neural network provided in an embodiment of the present invention. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. The embodiments of the present invention are not limited thereto.

[0061] Example 1

[0062] like Figure 1 As shown, the key decoupling scene recognition method based on EEG data and fuzzy neural networks includes:

[0063] Step S1: Design a driving scenario and collect the driver's EEG data;

[0064] Step S2: The EEG data is first preprocessed to clean the data, and then the EEG features are extracted using time-frequency analysis.

[0065] Step S3: Construct a fuzzy convolutional neural network model to process EEG features;

[0066] Step S4: Based on the EEG feature processing results, identify the detachment scenarios involved in the experiment.

[0067] Specifically, it includes the following:

[0068] Experiment setup

[0069] 1.1: Experimental Equipment

[0070] This invention's driving simulation experiment was conducted using a highly realistic Ford driving simulator developed by the American company Realtime Technologies. The simulator's hardware primarily includes: a driving simulation vehicle modeled after a standard Ford passenger car, a projector and projection screen, a linear motion control platform, sound and vibration simulation facilities, various sensors, and in-vehicle video monitoring equipment. In addition, it is equipped with two core software programs for designing simulated driving scenarios and operational methods: SimVista software, used for driving simulation scenario and situation design, which allows users to create driving simulation scenarios based on different experimental needs using its built-in modules; and SimCreator software, used for graphical simulation and modeling, which designs the connection methods and paths between components to realize the operation of the virtual scene.

[0071] The EEG acquisition device used in this invention is a 64-channel EEG acquisition system developed by Neuroscan, Inc., USA. The hardware of this system mainly includes the SynAmps2 EEG signal amplifier and the Quik-Cap electrode cap. The SynAmps2 amplifier's headbox consists of 64 monopolar leads (scalp EEG acquisition), 4 bipolar leads (ECG and EMG), and 2 high-level inputs (reference electrodes), for a total of 70 lead channels, simultaneously acquiring signals at a sampling rate of up to 1000Hz for each lead channel. The Quik-Cap electrode cap directly contacts the subject's scalp, using conductive gel as the conductive material to acquire scalp EEG signals. The electrode cap contains 66 electrodes, including two mastoid electrodes and vertical and horizontal electrooculography (EOG) electrodes. The distribution of the channels follows the 10-20 international EEG recording system established by the International Society for Clinical Neurophysiology.

[0072] 1.2: Experimental Design

[0073] This invention designs classic scenarios for critical disengagement in autonomous driving, including lane changing to avoid a disabled vehicle, road construction, intersection maneuvering, and other vehicles cutting in. Lane changing to avoid a disabled vehicle: The auxiliary vehicle stops in the center of the right lane, with a warning triangle placed behind it; Road construction: Cones and construction signs are placed at designated locations to simulate road construction; Intersection maneuvering (intersection straight-ahead encounter with continuous left-turning vehicles from the opposite lane): As the main vehicle approaches and passes through the intersection, continuous auxiliary vehicles making left turns are placed in the opposite lane; Other vehicles cutting in: The main vehicle is traveling on a straight road, and an auxiliary vehicle traveling along the center line of the adjacent lane is placed in front of it. The auxiliary vehicle maintains its speed and changes lanes into the main lane. After completing the lane change, the auxiliary vehicle continues to travel along the center line of the main lane.

[0074] This experiment recruited 50 drivers, all of whom were college students aged 22-27, possessed driver's licenses, and had some driving experience. None of them had any brain disorders. The drivers received pre-experiment training to familiarize themselves with the experimental procedures.

[0075] Driving data analysis

[0076] For the above experimental scenarios, drivers need to make decisions based on their perception and judgment. Specifically, avoiding a disabled vehicle while changing lanes, road construction, and other vehicles cutting in require drivers to make lane-changing decisions, while intersection maneuvers require drivers to make turning decisions after hearing voice commands before the intersection. This embodiment selects 3 seconds of data for analysis and research; it collects 3 seconds of EEG signals from the driver making lane-changing and turning decisions, and processes and analyzes these 3 seconds of EEG signals. For lane-changing decisions, the driver's turn signal (up or down) is used to determine the timing of the lane change; for turning decisions, the driver's turn signal (up or down) is used to determine the timing of the turn.

[0077] EEG data analysis

[0078] 1.1: Preprocessing

[0079] 1) Preview of EEG waveform

[0080] A preliminary examination of the EEG waveforms is conducted, and each group of EEG data records is manually checked to ensure completeness and the absence of large-scale abnormalities in the waveforms. Ultimately, the validity of the EEG signal recording is determined.

[0081] 2) Reduce sampling rate

[0082] The main brainwave frequency range studied in this experiment is 0.5Hz-70Hz. In order to improve the computer's computing efficiency, the sampling rate was reduced to 256Hz or 512Hz.

[0083] 3) Filtering

[0084] The frequency range studied in this experiment is 0.5Hz-70Hz. Therefore, a finite-length unit impulse response filter is used for filtering. The cutoff frequency of the low-pass filter is 0.5Hz, and the cutoff frequency of the high-pass filter is 70Hz.

[0085] 4) Remove invalid electrodes

[0086] In this experiment, the electrodes on the electrode caps are connected to the subject's scalp via conductive gel. If too much conductive gel is injected, or if there is poor contact between the scalp and the electrodes, the collected EEG signals from the corresponding electrodes will exhibit large-scale drift and abnormal fluctuations, rendering them unusable. Such electrodes are considered invalid. The Automatic Channel Rejection function in EEGLAB can automatically eliminate invalid electrodes.

[0087] 5) Segmentation of EEG data

[0088] This experiment mainly uses Independent Component Analysis (ICA) for artifact removal. This method requires segmenting the EEG data before running ICA. In order to meet the needs of subsequent EEG data analysis, the raw EEG data is segmented at 1-second intervals.

[0089] 6) Eliminate forgeries

[0090] The artifact removal in this experiment mainly involves two operations: first, running ICA to obtain independent components (ICs) consistent with the number of electrodes; and then using the EEGLAB ADJUST plugin to automatically identify artifacts and manually remove them.

[0091] 7) Baseline calibration

[0092] Baseline calibration is primarily used to correct data skew and can improve baseline discrepancies caused by low-frequency drift or artifacts. Baseline calibration can be performed automatically using the Remove Baseline function in EEGLAB.

[0093] 1.2: Time-Frequency Analysis

[0094] Wavelet transform is an important tool for time-frequency research. It enables frequency localization analysis, overcoming the limitations of Fourier transform and making it suitable for processing non-stationary signals. It features multi-resolution analysis and the ability to characterize local signal features in both the time and frequency domains. It's a time-frequency localization analysis method with a fixed window size but a changeable shape, allowing for adjustments to both the time and frequency windows, making it well-suited for processing EEG signals. Wavelet transform is a time-frequency domain localization analysis method that refines signals at multiple scales through scaling and translation operations, allowing for frequency subdivision at low frequencies and time subdivision at high frequencies. It can effectively decompose various time-varying signals and effectively separate signal from noise. Commonly used wavelet basis functions include wavelet functions, wavelet function systems, and complex wavelets. The wavelet basis is chosen based on which set of wavelet bases yields the largest number of wavelet coefficients and the smallest variance. The wavelet basis in the Daubechies wavelet function system is denoted as dbN, where N is a sequence N = 1, 2, ..., 10. The basic principle of wavelet transform is to approximate a signal using a system of wavelet functions. The signal is obtained by shifting or scaling a basic wavelet function.

[0095] Fuzzy Neural Network Model

[0096] Fuzzy neural networks combine fuzzy systems and neural networks. Essentially, they process the neural network's input through a fuzzy system, transforming it into fuzzy input signals and fuzzy weights, and then defuzzify the neural network's output, resulting in intuitive, effective numerical values. Specifically, in a fuzzy neural network, the neural network's input and output represent the fuzzy system's input and output. The membership functions and fuzzy rules of the fuzzy system are incorporated into the neural network's hidden nodes, fully leveraging the parallel processing capabilities of the neural network and the reasoning abilities of the fuzzy system.

[0097] A standard fuzzy neural network has five layers: an input layer, a fuzzification layer, a rule layer, an inference layer, and a defuzzification layer.

[0098] 1) The input layer is used to convert the EEG features extracted by wavelet decomposition into a data format that the neural network can understand as input data; that is, to convert it into a vector or matrix as input to the input layer.

[0099] 2) The fuzzing layer fuzzifies the data from the input layer (converting EEG features into data that the neural network can understand), transforming the input into a set of membership functions; specifically, each feature is converted into a fuzzy set, and each fuzzy set is described by a set of membership functions.

[0100] Suppose the input vector is X = (x1, x2, ..., x...). n ) T , where x i It is the value of the i-th input variable, using the fuzzy set F. j (x) represents the fuzzification result of the j-th input variable, where j = 1, 2, ..., n;

[0101] In fuzzy set theory, fuzzification operations are typically defined using membership functions; for each input variable x i and each membership function μ j (x), fuzzification result F j (x i It can be calculated using the following formula:

[0102] F j (x i )=μ j (x i )

[0103] 3) Rule layer: Defines the fuzzy rules of the fuzzy inference system, which describe the relationship between input features and output;

[0104] In the rule layer, each rule typically contains a condition part and a conclusion part. The condition part uses the membership degree of the input features to match the rule, and the conclusion part gives the corresponding output.

[0105] The rules include the following: Rule 1: If the energy of a certain frequency band in the feature vector exceeds a threshold, it is judged as an active brainwave state; Rule 2: If the energy of all frequency bands in the feature vector is below a threshold, it is judged as a resting brainwave state; Rule 3: If the energy of multiple frequency bands in the feature vector is close to the threshold but does not exceed it, it is judged as a moderately active brainwave state.

[0106] The main function of the rule layer is to perform logical judgments or transformations on the input variables according to predefined rules or conditions in order to obtain the required output, which is a label representing the EEG state, including active state, resting state or moderately active state.

[0107] The premises in the rule layer are usually composed of the membership degrees of the input features. The membership degree of each input feature represents the degree of membership of that feature under a certain membership function. The activation degree represents the activation degree or confidence of the rule, reflecting the degree of influence of the rule on the output. The activation degree is usually represented by a value between 0 and 1. The larger the value, the stronger the rule activation.

[0108] The specific rule operation process is as follows: Assume there are m rules, each rule consists of a condition and a conclusion; for each rule i, the condition can be expressed as:

[0109] IF x1 is A i1 and x2 is A i2 and ... and x n is A in THEN y is B i

[0110] Among them, A ij The input variable x i The membership value of B under the j-th membership function i This is the corresponding output;

[0111] Based on the intersection operation of fuzzy sets, we can calculate the activation degree of each rule, representing the degree to which the rule matches the input; the activation degree of rule i can be calculated as:

[0112] w i =min{u i1 (x1),μ i2 (x2),...,μ in (x n )}

[0113] Where, μ ij (x i ) is the input variable x i The j-th membership function μ j The membership value under (x) indicates that there are n input variables;

[0114] 4) Inference layer: Using the activation degree and conclusion of the rule, the conclusion is the label of the output EEG state, and the value of each output variable is calculated; for each output variable y, the fuzzy output result G(y) can be calculated as:

[0115]

[0116] Among them, w i B is the activation degree of rule i. i This is the corresponding output, where m represents the total number of rules;

[0117] 5) Defuzzification layer: This layer converts the fuzzy values ​​of the output variables into precise values, i.e., by mapping the fuzzy values ​​of the output variables to a precise value range; the specific calculation process is as follows:

[0118] Suppose there is an output variable y, which is the output of the inference layer. Its fuzzification result is G(y), and the defuzzification result can be expressed as:

[0119]

[0120] Among them, y i μ is the i-th element in the blurred result of y. i It is y i The membership value is k, where k is the size of the fuzzy set.

[0121] Identify the detached scenarios involved in the experiment

[0122] 1) The performance and effectiveness of the trained model are evaluated using accuracy, recall, and F1 score as evaluation metrics.

[0123] 2) The trained model is used to classify the output data to identify four departure scenarios: lane change to avoid a disabled vehicle, road construction, intersection game, and social vehicles cutting in.

[0124] 3) Based on the output of the classifier, it can be determined whether the driver is out of the scene.

[0125] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the processes depicted in the drawings are not necessarily essential for implementing the present invention.

Claims

1. A method for key disengagement scene recognition based on electroencephalogram data and fuzzy neural network, characterized in that, The method comprises the following steps: Step S1, designing a driving scene and collecting brain electrical data of a driver; Step S2, pre-processing the brain electrical data to clean the data, and then extracting brain electrical characteristics by using a time-frequency analysis method; Step S3, constructing a fuzzy convolutional neural network model to process the brain electrical characteristics; Step S4, identifying the disengagement scenarios involved in the experiment according to the brain electrical characteristic processing results; In step S3, the fuzzy convolutional neural network comprises five layers: an input layer, a fuzzification layer, a rule layer, an inference layer, and a defuzzification layer; wherein 1) the input layer is used to convert the brain electrical characteristics extracted by wavelet decomposition into a data format that can be understood by the neural network as input data; that is, the brain electrical characteristics are converted into a vector or a matrix as the input of the input layer; 2) the fuzzification layer performs fuzzification processing on the data of the input layer (the brain electrical characteristics converted into data that can be understood by the neural network), and converts the input into a set of membership functions; specifically, each characteristic is converted into a fuzzy set, and each fuzzy set is described by a set of membership functions; Let X = (x1, x2,..., xn) be the input vector. n ) T where xi is the value of the ith input variable, and Fj(x) denotes the fuzzification result of the jth input variable, where j = 1, 2,..., n. i j ​​ In fuzzy set theory, the fuzzification operation uses membership functions to define it; for each input variable x i and each membership function μ j (x), the fuzzified result F j (x i ) can be computed by the following formula: F j (x i )=μ j (x i ) 3) the rule layer defines the fuzzy rules of the fuzzy inference system, which describes the relationship between the input characteristics and the output; 4) the inference layer calculates the value of each output variable using the activation degree of the rule and the conclusion of the rule, and the conclusion is the label of the output brain electrical state; for each output variable y, the fuzzy output result G(y) can be calculated as: where w i is the activation of rule i, B i is the corresponding output, and m indicates that there are m rules in common; 5) the defuzzification layer converts the fuzzy value of the output variable into an accurate value, that is, by mapping the fuzzy value of the output variable to an accurate value domain; the specific calculation process is: Assuming that there is an output variable y, which is the output result of the inference layer, and the fuzzy result of y is G(y), the defuzzification result can be represented as: wherein y i is the i-th element in the fuzzification result of y, μ i is the membership value of y i , and k is the size of the fuzzy set.

2. The method of claim 1, wherein the method is characterized by, In step S1, The driving scene is a classic scene related to key disengagement of autonomous driving and the main reason for key disengagement, including: changing lanes to avoid a broken-down vehicle, road construction, intersection game, and social vehicle cutting in; Collecting brain electrical data of a driver includes collecting 3-second brain electrical data of the driver when making lane changing and turning decisions, and processing and analyzing the 3-second brain electrical data to obtain the brain electrical data of the driver. 3.The key disengagement scenario recognition method based on electroencephalogram data and fuzzy neural network according to claim 1, characterized in that, Step S2 specifically comprises: Step S21, pre-processing the brain electrical data; specifically comprising: Step S211, filtering the collected brain electrical data by a frequency range of 0.5Hz-70Hz; Step S212, removing electrodes with large-scale drift and abnormal fluctuations in brain electrical data, i.e. invalid electrodes; Step S213, segmenting the brain electrical data according to subsequent analysis requirements; Step S214, using independent component analysis to remove artifacts, obtaining independent components consistent with the number of electrodes, and then automatically identifying artifacts using a plug-in and manually removing them; Step S22, performing time-frequency analysis on the pre-processed brain electrical data; specifically: Step S221, selecting wavelet analysis as the time-frequency analysis method; Step S222, selecting a Daubechies wavelet basis function for analyzing brain electrical data; Step S223, wavelet transform is performed on the preprocessed electroencephalogram data using a Daubechies wavelet basis function; the wavelet transform decomposes the signal into sub-signals of different frequency ranges, and information on the frequency and time scale is obtained by calculating the wavelet coefficients of each sub-signal; Step S224, the number of wavelet decomposition layers is determined; Step S225, wavelet coefficients or wavelet energy are selected as electroencephalogram features according to the analysis purpose; The frequency localization analysis is performed by wavelet transform, and the electroencephalogram features are extracted. 4.The key disengagement scenario recognition method based on electroencephalogram data and fuzzy neural network according to claim 1, characterized in that, The fuzzy convolutional neural network in step S3 includes five layers: an input layer, a fuzzification layer, a rule layer, an inference layer, and a defuzzification layer; wherein, 3) The rule layer defines the fuzzy rules of the fuzzy inference system, and describes the relationship between the input features and the output; Each rule in the rule layer includes a condition part and a conclusion part, the condition part uses the membership degree of the input features to match the rule, and the conclusion part gives the corresponding output; The rules include the following: rule 1: if the energy of a certain frequency band in the feature vector exceeds a threshold value, the electroencephalogram is judged to be in an active state; rule 2: if the energy of all frequency bands in the feature vector is lower than a threshold value, the electroencephalogram is judged to be in a resting state; rule 3: if the energy of multiple frequency bands in the feature vector is close to the threshold value but does not exceed it, the electroencephalogram is judged to be in a moderately active state; The main function of the rule layer is to make logical judgments or conversions on the input variables according to the predefined rules or conditions, so as to obtain the required output, which is a label representing the electroencephalogram state, including active state, resting state or moderately active state; The specific rule operation process is as follows: assuming that there are m rules, each rule consists of a condition and a conclusion; for each rule i, the condition can be expressed as: IF x1 is A i1 and x2 is A i2 and...and x n is A in THEN y is B i where A ij is the input variable x i the membership value under the jth membership function, B i is the corresponding output; According to the intersection operation of fuzzy sets, the activation degree of each rule can be calculated, which represents the matching degree of the rule to the input; the activation degree of rule i can be calculated as: w i = min{u i1 (x1), μ i2 (x2),..., μ in (x n )} where μ ij (x i ) is the input variable x i The membership value under the jth membership function μ j (x) is denoted by n, which represents the total number of input variables.

5. The electroencephalogram data and fuzzy neural network based key disengagement scenario identification method of claim 1, wherein, In step S4, Step S41, the accuracy, recall rate and F1 value are used as evaluation indexes to evaluate the performance and effect of the trained model; Step S42, the trained model is used to classify the output data to identify four types of escape scenes: lane changing of avoidance failure vehicle, road construction, intersection game and social vehicle cutting in; Step S43, according to the output result of the classifier, it can be judged whether the driver is in the state of escape scene.

Images