A method for decoding and continuous assessment of motion sickness electroencephalogram features for automatic driving

Abstract

The application is suitable for the field of intelligent networked vehicles and physiological signal processing technology, and provides a method for decoding and continuous evaluation of motion sickness electroencephalogram features for automatic driving, which comprises: acquiring multi-channel electroencephalogram data and performing event-locked segmentation; using a parameterized Sinc filter bank, combining physiological priors to mine specific band features; capturing dynamic timing features under different time windows through a multi-time scale time domain encoder, and strengthening key brain region spatial interaction features through a graph-enhanced brain region attention module; finally, performing spatio-temporal-frequency feature fusion and multi-task decoding, and synchronously outputting discrete motion sickness grades and continuous motion sickness severity scores. Through the construction of an end-to-end spatio-temporal-frequency joint modeling and multi-task learning framework, the application realizes continuous evaluation of motion sickness with high precision, high robustness and high interpretability, and the output can be directly used as a comfort cost for automatic driving closed-loop control.

Description

Technical Field

[0001] This invention belongs to the field of intelligent connected vehicle and physiological signal processing technology, and in particular relates to a method for decoding and continuously evaluating EEG features of motion sickness for autonomous driving. Background Technology

[0002] With the rapid development of intelligent connected vehicles and autonomous driving technologies, the role of passengers is shifting from traditional drivers to supervisors or passive passengers. In such scenarios, passengers are more prone to motion sickness due to activities such as reading and using electronic devices. The dizziness, nausea, and other discomfort caused by motion sickness severely impact riding comfort and reduce passenger acceptance of autonomous driving systems. Therefore, how to objectively, in real-time, and accurately identify passenger motion sickness in vehicle scenarios has become a key technical issue in intelligent cockpit comfort control and human-vehicle collaborative systems.

[0003] Existing motion sickness assessment technologies can be mainly categorized as follows: The first category is subjective assessment. This method obtains the passenger's subjective degree of discomfort through motion sickness scales, verbal reports, etc. While simple to implement, this method heavily relies on the passenger's active expression, making continuous monitoring impossible; it is highly subjective and lacks objectivity; and feedback is delayed, making it difficult to meet the needs of online status perception and closed-loop control. The second category is physiological signal monitoring methods, such as heart rate, skin conductance, and electroencephalography (EEG). Compared to subjective assessment, these methods are more objective. However, peripheral physiological signals are easily interfered with by various factors, resulting in an indirect reflection of motion sickness. Although EEG signals can more directly reflect brain neural activity, existing EEG-based motion sickness identification methods still have significant shortcomings.

[0004] First, in sample construction, many methods use fixed time windows, failing to accurately synchronize with the actual changes in motion sickness levels, leading to biased sample labels. Second, in frequency domain feature extraction, many methods use preset fixed frequency bands, lacking adaptive learning capabilities for motion sickness tasks and failing to capture the dynamic changes of key frequency components. Third, in temporal feature modeling, a single-length time window is typically used, making it difficult to simultaneously consider both short-term fluctuations and long-term cumulative information of motion sickness. Furthermore, in spatial feature modeling, most methods simply process multi-channel EEG signals in parallel, failing to fully utilize the topological connections between electrodes and the differential roles of different brain regions in motion sickness formation. Finally, in the output task, existing methods are mostly binary classification-based, which is insufficient to meet the practical needs of graded perception and differentiated intervention for motion sickness severity.

[0005] In summary, existing motion sickness recognition technologies still have shortcomings in terms of methodological structure, implementation process, and output functions. There is a lack of a technical solution that is geared towards vehicle scenarios, can jointly model the spatial, temporal, and frequency characteristics of motion sickness EEG signals, and can achieve stable classification and continuous assessment of motion sickness levels. Summary of the Invention

[0006] The purpose of this invention is to provide a method for decoding and continuously evaluating EEG features of motion sickness in autonomous driving, aiming to solve the problems mentioned in the background art.

[0007] The present invention is implemented as follows: a method for decoding and continuously evaluating EEG features of motion sickness for autonomous driving includes the following steps:

[0008] Step 1: Data input and preprocessing;

[0009] Acquire vehicle-mounted multichannel EEG data that has been denoised and segmented by event locking, extract the signal tensor of the current evaluation window, and use it as the input feature tensor;

[0010] Step 2: Frequency domain adaptive filtering extraction;

[0011] A parameterized Sinc filter bank is constructed, in which the lower and upper cutoff frequencies of each frequency band are used as learnable parameters for end-to-end optimization, and the amplitude is constrained within the legal physiological range by a limiting operation; a temporal impulse response is constructed based on a Hamming window and a low-pass Sinc kernel; the input feature tensor is passed through the filter bank to obtain a multi-band tensor, and then modulated by a channel attention mechanism to generate the final frequency domain embedding;

[0012] Step 3: Feature refinement and spatiotemporal structure generation;

[0013] Brain topology structural features and multi-scale temporal features are extracted using a parallel network architecture, including:

[0014] Graph-enhanced brain region attention: The physiological prior baseline weights, sample dynamic activation weights, and topological graph connection weights extracted by the graph attention network are multiplicatively modulated to generate a comprehensive spatial attention graph. The input channels are reweighted using the calibrated weights to obtain the EEG feature tensor with spatial feature enhancement. Then, multi-head self-attention is used to capture the dynamic sequence evolution across time steps.

[0015] Multi-scale temporal coding: Construct parallel convolutional branches at multiple time scales, calculate multi-scale attention weight matrices through sample-level pooling and nonlinear mapping, recalibrate features at each scale, concatenate the weighted features and fuse them through convolution, and output the final spatiotemporal fusion representation.

[0016] Step 4: Multi-task decoding and joint supervised loss function;

[0017] The spatiotemporal embedding vector output from the time-domain branch is concatenated with the spectral embedding vector output from the frequency-domain branch and subjected to a nonlinear transformation to form a time-frequency fusion feature representation. The time-frequency fusion feature representation is then decoded in parallel using a multi-task head, including: continuous severity regression decoding, which outputs a physically meaningful continuous motion sickness severity score; and motion sickness level classification decoding, which outputs the probability distribution of discrete levels. The model is then trained using a joint supervised loss function.

[0018] The present invention provides a method for decoding and continuously evaluating electroencephalogram (EEG) features of motion sickness for autonomous driving, which has the following beneficial effects:

[0019] (1) High precision and strong robustness: This invention unifies graph network spatial topology modeling and multi-timescale time coding into an end-to-end framework, achieving high precision motion sickness severity assessment under real road conditions. The accuracy of four-class classification can reach more than 90%, and the Kappa coefficient is significantly better than existing methods, proving its robustness in handling individual differences and complex noise environments.

[0020] (2) High interpretability: The learnable filter bank based on the parameterized Sinc function ensures the physiological basis of frequency domain extraction and can adaptively discover task-related frequency bands (such as adaptively enhancing the key Delta and Theta low-frequency feature responses), which greatly improves the spectral interpretability of the model. At the same time, the graph-enhanced brain region attention mechanism explicitly characterizes the functional connectivity topology between key brain networks such as the vestibular-parietal lobe and visual-occipital lobe, providing a neurophysiological basis for model decision-making.

[0021] (3) High temporal resolution and closed-loop application value: This invention realizes unified multi-task spatiotemporal-frequency joint decoding, and outputs discrete classification and accurate continuous severity objective score, which significantly enhances the temporal resolution of the evaluation, so that the output results can be directly introduced into the autonomous driving trajectory planning and control closed loop as comfort cost, providing a new paradigm for active motion control of "anti-motion sickness". Attached Figure Description

[0022] Figure 1 A flowchart of a method for decoding and continuously evaluating EEG features of motion sickness for autonomous driving provided by an embodiment of the present invention;

[0023] Figure 2 This is a framework diagram of a method for decoding and continuously evaluating EEG features of motion sickness for autonomous driving, provided by an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0026] like Figure 1 and Figure 2 ( Figure 2 In the figure, d represents the feature representation dimension, which is used to characterize the number of hidden feature channels within the model at each time step. This dimension can be preset according to the task complexity, computing resources, and model expressive power requirements, for example, it can be set to 32, 64, 128, etc. (This embodiment does not limit this). As shown, this is an embodiment of the present invention providing a method for decoding and continuously evaluating motion sickness EEG features for autonomous driving (hereinafter referred to as the MS-GTFNet method), which includes the following steps:

[0027] Step 1: Data input and preprocessing;

[0028] First, the signal tensor of the current evaluation window is extracted from the denoised and event-locked segmented multi-channel EEG data from the vehicle. For each motion sickness button event, the truncated duration is [duration value missing]. Each sampling point, including EEG segments from each electrode channel form the input feature tensor. ,and ,in For batch size, This represents the set of real numbers. The motion sickness button event refers to a labeled event generated during the experiment where the subject, referring to the internationally used MISC (Mild Motion Sickness Subjective Assessment Scale) description of motion sickness severity, subjectively assesses their current motion sickness state in real time and inputs four severity levels—no motion sickness, mild motion sickness, moderate motion sickness, and severe motion sickness—using a self-made four-level button device. Each motion sickness button event is simultaneously recorded with a trigger timestamp and corresponding level label for event locking and supervision label construction of EEG segments.

[0029] Step 2: Frequency domain adaptive filtering extraction;

[0030] To address the key challenge of balancing heterogeneous individual differences with rigid frequency band division, this method innovatively proposes a parameterized Sinc filter bank. For the... Each frequency band, with its lower and upper cutoff frequencies and End-to-end optimization is performed using learnable parameters, and the parameters are constrained within legal physiological ranges through amplitude limiting operations.

[0031] (1)

[0032] (2)

[0033] in, For the amplitude limiting function, To parameterize the number of frequency bands in the Sinc filter bank, i.e., the number of filters, and . To control the first Learnable parameters of cutoff frequency in each frequency band This is a preset lower limit for effective physiological frequency. The upper limit of the preset effective physiological frequency, To prevent the upper and lower cutoff frequencies from coinciding or overflowing, a very small positive constant is used. To control the first Learnable parameters for each bandwidth. As a preferred implementation, 0.5Hz is acceptable. 100Hz is acceptable. 0.1Hz is acceptable.

[0034] Subsequently, a Hamming window-based system was built. and low-pass Sinc core time-domain impulse response :

[0035] (3)

[0036] in, The discrete-time sampling point number. For the first The bandwidth of each frequency band;

[0037] Input feature tensor By incorporating physiological prior weights The filter bank obtains multiband tensors Then, the multi-band tensor is weighted using a channel-band gated modulation mechanism to generate the final frequency domain embedding containing task-specific spectral features. :

[0038] (4)

[0039] (5)

[0040] in, For one-dimensional convolution computation in the time domain, One-dimensional grouping Convolution operations are used to generate channel-band gated weights that match the dimensions of the multi-band feature tensor. This is element-wise multiplication.

[0041] Step 3: Feature refinement and spatiotemporal structure generation;

[0042] This step aims to extract brain topology structural features and multi-scale temporal features closely related to motion sickness using a parallel network architecture:

[0043] 1) Image-enhanced brain region attention: Weighting of physiological prior baselines Sample dynamic activation weights Connection weights of the topological graph extracted by the graph attention network Perform multiplicative modulation to generate a comprehensive spatial attention map. :

[0044] (6)

[0045] Using calibrated weights The input channels are reweighted to obtain the EEG feature tensor enhanced with spatial features. :

[0046] (7)

[0047] (8)

[0048] in, This represents a weight calibration function (such as a normalization or activation function). This indicates an element-wise multiplicative weighted operation.

[0049] Then, multi-head self-attention is used to capture the dynamic sequence evolution across time steps:

[0050] (9)

[0051] in, For the first The output matrix of each attention head, where... This is the attention head index in a multi-head self-attention mechanism, and , The number of attention heads can be set as a preset hyperparameter. In this embodiment, A value of 4 or 8 can be chosen to strike a balance between feature representation capability and computational complexity. In other implementations, The specific values ​​can be adjusted according to the input feature dimension, model size, computing resources and task complexity, and do not constitute a limitation on the scope of protection of this invention. For normalized exponential functions, For the first A query matrix with attention heads For the first The key matrix of each attention head. For the first The value matrix of each attention head, Let be the feature dimension of the attention head, and ⊤ denote the matrix transpose;

[0052] 2) Multi-timescale temporal coding: Construct four timescales (as a preferred implementation, when the EEG sampling rate is 500Hz and the input window is 4s, the equivalent timescales of the four branches cover approximately 0.1s, 1s, 2s and 4s respectively, and their one-dimensional convolution kernel sizes can be set to 51, 501, 1001 and 2001 respectively, with a stride of 1, and corresponding padding sizes of 25, 250, 500 and 1000 respectively).

[0053] The aforementioned kernel size is used to approximate the receptive field at different time scales. Its specific value can be adjusted according to the sampling rate, input window length, and model complexity, and does not constitute a limitation on the scope of protection of this invention. Output characteristics of each branch The multi-scale attention weight matrix is ​​calculated through sample-level pooling and nonlinear mapping. Recalibrate features at each scale:

[0054] (10)

[0055] in, The scale features are after attention recalibration. For the first In the Scale Branch, the th The original output feature map of each sample. For scale branch index, ; For sample index, and , The batch size mentioned above can be adjusted based on available GPU memory, data volume, and training strategy, such as 8, 16, or 32. In practical applications, the total number of valid samples is denoted as... , The duration of EEG data collection, the number of subjects, the number of motion sickness button events, and the sample segmentation method were all determined together. . and The values ​​can be adjusted according to the actual scale of collected data, training strategy, and computing resources, and the specific values ​​do not constitute a limitation on the scope of protection of this invention.

[0056] The weighted features are concatenated along the channel dimension and then fused using a 1x1 convolution to output the final spatiotemporal fusion representation. ,in This represents the uniform sequence length after adaptive average pooling.

[0057] Step 4: Multi-task decoding and joint supervised loss function;

[0058] Spatiotemporal embedding vector output by the time-domain branch Spectral embedding vector of frequency domain branch output Decoding and mapping are then performed. These outputs are not only used for traditional discrete classification, but more importantly, they supplement the continuous quantization parameters required for autonomous driving closed-loop control: the two feature streams are concatenated and subjected to a nonlinear transformation to form a time-frequency fusion feature representation. :

[0059] (11)

[0060] in, The set of learnable parameters for the feature fusion layer. For mapping operations that include batch normalization and non-linear activation functions, This is an element-wise addition;

[0061] Subsequently, using multi-task head to Perform parallel decoding:

[0062] 1) Continuous Severity Regression Decoding: Regression Head Output a physically meaningful continuous motion sickness severity score (0-100 points):

[0063] (12)

[0064] in, The continuous motion sickness severity score is predicted by the regression branch. Use the Sigmoid activation function;

[0065] 2) Motion sickness level classification decoding: classification head Output the probability distribution of four severity levels of motion sickness: no motion sickness, mild motion sickness, moderate motion sickness, and severe motion sickness.

[0066] (13)

[0067] in, The probability distribution of the motion sickness severity level predicted for the classification branch output. It is a normalized exponential function;

[0068] To ensure the accuracy and robustness of the evaluation, a joint supervised loss function was tailored for each decoding task. Total loss function. It includes two categories: 1) Huber regression constraints used to resist subjective scoring noise. ;2) Cross-entropy constraints used for hierarchical boundary classification And ensure the generalization ability of the network weight perspective. Regular terms:

[0069] (14)

[0070] in, , and To balance the hyperparameters of gradient optimization across multiple tasks, they can typically be selected within the range of [0,1]; as a preferred implementation, they can be set to... =1、 =0.5、 =1×10 -4 . To inject a continuous severity pseudo-true label into the Gaussian perturbation, the standard deviation of the Gaussian perturbation can be selected in the range of 0.01-0.05, preferably 0.02; This provides a true label for the discrete severity level of motion sickness. This represents the set of all optimizable weight parameters in the network model. Through this joint constraint, this method greatly enhances the possibility of continuous and smooth scoring of the model output while taking into account the lower bound of the ranking accuracy.

[0071] In a preferred embodiment of the present invention, in step 1, a 32-channel EEG acquisition system is used to acquire scalp potential signals under real vehicle motion conditions (such as longitudinal acceleration and deceleration, and lateral curve rotation). After acquiring the raw data, the system performs bandpass filtering and independent component analysis (ICA) for noise reduction. Then, using the occupant's subjective button feedback as anchor points, it extracts continuous EEG data segments of fixed length (such as a 4-second window with a sampling rate of 500Hz) to form a multi-channel temporal sample tensor, which serves as the input to a deep neural network.

[0072] The architecture and model training of this method include: 1) a 4-second continuous EEG segment with an input size of 32 channels × 2000 sampling points; 2) a frequency domain branch using a parameterized Sinc filter bank to extract an adaptive frequency band within 0.5-100Hz, and initializing physiological weights including classic EEG frequency bands such as Delta and Theta; 3) a graph-enhanced spatial branch using a graph attention network (GAT) to calculate the topological connectivity strength of scalp electrodes, and combining multi-head attention mechanisms for brain region reweighting; 4) a multi-timescale temporal encoder using a parallel one-dimensional convolutional kernel design (covering differentiated receptive fields from 0.1 seconds to 4 seconds) to extract temporal fluctuations. The training scheme employs end-to-end joint optimization for all branches, using the AdamW optimizer, with the total loss function being a weighted sum of Huber regression loss and cross-entropy classification loss. The specific model implementation is deployed on a workstation equipped with an NVIDIA GeForce RTX4060 (8GB) GPU, utilizing a memory optimization strategy to complete multi-epoch iterative training with a preset batch size.

[0073] The evaluation framework includes two motion sickness induction benchmark scenarios based on real vehicle motion: the longitudinal straight line (LS) benchmark and the curve lateral rotation (LR) benchmark, which are used to comprehensively evaluate the model's ability to quantify motion sickness severity under different physical motion stimuli.

[0074] Longitudinal Linear (LS) Benchmark: Subjects experience periodic starting, acceleration, deceleration and braking (up to 30 km / h) in a closed test area, mainly inducing motion sickness caused by abrupt changes in linear acceleration and low-frequency forward and backward pushing and pulling sensations, focusing on evaluating the model's ability to represent typical daily driving scenarios such as stop-and-go traffic and congestion.

[0075] Lateral Rotation (LR) Benchmark: Subjects experience continuous curves and figure-eight trajectories, with superimposed lateral acceleration and yaw rate, so that the occupants are continuously in a state of combined excitation of lateral load and rotational change, focusing on evaluating the model's ability to represent and evaluate high-intensity somatic loads such as complex steering and continuous changes in direction over long periods of time.

[0076] The comprehensive evaluation employs two core metrics: 1) Classification performance metrics, including Balanced Accuracy (BA), Macro Recall (Macro R), Macro Precision (Macro P), Macro F1 Score (Macro F1), and Kappa Coefficient (Ka), used to assess the discrimination accuracy for four severity levels of motion sickness: no motion sickness, mild motion sickness, moderate motion sickness, and severe motion sickness; 2) Regression performance metrics, including Mean Absolute Error (MAE) and Coefficient of Determination (CDE). This is used to quantify the fit of the continuous motion sickness score (0-100 points). The comparison baselines include top-level methods in several EEG decoding and state assessment fields (such as VIMSNET, EEGNet, TSception, etc.):

[0077] Table 1. Classification Results of Longitudinal Straight Line (LS) and Lateral Rotation of Curve (LR) Benchmark Tests

[0078]

[0079] Detailed evaluation results for the longitudinal and lateral benchmarks are shown in Table 1. In the LS and LR benchmark tests, our method achieved the current best four-class balanced accuracy (90.14% and 89.64%, respectively), surpassing the previous best-performing VIMSNET model by 2.36% in the LS scenario. Notably, our method demonstrates a significant advantage over traditional shallow and deep convolutional networks (such as Shallow ConvNet and DeepConvNet) in the consistency metric Kappa coefficient, proving its robust discrimination ability in handling individual differences and complex real-world vehicle noise environments.

[0080] To quantify the quality of continuous severity assessment and verify the effectiveness of the architecture design, an ablation assessment criterion was established for continuous score prediction.

[0081] Table 2 Ablation test results of core architecture components on the motion sickness continuous scoring task.

[0082]

[0083] The quantitative analysis in Table 2 reveals a fundamental improvement in the quality of continuous-state quantization. Compared to variants lacking specific modules, the complete architecture of this invention not only improves the overall classification accuracy to 90.54%, but also reduces the mean absolute error (MAE) of regression prediction to 4.49 (0-100 range), while achieving a high coefficient of determination of 0.9221. The lack of frequency band priors leads to... The error decreased significantly to 0.8897, while the removal of the short-time-scale branch caused the error to rise to 5.14. This directly confirms the key role of the synergistic effect of multi-scale temporal coding and learnable Sinc filtering in accurately fitting the long-term accumulation and short-term fluctuations of motion sickness in this invention.

[0084] Experimental results from two real-world driving-induced benchmarks comprehensively validated the absolute advantages of MS-GTFNet in joint extraction of spatiotemporal frequency multidimensional features and multi-task decoding. Quantitative indicators and ablation analysis consistently verified the core innovations of this invention: 1) Feature extraction based on learnable Sinc effectively bridges the individual differences caused by rigid frequency bands; 2) Multi-timescale feature recalibration ensures comprehensive perception of neural transient and cumulative signals; 3) Graph-enhanced brain region topological modeling endows the model with a high degree of physiological connectivity prior, greatly enhancing the interpretability of the neural mechanisms of model decision-making. These advancements collectively provide a continuous, objective, and smooth occupant comfort value for end-to-end autonomous driving control systems.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for decoding and continuously evaluating EEG features of motion sickness in autonomous driving, characterized in that, Includes the following steps: Step 1: Data input and preprocessing; Acquire vehicle-mounted multichannel EEG data that has been denoised and segmented by event locking, extract the signal tensor of the current evaluation window, and use it as the input feature tensor; Step 2: Frequency domain adaptive filtering extraction; A parameterized Sinc filter bank is constructed, in which the lower and upper cutoff frequencies of each frequency band are used as learnable parameters for end-to-end optimization, and the amplitude is constrained within the legal physiological range by a limiting operation; a temporal impulse response is constructed based on a Hamming window and a low-pass Sinc kernel; the input feature tensor is passed through the filter bank to obtain a multi-band tensor, and then modulated by a channel attention mechanism to generate the final frequency domain embedding; Step 3 Feature refinement and spatiotemporal structure generation; Brain topology structural features and multi-scale temporal features are extracted using a parallel network architecture, including: Graph-enhanced brain region attention: The physiological prior baseline weights, sample dynamic activation weights, and topological graph connection weights extracted by the graph attention network are multiplicatively modulated to generate a comprehensive spatial attention graph. The input channels are reweighted using the calibrated weights to obtain the EEG feature tensor with spatial feature enhancement. Then, multi-head self-attention is used to capture the dynamic sequence evolution across time steps. Multi-scale temporal coding: Construct parallel convolutional branches at multiple time scales, calculate multi-scale attention weight matrices through sample-level pooling and nonlinear mapping, recalibrate features at each scale, concatenate the weighted features and fuse them through convolution, and output the final spatiotemporal fusion representation. Step 4: Multi-task decoding and joint supervised loss function; The spatiotemporal embedding vector output from the time-domain branch is concatenated with the spectral embedding vector output from the frequency-domain branch and subjected to a nonlinear transformation to form a time-frequency fusion feature representation. The time-frequency fusion feature representation is then decoded in parallel using a multi-task head, including: continuous severity regression decoding, which outputs a physically meaningful continuous motion sickness severity score; and motion sickness level classification decoding, which outputs the probability distribution of discrete levels. The model is then trained by optimizing the joint supervised loss function.

2. The method for decoding and continuously evaluating EEG features of motion sickness for autonomous driving according to claim 1, characterized in that, In step 1, for each motion sickness button event, the duration is [duration value missing]. Each sampling point, including EEG segments from each electrode channel form the input feature tensor. ,and ,in For batch size, The set of real numbers is represented; the motion sickness button event refers to the event in which the subject, during the experiment, subjectively assesses their current motion sickness state in real time by referring to the motion sickness description of the MISC motion sickness subjective evaluation scale, and inputs four levels of motion sickness severity—no motion sickness, mild motion sickness, moderate motion sickness, and severe motion sickness—through a self-made four-level button device; each motion sickness button event is synchronously recorded with a trigger timestamp and corresponding level label, which is used for event locking and interception of EEG segments and supervision label construction.

3. The method for decoding and continuously evaluating EEG features of motion sickness for autonomous driving according to claim 2, characterized in that, In step 2, for the first Each frequency band, with its lower and upper cutoff frequencies and End-to-end optimization is performed using learnable parameters, and the parameters are constrained within legal physiological ranges through amplitude limiting operations. （1） （2） in, For the amplitude limiting function, To parameterize the number of frequency bands in the Sinc filter bank, i.e., the number of filters, and ; To control the first Learnable parameters of cutoff frequency in each frequency band This is a preset lower limit for effective physiological frequency. The upper limit of the preset effective physiological frequency, To prevent the upper and lower cutoff frequencies from coinciding or overflowing, a very small positive constant is used. To control the first Learnable parameters for each bandwidth; Subsequently, a Hamming window-based system was built. and low-pass Sinc core time-domain impulse response : （3） in, The discrete-time sampling point number. For the first The bandwidth of each frequency band; Input feature tensor By incorporating physiological prior weights The filter bank obtains multiband tensors Then, the multi-band tensor is weighted using a channel-band gated modulation mechanism to generate the final frequency domain embedding containing task-specific spectral features. : （4） （5） in, For one-dimensional convolution computation in the time domain; For batch normalization operation; One-dimensional grouping Convolutional operations are used to generate channel-band gated weights that match the dimensions of the multi-band feature tensor; This is element-wise multiplication.

4. The method for decoding and continuously evaluating EEG features of motion sickness for autonomous driving according to claim 3, characterized in that, In step 3, for graph-enhanced brain region attention, the physiological prior baseline weights are adjusted. Sample dynamic activation weights Connection weights of the topological graph extracted by the graph attention network Perform multiplicative modulation to generate a comprehensive spatial attention map. : （6） Using calibrated weights The input channels are reweighted to obtain the EEG feature tensor enhanced with spatial features. : （7） （8） in, This represents the weight calibration function. This represents an element-wise multiplicative weighted operation; Then, multi-head self-attention is used to capture the dynamic sequence evolution across time steps: （9） in, For the first The output matrix of each attention head, where... This is the attention head index in a multi-head self-attention mechanism, and , For the number of heads; For normalized exponential functions, For the first A query matrix with attention heads For the first The key matrix of each attention head. For the first The value matrix of each attention head, Let be the feature dimension of the attention head, and ⊤ denote the matrix transpose; For multi-timescale temporal coding, construct parallel convolutional branches at four timescales. Output characteristics of each branch The multi-scale attention weight matrix is ​​calculated using sample-level pooling and nonlinear mapping. Recalibrate features at each scale: （10） in, The scale features are after attention recalibration. For the first In the Scale Branch, the th The original output feature map of each sample. It is a scale branch index, and ; For sample index, and ; The weighted features are concatenated along the channel dimension and then fused using a 1x1 convolution to output the final spatiotemporal fusion representation. ,in This represents the uniform sequence length after adaptive average pooling.

5. The method for decoding and continuously evaluating EEG features of motion sickness for autonomous driving according to claim 4, characterized in that, In step 4, the spatiotemporal embedding vector output by the time-domain branch is... Spectral embedding vector of frequency domain branch output The concatenation and nonlinear transformation are performed to form a shared time-frequency fusion feature representation. : （11） in, The set of learnable parameters for the feature fusion layer. For mapping operations that include batch normalization and non-linear activation functions, This is an element-wise addition; Subsequently, using multi-task head to Perform parallel decoding: Continuous Severity Regression Decoding: Regression Head Output a physically meaningful continuous motion sickness severity score: （12） in, The continuous motion sickness severity score is predicted by the regression branch. Use the Sigmoid activation function; Motion sickness level classification decoding: classification head Output the probability distribution of four severity levels of motion sickness: no motion sickness, mild motion sickness, moderate motion sickness, and severe motion sickness. （13） in, The probability distribution of the motion sickness severity level predicted for the classification branch output. It is a normalized exponential function.

6. The method for decoding and continuously evaluating EEG features of motion sickness for autonomous driving according to claim 5, characterized in that, In step 4, for the joint supervised loss function, the total loss function... It includes two categories: one is Huber regression constraints used to resist subjective scoring noise. ; The second is the cross-entropy constraint used for hierarchical boundary classification. And ensure the generalization ability of the network weight perspective. Regular terms: （14） in, , and To optimize gradients for multi-task optimization; For continuous severity pseudo-real labels injected with Gaussian perturbations; This provides a true label for the discrete severity level of motion sickness. This is the set of all optimizable weight parameters in the network model.

Images