Multi-scene driving risk assessment method based on transfer learning

Abstract

The invention provides a multi-scene driving risk assessment method based on transfer learning. The method is suitable for real-time assessment and short-time early warning of driving risks in tunnel and non-tunnel environments. According to the method, a multi-source data set collected by multiple sensors is constructed based on a natural driving test, static and dynamic characteristics are extracted by using a sliding time window technology, and the driving risk is evaluated through a multi-scene driving risk evaluation model. The scene driving risk assessment model integrates a static information encoder, a variable selection network and an interpretable multi-head self-attention mechanism, has strong feature selection and time sequence modeling capabilities, and improves the adaptability and generalization capability to scarce tunnel data through a transfer learning strategy. Test results show that the method is superior to the existing mainstream model in the aspects of alarm rate, accuracy rate and false alarm rate, and is suitable for a risk assessment and early warning module in an intelligent traffic system.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation, deep learning, and driving risk prediction, and in particular to a multi-scenario driving risk assessment method based on transfer learning. Background Technology

[0002] With the continuous development of Intelligent Transportation Systems (ITS) and Advanced Driver Assistance Systems (ADAS), real-time assessment and short-term warning of driving risks have become crucial for improving road traffic safety and emergency response efficiency. Numerous studies have shown that approximately 90% of traffic accidents are related to human factors, including fatigue, inattention, operational errors, and risky behavior. Tunnels, due to their unique enclosed environment, sudden changes in lighting, and visual adaptation issues, are high-risk areas for traffic accidents. Once a traffic accident occurs in a tunnel, it can not only cause serious injuries or fatalities but also easily trigger traffic congestion and secondary accidents, impacting the overall operational efficiency of the tunnel.

[0003] To address the above issues, a certain degree of research on driving risk assessment has been carried out. The main methods can be divided into three categories: (1) Manual assessment method: Professionals manually label driving behavior. However, this method is labor-intensive, time-consuming, and easily affected by subjective factors, resulting in low efficiency and consistency. (2) Threshold setting method: Fixed behavioral indicator thresholds are set based on experience to judge the level of driving risk. However, this method lacks adaptability, is difficult to cover complex road conditions and individual driving differences, and has poor robustness. (3) Cluster analysis method: Unsupervised clustering is performed based on driving data to classify risky behaviors. However, the clustering model is difficult to handle multi-source data with high dimensionality and cannot model the evolution relationship in the time series, which limits the ability to identify and warn of risk evolution trends.

[0004] In addition, existing methods generally have the following key problems: (1) Insufficient data expression: They mostly use single behavioral variables (such as speed and acceleration) as core indicators, and fail to integrate multi-dimensional and multi-modal data such as driver's physiological characteristics, eye movement information, and road environment, lacking a comprehensive understanding of the causes of driving behavior; (2) Low utilization of time series information: It is difficult to model the dynamic evolution characteristics of driving behavior, and it is impossible to accurately capture the "precursor" information of risks, resulting in insufficient early warning capabilities; (3) Weak model generalization ability: Most methods are trained based on a certain road condition, and their transfer adaptability across scenarios (such as from non-tunnel to tunnel) is poor, and their performance in high-risk scenarios such as tunnels is greatly reduced; (4) Lack of interpretability and visualization ability: Although deep learning models improve prediction accuracy, they are usually "black boxes", making it difficult to explain the basis for risk judgment, which limits their trust and deployment in traffic safety management and assisted driving systems.

[0005] Therefore, a technical solution with the following characteristics is urgently needed to address the above challenges: (1) integrate multi-source and multi-modal driving data to achieve comprehensive perception and modeling of complex interactive behaviors of "human-vehicle-road"; (2) have the ability to model long-term dependence on historical behaviors, accurately identify potential risks and provide forward-looking warnings; (3) have good cross-scene transfer capabilities, especially the ability to transfer the learning results of non-tunnel samples to tunnel driving risk prediction, alleviating the problem of scarce tunnel data; (4) have a highly interpretable structure and feature selection mechanism to support the tracing of risk identification basis and visualization of feature importance. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a multi-scenario driving risk assessment method that combines transfer learning mechanisms with interpretable deep learning models. This method aims to improve the model's risk identification capabilities in complex driving environments and is particularly suitable for real-time risk assessment and short-term early warning tasks in tunnel driving scenarios.

[0007] The technical solution adopted in this invention is as follows: A multi-scenario driving risk assessment method based on transfer learning is proposed. This method is implemented using a multi-scenario driving risk assessment model, which can predict driving risks, specifically including the risk level at the current moment and within multiple future time steps. It enables real-time identification and early warning of high-risk driving states. The multi-scenario driving risk assessment model integrates a static information encoder, a variable selection network, and an interpretable multi-head self-attention mechanism, possessing strong feature selection and temporal modeling capabilities. Furthermore, it enhances adaptability and generalization ability to scarce tunnel data through a transfer learning strategy. The training process of the multi-scenario driving risk assessment model includes the following steps: c1. Acquisition of multi-source driving data: Conduct natural driving experiments to acquire multi-source driving data generated by drivers during real-world driving scenarios. Specifically, this data is collected through in-vehicle equipment, including driver behavior data (steering wheel angle and its rate of change, accelerator pedal opening and its rate of change, brake pedal opening and its rate of change, obtained using an in-vehicle 5G unit), vehicle operating status data (vehicle lateral and longitudinal speeds, lateral and longitudinal acceleration, obtained using a GNSS positioning system), eye movement data (gaze duration, saccade duration, pupil area and region of interest, obtained using an eye tracker), physiological data (electroencephalogram (EEG) and blood volume pulse (BVP), obtained using physiological monitoring equipment), and road condition data (surrounding road ambient brightness and number of vehicles, obtained using in-vehicle and external cameras). c2. Construction of multi-source datasets The collected multi-source driving data is subjected to frequency unification (to align global time) and sliding time window processing to extract multi-dimensional input features, including static features (driver ID, gender, driving experience), dynamic time series features (driver behavior, vehicle operating status, road conditions, eye movement features, and physiological features), and statistical features (mean, standard deviation, maximum value, minimum value, variability, peak-to-peak value, margin factor, waveform factor, and impulse factor), to construct a multi-source dataset; c3. Construction and labeling of driving risk units: Based on the multi-source dataset, the driving process is divided into multiple driving risk units according to the time series. Each driving risk unit contains multi-source driving data within a fixed time window (e.g., 3 seconds). Clustering methods are used to classify driving risk units into risk levels and form multi-level driving risk labels. Specifically, unsupervised clustering algorithms (such as fuzzy C-means) are used to perform dimensionality reduction clustering on driver behavior data to generate three levels of driving risk labels: safe level, low-risk video, and high-risk level, and the proportion of risk units at different levels is obtained. c4. Construction of a multi-scenario driving risk assessment model: A multi-scenario driving risk assessment model is constructed, which includes the following key modules: Static encoder: Used for deep representation of static features, it uses a separate gated residual network (GRN) as the encoder to produce three distinct vectors. c s , c e and c h ,in, c s Used for variable selection networks c h For dynamic encoders c e Used in temporal fusion decoders; among them, gated residual networks can be used to improve the model's nonlinear modeling ability and training stability; Dynamic encoder: Used to encode dynamic temporal features, it uses LSTM as the encoder and... c h The initial hidden state vector used for the LSTM encoder; Variable selection network: It realizes the importance assessment and screening of input features. There is a dedicated variable selection network for each type of input variable. This represents the input after the transformation of the j-th variable at time t. It is the flattened vector of all past inputs at time t. After nonlinear processing by GRN, Flattened vector ( Together with c s The GRN is fused with the common input, and the parameter selection weights are further calculated through a Softmax layer. Finally, By weight The selected vector is obtained by weighting the vectors. Temporal fusion decoder: This includes a temporal enhancement layer, a multi-head attention layer, and a forward propagation layer to achieve risk prediction across multiple time steps, both current and future. The temporal enhancement layer is used to fuse the vectors output by the dynamic encoder and the output by the static encoder. c e The output of this layer is fed into the subsequent multi-head attention layer. The multi-head attention layer employs a Transformer-based multi-head attention mechanism to extract long-term dependencies in the historical time series (input features) and improve the model's interpretability in the time dimension. The forward propagation layer processes the computation results of the multi-head attention layer, performing a non-linear transformation before the model output, which is then passed through an LSTM decoder. The fully connected layer and the Softmax layer output probability prediction results for different risk levels.

[0008] c5. In the multi-scenario driving risk assessment model, a transfer learning mechanism is adopted. Based on the data of non-tunnel road sections and tunnel road sections in the multi-source dataset, a source dataset (mainly non-tunnel data, supplemented by some tunnel data) and a target dataset (tunnel scenario data) are constructed, which are used for model pre-training and model fine-tuning, respectively. The parameters of the static encoder, dynamic encoder, temporal enhancement layer and forward propagation layer in the model are frozen, and the parameters of the remaining structure are fine-tuned, thereby enhancing the model's adaptability in the target scenario and improving the convergence speed. c6. Model Output and Evaluation: The model ultimately outputs the driving risk level at the current and multiple future time steps through a Softmax classifier; The model's ability to identify different levels of driving risks is evaluated by analyzing multiple assessment indicators such as accuracy, alarm rate, precision, and false alarm rate through a multi-class confusion matrix. Furthermore, the model's interpretability and practicality are enhanced by combining attention weight output to visualize features.

[0009] Further, in step c2, driver behavior includes steering wheel angle and its rate of change, accelerator and brake pedal opening and their rates of change; vehicle operating status includes lateral and longitudinal speed and acceleration during vehicle operation; road conditions include ambient brightness and number of surrounding vehicles; eye movement features include gaze duration, saccade duration, pupil area and region of interest; physiological features include electroencephalogram and blood volume pulse; and the specific method for extracting the statistical features is as follows: the dynamic temporal features are divided into segments of data for each time window using the sliding time window method, and the mean, standard deviation, maximum value, minimum value, variability, peak-to-peak value, margin factor, waveform factor and impulse factor of each time window segment are calculated.

[0010] Further, in step c5, during the transfer learning step, a risk stratified sampling method is used to construct source and target datasets based on data from non-tunnel and tunnel road sections in the multi-source dataset. Specifically, tunnel and non-tunnel road section samples are extracted from the multi-source dataset, and the proportion of the three driving risk units in the tunnel road section samples is determined. Samples of each risk level in the tunnel road section samples are divided in a 6:4 ratio and used in the target and source datasets respectively to ensure that the distribution of samples of different risk levels conforms to the actual tunnel driving risk pattern. The source dataset is expanded using non-tunnel road section samples so that the ratio of non-tunnel samples to tunnel samples in the constructed source dataset is 5:1, thus obtaining the final source dataset. Different learning rates are set during model pre-training and fine-tuning, and a parameter freezing strategy is applied during fine-tuning to accelerate the process.

[0011] Furthermore, the driving risk prediction result includes the risk level of each driving risk unit and its corresponding probability value, and the output is a multi-classification result after activation by the Softmax function.

[0012] Furthermore, the multi-scenario driving risk assessment model is visualized by means of: outputting the feature dimension weight distribution through the variable selection network and outputting the time dimension weight distribution through the multi-head attention layer, thereby realizing the interpretability of the model in the feature dimension and time dimension.

[0013] Furthermore, during model training, the evaluation metrics include accuracy, alarm rate, precision, and false alarm rate. A multi-class confusion matrix is ​​used to analyze each evaluation metric to assess the model's ability to identify different levels of driving risks.

[0014] Compared with the prior art, the present invention has the following significant technical advantages and practical value: First, it adapts to multiple scenarios and focuses on risk identification in tunnel scenarios: This invention is specifically designed for mixed tunnel and non-tunnel scenarios. It comprehensively captures driver behavior and risk characteristics through high-dimensional, multi-source data, and can be widely applied to risk identification tasks in complex road environments. It has good versatility and deployment potential.

[0015] Second, the introduction of transfer learning can alleviate the problem of data scarcity. By pre-training with non-tunnel samples and fine-tuning with a small amount of tunnel data, the model can reduce its large-scale dependence on tunnel driving data while ensuring accuracy, thus improving its practicality under small sample conditions.

[0016] Third, deep models have strong interpretability and feature importance analysis capabilities: By using a variable selection network and a multi-head attention layer, the interpretability of the model in both the feature and time dimensions is improved, which can help developers and traffic managers understand the sources of risk.

[0017] Fourth, integrated risk assessment and early warning output with strong real-time performance: The model supports risk prediction for the current and future time steps and has "short-term early warning capability," which can provide early warning before potential risks occur, thus assisting in the design of advanced driver assistance systems (ADAS) and autonomous driving technologies.

[0018] Fifth, the evaluation indicators are comprehensive and the performance is excellent: Experiments show that the method has an accuracy rate of 91.74%, an alarm rate of 95.34%, a risk identification accuracy rate of 95.02%, and a false alarm rate of 5.97% for safe driving. Its performance is superior to traditional machine learning and deep learning models in all aspects.

[0019] In summary, the multi-scenario driving risk assessment method based on transfer learning proposed in this invention has good practicality, accuracy and scalability, and is suitable for application in many key fields such as intelligent transportation, driving behavior analysis, and tunnel safety monitoring. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the training process of the multi-scenario driving risk assessment model in this invention; Figure 2 This is a diagram illustrating the data partitioning for transfer learning. Figure 3 It is a heatmap of attention distribution over time extracted by a multi-head attention mechanism; Figure 4 This is a schematic diagram of the weight distribution of input features extracted by the variable selection network. Detailed Implementation

[0021] The method of the present invention will be further described below with reference to the accompanying drawings and specific examples.

[0022] A multi-scenario driving risk assessment method based on transfer learning is proposed. This method is implemented based on a multi-scenario driving risk assessment model and is applicable to real-time assessment and short-term early warning of driving risks in tunnel and non-tunnel environments.

[0023] like Figure 1 The diagram shows the training process for a multi-scenario driving risk assessment model, which includes the following steps: I. Natural Driving Data Collection This invention utilizes multi-source driving data acquired through various sensors in a natural driving experiment. The experiment involved 35 participants with varying driving experience, who drove the test vehicle in real-world road conditions. The test route included suburban roads and highway tunnel sections. The vehicle was equipped with the following components both inside and outside the vehicle: GNSS positioning equipment is used to collect operational data such as vehicle position, speed, and acceleration; The 5G vehicle unit and CAN interface module are used to collect behavioral data such as throttle, braking, and steering wheel angle. Four high-definition cameras are used to capture the front, rear, left and right views of the vehicle; Eye trackers and physiological monitors collect data such as pupil area, saccade time, EEG brain signals, and BVP pulse signals.

[0024] II. Construction of Multi-Source Datasets All data are processed to a uniform frequency, downsampled to 10 Hz, and combined with OpenCV image processing and object detection algorithms to extract information such as ambient brightness and the number of surrounding vehicles.

[0025] After cleaning and feature processing, dynamic temporal features (including driver behavior, vehicle operating status, road conditions, eye-tracking features, etc.) and their statistical features (including mean, standard deviation, and variation rate, etc.) are extracted and combined with static features (such as driver information, road type, etc.) to form a multi-source dataset.

[0026] III. Construction of Driving Risk Units and Classification of Risk Levels The driving data is divided into driving risk units by sliding time windows, and the duration of each driving risk unit is 3 seconds.

[0027] After using principal component analysis (PCA) to reduce the dimensionality of driver behavior data, a fuzzy C-means (FCM) clustering algorithm was introduced to divide all samples into: risk level 0; low risk level 1; and high risk level 2.

[0028] Cluster quality was assessed using the silhouette coefficient and the Davies-Bouldin index. The final result was three categories of labeled risk samples, totaling approximately 470,000 samples.

[0029] IV. Design of Multi-Scenario Driving Risk Assessment Model This invention constructs an interpretable deep neural network model as a multi-scenario driving risk assessment model architecture, which mainly includes the following modules: Static encoder: A separate gated residual network is used as the encoder to vectorize static features such as driver ID and road type, generating multiple context vectors for subsequent modules. Specifically, this includes... c s , c e and c h ,in, c s Used for variable selection networks c h For dynamic encoders c e Used in time fusion decoders. Gated residual networks provide controllable nonlinear modeling capabilities, ensuring smooth information flow and avoiding performance degradation caused by excessive network depth.

[0030] Dynamic encoder: Uses LSTM as the encoder to encode dynamic temporal features such as vehicle operating status; c h The initial hidden state vector used for the LSTM encoder.

[0031] Variable selection network: This network automatically selects key features from the input features and assigns them weights, achieving feature selection and noise reduction. There is a dedicated variable selection network for each type of input variable. This represents the input after the transformation of the j-th variable at time t. It is the flattened vector of all past inputs at time t. After nonlinear processing by GRN, Flattened vector ( Together with c s The GRN is fused using a common input, and the parameter selection weights are further calculated through a Softmax layer. Finally, By weight The selected vector is obtained by weighting.

[0032] The temporal fusion decoder comprises a temporal enhancement layer, a multi-head attention layer, and a forward propagation layer to achieve risk prediction across multiple time steps, both current and future. The temporal enhancement layer fuses the vectors output by the dynamic encoder and the vectors output by the static encoder. c eThe output of this layer is fed into the subsequent multi-head attention layer. The multi-head attention layer employs a Transformer-based multi-head attention mechanism to extract long-term dependencies in the historical time series (input features) and improve the model's interpretability in the time dimension. The forward propagation layer processes the computation results of the multi-head attention layer, performing a non-linear transformation before the model output, which is then passed through an LSTM decoder. The fully connected layer and the Softmax layer output probability prediction results for different risk levels.

[0033] The model output is a sequence of length 7, with each value being one of 0 (safe), 1 (low risk), or 2 (high risk), representing the driving risk level within the next 7 time windows from the current time.

[0034] V. Design of Transfer Learning Strategies To improve the model's adaptability in tunnel scenarios, this invention employs a transfer learning framework, specifically as follows: A source dataset (Source Data) for model pre-training and a target dataset (Domain Data) for model fine-tuning and testing are constructed by combining samples from tunnel and non-tunnel road sections. The datasets are divided as follows: Figure 2 The tunnel section samples were extracted hierarchically according to the ratio of three driving risk units of 5.3:3.5:1, and then divided into the target dataset and the source dataset at a ratio of 6:4. Next, the source dataset was expanded by adding non-tunnel samples to tunnel samples at a ratio of 5:1. The final source dataset has two characteristics: ① The 5:1 distribution of non-tunnel samples to tunnel samples, and the retention of tunnel samples, improves the robustness and generalization ability of the pre-trained model during transfer learning; ② The three driving risks conform to the distribution pattern of driving risks in tunnel sections, improving the predictive performance of risk identification in tunnel sections.

[0035] Then, a parameter freezing strategy was adopted: the parameters of the static encoder, dynamic encoder, temporal enhancement layer and forward propagation layer of the model were frozen, and only the parameters of the remaining structures were fine-tuned; the fine-tuning learning rate was set to 0.0001 to maintain the stability of the overall structure, reduce the number of tunnel samples required, and accelerate the convergence speed of the model; the learning rate of the model pre-training was set to 0.001.

[0036] VI. Model Training and Performance Validation The model of this invention is trained on an RTX 3090 GPU using the PyTorch framework, employing a cross-entropy loss function and L2 regularization. Furthermore, it proposes model evaluation metrics including accuracy (Accuracy, Ac), alarm rate (Ar), precision rate (Pr), and false alarm rate (Fr). A multi-class confusion matrix is ​​used to analyze each evaluation metric, assessing the model's ability to identify different levels of driving risks. The corresponding multi-class confusion matrix is ​​shown in Table 1 below. Table 1 The corresponding indicator is calculated as follows: To evaluate the performance of the proposed model in driving risk assessment, representative machine learning models and deep learning models were selected for comparison. The models compared included: (0) the model proposed in this invention after pre-training with transfer learning; (1) the model trained and tested directly on the target dataset without transfer learning; (2) the Transformer model; (3) the Long Short-Term Memory (LSTM) neural network model; (4) the Random Forest (RF) model; (5) the XGBoost model; and (6) the Support Vector Machine (SVM). All models were trained and tested on the target dataset constructed in this invention. The results of the four evaluation metrics for different models at each time step are shown in Table 2 below. Table 2 Comparison of evaluation results from different models shows that, compared to other deep learning and machine learning models, the model used in this invention can more accurately assess driving risk levels in real time, provide short-term warnings for driving risk levels in multiple future units, and has a high ability to identify risk types. Considering all factors, this model demonstrates strong superiority and competitiveness.

[0037] VII. Input Feature Visualization Support The variable selection network can extract attention weights for different input features, which, combined with the attention heatmap, can be used to... Figure 3 Feature weight diagram as follows Figure 4 It can achieve transparent identification of driving behavior, providing decision support for traffic management.

Claims

1. A multi-scenario driving risk assessment method based on transfer learning, characterized in that, This is achieved through a multi-scenario driving risk assessment model. The driving risk prediction results obtained based on the multi-scenario driving risk assessment model include the risk level at the current moment and in multiple future time steps. The training process of the multi-scenario driving risk assessment model includes the following steps: c1. Acquire multi-source driving data generated by the driver during driving in real road scenarios. The multi-source driving data includes driver behavior data, vehicle operating status data, physiological data, eye movement data, and road condition data. c2. Perform frequency unification processing on the multi-source driving data, extract static features and dynamic time-series features, and further extract statistical features through a sliding time window method to construct a multi-source dataset; c3. Based on the multi-source dataset, the driving process is divided according to the time series, driving risk units are constructed, and the risk level of the driving risk units is classified by clustering method to form multi-level driving risk labels and obtain the proportion of risk units of different levels. c4. Construct a multi-scenario driving risk assessment model, which includes a static encoder, a dynamic encoder, a variable selection network, and a temporal fusion decoder, to assess and predict driving risks within the current and future time windows. The static encoder encodes static features; the dynamic encoder encodes dynamic temporal features; the variable selection network evaluates and filters the importance of static and dynamic temporal features; and the temporal fusion decoder includes a temporal enhancement layer, a multi-head attention layer, and a forward propagation layer, to predict risks at multiple time steps in the current and future. c5. In the multi-scenario driving risk assessment model, a transfer learning mechanism is adopted to construct source datasets and target datasets based on data from non-tunnel road sections and tunnel road sections in the multi-source dataset, which are used for model pre-training and model fine-tuning, respectively; the parameters of the static encoder, dynamic encoder, temporal enhancement layer and forward propagation layer in the model are frozen, and the parameters of the remaining structures are fine-tuned; c6. Model Output and Evaluation: The model ultimately outputs the driving risk level for the current and future multiple time steps; the model is evaluated through multiple evaluation indicators, and the feature visualization is achieved by combining the attention weight output.

2. The multi-scenario driving risk assessment method based on transfer learning as described in claim 1, characterized in that: In step c1, the multi-source data is collected through the following devices: an in-vehicle 5G unit for collecting driver behavior data; a GNSS positioning system for acquiring vehicle operating status data; an eye tracker for collecting eye movement data; a physiological monitoring device for acquiring physiological data, including electroencephalogram and blood volume pulse; and multiple in-vehicle and external cameras for acquiring road condition data, specifically including ambient brightness and number of vehicles.

3. The multi-scenario driving risk assessment method based on transfer learning as described in claim 1, characterized in that: In step c2, the static features include driver ID, gender, and driving experience; the dynamic temporal features include driver behavior, vehicle operating status, road conditions, eye movement features, and physiological features. Driver behavior includes steering wheel angle and its rate of change, accelerator and brake pedal openings and their rates of change; vehicle operating status includes lateral and longitudinal speeds and accelerations during vehicle operation; road conditions include ambient brightness and the number of surrounding vehicles; eye movement features include gaze duration, saccade duration, pupil area, and region of interest; and physiological features include electroencephalogram (EEG) and blood volume pulse. The specific method for extracting the statistical features is as follows: the dynamic temporal features are divided into segments of data for each time window using a sliding time window method, and the mean, standard deviation, maximum value, minimum value, variability, peak-to-peak value, margin factor, waveform factor, and impulse factor for each time window segment are calculated.

4. The multi-scenario driving risk assessment method based on transfer learning as described in claim 1, characterized in that: In step c3, the driving risk unit includes multi-source driving data within a fixed time window. The driving risk unit is then subjected to unsupervised clustering based on driver behavior data to obtain three driving risk levels, including a safe level, a low-risk level, and a high-risk level.

5. The multi-scenario driving risk assessment method based on transfer learning as described in claim 1, characterized in that: In step c5, during the transfer learning step, source and target datasets are constructed using risk stratified sampling based on data from non-tunnel and tunnel road sections in the multi-source dataset. Specifically, tunnel and non-tunnel road section samples are extracted from the multi-source dataset, and the proportion of the three driving risk units in the tunnel road section samples is determined. Samples of each risk level in the tunnel road section samples are divided into two groups at a ratio of 6:4 and used in the target and source datasets respectively to ensure that the distribution of samples of different risk levels conforms to the actual tunnel driving risk pattern. The source dataset is expanded using non-tunnel road section samples so that the ratio of non-tunnel samples to tunnel samples in the constructed source dataset is 5:1, resulting in the final source dataset. Different learning rates are set during model pre-training and fine-tuning.

6. The multi-scenario driving risk assessment method based on transfer learning as described in claim 1, characterized in that: The driving risk prediction results include the risk level of each driving risk unit and its corresponding probability value, and the output is a multi-classification result after the Softmax function is activated.

7. The multi-scenario driving risk assessment method based on transfer learning as described in claim 1, characterized in that: The multi-scenario driving risk assessment model is visualized by means of: outputting the feature dimension weight distribution through the variable selection network and outputting the time dimension weight distribution through the multi-head attention layer, thereby realizing the interpretability of the model in the feature dimension and time dimension.

8. The multi-scenario driving risk assessment method based on transfer learning as described in claim 1, characterized in that: During model training, the evaluation metrics include accuracy, alarm rate, precision, and false alarm rate. A multi-class confusion matrix is ​​used to analyze each evaluation metric to assess the model's ability to identify different levels of driving risks.

Images