Intelligent traffic decision-making method based on multi-source sensor data

Abstract

The invention provides an intelligent traffic decision-making method based on multi-source sensor data, and the method comprises the steps: obtaining vehicle sensor data and road information, and obtaining a standardized traffic flow feature vector through data synchronization, fusion, feature extraction and dimension reduction processing; inputting the standardized traffic flow feature vector into a traffic flow prediction model, and predicting the traffic flow through a vehicle dynamics model constraint state transition matrix; according to the predicted traffic flow and the real-time road condition information, the road congestion state is judged, and a real-time congestion map is generated; in combination with the real-time congestion map, the road network speed limit rule and the starting and ending point position information, an optimal path is planned, and path smoothing processing is performed through a driving behavior mode; and constructing a vehicle intelligent decision-making system, integrating traffic flow prediction, congestion detection and a path planning model, and outputting a vehicle driving strategy according to real-time vehicle sensor data.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an intelligent traffic decision-making method based on multi-source sensor data. Background Art

[0002] In artificial intelligence data analysis, the formats of raw data are diverse and need to be converted into a unified structure suitable for analysis. The problem is how to transform the format, specifically how to convert unstructured data such as text, images, audio, etc. through various algorithms to make it into structured data.

[0003] In the field of intelligent transportation, it is necessary to analyze various sensor data of vehicles, including vehicle speed, acceleration, steering angle, GPS location, etc. These data formats are different. How to convert them into unified time series data is a challenging problem.

[0004] This application studies how to integrate multi-source heterogeneous data of vehicles into structured spatiotemporal feature sequences through data synchronization, data cleaning, feature engineering and other technologies, so as to facilitate subsequent traffic flow prediction, congestion detection and route planning. Summary of the invention

[0005] The present invention provides an intelligent traffic decision-making method based on multi-source sensor data, which mainly includes: Acquire vehicle sensor data and road information, and obtain a standardized traffic flow feature vector through data synchronization, fusion, feature extraction and dimensionality reduction processing; input the standardized traffic flow feature vector into the traffic flow prediction model, and predict the traffic flow through the vehicle dynamics model constraint state transfer matrix; judge the road congestion status based on the predicted traffic flow and real-time road condition information, and generate a real-time congestion map; plan the optimal path based on the real-time congestion map, road network speed limit rules and start and end point location information, and perform path smoothing through driving behavior patterns; build a vehicle intelligent decision-making system that integrates traffic flow prediction, congestion detection and path planning models, and outputs vehicle driving strategies based on real-time vehicle sensor data. Furthermore, the acquisition of vehicle sensor data and road information, and obtaining a standardized traffic flow feature vector through data synchronization, fusion, feature extraction and dimensionality reduction processing, include: acquiring vehicle sensor data of different sampling frequencies, arranging them in time series through data packet sequence numbers, and synchronizing the data to a unified time base using a linear interpolation or cubic spline interpolation algorithm to obtain synchronized sensor data; combining the road network topology and real-time road condition information, and constructing a spatiotemporal feature matrix reflecting traffic flow changes based on historical path trajectory data, vehicle speed distribution and driving behavior patterns; using a principal component analysis method to reduce the dimensionality of the spatiotemporal feature matrix and extract key traffic flow feature vectors; and normalizing the feature vectors in combination with road network speed limit rules to obtain the standardized traffic flow feature vector. Furthermore, the method of inputting the standardized traffic flow feature vector into a traffic flow prediction model and predicting traffic flow by constraining a state transfer matrix through a vehicle dynamics model includes: taking the standardized traffic flow feature vector that has been preprocessed and feature extracted as input; constructing a long short-term memory network model and inputting the standardized traffic flow feature vector into the model; introducing a vehicle dynamics model into the long short-term memory network model and constraining the network state transfer matrix through a vehicle dynamics equation; training the long short-term memory network model using a stochastic gradient descent algorithm, calculating the loss function gradient through a back propagation algorithm, and updating the network weights; and optimizing the model hyperparameters using a cross-validation method and a grid search strategy to obtain an optimized traffic flow prediction model, and outputting the predicted traffic flow according to the input feature vector. Furthermore, judging the road congestion status and generating a real-time congestion map based on the predicted traffic flow and real-time road condition information includes: integrating the predicted traffic flow and real-time road condition information; judging the congestion status of each road section by threshold comparison according to a preset vehicle speed threshold and occupancy threshold, and generating a real-time congestion map; obtaining historical congestion data for comparative verification, and calculating the difference between the current real-time congestion map and the historical congestion data; if the difference exceeds a preset threshold, obtaining real-time user feedback data for optimization; adjusting the key parameters for congestion judgment according to the historical data verification results and the user feedback data; optimizing the congestion judgment threshold by training a regression model, and generating a corrected real-time congestion map. Furthermore, the real-time congestion map, road network speed limit rules and starting and ending point location information are combined to plan the optimal path, and the path is smoothed through driving behavior patterns, including: obtaining the latitude and longitude coordinates of the starting point and the ending point input by the user; obtaining real-time traffic data from the map API, including the road congestion index and section speed limit information; updating the road network data according to the real-time traffic data to generate a real-time traffic map; according to the real-time traffic map, using the A search algorithm for path planning, using the section travel time as the cost function, and searching for the optimal path from the starting point to the end point; combining the user's driving behavior pattern, using different path optimization strategies, and using the Bezier curve algorithm for path smoothing. Furthermore, the vehicle intelligent decision-making system is constructed, integrating traffic flow prediction, congestion detection and path planning models, and outputting vehicle driving strategies according to real-time vehicle sensor data, including: obtaining real-time traffic data collected by vehicle sensors, and performing preprocessing, including outlier detection and noise filtering, to obtain a normalized traffic data set; inputting the normalized traffic data set into a traffic flow prediction model to predict the traffic flow change trend in the future; according to the traffic flow prediction results, using a congestion detection model based on a speed threshold and a vehicle density threshold to determine the road congestion status; passing the congested section information to the path planning model, and obtaining the starting and ending positions of the vehicle, and searching for the optimal travel path that avoids the congested section by constructing a road network topology map. Furthermore, it also includes: making adaptive adjustments to different road scenarios and driving behavior patterns; acquiring road scene information and driving behavior data, using long short-term memory neural networks to adaptively adjust the intelligent decision-making system, training decision models under different scenarios, and obtaining a decision model library; optimizing and updating the models in the decision model library using incremental learning algorithms for newly acquired road scene information and driving behavior data; selecting the most matching model from the decision model library based on the vehicle's current road scenario and driving behavior pattern; combining real-time traffic flow data, using dynamic programming algorithms to adjust decision parameters and generate the optimal decision plan. Furthermore, the adaptive adjustment for different road scenarios and driving behavior patterns also includes: using a rule-based reasoning engine to verify the rationality of decision-making plans; setting a series of safety and reliability rules, using forward-linking reasoning to determine whether the decision meets the requirements, and adjusting decision plans that do not meet the requirements; using an evaluation function to evaluate and score the decision effect based on the vehicle's driving trajectory and decision execution; through comparative analysis of historical decision data, using a heuristic search algorithm to find an optimizable decision path, using the evaluation results as a feedback signal, and iteratively optimizing the decision model; establishing a driving behavior profile and a road scene feature library, using a clustering algorithm to classify driver behavior and scene features, forming personalized decision preferences, and providing a reference basis for the decision model.

[0006] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: The present invention discloses an intelligent traffic decision-making method based on multi-source sensor data. The method first synchronously processes vehicle sensor data with different sampling frequencies to eliminate time deviation. Then, a spatiotemporal feature matrix reflecting traffic flow changes is constructed by combining road network topology and real-time road conditions. Key features are extracted by principal component analysis and input into a long short-term memory network model for traffic flow prediction. A real-time congestion map is generated based on the prediction results, and the optimal path is planned using an A search algorithm. Finally, traffic flow prediction, congestion detection and path planning models are integrated to construct an intelligent decision-making system to provide vehicles with future traffic condition predictions and optimal travel path recommendations.

[0007] The present invention realizes accurate modeling and prediction of complex traffic scenarios through multi-source data fusion and deep learning methods, and can provide reliable decision support for vehicle scheduling and traffic management. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a flow chart of the intelligent traffic decision-making based on multi-source sensor data of the present invention. DETAILED DESCRIPTION

[0009] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0010] like Figure 1 The intelligent traffic decision-making method based on multi-source sensor data in this embodiment may specifically include: Step S101, obtain vehicle sensor data of different sampling frequencies, arrange the data in time sequence by data packet sequence number according to the differences in data delay time and timestamp accuracy, determine the data validity using sensor working mode information, use linear interpolation algorithm and spline interpolation algorithm to synchronize the data to a unified time base, eliminate the time deviation of the data, and obtain synchronized sensor data.

[0011] The raw data of the vehicle sensor is obtained, and the data is arranged in time sequence according to the sequence number information of the data packet. By analyzing the working mode information of the sensor, it is determined whether the arranged data is valid and invalid data is eliminated. The valid data after the invalid data is eliminated are integrated to obtain the complete data sequence of each sensor. In view of the problem of inconsistent sampling frequencies of different sensors, the average sampling interval of each sensor data is first calculated, and the minimum sampling interval is used as a unified time base. Then, the linear interpolation algorithm is used to interpolate the data of each sensor to a unified time base. The accuracy of linear interpolation is evaluated by calculating the mean square error of the data points before and after interpolation. If the mean square error exceeds the preset threshold, the cubic spline interpolation algorithm is used to improve the data interpolation accuracy. Through the interpolation operation, the synchronized data of different sensors under the unified time base is obtained. The absolute value average of the timestamp difference of the data before and after synchronization of each sensor is calculated as the synchronization deviation of the sensor. If the synchronization deviation exceeds the preset threshold, the interpolation parameters of the sensor are adjusted, such as increasing the number of interpolation points or adjusting the interpolation method, to further reduce the synchronization deviation. The synchronized sensor data is used as input, and the Kalman filter algorithm is applied for data fusion to obtain the estimated values ​​of the vehicle's motion state, such as position, speed, acceleration, etc. Based on the estimated values ​​of the motion state, a vehicle dynamics model is constructed to achieve vehicle state monitoring and fault diagnosis. At the same time, the synchronized sensor data is stored in a time series database to provide support for subsequent data analysis and mining.

[0012] Specifically, obtaining raw data from vehicle sensors is the first step in vehicle status monitoring. Take a car equipped with GPS, accelerometers, and gyroscopes as an example. These sensors continuously output data packets. Each data packet contains serial number information. For example, a GPS data packet may be output once per second, with the serial number starting from 1 and increasing; while an accelerometer may output once every 0.01 seconds. By arranging the data in time sequence according to the serial number, the time sequence of the sensor data can be restored. Sensor operating mode information is crucial to judging the validity of the data. For example, GPS may lose signal when entering a tunnel, and the output data at this time is invalid. By analyzing indicators such as GPS signal strength, these invalid data can be identified and eliminated. Similarly, when the vehicle is stationary, the slight fluctuations in the accelerometer output may be just noise and should also be regarded as invalid data. Integrating valid data to obtain a complete sequence lays the foundation for subsequent processing. For GPS, a series of longitude and latitude coordinate points may be formed; for accelerometers, it is a time series of three-axis acceleration values. These sequences reflect the motion trajectory and dynamic characteristics of the vehicle. The sampling frequencies of different sensors are often inconsistent, which brings challenges to data fusion. Calculating the average sampling interval can quantify this difference. Assuming that the sampling interval of GPS is 1 second, that of accelerometer is 0.01 second, and that of gyroscope is 0.05 second, 0.01 second is used as a unified time base. Using a linear interpolation algorithm, the data of GPS and gyroscope can be interpolated to this unified base. Evaluating the interpolation accuracy is important to ensure data quality. By calculating the mean square error of the data points before and after interpolation, the error caused by interpolation can be quantified. If the error is too large, such as exceeding the preset 0.1 meter (for GPS position data), a more complex interpolation method is required. Cubic spline interpolation usually provides higher accuracy, especially when dealing with nonlinear data. Synchronization deviation is an important indicator of data synchronization quality. This indicator is calculated by calculating the mean absolute difference of the data timestamps before and after synchronization for each sensor. For example, if the synchronization deviation of GPS data is 0.05 seconds, and the preset threshold is 0.01 seconds, the interpolation parameters need to be adjusted. The synchronization deviation can be reduced by increasing the number of interpolation points or using a higher order interpolation method. Kalman filtering is a powerful data fusion algorithm, especially suitable for processing noisy sensor data. By feeding synchronized GPS, accelerometer, and gyroscope data into a Kalman filter, a more accurate estimate of the vehicle's position, velocity, and attitude can be obtained. For example, even if the GPS signal is temporarily lost, the Kalman filter can still provide a relatively accurate position estimate in combination with the accelerometer and gyroscope data. Building a vehicle dynamics model based on the fused data allows for more advanced condition monitoring and fault diagnosis. For example, by comparing the actual acceleration with the engine output power, a drop in engine efficiency can be detected; by analyzing the relationship between steering angle and lateral acceleration, tire wear can be monitored.Storing processed data in a time series database, such as InfluxDB, can support efficient time series analysis. This storage method facilitates long-term trend analysis, such as changes in vehicle performance over time, seasonal fluctuations in fuel efficiency, etc. At the same time, the query performance optimization of the time series database makes real-time monitoring and alarming possible, such as quickly triggering an alarm when a certain indicator exceeds the normal range. This series of data processing and analysis steps, from raw sensor data acquisition to final status monitoring and fault diagnosis, constitutes a complete vehicle health management system. With such a system, vehicle reliability can be improved, maintenance plans can be optimized, and even potential failures can be predicted, thereby reducing operating costs and improving safety. In addition, the large amount of accumulated data can also be used to improve vehicle design and provide valuable input for future intelligent transportation systems.

[0013] Step S102, based on the road network topology and real-time traffic information, combined with the synchronized sensor data, obtain historical path trajectory data. Through the vehicle speed distribution data and driving behavior patterns, a spatiotemporal feature matrix reflecting traffic flow changes is constructed. The spatiotemporal feature matrix is ​​cleaned, normalized and standardized to obtain a pre-processed spatiotemporal feature matrix.

[0014] Obtain road network topology data, real-time traffic information data, and vehicle sensor data, perform time synchronization processing on these data, and then merge them together to construct a complete road traffic data set. By analyzing the historical path trajectory data in the road traffic data set, the vehicle speed distribution characteristics and driving behavior pattern characteristics in different sections are extracted, and the spatiotemporal feature matrix reflecting the changes in traffic flow is constructed based on the extracted features. The constructed spatiotemporal feature matrix is ​​preprocessed, including removing noise data such as outliers and missing values, normalizing and standardizing the data to improve data quality and consistency. The preprocessed spatiotemporal feature matrix is ​​input into the machine learning model for training, and the spatiotemporal evolution law of traffic flow in the road network is learned through the model. During the training process, the hyperparameters of the model are optimized using methods such as grid search to improve the prediction performance of the model. The trained machine learning model is used to predict the real-time road condition data to obtain the traffic flow distribution of the road network in the future. By comparing and analyzing historical data, a reasonable traffic flow warning threshold is set. When the predicted traffic flow exceeds the threshold, it is judged that traffic congestion may occur. Automatically generate traffic diversion plans based on traffic congestion warning results, guide vehicles to bypass congested sections or take traffic restriction measures on congested sections to ease traffic pressure. Continuously collect road traffic data, compare traffic flow prediction results with actual road conditions, and calculate prediction errors. When the prediction error exceeds a certain range, trigger model retraining, use the latest data to fine-tune the model, improve the accuracy and real-time performance of traffic flow prediction, and form a closed-loop control with dynamic update iterations.

[0015] Specifically, the road network topology data contains information such as the connection relationship, length, and number of lanes of the road, while the real-time road condition information reflects the current traffic conditions, such as traffic volume and average speed. The on-board sensor data provides microscopic information such as the movement trajectory and speed of a single vehicle. By synchronizing and fusing these data in time, a comprehensive road traffic data set can be constructed. For example, the traffic flow data of a city's main road during the morning rush hour may show that 5,000 vehicles pass through per hour, while the on-board sensor data reflects that the average speed of individual vehicles on this section of road is only 20 kilometers per hour. The combination of these two types of data can more accurately describe the congestion situation. By analyzing the historical path trajectory, the speed distribution characteristics and driving behavior patterns of different sections can be extracted. For example, on roads near schools, the speed distribution may be concentrated at 30-40 kilometers per hour, and the braking frequency is high; on highways, the speed may be distributed between 80-120 kilometers per hour, and the acceleration changes are small. These features can be used to construct a spatiotemporal feature matrix that reflects changes in traffic flow. Preprocessing the spatiotemporal feature matrix is ​​a key step to improve data quality. Outliers, such as data points with instantaneous speeds exceeding 200 km / h due to sensor failure, need to be removed. For missing values, interpolation can be performed using data from nearby time points or similar sections. Data normalization can unify features of different dimensions (such as vehicle speed and traffic volume) to the same scale, which is convenient for subsequent modeling. The preprocessed data is input into a machine learning model, such as a long short-term memory network (LSTM), to learn the spatiotemporal evolution of traffic flow. During model training, the grid search method can be used to optimize hyperparameters, such as the number of hidden layers and learning rate of LSTM, to improve prediction performance. For example, experiments may find that for the traffic data of the city, the model performs best when a 3-layer LSTM with 128 neurons per layer and a learning rate of 0.001 is used. Using the trained model to predict real-time traffic conditions, the expected traffic volume of each section in the next few hours can be obtained. By comparing historical data, a reasonable warning threshold can be set. For example, the normal traffic volume at a certain intersection is 2,000 vehicles per hour. When the predicted value exceeds 3,000 vehicles, the system will issue a congestion warning. Based on the early warning results, the traffic management system can automatically generate a diversion plan. For example, when it is predicted that a section of a road in the city center is about to be congested, the system can suggest that the driver choose an alternative route, or temporarily adjust the signal light timing plan to increase the green light time in the congested direction. These measures can effectively alleviate traffic pressure and improve road utilization efficiency. To ensure the accuracy of the prediction, the system needs to continuously collect new traffic data and compare it with the prediction results. When the prediction error exceeds the preset threshold (such as the actual traffic volume differs from the predicted value by more than 20%), the model will be triggered to retrain. This dynamic update mechanism ensures that the model can adapt to long-term changes in traffic patterns, such as changes in traffic flow caused by the opening of new roads or the completion of large buildings.Through this complete traffic flow prediction and management system, cities can better deal with traffic congestion problems, improve road utilization efficiency, and reduce citizens' travel time, while also laying the foundation for the construction of future smart cities.

[0016] Step S103, using the principal component analysis method to reduce the dimension of the preprocessed spatiotemporal feature matrix, extracting the key traffic flow feature vector, and normalizing the feature vector in combination with the road network speed limit rules to obtain a standardized traffic flow feature vector.

[0017] According to the historical data of traffic flow, a spatiotemporal feature matrix is ​​constructed, which contains the characteristic values ​​of traffic flow at different time and space positions. Before constructing the spatiotemporal feature matrix, the historical data of traffic flow needs to be preprocessed, including data cleaning, missing value processing and outlier processing. Singular value decomposition (SVD) is used to reduce the dimension of the spatiotemporal feature matrix. Through matrix decomposition and eigenvalue extraction, the key feature vector that can characterize the law of traffic flow change is obtained. SVD can effectively extract the main features of the matrix, reduce the data dimension, and retain the key information of the data. The speed limit rule information of the road network is obtained, and the speed limit values ​​of different sections are determined according to factors such as road grade and section attributes. The extracted traffic flow feature vector is compared with the speed limit value of the corresponding section. If the traffic flow represented by the feature vector exceeds the speed limit value, it is judged that there is a risk of traffic congestion in the section. The maximum and minimum values ​​of the traffic flow feature vector are normalized, and the traffic flow feature values ​​of different sections and different times are unified to the range of [0, 1] to obtain a standardized traffic flow feature vector. Normalization can eliminate the dimensional differences between different features, making them comparable. Based on the standardized traffic flow feature vector, the support vector regression (SVR) algorithm is used to establish a traffic flow prediction model. The historical traffic flow data is divided into training sets and test sets. The SVR kernel function type, penalty coefficient and other parameters are optimized using grid search and cross-validation methods to improve the accuracy of traffic flow prediction. The predicted traffic flow information for a period of time in the future is compared with the road network speed limit rules to determine the sections and time periods where traffic congestion may occur. If the predicted traffic flow exceeds the speed limit value of the corresponding section, it indicates that the section may experience traffic congestion in the future, and corresponding traffic control measures need to be taken in advance to provide a basis for the dispatching decision of the traffic management department.

[0018] Specifically, constructing a spatiotemporal feature matrix is ​​a key step in traffic flow prediction. Taking a city's main road as an example, a day can be divided into 24 hours, with each hour as a time dimension, and the spatial dimension includes multiple sections on the main road. For each spatiotemporal point, characteristic values ​​such as traffic volume and average speed are recorded. In the preprocessing stage, outliers and missing values ​​need to be processed. For example, the high traffic volume data that suddenly appears on a certain section at 3 a.m. may be an outlier and needs to be removed; while the missing data due to equipment failure can be interpolated and supplemented with adjacent time points or historical data of the same period. Singular value decomposition (SVD) is an effective dimensionality reduction method. The assumed spatiotemporal feature matrix size is 1000×100, representing 1000 time points and 100 spatial locations. Through SVD, it may be found that the first 10 singular values ​​can explain 95% of the data variance. This means that the original 1000×100 matrix can be compressed into a 1000×10 matrix, greatly reducing the data dimension while retaining key information. The acquisition of speed limit rule information is crucial for determining the risk of traffic congestion. For example, the speed limit on a city expressway may be 80 km / h, while the speed limit on a road near a school may be 30 km / h. By comparing the extracted traffic flow features with these speed limit values, potential congested sections can be identified. If the average speed of a certain expressway section is continuously lower than 50% of the speed limit, that is, 40 km / h, it can be determined that the section is at risk of congestion. Normalization can make different features comparable. For example, the traffic volume may vary between 0-5000 vehicles / hour, while the average speed may vary between 0-100 km / h. Through maximum and minimum value normalization, these features can be mapped to between 0-1. In this way, whether it is traffic volume or speed, 0.8 represents a value close to the maximum value of the feature, which is convenient for subsequent model training and comparison. The support vector regression (SVR) algorithm performs well in traffic flow prediction. Traffic data from the past 4 hours can be used to predict traffic conditions for the next hour. During the parameter optimization process, it may be found that when the RBF kernel function is used, the penalty coefficient C=10, and epsilon=0.1, the root mean square error of the model on the test set is the smallest, achieving a good prediction effect. Finally, by comparing the prediction results with the speed limit rules, the sections of the road that may be congested can be identified in advance. For example, if it is predicted that the average speed of a section of road will drop to 20 kilometers per hour between 5 and 6 pm, which is much lower than the speed limit of 40 kilometers per hour on that section of road, the traffic management department can take measures in advance, such as adjusting the timing of traffic lights and guiding vehicle diversion, to alleviate potential traffic pressure. This data-driven prediction and management method can significantly improve the operating efficiency of urban traffic, reduce citizens' travel time, and provide strong support for the construction of smart cities.

[0019] Step S104: input the standardized traffic flow feature vector into the long short-term memory network model, train the traffic flow prediction model by constraining the state transfer matrix of the network through the vehicle dynamics model, and use the cross-validation method to optimize the model hyperparameters to obtain the optimized traffic flow prediction model.

[0020] The original traffic flow data is obtained, and the data is cleaned and preprocessed to remove outliers and missing values ​​to obtain a standardized traffic flow data set. According to the characteristics of the traffic flow data, a multi-dimensional traffic flow feature vector is constructed, including vehicle flow, vehicle speed, vehicle density, etc., and key features are extracted through feature engineering methods. The traffic flow feature vector is standardized by the Min-Max standardization method, and the feature data is mapped to the interval [0, 1] to eliminate the dimension effect. The standardized traffic flow feature vector will be used as the input of the subsequent model. A long short-term memory (LSTM) network model is constructed, and the standardized traffic flow feature vector is used as the model input. In the LSTM model, the state transfer matrix of the network is constrained by introducing the prior knowledge of the vehicle dynamics model. Specifically, the vehicle dynamics model is used to describe the transfer relationship between the vehicle motion states, which is used as the prior constraint of the state transfer matrix in the LSTM model to improve the physical interpretability and prediction accuracy of the model. The LSTM model is trained using the stochastic gradient descent algorithm, the gradient of the loss function to the model parameters is calculated by the back propagation algorithm, and the network weight parameters are updated using the gradient descent method to continuously optimize the model to fit the training data. In order to optimize the hyperparameters of the LSTM model, a 5-fold cross-validation method was used. The data set was randomly divided into 5 subsets, 4 of which were selected as training sets each time, and the remaining 1 subset was selected as the validation set. This was repeated 5 times, and the average performance was taken as the evaluation index of the model. Through grid search, different hyperparameter combinations, such as the number of hidden layers, the number of hidden units, and the learning rate, were traversed, and the hyperparameter combination with the best performance on the validation set was selected as the final model hyperparameter. The LSTM model was retrained using the optimized hyperparameters to obtain the final traffic flow prediction model. This model can be used to predict and analyze traffic flow in the future and provide support for traffic management and decision-making.

[0021] Specifically, the acquisition and preprocessing of traffic flow data are the basis for building a prediction model. Taking a main road in a city as an example, the traffic flow, average speed and other data can be collected every 5 minutes through roadside sensors. In the data cleaning stage, it is necessary to identify and process outliers and missing values. For example, the traffic flow of 5,000 vehicles / hour that suddenly appeared on a certain road section at 3 a.m. is obviously unreasonable. It may be caused by equipment failure and should be eliminated. For data missing due to temporary equipment failure, linear interpolation can be used to fill it. Feature engineering is a key step to improve model performance. In addition to basic traffic flow and speed, derived features can also be constructed. For example, the vehicle density can be obtained by dividing the traffic flow by the speed, which reflects the degree of road congestion. In addition, considering the periodicity of traffic flow, time can be encoded as a periodic feature, such as encoding 24 hours as sine and cosine values ​​to better capture the time pattern. Standardization can eliminate the dimensional differences between different features. Taking traffic flow and speed as examples, the traffic flow may vary between 0-5,000 vehicles / hour, while the speed varies between 0-100 kilometers / hour. Through Min-Max normalization, both can be mapped to the interval [0,1], so that the model can fairly consider the influence of each feature. The long short-term memory (LSTM) network model performs well in processing time series data. In traffic flow prediction, LSTM can effectively capture long-term dependencies, such as the pattern of morning and evening peaks. By introducing prior knowledge of the vehicle dynamics model, the physical interpretability of the model can be further improved. For example, the vehicle following model is used to describe the speed and distance relationship between adjacent vehicles, which is used as a constraint condition for LSTM state transition, so that the model prediction results are more in line with the actual traffic operation law. During model training, the stochastic gradient descent algorithm can efficiently optimize large-scale data sets. Assume that there is one year of historical data, about 525,600 samples (one sample every 5 minutes). The batch size can be set to 128, the initial value of the learning rate is 0.01, and the learning rate decay strategy is used to halve the learning rate every 50 epochs to ensure that the model converges to a better solution. Hyperparameter optimization is an important means to improve model performance. Through 5-fold cross validation, model performance can be evaluated more robustly. For example, you can try different combinations of the number of hidden layers (1-3 layers), the number of hidden units (32, 64, 128, 256), and the learning rate (0.001, 0.01, 0.1). Assume that after grid search, it is found that the model has the best performance when there are 2 layers of LSTM, 128 hidden units in each layer, and an initial learning rate of 0.01, which can be used as the final hyperparameter configuration. The final LSTM model can be used to predict traffic flow in the next few hours. For example, the model can predict that the average speed of a certain section of road will drop to 25 km / h during the morning rush hour from 7:30 to 8:30 tomorrow, which is much lower than the normal level of 50 km / h on this section.Based on this prediction, traffic management departments can adjust the timing of signal lights in advance, extend the green light time on main roads, or guide some vehicles to choose alternative routes through variable message boards, thereby alleviating potential traffic congestion.

[0022] Step S105: Based on the prediction results of the optimized traffic flow prediction model and the real-time road condition information, the speed threshold and the occupancy threshold are set to determine the road congestion status and generate a real-time congestion map. The generated congestion map is verified and optimized using historical data and real-time feedback to obtain a corrected real-time congestion map.

[0023] Obtain the prediction result data and real-time road condition information data of the optimized traffic flow prediction model, fuse the two types of data, and obtain the fused traffic data. Analyze the fused traffic data according to the preset vehicle speed threshold and occupancy threshold, judge the congestion status of each road section by threshold comparison, and obtain the congestion status data. Generate a real-time congestion map according to the congestion status data, match the congestion status with the map road section, and mark the road sections with different congestion levels with different colors or icons on the map to form an intuitive congestion map. Obtain the historical congestion data of the time period corresponding to the current congestion map, compare and verify the real-time congestion map with the historical congestion data, and calculate the difference between the two. If the difference exceeds the preset threshold, it is determined that the current congestion map may be biased and needs to be optimized and adjusted. Obtain real-time feedback data from users, including information such as congestion reported by users and changes in road conditions. Compare the user feedback data with the historical congestion data. If it is found that there is a large difference between the user feedback and the historical data, it indicates that the real-time congestion map may be biased and needs to be further optimized. Dynamically optimize the real-time congestion map based on historical data verification results and user feedback data. Use the logistic regression algorithm to adjust the key parameters that affect congestion judgment. Through training sample data, establish a regression model between congestion status and parameters such as vehicle speed and occupancy rate to obtain the optimized congestion judgment threshold. Use the optimized threshold to regenerate the corrected congestion map. Publish the corrected real-time congestion map to provide users with accurate and real-time traffic congestion information services. At the same time, apply the optimized congestion judgment threshold to the subsequent congestion map generation process, and continuously improve the accuracy of congestion judgment and map generation through continuous iterative optimization.

[0024] Specifically, the fusion of the output results of the traffic flow prediction model and real-time road condition information is the key to improving the accuracy of congestion judgment. For example, the prediction model of a certain intersection shows that the traffic flow will reach 2,000 vehicles / hour in the next hour, while the real-time monitoring data shows that the current traffic flow is 1,800 vehicles / hour. A more accurate traffic flow estimate can be obtained by weighted average. This fusion method can balance historical rules and real-time changes and improve the reliability of congestion judgment. The judgment of congestion status is usually based on two key indicators: vehicle speed and occupancy rate. Taking a certain city main road as an example, the vehicle speed threshold can be set to 20 kilometers / hour and the occupancy threshold can be set to 30%. When the actual vehicle speed is lower than the threshold and the occupancy rate is higher than the threshold, the road section is judged to be in a congested state. This multi-dimensional judgment can more comprehensively reflect the road traffic conditions and avoid misjudgments that may be caused by a single indicator. The visualization of congestion maps is crucial for users to understand road conditions. A red, yellow, and green color system can be used to indicate the degree of congestion, where red indicates severe congestion (vehicle speed <10 km / h), yellow indicates mild congestion (10-20 km / h), and green indicates smooth traffic (>20 km / h). This intuitive color coding scheme allows users to quickly identify congested sections and plan their trips appropriately. Comparing the real-time congestion map with historical data is an effective way to verify its accuracy. For example, a section of road is usually slightly congested during the morning rush hour (7:30-9:00) on weekdays. If the real-time map shows that the road is completely smooth during this period, there may be a deviation. The difference threshold can be set to 20%, that is, when the consistency between the real-time congestion status and the historical data is less than 80%, the optimization process is triggered. This mechanism can detect abnormal situations in a timely manner and ensure the reliability of the congestion map. User feedback is an important supplement to the optimization of the congestion map. Assume that there is a mobile application that allows users to report real-time traffic conditions. For example, if a user reports that a traffic accident has caused severe congestion on a certain section of road, but the system shows that the section is smooth, this obvious difference will trigger further investigation and optimization. User feedback can not only capture emergencies, but also make up for the problem of insufficient coverage of fixed sensors and improve the comprehensiveness of congestion judgment. Dynamic optimization is the key to improving the accuracy of congestion maps. Through the logistic regression model, the congestion judgment threshold can be dynamically adjusted based on historical data and user feedback. For example, the analysis found that the congestion characteristics of a commercial district on weekends are significantly different from those on weekdays, and differentiated thresholds can be set for different time periods. Assuming that the speed threshold on weekdays is 20 kilometers per hour, it may need to be adjusted to 15 kilometers per hour on weekends to more accurately reflect the actual congestion situation. This data-driven dynamic optimization can make congestion judgment more in line with actual traffic characteristics. Continuous iterative optimization is the key to maintaining the long-term accuracy of congestion maps. By regularly (such as monthly) analyzing the consistency of user feedback, actual traffic data and congestion maps, systematic deviations can be identified and corrected. For example, it is found that the congestion prediction of a newly developed area is generally low, which may be due to changes in traffic patterns in the area. The prediction model needs to be retrained and the congestion judgment parameters need to be adjusted.This closed-loop optimization mechanism ensures that the congestion map can adapt to the dynamic changes in urban traffic and continue to provide users with accurate and reliable traffic information.

[0025] Get the coordinates of the starting point and the end point, and use the A search algorithm to search for the optimal path based on the corrected real-time congestion map and road network speed limit rules. Combined with the driving behavior pattern, the path is smoothed to generate a humanized navigation path.

[0026] Get the starting and ending addresses entered by the user, and convert the addresses into longitude and latitude coordinates through the Amap API. Get real-time traffic data from the Amap API, including road congestion index and road speed limit information, dynamically update the original road network data, and generate a real-time traffic map. According to the road network structure and real-time traffic conditions, use the A search algorithm for path planning. Take the road section travel time as the cost function, comprehensively consider factors such as road section length, speed limit, congestion, etc., and search for the optimal path from the starting point to the end point. Among them, the road section travel time is dynamically calculated based on the road section length and real-time traffic conditions. Get the user's historical trajectory data, use the DBSCAN clustering algorithm to cluster the trajectory points, and identify different driving behavior patterns. According to the clustering results, count the distribution of characteristic parameters such as speed, acceleration, steering angle, etc. in each driving behavior mode. According to the user's current location and driving speed, determine which driving behavior mode he belongs to. Different path optimization strategies are used for different driving behavior modes. For the robust driving mode, a path with a smaller curvature is selected to avoid sudden changes such as sharp turns; for the aggressive driving mode, the curvature of the path is allowed to be appropriately increased to shorten the travel distance. The path obtained in the previous step is smoothed using the Bezier curve algorithm to generate a smooth and natural navigation path. The path smoothing effect under different driving behavior modes is achieved by adjusting the control point position of the Bezier curve. After smoothing, abnormal driving behaviors such as sharp turns and sudden acceleration can be reduced, and driving comfort can be improved. The smoothed navigation path is segmented, and the ID3 decision tree algorithm is used to generate the optimal navigation prompt strategy for each segment of the path according to the characteristics of the road type, congestion, weather conditions, etc., including the timing of voice broadcast, broadcast frequency, and road condition broadcast content. The training samples of the decision tree come from user feedback and driving behavior analysis results. The planned navigation path and the generated navigation prompt strategy are returned to the client, and the navigation information is visualized through the Android App. According to the shape characteristics of the navigation path, a smooth colored path curve is drawn on the map, and text annotations such as direction arrows, road names, and distances are superimposed on the path. During navigation, the user's real-time location reported by the GPS device is obtained at a frequency of 1 second, matched with the planned route, the projection point of the user's current location on the route is calculated, and the map perspective is updated with the projection point as the center. At the same time, according to the user's current speed and the road speed limit, the voice navigation prompt is dynamically triggered to guide the user to drive along the planned route. When the user reaches the destination, the navigation ends.

[0027] Specifically, the core of the intelligent navigation system lies in real-time traffic analysis and personalized route planning. First, the address entered by the user is converted into longitude and latitude coordinates through the Amap API. This step ensures the accuracy of subsequent route planning. For example, if the user enters "Beijing West Railway Station" as the starting point, the system will convert it into coordinates of longitude 116.321592 and latitude 39.894806. The acquisition of real-time traffic data is crucial for dynamically updating the road network. Assuming that the congestion index of a main road is 0.8 (between 0 and 1, the closer to 1, the more congested), and the speed limit is 60 kilometers per hour, the system will adjust the travel time estimate of the road section accordingly. This dynamic update mechanism can reflect the impact of emergencies on traffic, such as temporary congestion caused by traffic accidents. The A* algorithm is used for route planning, and its core idea is to guide the search direction through the evaluation function. In this system, the evaluation function comprehensively considers the length of the road section, the real-time speed, and the estimated time to reach the destination. For example, for a 5-kilometer section, if the current average speed is 40 kilometers per hour, the travel time for the section is about 7.5 minutes. By comparing the total travel time of different paths, the algorithm can find the optimal route. User driving behavior analysis is the key to personalized navigation. The DBSCAN clustering algorithm can identify different driving patterns based on the density of trajectory points. For example, frequent acceleration and deceleration and sharp turns may be classified as "aggressive" driving, while steady speed and slow steering may be identified as "robust" driving. For robust drivers, the system will give priority to paths with less curvature, even if this may slightly increase the driving distance; for aggressive drivers, the system may choose a more direct route that requires more turns. Path smoothing uses the Bezier curve algorithm, which can significantly improve the navigation experience. For example, at a sharp turn, the original path may appear as a sharp corner, but after Bezier curve processing, the turn will become smooth and more in line with the actual driving trajectory. This not only improves the visual comfort of navigation, but also helps drivers better predict the timing of turning. The generation of navigation prompt strategy uses ID3 decision tree algorithm, considering multiple factors to determine the best prompt method. For example, on congested roads, the system may increase the prompt frequency to remind the driver to pay attention to the road conditions ahead; on highways, the prompt interval may be extended to avoid excessive interference. Weather factors also affect the prompt strategy. For example, on rainy days, the system may additionally remind the driver to slow down. Finally, the visualization and real-time update of navigation information are key to user experience. The system updates the user's location every second and calculates the deviation from the planned path. If the deviation exceeds a preset threshold (such as 50 meters), the system automatically replans the path. At the same time, the triggering timing of the voice prompt is dynamically adjusted according to the ratio of the user's speed to the road speed limit. For example, when approaching a turning point, if the user's speed is significantly higher than the safe turning speed, the system will issue a deceleration reminder in advance.This intelligent navigation system, which takes into account real-time traffic conditions, personal driving habits and safety factors, can not only provide the best route, but also significantly improve driving experience and safety. Through continuous learning and optimization, the system can adapt to the needs of different users and provide personalized navigation services for every trip.

[0028] The optimized traffic flow prediction model, congestion detection model and path planning model are integrated to build a vehicle intelligent decision-making system. The real-time collected vehicle sensor data is analyzed to output traffic flow prediction, congestion status and optimal travel path for a period of time in the future to guide vehicle scheduling and management.

[0029] Real-time traffic data collected by vehicle sensors is obtained, and the data is preprocessed, including outlier detection and noise filtering. Outliers are removed using methods such as median filtering, and noise is removed using methods such as Kalman filtering to obtain a standardized traffic data set. The preprocessed traffic data is input into the traffic flow prediction model based on the LSTM neural network. By learning and training historical traffic flow data, the trend of traffic flow changes in the future is predicted to obtain traffic flow prediction results. According to the traffic flow prediction results, a congestion detection model based on speed threshold and vehicle density threshold is used to judge the road congestion state. Through statistical analysis of historical traffic data, reasonable speed threshold and vehicle density threshold are set. When the real-time traffic data exceeds the threshold range, it is judged as a congested state and the congested road section information is obtained. The congested road section information is passed to the path planning model based on the Dijkstra algorithm, and the starting and ending positions of the vehicle are obtained at the same time. By constructing a road network topology map, the Dijkstra shortest path algorithm is used to search for the optimal travel path that avoids congested sections with the optimization goal of minimizing travel time. Integrate traffic flow prediction results, congestion status information and optimal travel routes into the vehicle intelligent decision-making system to form a real-time updated vehicle driving strategy. According to the driving strategy, the intelligent decision-making system automatically generates detailed vehicle dispatching instructions, such as acceleration, deceleration, lane change, etc. through preset decision rules, and sends them to the vehicle terminal through the vehicle communication network to guide the vehicle to execute the optimal driving route. While the vehicle is driving along the optimal path, the on-board sensors continuously collect data such as vehicle speed and position and upload them to the intelligent decision-making system. The system uses GPS positioning and trajectory matching algorithms to achieve real-time monitoring and trajectory tracking of the vehicle's operating status, and feeds back the monitoring information to the decision-making system, forming a closed-loop control of vehicle dispatching management, and realizing dynamic adjustment and optimization.

[0030] Specifically, the core of the intelligent transportation system lies in real-time data processing and predictive analysis. First, the raw data collected by vehicle sensors often contains noise and outliers, which need to be preprocessed. For example, the speed values ​​collected continuously by a vehicle speed sensor within 1 second are: 50, 51, 49, 150, 52 km / h, of which 150 km / h is obviously an outlier. The use of median filtering can effectively remove such outliers and retain the real speed data. Kalman filtering can smooth random fluctuations in the data. Assume that the actual traffic volume of a certain section of road is 500 vehicles per hour, but due to various factors, the data collected by the sensor may fluctuate between 480-520 vehicles. Kalman filtering can obtain an estimation result that is closer to the true value through two steps of prediction and update. The preprocessed data is input into the traffic flow prediction model based on LSTM. The advantage of LSTM network is that it can capture long-term dependencies and is particularly suitable for processing time series data. For example, by learning historical data, the model can find the pattern that the traffic volume of a certain section of road will increase significantly from 4 to 6 pm every Friday, so as to predict the possible congestion in advance on the next Friday. The congestion detection model uses preset thresholds to determine the road status. For example, the speed threshold of a city's main road is set to 30km / h, and the vehicle density threshold is set to 50 vehicles per kilometer. When real-time data shows that the average speed is less than 30km / h and the vehicle density exceeds 50 vehicles per kilometer, the system will determine that the road section is in a congested state. Based on the congestion information, the Dijkstra algorithm is used for path planning. Suppose there are multiple paths from point A to point B, and the shortest path passes through the congested section C, which is expected to take 30 minutes. The Dijkstra algorithm will consider other possible paths, such as the path around point D, which is longer, but due to smooth traffic, the actual time may only be 25 minutes. The algorithm will choose the latter as the optimal path. The intelligent decision-making system integrates the prediction results and path planning to generate specific driving instructions. For example, the system may instruct the vehicle to slow down to 40km / h 500 meters before reaching the congested section, and recommend changing to the right lane 200 meters away to facilitate subsequent turns. These precise instructions help smooth traffic flow and reduce congestion. Real-time monitoring and trajectory tracking are important feedback mechanisms of the system. Through GPS positioning, the system can detect whether the vehicle deviates from the planned path. If the deviation exceeds a preset threshold (such as 50 meters), the system will immediately recalculate the path. This dynamic adjustment capability enables the system to respond to emergencies, such as road construction or traffic accidents. The whole process forms a closed-loop control system that continuously optimizes decisions. For example, if the system finds that most vehicles choose to turn right instead of going straight at a certain intersection, it will analyze the reason for this phenomenon, which may be due to a blockage on the road ahead that the system has not yet known. The system will update the road network status based on this information and provide a better path selection for subsequent vehicles. This intelligent transportation system can not only improve the driving efficiency of individual vehicles, but also optimize the traffic flow distribution as a whole.Through predictive traffic management, the system can proactively disperse traffic flow and avoid congestion, thereby significantly improving the traffic operation efficiency of the entire city.

[0031] The intelligent decision-making system is adaptively adjusted for different road scenarios and driving behavior patterns. Through incremental learning, the prediction performance of the model is continuously optimized, the robustness and real-time response capabilities of the system are improved, and reliable decision-making support is provided for safe and efficient travel of vehicles.

[0032] According to the acquired road scene information and driving behavior data, the long short-term memory neural network (LSTM) is used to adaptively adjust the intelligent decision-making system. By analyzing historical data and real-time data, the decision models in different scenarios are trained using the TensorFlow framework to obtain a decision model library. For the newly acquired road scene information and driving behavior data, the incremental learning algorithm is used to optimize and update the models in the decision model library. The gradient descent method is used to adjust the model parameters, and the prediction performance of the model is improved by continuously accumulating and learning new data. According to the current road scene and driving behavior mode of the vehicle, the most matching model is selected from the decision model library. Combined with real-time traffic flow and other data, the dynamic programming algorithm is used to adjust the decision parameters to generate the optimal decision plan and improve the real-time response capability of the system. The rule-based reasoning engine is used to verify the rationality of the decision plan. By setting a series of safety and reliability rules, forward-linking reasoning is used to determine whether the decision meets the requirements. The decision plans that do not meet the requirements are adjusted to enhance the robustness of the system. According to the vehicle's driving trajectory and decision execution, the evaluation function is used to evaluate and score the decision effect. By comparing and analyzing historical decision data, using heuristic search algorithms to find optimizable decision paths, using evaluation results as feedback signals, iteratively optimizing decision models, and forming closed-loop control. Establish driving behavior portraits and road scene feature libraries, and use K-means clustering algorithms to classify driver behaviors and scene features. By analyzing the differences in behavior patterns of different drivers in similar scenarios, form personalized decision preferences, provide references for decision models, and improve the adaptability of decisions. Considering multi-dimensional factors such as road scenes, traffic flow, and weather conditions, use genetic algorithms to build multi-objective optimization models. By solving the Pareto optimal solution set, formulate vehicle operation decision plans that take into account safety, efficiency, and comfort. Pass the optimal decision plan to the vehicle control system for execution, providing users with reliable travel decision support.

[0033] Specifically, the adaptive adjustment of intelligent decision-making systems is the key to improving the level of vehicle intelligence. Long short-term memory neural networks (LSTM) perform well in processing time series data and can capture long-term dependencies. For example, by analyzing historical traffic flow data, LSTM can learn traffic patterns during rush hours in the morning and evening on weekdays, so as to make more accurate predictions for similar periods in the future. The TensorFlow framework provides strong support for model training. Assuming that driving data under different weather conditions are collected, decision models that adapt to sunny, rainy, snowy and other scenarios can be trained. These models constitute a decision model library that provides a basis for decision-making in different situations. Incremental learning algorithms allow models to learn new information without losing existing knowledge. For example, when the system encounters a new road construction scenario, it can quickly adapt to this situation through incremental learning without retraining the entire model. The gradient descent rule is used to fine-tune model parameters and continuously improve prediction accuracy. Dynamic programming algorithms play an important role in path planning. Assuming that there are multiple paths from point A to point B, dynamic programming can calculate the optimal route based on real-time traffic conditions. If the main road is suddenly congested, the algorithm can quickly adjust and choose a solution to bypass the secondary road. The rule-based reasoning engine ensures the safety and rationality of decisions. For example, the system may generate a time-saving route, but if the route contains dangerous sections, the reasoning engine will reject this option based on safety rules and choose a safer alternative route. The evaluation function quantitatively evaluates the effectiveness of the decision. Assuming that the scoring criteria for multiple indicators such as driving time, fuel consumption, and comfort are set, the system will calculate the score based on the actual driving data. If a decision results in a long waiting time, the system will lower the score of the decision plan and adjust the strategy in similar situations in the future. The K-means clustering algorithm is used to analyze driving behavior patterns. By clustering a large amount of driving data, different driving styles such as aggressive, conservative, and stable can be identified. Based on the identification results, the system can provide personalized suggestions for different types of drivers, such as recommending a more conservative driving strategy for aggressive drivers in congested sections. Genetic algorithms perform well in multi-objective optimization. Assuming that driving time, energy efficiency, and passenger comfort need to be considered at the same time, genetic algorithms can generate a series of solutions that balance these goals by simulating the evolutionary process. The system can choose the most suitable solution for the current situation, such as focusing more on energy efficiency when driving long distances and more on comfort when driving in urban areas. Through the comprehensive application of these technologies, the intelligent decision-making system can continuously learn and adapt to the complex and changing traffic environment, providing users with a safe, efficient and comfortable travel experience. The system's adaptive ability enables it to respond to various emergencies, such as road construction, weather changes or traffic accidents, and always maintain the best decision-making level.

[0034] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent traffic decision-making method based on multi-source sensor data, characterized in that: include: Obtain vehicle sensor data and road information, and obtain standardized traffic flow feature vectors through data synchronization, fusion, feature extraction and dimensionality reduction processing; Inputting the standardized traffic flow feature vector into a traffic flow prediction model, and predicting the traffic flow through a vehicle dynamics model constrained state transfer matrix; According to the predicted traffic flow and real-time road condition information, the road congestion status is judged and a real-time congestion map is generated; Combining the real-time congestion map, road network speed limit rules and start and end point location information, planning the optimal route, and performing path smoothing through driving behavior patterns; Build a vehicle intelligent decision-making system that integrates traffic flow prediction, congestion detection and path planning models, and outputs vehicle driving strategies based on real-time vehicle sensor data.

2. The intelligent traffic decision-making method based on multi-source sensor data as claimed in claim 1, characterized in that: The vehicle sensor data and road information are obtained, and a standardized traffic flow feature vector is obtained through data synchronization, fusion, feature extraction and dimensionality reduction processing, including: Acquire vehicle sensor data with different sampling frequencies, arrange them in time sequence by data packet sequence number, and synchronize the data to a unified time base using linear interpolation to obtain synchronized sensor data; Combining the road network topology and real-time traffic information, a spatiotemporal feature matrix reflecting traffic flow changes is constructed based on historical path trajectory data, vehicle speed distribution, and driving behavior patterns; The principal component analysis method is used to reduce the dimension of the spatiotemporal feature matrix and extract the key traffic flow feature vector; The characteristic vector is normalized in combination with the speed limit rule of the road network to obtain the standardized traffic flow characteristic vector.

3. The intelligent traffic decision-making method based on multi-source sensor data as claimed in claim 1, characterized in that: The step of inputting the standardized traffic flow feature vector into a traffic flow prediction model and predicting the traffic flow by constraining a state transfer matrix of a vehicle dynamics model comprises: Taking the normalized traffic flow feature vector after preprocessing and feature extraction as input; Constructing a long short-term memory network model, and inputting the standardized traffic flow feature vector into the model; In the long short-term memory network model, the vehicle dynamics model is introduced, and the network state transfer matrix is ​​constrained by the vehicle dynamics equation; The long short-term memory network model is trained using a stochastic gradient descent algorithm, the loss function gradient is calculated using a back-propagation algorithm, and the network weights are updated; The cross-validation method and grid search strategy are used to optimize the model hyperparameters to obtain the optimized traffic flow prediction model, which outputs the predicted traffic flow based on the input feature vector.

4. The intelligent traffic decision-making method based on multi-source sensor data as claimed in claim 3, characterized in that: Also includes: The Min-Max standardization method is used to standardize the traffic flow feature vector, mapping the feature data to the interval [0, 1] to eliminate the dimension effect; Construct a long short-term memory (LSTM) network model and use the standardized traffic flow feature vector as the model input; In the LSTM model, the state transfer matrix of the network is constrained by introducing the prior knowledge of the vehicle dynamics model.

5. The intelligent traffic decision-making method based on multi-source sensor data as claimed in claim 4, characterized in that: In the LSTM model, by introducing the prior knowledge of the vehicle dynamics model, the state transfer matrix of the constrained network includes: The vehicle dynamics model is used to describe the transfer relationship between vehicle motion states, and it is used as a priori constraint of the state transfer matrix in the LSTM model to improve the physical interpretability and prediction accuracy of the model. The LSTM model is trained using the stochastic gradient descent algorithm, the gradient of the loss function to the model parameters is calculated through the back propagation algorithm, and the network weight parameters are updated using the gradient descent method to continuously optimize the model to fit the training data.

6. The intelligent traffic decision-making method based on multi-source sensor data as claimed in claim 1, characterized in that: The method of judging the road congestion status and generating a real-time congestion map based on the predicted traffic flow and real-time road condition information includes: integrating the predicted traffic flow with real-time road condition information; According to the preset vehicle speed threshold and occupancy rate threshold, the congestion status of each road section is judged by threshold comparison, and a real-time congestion map is generated; Obtain historical congestion data for comparison and verification, and calculate the difference between the current real-time congestion map and historical congestion data; If the difference exceeds a preset threshold, real-time user feedback data is obtained for optimization; Adjusting the key parameters for congestion determination based on historical data verification results and the user feedback data; The congestion judgment threshold is optimized by training the regression model to generate a corrected real-time congestion map.

7. The intelligent traffic decision-making method based on multi-source sensor data as claimed in claim 1, characterized in that: The method combines the real-time congestion map, the road network speed limit rules and the starting and ending point location information to plan the optimal path, and performs path smoothing processing through the driving behavior model, including: Get the latitude and longitude coordinates of the starting point and the end point entered by the user; Get real-time traffic data from the map API, including road congestion index and road speed limit information; Update road network data according to the real-time traffic data to generate a real-time traffic map; According to the real-time traffic map, the A search algorithm is used for path planning, and the road section travel time is used as the cost function to search for the optimal path from the starting point to the end point; Different path optimization strategies are adopted in combination with the user's driving behavior patterns, and the Bezier curve algorithm is used for path smoothing.

8. The intelligent traffic decision-making method based on multi-source sensor data as claimed in claim 1, characterized in that: The vehicle intelligent decision-making system is constructed to integrate traffic flow prediction, congestion detection and path planning models, and output vehicle driving strategies based on real-time vehicle sensor data, including: Obtain real-time traffic data collected by vehicle sensors and perform preprocessing, including outlier detection and noise filtering, to obtain a normalized traffic data set; Inputting the normalized traffic data set into a traffic flow prediction model to predict the traffic flow change trend in the future; According to the traffic flow prediction results, a congestion detection model based on vehicle speed threshold and vehicle density threshold is used to determine the road congestion status; The congested road section information is passed to the path planning model, and the starting and ending positions of the vehicle are obtained. By constructing a road network topology map, the optimal travel path that avoids congested sections is searched.

9. The intelligent traffic decision-making method based on multi-source sensor data as claimed in claim 1, characterized in that: Also includes: Adaptive adjustments are made to different road scenarios and driving behavior patterns; Obtain road scene information and driving behavior data, use long short-term memory neural network to adaptively adjust the intelligent decision-making system, train decision models under different scenarios, and obtain a decision model library; Based on the newly acquired road scene information and driving behavior data, the incremental learning algorithm is used to optimize and update the models in the decision model library; According to the current road scene and driving behavior mode of the vehicle, the most matching model is selected from the decision model library; Combined with real-time traffic flow data, dynamic programming algorithm is used to adjust decision parameters and generate the optimal decision plan.

10. The intelligent traffic decision-making method based on multi-source sensor data as claimed in claim 9, characterized in that: The adaptive adjustment according to different road scenes and driving behavior modes also includes: Use rule-based reasoning engines to verify the rationality of decision-making solutions; Set a series of safety and reliability rules, use forward-chaining reasoning to determine whether the decision meets the requirements, and adjust the decision plans that do not meet the requirements; According to the vehicle's driving trajectory and decision execution, an evaluation function is used to evaluate and score the decision effect; By comparing and analyzing historical decision data, using heuristic search algorithms to find optimizable decision paths, using evaluation results as feedback signals, and iteratively optimizing decision models; Establish a driving behavior profile and road scene feature library, use clustering algorithms to classify driver behavior and scene features, form personalized decision-making preferences, and provide a reference basis for the decision-making model.

Images