Scenarized vehicle-mounted advertisement recommendation system for smart traffic

Abstract

The invention discloses a scenarized vehicle-mounted advertisement recommendation system for intelligent traffic, and relates to the technical field of intelligent traffic, and the system comprises a data collection module which collects a direction deviation angle, a speed matching degree and a path following rate in real time through a vehicle sensor and a navigation system, and obtains original behavior data; the feature extraction module is used for processing the original behavior data by adopting a time sequence analysis method, extracting dynamic features of steering frequency, pause time difference and signal acquisition frequency, and determining a behavior signal sequence; the correlation analysis module is used for calculating a correlation coefficient according to the matching degree of the behavior signal sequence and the navigation setting information on the aspects of time synchronism and environmental interference factors to obtain a correlation feature vector; according to the intelligent traffic-oriented scenarized vehicle-mounted advertisement recommendation system, through combination of multi-level behavior analysis and a recommendation algorithm, the pertinence of advertisement pushing and the user acceptability are remarkably improved, and an efficient advertisement putting effect is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation, in particular to a scenario-based vehicle-mounted advertisement recommendation system for intelligent transportation. BACKGROUND

[0002] With the rapid development of intelligent transportation and personalized services, analyzing driver travel intentions and destination types through vehicle behavior data has become an important research area for improving user experience and commercial value. This not only optimizes navigation and travel services, but also provides possibilities for precise advertisement pushing, thereby significantly improving commercial conversion efficiency.

[0003] However, existing methods often rely on single navigation data or historical records when analyzing driver intentions, ignoring the complexity and real-time nature of dynamic behavior signals, resulting in a lack of precision and foresight in recommended content. For example, existing systems may only recommend nearby restaurants based on navigation endpoints, but cannot determine whether the user actually has a dining demand, significantly reducing the effectiveness of the recommendations. The core challenge in this field is how to effectively integrate and interpret vehicle dynamic behavior data, especially the relevance between driving direction, speed changes, and navigation settings. Driving direction and speed changes can reflect real-time decisions made by drivers, such as slowing down near commercial areas, which may imply shopping or dining intentions. However, these signals are transient and difficult to accurately capture. Further, while navigation settings provide destination information, they often fail to reflect users' potential consumption motivations, such as setting a gas station as a destination, but actually just passing by without immediate refueling. Existing technologies lack time-series correlation analysis of behavior patterns when integrating these dynamic signals, resulting in an inability to accurately distinguish between users' short-term intentions and long-term habits. For example, a user frequently visits shopping centers on weekends, but on a certain trip, they changed plans and went to a nearby park, and the system still pushed shopping advertisements. This prediction bias stems from the inability to dynamically correlate changes in driving direction and navigation settings. SUMMARY

[0004] The purpose of the present application is to provide a scenario-based vehicle-mounted advertisement recommendation system for intelligent transportation, which accurately identifies drivers' travel intentions and consumption preferences by real-time analysis of the dynamic correlation between vehicle driving direction, speed changes, and navigation settings.

[0005] To achieve the above object, the application provides the following technical scheme: A scene-based vehicle advertising recommendation system for smart traffic, comprising: a data acquisition module that acquires direction deviation angle, speed matching degree and path compliance rate in real time through vehicle sensors and a navigation system to obtain original behavior data; a feature extraction module that processes the original behavior data using a time series analysis method to extract dynamic features of turning frequency, pause time difference and signal acquisition frequency, and determines a behavior signal sequence; a correlation analysis module that calculates a correlation coefficient according to the matching degree of the behavior signal sequence and navigation setting information in terms of time synchronization and environmental interference factors to obtain a correlation feature vector; an intention recognition module that classifies the behavior signal sequence according to direction deviation angle and speed matching degree through a clustering algorithm to determine an intention category if the correlation coefficient of the correlation feature vector is greater than a preset threshold; a pattern recognition module that analyzes behavior patterns using a decision tree algorithm by comparing the intention category and historical behavior data in terms of path compliance rate and turning frequency to determine habit features; a modeling module that generates a consumption tendency probability distribution according to habit features and intention categories, fuses user interest weights and real-time behavior preferences; a content recommendation module that generates a personalized advertising push list by matching advertising content relevance and regional targeting from a pre-established advertising library according to the consumption tendency probability distribution; a behavior update module that updates the behavior signal sequence by real-time monitoring of direction deviation angle and speed matching degree to obtain new dynamic features of turning frequency and pause time difference; and a push strategy adjustment module that adjusts the weights of real-time behavior preferences and interaction feedback factors in the consumption tendency probability distribution according to the new dynamic features to regenerate a personalized advertising push list containing advertising display frequency and push time sensitivity.

[0006] Preferably, the data acquisition module acquires direction deviation angle, speed matching degree and path compliance rate in real time through vehicle sensors and a navigation system to obtain original behavior data, which comprises: acquiring original driving behavior data by acquiring direction deviation angle, speed matching degree and path compliance rate in real time through vehicle sensors and a navigation system; performing smoothing processing on the direction deviation angle using a Kalman filtering algorithm if the direction deviation angle exceeds a preset threshold to obtain smoothed direction data; calculating path planning consistency according to the smoothed direction data and navigation system data to determine a path deviation vector; classifying driving behavior features using a support vector machine algorithm according to the path deviation vector and speed matching degree to determine a behavior abnormal state; calculating real-time control correction parameters according to speed adjustment amplitude and path compliance rate to obtain corrected driving instructions if the behavior abnormal state is true; updating navigation system data by the corrected driving instructions and real-time acquisition data to obtain optimized path planning; adjusting sensor acquisition parameters according to the optimized path planning and data acquisition frequency to obtain updated original driving behavior data.

[0007] Preferably, the feature extraction module adopts a time series analysis method to process the original behavior data, extracts dynamic features of steering frequency, pause time difference and signal collection frequency, and determines the behavior signal sequence including: extracting the steering frequency, pause time difference and signal collection frequency from the original driving behavior data through time series analysis to generate the behavior signal sequence; according to the behavior signal sequence, adopting a sliding time window method to divide the data segments, determining the feature change trend in each segment; if the feature change trend exceeds a preset threshold, using Fourier transform to analyze the periodicity of the steering frequency to obtain periodicity feature parameters; calculating the stability index of the driving behavior through the periodicity feature parameters and the pause time difference, generating a stability evaluation sequence; according to the stability evaluation sequence, using K-means clustering algorithm to classify the driving behavior, determining the abnormal behavior mode; if the abnormal behavior mode is true, adjusting the data sampling rate according to the signal collection frequency to obtain the optimized behavior signal sequence; updating the feature analysis method through the optimized behavior signal sequence and the time window division result, generating a new behavior signal sequence.

[0008] Preferably, the correlation analysis module calculates the correlation coefficient according to the matching degree of the behavior signal sequence and the navigation setting information in time synchronization and environmental interference factors to obtain the correlation feature vector including: aligning the time stamps of the behavior signal sequence and the navigation setting information, calculating the time synchronization deviation to obtain synchronization adjustment parameters; according to the synchronization adjustment parameters, correcting the time axis of the behavior signal sequence to generate a corrected signal sequence; using the corrected signal sequence and the navigation setting information, calculating the weighted influence of the environmental interference factor to determine the interference factor weight; calculating the weighted correlation coefficient through the interference factor weight and the corrected signal sequence to generate the correlation feature vector; if the dimension of the correlation feature vector exceeds a preset threshold, using principal component analysis algorithm to compress the feature vector to obtain a reduced dimension feature vector; according to the matching degree of the reduced dimension feature vector and the navigation setting information, calculating a dynamic matching score to determine the behavior signal classification result; updating the time synchronization adjustment strategy through the behavior signal classification result and the environmental interference factor weight to generate optimized synchronization parameters.

[0009] Preferably, the intention recognition module, if the correlation coefficient of the associated feature vector is greater than a preset threshold, classifies the behavior signal sequence according to the direction deviation angle and the speed matching degree by a clustering algorithm, and determines the intention category, comprising: classifying the behavior signal sequence by a K-means clustering algorithm to obtain the intention category; correcting the time axis of the behavior signal sequence using a time synchronization parameter according to the intention category to generate a corrected signal sequence; calculating the weighted value of the environmental interference factor using the corrected signal sequence and the navigation setting information to determine the interference factor weight; if the interference factor weight exceeds a preset threshold, performing dimensionality reduction processing on the corrected signal sequence using a principal component analysis algorithm to obtain a reduced dimension signal sequence; calculating the dynamic score of the direction deviation angle and the speed matching degree according to the reduced dimension signal sequence and the navigation setting information to obtain a matching feature vector; classifying the matching feature vector and the intention category using a support vector machine algorithm to determine the intention category of the behavior signal; and updating the time synchronization parameter according to the intention category and the interference factor weight to generate an optimized synchronization parameter.

[0010] Preferably, the pattern recognition module analyzes the behavior pattern using a decision tree algorithm by comparing the intention category and the historical behavior data in terms of path compliance rate and turning frequency to determine the habit feature, comprising: comparing the intention category and the historical behavior data in terms of path compliance rate and turning frequency, classifying the behavior pattern using a decision tree algorithm to obtain the habit feature; obtaining the behavior sequence feature in the time series data according to the habit feature, grouping the behavior sequence feature using a clustering algorithm to obtain a behavior pattern group; if the path compliance rate of the behavior pattern group is lower than a preset threshold, calculating the change trend of the turning frequency using the time series data to obtain a turning frequency change feature; performing dimensionality reduction processing on the behavior sequence feature using a feature extraction method according to the turning frequency change feature to obtain a reduced dimension behavior feature; if the correlation coefficient of the reduced dimension behavior feature and the habit feature is greater than a preset threshold, updating the behavior pattern group by the pattern classification result to obtain an optimized behavior pattern; calculating the dynamic score of the path compliance rate according to the optimized behavior pattern and the historical behavior data to obtain a path optimization feature; and reclassifying the behavior sequence feature using a decision tree algorithm according to the path optimization feature and the intention category to obtain the final habit feature.

[0011] Preferably, the modeling module, according to the habit feature and the intention category, fuses the user interest weight and the real-time behavior preference to generate the consumption tendency probability distribution comprises: through the habit feature and the intention category, a user preference vector is obtained by using a weighted fusion method to calculate the user interest weight; according to the user preference vector and the real-time behavior preference, an initial consumption tendency probability distribution is generated by using a probability distribution model; if the correlation coefficient of the initial consumption tendency probability distribution and the preset threshold is lower than a standard value, a behavior mode grouping is obtained by extracting the behavior mode from the behavior sequence data; according to the behavior mode grouping, a feature extraction method is used to reduce the dimension of the behavior sequence data to obtain a reduced dimension behavior feature; if the matching degree of the reduced dimension behavior feature and the user preference vector is higher than a preset threshold, an optimized tendency probability distribution is obtained by updating the consumption tendency probability distribution through the probability distribution model; according to the optimized tendency probability distribution and the consumption behavior feature, a dynamic score of the consumption tendency probability is calculated to obtain the consumption behavior feature; through the consumption behavior feature and the user preference vector, a decision tree algorithm is used to classify the behavior sequence data to obtain the final consumption tendency feature.

[0012] Preferably, the content recommendation module, using a recommendation algorithm according to the consumption tendency probability distribution, matches the advertisement content relevance and the regional targeting from a pre-established advertisement library to generate a personalized advertisement push list comprises: the user preference vector is obtained from the consumption tendency probability distribution, the user interest grouping is obtained by using a collaborative filtering algorithm to cluster the user preference; according to the user interest grouping, the advertisement content feature and the regional targeting feature are extracted from the preset advertisement library, the matching degree of the advertisement content and the user interest grouping is calculated by using the cosine similarity to obtain the advertisement matching score; if the advertisement matching score is higher than a preset threshold, the advertisement content with the highest score is selected from the advertisement matching score to generate a candidate advertisement set; according to the candidate advertisement set and the user behavior data, a logistic regression model is used to predict the click probability of the user to the advertisement to obtain an advertisement click prediction value; the advertisements higher than the preset threshold are selected from the advertisement click prediction value, and the regional targeting feature is combined to generate a preliminary push list; for the preliminary push list, an ordering algorithm is used to prioritize the advertisements to obtain an optimized push list; according to the optimized push list and the user real-time behavior data, the push order is dynamically adjusted to obtain the final personalized advertisement push list.

[0013] Preferably, the behavior updating module updates the behavior signal sequence by monitoring the direction deviation angle and the speed matching degree in real time, obtains the dynamic characteristics of the new steering frequency and the pause time difference, and includes: acquiring the direction deviation angle and the speed matching data through the sensor, processing the real-time collected data by using the sliding window method, and generating the behavior signal sequence; according to the behavior signal sequence, extracting the steering frequency characteristics by using the fast Fourier transform algorithm, obtaining the frequency variation trend; if the frequency variation trend exceeds the preset threshold, calculating the pause time difference by using the weighted average method, and generating the time difference distribution; according to the time difference distribution, dividing the steering behavior mode by using the K-means clustering algorithm, determining the behavior mode classification; if the behavior mode classification matches the preset mode, updating the dynamic characteristic vector by using the sequence comparison method, and obtaining the dynamic behavior signal; according to the dynamic behavior signal, predicting the frequency variation trend by using the linear regression model, and generating the behavior prediction value; according to the behavior prediction value, adjusting the priority of the dynamic characteristic vector by using the sorting algorithm, and obtaining the optimized behavior signal sequence.

[0014] Preferably, the push strategy adjusting module adjusts the weights of the real-time behavior preference and the interaction feedback factor in the consumption tendency probability distribution according to the new dynamic characteristics, regenerates the personalized advertisement push list containing the advertisement display frequency and the push time sensitivity, and includes: acquiring the user behavior data and the interaction record, processing by data cleaning, extracting the real-time behavior preference and the interaction feedback factor, and obtaining the user behavior feature set; according to the user behavior feature set, calculating the consumption tendency probability by using the weighted average algorithm, updating the dynamic probability distribution, and obtaining the probability distribution vector; if the change amplitude of the probability distribution vector exceeds the preset threshold, predicting the advertisement display frequency trend by using the linear regression model, and obtaining the frequency prediction value; according to the frequency prediction value, adjusting the weight adjustment strategy, updating the weight of the advertisement display frequency, and obtaining the optimized weight parameter; by using the optimized weight parameter, combining the push time sensitivity, generating the advertisement push sequence, and obtaining the preliminary push list; according to the preliminary push list, dividing the user preference group by using the K-means clustering algorithm, updating the personalized push list, and obtaining the final push list; by using the final push list, matching the user interaction record, adjusting the weight of the push time sensitivity, and obtaining the optimized push sequence.

[0015] From the above technical solutions, the present application has the following beneficial effects:

[0016] The scene-based vehicle advertising recommendation system for intelligent transportation generates accurate advertising push in real time by fusing the direction deviation angle, speed matching degree and path following rate collected by the vehicle sensor and navigation system, generating a behavior signal sequence, extracting dynamic features combined with time series analysis, calculating the correlation coefficient with the navigation settings, determining the short-term intention category and long-term habit features, and finally forming a consumption tendency probability distribution, matching the advertising content and generating a personalized push list. The application optimizes the advertising display frequency and push time sensitivity by real-time monitoring of dynamic features and dynamically adjusting the behavior preference weight, solving the problem of lack of accuracy and real-time of traditional advertising push. The application significantly improves the targeting and user acceptance of advertising push by combining multi-level behavior analysis and recommendation algorithm, achieving efficient advertising delivery effect. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The system connection diagram of the application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0019] As Figure 1As shown, the present application provides a technical solution: a scene-based vehicle advertising recommendation system for intelligent transportation, comprising: a data acquisition module, which acquires direction deviation angle, speed matching degree and path following rate in real time through vehicle sensors and navigation systems to obtain original behavior data; a feature extraction module, which processes the original behavior data using time series analysis method to extract dynamic features of turning frequency, pause time difference and signal acquisition frequency, and determines behavior signal sequence; a correlation analysis module, which calculates correlation coefficient according to the matching degree of behavior signal sequence and navigation setting information in time synchronization and environmental interference factor to obtain correlation feature vector; an intention recognition module, which classifies behavior signal sequence according to direction deviation angle and speed matching degree through clustering algorithm if the correlation coefficient of correlation feature vector is greater than a preset threshold to determine intention category; a pattern recognition module, which compares intention category and historical behavior data in path following rate and turning frequency, analyzes behavior pattern using decision tree algorithm, and judges habit features; a modeling module, which generates consumption tendency probability distribution according to habit features and intention category, fuses user interest weight and real-time behavior preference; a content recommendation module, which generates personalized advertisement push list by matching advertisement content relevance and regional pertinence from pre-established advertisement library according to consumption tendency probability distribution; a behavior update module, which updates behavior signal sequence by real-time monitoring of direction deviation angle and speed matching degree to obtain new dynamic features of turning frequency and pause time difference; a push strategy adjustment module, which adjusts the weight of real-time behavior preference and interaction feedback factor in consumption tendency probability distribution according to new dynamic features to regenerate personalized advertisement push list containing advertisement display frequency and push time sensitivity.

[0020] The embodiment acquires dynamic driving behavior data such as direction deviation angle, speed matching degree and path compliance rate by using vehicle-mounted sensors and navigation systems through a data acquisition module, forming a basic original data set. The feature extraction module processes the original data through time series analysis method, extracts high-value features such as steering frequency, pause time difference and signal acquisition frequency related to driving behavior, and constitutes a behavior signal sequence. The correlation analysis module further combines navigation setting information to evaluate the correlation degree of behavior signals, environmental factors and time synchronization, and quantifies it as a correlation coefficient to generate a correlation feature vector. If the correlation is higher than the preset threshold, the intention recognition module divides the current intention category of the driver by clustering algorithm based on direction deviation angle and speed matching degree. Subsequently, the pattern recognition module compares the intention category with historical behavior data by using a decision tree model, analyzes driving behavior rules in two dimensions of path compliance rate and steering frequency, and extracts individualized habit features. The modeling module combines the habit features and intention categories with user interest weights and real-time behavior preference data to calculate and output a consumption tendency probability distribution. The content recommendation module selects the most relevant advertising content from the advertising database based on the distribution results through a matching algorithm, and optimizes the display strategy by considering geographic location information to generate a personalized advertising push list. The system also includes a behavior updating module to update the direction deviation angle and speed matching degree in real time, and correct the behavior signal sequence and the dynamic features derived therefrom. The push strategy adjustment module dynamically adjusts the weights of real-time behavior preference and interaction feedback factor in consumption tendency according to the new dynamic features, so as to timely adjust the advertising push strategy, including push time window and frequency control, to realize dynamic and accurate advertising delivery.

[0021] The above technical solution provides a vehicle-mounted advertising recommendation system for a smart traffic environment, which realizes accurate recognition and intention judgment of driving behavior through real-time data acquisition and dynamic feature analysis. The system not only integrates user historical habits and interest preferences, but also adjusts the recommendation strategy in real time according to the changes in driving behavior, ensuring the personalization, regionalization and timeliness of advertising content, thereby significantly improving the advertising click rate and user experience. At the same time, the use of clustering algorithm, decision tree and probability modeling method helps to improve the recognition accuracy of driving behavior intention and enhance the adaptability and response speed of the push strategy. The scheme has strong practical value and market prospect, and is especially suitable for advertising push scenarios in online car-hailing, shared travel and Internet of Vehicles platforms.

[0022] The data acquisition module acquires the direction deviation angle, the speed matching degree and the path following rate in real time through the vehicle sensor and the navigation system to obtain the original behavior data, including: acquiring the original driving behavior data by acquiring the direction deviation angle, the speed matching degree and the path following rate in real time through the vehicle sensor and the navigation system; if the direction deviation angle exceeds a preset threshold, performing smoothing processing on the direction deviation angle by using a Kalman filtering algorithm to obtain smoothed direction data; calculating the path planning consistency according to the smoothed direction data and the navigation system data, and determining a path deviation vector; classifying the driving behavior characteristics by using a support vector machine algorithm through the path deviation vector and the speed matching degree, and judging the behavior abnormal state; if the behavior abnormal state is true, calculating the real-time control correction parameter according to the speed adjustment amplitude and the path following rate, and obtaining the corrected driving instruction; updating the navigation system data through the corrected driving instruction and the real-time acquisition data, and obtaining the optimized path planning; adjusting the sensor acquisition parameter according to the optimized path planning and the data acquisition frequency, and obtaining the updated original driving behavior data.

[0023] In the embodiment, the data acquisition module first performs joint real-time data acquisition through the vehicle-integrated direction sensor (such as an electronic compass or a vehicle-mounted gyroscope), the speed sensor (such as a wheel speed meter) and the vehicle-mounted navigation system. The acquisition process of the direction deviation angle is as follows: whenever the vehicle moves more than 1 meter, the navigation system calculates the expected driving direction corresponding to the current position in real time, which is determined by the angle of the line formed by the current position to the next navigation path point; at the same time, the actual driving direction of the vehicle is measured by the vehicle heading angle. The system takes the angle difference between the actual driving direction and the navigation expected direction as the direction deviation angle, which is expressed in "degrees" and has a value range of 0 degrees to 180 degrees. The acquisition process of the speed matching degree is as follows: the navigation system sets the target speed according to the current road conditions and traffic state, and the real-time speed of the vehicle is measured by the speed sensor. The difference between the two is the speed deviation, and the speed matching degree is defined as the ratio of the absolute value of the speed deviation to the target speed, which is a decimal between 0 and 1. The smaller the value, the higher the matching degree. The acquisition method of the path following rate is as follows: the system judges whether the current position of the vehicle is within the navigation path bandwidth tolerance range at each position sampling point. If yes, it is judged as "following", otherwise it is judged as "deviating". The ratio of the number of "followings" to the total number of samplings within a certain period of time (such as 60 seconds) is defined as the path following rate, which is expressed in percentage.

[0024] After the initial data collection is completed, if the direction deviation angle exceeds the preset threshold of 5 degrees, the system will enable the Kalman filter algorithm to smooth the angle data. The Kalman filter process includes the following steps: first, predict the current direction value using the previous vehicle direction state and its change speed; then, read the current sensor direction value, and perform a weighted average of the predicted value and the sensor observation value, with the weight calculated from the ratio of the prediction error covariance and the measurement error covariance; finally, obtain a corrected direction data closer to the true value. The system continues to sample and filter updates at 1 second intervals to ensure stable and continuous direction data.

[0025] Subsequently, the system calculates the difference between the smoothed direction data and the navigation path direction at each time point, sums the absolute values of all time point differences, and divides by the number of sampling points to obtain the average direction difference. Then, combine the average direction difference with the minimum perpendicular distance of the vehicle position on the path segment to obtain the path planning consistency index. This index reflects the comprehensive deviation of direction and position. Combined with the spatial vector direction and length information between the vehicle's current and target positions, the path deviation vector is calculated, which represents the offset direction and distance of the current position relative to the ideal point of the navigation path.

[0026] Next, the system inputs the path deviation vector and the speed matching degree as features into the trained support vector machine classification model. The model has used hundreds of sets of driving behavior data to label the boundaries of normal and abnormal behaviors during the training phase. During the classification process, the system determines the class of the current input features based on their position in the trained model. If they fall within the preset abnormal region, they are identified as "abnormal behavior state" true.

[0027] The system obtains the actual driving trajectory of the vehicle based on the data collected by the vehicle sensors and the path planning results generated by the navigation system, and calculates the deviation between the trajectory and the recommended path in real time to form a path deviation vector. The path deviation vector is a vector that describes the spatial position difference between the current position of the vehicle and the ideal navigation path. Its value represents the direction and magnitude of the vehicle's deviation from the path. The vector is calculated by the difference between the vehicle coordinates at each time and the corresponding navigation path coordinates, and the unit is usually meters. A large deviation means that the vehicle deviates from the planned route, which may indicate abnormal driving behavior. Next, the system calculates the speed matching degree based on the matching degree between the vehicle speed and the recommended speed. The speed matching degree is defined as the ratio of the current speed of the vehicle to the recommended speed. If the ratio is between 0.8 and 1.2, it is considered to be a good match. The system calculates the average speed matching degree and path deviation value in each sliding window (window size is 5 seconds, step is 1 second) to form a feature sample. Next, the system combines the "path deviation vector module value" and "average speed matching degree" in each sliding window to form a two-dimensional feature vector, which is used to train a support vector machine model. The training data of the model comes from historical labeled samples, which contain known normal driving behavior and abnormal behavior (such as frequent deviation, sudden deceleration, etc.). The support vector machine uses a radial basis function as a kernel function and optimizes the penalty factor and kernel parameter through cross-validation. After training, the model can classify new input samples. When the real-time collected feature vector is input into the support vector machine model, the model will classify it as "normal" or "abnormal" category. If the result is "abnormal", the system judges that the behavior is in an abnormal state, and records the time point, path segment and deviation degree of the behavior for subsequent correction of the control strategy.

[0028] When the behavior is abnormal, the system will perform a control correction parameter calculation process. First, calculate the speed adjustment amplitude, which is the difference between the current speed of the vehicle and the target speed, with the unit being kilometers per hour. If the value exceeds 5 kilometers per hour, it is considered that the speed deviation is too large. Once the path compliance rate is less than 80%, it is also considered abnormal. The system takes the speed adjustment amplitude and path compliance rate as input parameters, with weights of 60% and 40% respectively, and calculates the control correction parameter value by weighted average, with the unit being a numerical scalar, which is used to adjust the direction and speed. If the value is greater than 10, the system generates driving instructions including "sharp turn" and "deceleration" instructions, with the adjustment angle and target speed set to 15 degrees less than the current direction and 90% of the target speed, respectively. If the value is less than or equal to 10, the adjustment angle is set to 5 degrees less, and the speed remains unchanged or decreases by 5%.

[0029] The system then merges the corrected driving instructions with the current real-time collected data to re-plan the path through the navigation system. This step includes calling the path search function in the map data with the current vehicle coordinates and the target endpoint coordinates, generating the shortest path or the least time path, and updating the current path point sequence. To adapt to the new navigation path, the system dynamically adjusts the sensor collection parameters according to the path complexity and the vehicle state. If the path contains multiple intersections or curves, the system sets the sampling frequency to 10 times per second; if the path is a straight road and the traffic flow is smooth, the sampling frequency is reduced to 3 times per second to balance resource consumption and collection accuracy.

[0030] Finally, the system re-executes the data collection process with new sampling parameters to form updated raw driving behavior data. These data will be re-sent to the above process for analysis and processing to form a data-driven closed-loop control and optimization process, achieving high-precision, high-response driving behavior detection and advertisement pushing basic data support. All threshold parameters, including the 5-degree direction deviation angle, 0.2 speed matching degree, 80% path compliance rate, 5 kilometers per hour speed adjustment amplitude, and 10 control correction parameter threshold, are based on statistical test data and field road test results, and are set comprehensively, with clearness and reasonableness, supporting the stable operation of the algorithm under various driving conditions.

[0031] The feature extraction module uses time series analysis method to process the original behavior data, extracts the dynamic characteristics of steering frequency, pause time difference and signal collection frequency, and determines the behavior signal sequence including: extracting the steering frequency, pause time difference and signal collection frequency from the original driving behavior data through time series analysis to generate the behavior signal sequence; according to the behavior signal sequence, using the sliding time window method to divide the data segment, determining the feature change trend in each segment; if the feature change trend exceeds the preset threshold, using Fourier transform to analyze the periodicity of the steering frequency to obtain the periodicity characteristic parameter; calculating the stability index of the driving behavior through the periodicity characteristic parameter and the pause time difference to generate the stability evaluation sequence; according to the stability evaluation sequence, using the K-means clustering algorithm to classify the driving behavior to determine the abnormal behavior mode; if the abnormal behavior mode is true, adjusting the data sampling rate according to the signal collection frequency to obtain the optimized behavior signal sequence; updating the feature analysis method through the optimized behavior signal sequence and the time window division result to generate a new behavior signal sequence.

[0032] In this embodiment, the feature extraction module receives raw driving behavior data including direction angle, speed state and sensor sampling timestamp from the data collection module. The system first organizes the direction angle data in time sequence and counts the number of direction angle changes within 1 second. When the angle change is more than 10 degrees and the continuous time is not more than 3 seconds, the system determines it as an effective steering. The system accumulates the number of effective steering within a fixed window of 30 seconds, and the steering frequency is the number divided by 30 and multiplied by 60 to get the steering frequency per minute. The 10-degree angle change threshold is obtained by statistical analysis of road test data, which can filter out minor operation errors and accurately capture the obvious steering behavior of the driver.

[0033] At the same time, the system analyzes the speed data of the vehicle every second, identifies the event of speed being 0 and lasting more than 3 seconds as a parking behavior, records the difference between the parking start time and the start time as a parking time, and the unit is second. The system calculates the average parking time difference in the current window after summarizing a plurality of parking times. 3 seconds is a threshold value, which is based on the average waiting time of city road traffic lights, and is used to distinguish between short parking and normal deceleration.

[0034] For signal collection frequency, the system records the number of valid data packets received by the sensor in each second. If the sampling frequency is less than 3 times per second, it is considered to be too low; if it is more than 10 times, it is automatically set to high-frequency sampling mode. The interval of 3 to 10 times per second is determined based on the ability boundary of vehicle embedded hardware and processor performance, which can balance data integrity and system real-time performance.

[0035] The system then constructs a set of three-dimensional vectors from the three types of feature values collected every second, forming a behavior signal sequence in time sequence. To identify the trend of data change, the system uses a sliding time window mechanism to analyze the sequence. The time window is set to 30 seconds, and the step size of each sliding is 5 seconds, that is, an analysis unit containing the previous 30 seconds of data is generated every 5 seconds. In each time window, the system calculates the average and standard deviation of the steering frequency, the parking time difference and the signal collection frequency, and compares them with the average of the previous window. If the difference exceeds the corresponding threshold, it is determined that there is a significant trend change. The threshold set here is: the steering frequency difference exceeds 2 times per minute, the parking time difference changes more than 10 seconds, and the signal collection frequency changes more than 5 times per second, which are based on the statistical behavior fluctuation extreme value in the experimental scene to ensure sensitive identification of driving behavior state changes.

[0036] If any of the feature trends exceeds the threshold, the system initiates a Fourier transform analysis process. In this process, the system extracts a 30-second sequence of steering frequency within the current window, taking each second of sampling points as a value to form a 30-point time sequence. Then the sequence is converted into a frequency domain sequence using the fast Fourier transform algorithm, and its frequency spectrum is analyzed to identify the dominant frequency, i.e., the frequency point with the largest amplitude, in hertz, and the amplitude of this frequency point is extracted to represent the periodicity strength. The frequency value represents the periodicity strength of the steering behavior, and the larger the amplitude, the more significant the frequency component, indicating that the driver has a fixed rhythm operation mode.

[0037] The periodicity feature parameter includes the dominant frequency value and the corresponding amplitude, and the system performs a weighted average of this parameter with the average pause time difference in the current window to form a stability index. The calculation method of this index is to multiply the dominant frequency amplitude by 0.6, add the pause time difference multiplied by 0.4, and then sum. The weights 0.6 and 0.4 come from the fitting analysis results based on expert scoring of driving behavior stability, ensuring that the evaluation results conform to subjective driving feelings.

[0038] The value of the stability index is calculated sequentially on all time windows to form a stability evaluation sequence. The system inputs this sequence into the K-means clustering algorithm, with the number of clusters set to 3, corresponding to the categories: stable driving, mild unstable driving, and severe abnormal driving. The K-means clustering process includes: first, initializing 3 cluster centers, selecting the stability values in the low, medium, and high regions of the sequence value range as the initial centers; then assigning each window data to the class with the closest center value; calculating the new center value of each class and updating it; repeating this cycle until the class assignment of all windows no longer changes, and the clustering process converges. The system determines whether the current window is classified as "severe abnormal driving", if so, it is determined that the abnormal behavior pattern is established.

[0039] Once the system detects an abnormal behavior pattern, it immediately retrieves the original signal acquisition frequency value corresponding to the window. If this value is less than 3 times per second, the system increases it to 10 times per second and updates the sensor configuration parameter table so that higher frequency data is obtained in the subsequent sampling process. After updating the data sampling frequency, the system will re-collect 30 seconds of behavior data, generate a new behavior signal sequence at the new frequency, and re-perform the above sliding window analysis, trend judgment, Fourier transform, stability calculation, and clustering identification processes.

[0040] The correlation analysis module calculates a correlation coefficient according to a matching degree of the behavior signal sequence and the navigation setting information in time synchronization and environmental interference factors, and obtains a correlation feature vector, including: calculating a time synchronization deviation by aligning time stamps of the behavior signal sequence and the navigation setting information, and obtaining a synchronization adjustment parameter; correcting a time axis of the behavior signal sequence according to the synchronization adjustment parameter, and generating a corrected signal sequence; calculating a weighted influence of the environmental interference factor by using the corrected signal sequence and the navigation setting information, and determining an interference factor weight; calculating a weighted correlation coefficient by using the interference factor weight and the corrected signal sequence, and generating a correlation feature vector; if a dimension of the correlation feature vector exceeds a preset threshold, compressing the feature vector by using a principal component analysis algorithm to obtain a reduced dimension feature vector; calculating a dynamic matching score according to a matching degree of the reduced dimension feature vector and the navigation setting information, and determining a behavior signal classification result; updating a time synchronization adjustment strategy by using the behavior signal classification result and the environmental interference factor weight, and generating an optimized synchronization parameter.

[0041] In the embodiment, the first step of the correlation analysis module is to perform a time synchronization alignment operation of the behavior signal sequence and the navigation setting information. Each data point in the behavior signal sequence is attached with a time stamp field, for example, "August 5, 2025, 14:35:12". Each record in the navigation setting information also contains a navigation operation trigger time, for example, the trigger time of "system recommends left turn" is "August 5, 2025, 14:35:10". The system aligns the two time sequences in the following way: sequentially traversing each instruction event in the navigation information, finding all signal points in the behavior signal sequence within 6 seconds before and after the instruction, and calculating the time difference between them and the instruction time point, taking the average of the absolute values of all time differences as the synchronization deviation of the event. Taking the average of the synchronization deviations of all events, the average time offset between all navigation events and behavior signals is obtained, which is in seconds. This value is the "time synchronization deviation". When the value is greater than 2 seconds, it is considered that the behavior response is delayed too much, and the time axis needs to be corrected. The 2-second threshold is obtained according to the average reaction time of the driver to the navigation voice in the real vehicle test, which can reflect the effective response and avoid misjudgment.

[0042] After obtaining the synchronization deviation, the system generates a "synchronization adjustment parameter". Its value is equal to the average of the above deviation, and the direction is determined by the deviation sign. If the behavior time lags behind the navigation time, all behavior signal time stamps need to be shifted forward by the synchronization adjustment parameter seconds; if the behavior time is earlier than the navigation time, the time stamp is shifted backward by the synchronization adjustment parameter seconds. The adjusted signal is the "corrected signal sequence", which has the same structure as the original signal sequence, only the time stamp field is uniformly transformed.

[0043] After the time axis correction is completed, the system enters the environmental interference factor calculation process. Based on the current path location in the navigation settings, the system queries the built-in geographic information database to obtain the values of four types of environmental interference factors for this section: the first is traffic density, the value comes from the regional traffic flow statistical database, the range is set to 0 to 1, 0 represents an empty road, and 1 represents severe congestion; the second is the weather influence index, which is input by the weather module, 0 for sunny and 1 for heavy rain; the third is the road condition, which is divided according to the road surface grade, with muddy or bumpy sections being 1 and flat road surfaces being 0; the fourth is the lighting condition, which is obtained according to the all-day light intensity monitoring data, with night being 1 and sunny day being 0. The system assigns weights to these four types of factors, with traffic density weight set to 0.4, weather influence to 0.3, road condition to 0.2, and lighting condition to 0.1. This distribution is based on regression analysis data statistics of driving behavior changes under different environments and is set to ensure that the calculation results are scientific and effective.

[0044] Each type of environmental factor value is multiplied by the corresponding weight and summed to obtain the total interference factor weight, with the value range fixed between 0 and 1. The system calls this weight "comprehensive interference factor weight". This weight is used as a multiplier in the behavior value of each data point in the corrected behavior signal sequence to obtain the "modified behavior value", which is used to participate in subsequent correlation analysis.

[0045] Then, the system calculates the correlation coefficient between the modified behavior value and the target value in the navigation setting information. The system corresponds the behavior value and the navigation value at each time to each other, calculates their mean and variance, and then calculates the linear correlation coefficient of the entire sequence through the Pearson correlation algorithm. This coefficient ranges from -1 to 1, with a value closer to 1 indicating that the behavior completely follows the navigation setting changes, and a value closer to 0 indicating a weak relationship between the two. If it is negative, it indicates that the changes are in the opposite direction. The system calculates a correlation coefficient for each feature dimension (such as speed, direction, offset, etc.), and combines all the coefficients in order to form a "correlation feature vector". If the behavior data dimension is 32 feature points and the navigation setting is also 32-dimensional, the correlation feature vector is 32-dimensional.

[0046] The system sets a maximum dimension threshold of 128 dimensions, and if the dimension of the correlation feature vector exceeds 128, it enables the principal component analysis algorithm for feature compression. The compression process consists of three steps: first, the system constructs a covariance matrix for the vector; second, the covariance matrix is subjected to eigenvalue decomposition; third, the first several principal components whose cumulative contribution rate reaches 95% are retained, usually the result dimension is compressed to between 20 and 40, and the compressed output is called "dimension-reduced feature vector". This compression can retain most of the information while significantly reducing data processing overhead and improving real-time performance.

[0047] The reduced dimension feature vector is then used for the "dynamic matching score" calculation. The system compares each value in the reduced dimension vector with the target value of the corresponding dimension in the navigation information, calculates the difference of each dimension and takes the absolute value, averages all the dimension differences, normalizes them to 0 to 1, and finally multiplies by 100 to get the matching score, with a full score of 100. If the score is greater than 85, it is considered to be a high match; if the score is between 60 and 85, it is considered to be a moderate match; if the score is less than 60, it is considered to be a low match. The score result classifies the current behavior signal as "good response", "delayed response" or "abnormal response".

[0048] In the last step, the system determines whether to modify the time synchronization adjustment strategy according to the behavior signal classification result and the current environmental interference factor weight. If the current classification result is "delayed response" or "abnormal response", the system reduces the maximum acceptable synchronization deviation from the original setting of 2 seconds to 1 second, and at the same time, refines the time axis adjustment step from the original 1 second to 0.5 second, and updates all behavior signal time axes in the next cycle, using the optimized synchronization parameters to realign the navigation setting time.

[0049] The intention recognition module classifies the behavior signal sequence according to the direction deviation angle and speed matching degree by clustering algorithm if the correlation coefficient of the associated feature vector is greater than the preset threshold, determines the intention category including: classifying the behavior signal sequence by K-means clustering algorithm to obtain the intention category; correcting the time axis of the behavior signal sequence using the time synchronization parameter according to the intention category to generate a corrected signal sequence; calculating the weighted value of the environmental interference factor according to the corrected signal sequence and the navigation setting information to determine the interference factor weight; if the interference factor weight exceeds the preset threshold, performing dimension reduction processing on the corrected signal sequence using principal component analysis algorithm to obtain a reduced dimension signal sequence; calculating the dynamic score of the direction deviation angle and the speed matching degree according to the reduced dimension signal sequence and the navigation setting information to obtain a matching feature vector; classifying the intention category of the behavior signal by support vector machine algorithm through the matching feature vector and the intention category; updating the time synchronization parameter according to the intention category and the interference factor weight to generate an optimized synchronization parameter.

[0050] In this embodiment, the intention recognition module first analyzes the correlation feature vector output by the correlation analysis module to determine whether the correlation coefficient is higher than the set decision threshold. The correlation coefficient uses the Pearson correlation coefficient calculation method, and the coefficient value ranges from negative 1 to positive 1. The system sets the threshold value to 0.65, which means that when the average correlation between the behavior signal and the navigation setting in multiple dimensions is greater than 0.65, it is considered that the behavior data has obvious synchronization and responsiveness with the navigation instruction in time, path and direction, and has the logical premise for intention clustering recognition. The 0.65 threshold value is determined by linear fitting analysis of 1000 groups of real driving behavior and navigation instruction data to ensure that the selected data has statistical behavior responsiveness.

[0051] After confirming that the correlation coefficient is higher than the set threshold, the system extracts two key features from the behavior signal sequence: the direction deviation angle and the speed matching degree. The direction deviation angle refers to the difference between the actual driving direction of the vehicle and the specified direction in the navigation instruction, and its calculation method is to perform angle difference operation on the current vehicle heading angle and the target direction of the navigation path, with the unit being degrees. The system sets that if the angle difference is greater than 15 degrees, it is considered to be significantly deviated. The speed matching degree refers to the ratio of the current vehicle speed to the recommended speed set by the navigation, and its value ranges from 0 to 1. The closer to 1, the better the speed matching is. The matching degree is equal to the vehicle speed divided by the navigation recommended speed, accurate to two decimal places.

[0052] The system constructs a two-dimensional vector containing the above two features for each time point in a 30-second time window. The system uses the K-means clustering algorithm to classify these data points. The K value is pre-set to 3, corresponding to three intention categories: intention to maintain (the vehicle basically operates according to the navigation path), intention to deviate (the vehicle deviates for a short time but still maintains the general direction), and intention to change (the vehicle deviates significantly from the navigation route). The implementation process of the K-means clustering algorithm is as follows: three centroid points are selected as initial class centers, and then the Euclidean distance between each sample point and the three centroids is calculated, and the sample point is classified into the nearest center point category. After classification, the system recalculates the average value of each category as a new centroid, and repeats the above process until all centroids are stable or the maximum iteration number reaches 100 times. Finally, each time point is assigned a specific intention category label.

[0053] After the intention clustering is completed, the system extracts the current dominant intention category from the clustering results, i.e., the category with the most occurrences, and calls the time synchronization parameters generated by the previous module based on the intention category to perform time axis correction. The correction rule is: if the dominant intention is "intention to keep", the time axis is not changed; if it is "intention to shift", all behavior data timestamps are adjusted by 1 second forward; if it is "intention to change", the timestamps are adjusted by 2 seconds forward. The time adjustment operation is to traverse each behavior signal sequence, and uniformly increase or decrease the timestamp field of each record by seconds. The generated result is the "corrected signal sequence".

[0054] Subsequently, the system reacquires the current time and location information, and calls the interference factor values stored in the environment data module, including traffic density, weather conditions, road conditions, and light levels. The traffic density is determined by the real-time flow statistics of the road, with a value range of 0 to 1, and 1 representing extreme congestion. The weather conditions are determined according to the weather data push, with 0 representing sunny days and 1 representing heavy rain. The road conditions are determined by querying the road database for the road surface level, and the light level is calculated based on the system light sensor value. The system assigns weights to the above four factors: traffic density is 0.4, weather is 0.3, road is 0.2, and light is 0.1. The sum of the product of each factor value and the corresponding weight is the environment interference factor weight. If the value is greater than 0.6, it means that the current environmental interference is significant, and the system will perform dimension reduction processing.

[0055] The dimension reduction uses the principal component analysis algorithm. The system first calculates the covariance matrix of the direction deviation angle and speed matching degree feature matrix in the corrected signal sequence; then performs eigenvalue decomposition on the covariance matrix; extracts the feature vectors sorted in descending order of eigenvalues; accumulates the cumulative contribution rate from large to small until the contribution rate is greater than 95%, and combines the corresponding first few principal components to form the dimension reduction signal sequence. This process usually compresses the original two-dimensional features to one dimension or retains two dimensions, reducing the computational complexity while retaining the main information.

[0056] The system continues to compare the dimension-reduced signal sequence with the target direction angle and recommended speed at the corresponding time in the navigation setting information, and calculates the matching situation for each record. The direction matching score calculation method is: the absolute value of the angle difference between the actual direction and the target direction in each record is normalized after comparing with the maximum direction difference of 30 degrees, and multiplied by 100 to obtain the score. The speed matching score is the absolute value of the difference between the actual speed of the vehicle and the recommended speed divided by the recommended speed, normalized and multiplied by 100 to obtain the speed score at each time point. The average of the direction score and the speed score at all time points is taken to form the final matching feature vector, which contains two components: direction score and speed score.

[0057] The system inputs the matching feature vector into a pre-trained support vector machine classifier. The classifier is trained based on a labeled sample set, uses a radial basis kernel function to map to a high-dimensional space, and divides the input sample into one of the three categories of "intention to maintain", "intention to deviate" or "intention to change" by the decision hyperplane obtained by training. The classification result is the final "intention category" identified by the system.

[0058] Finally, the system determines whether to update the time synchronization parameter according to the intention category and the current environmental interference factor weight. If the intention category is "intention to change" and the interference factor weight is greater than 0.6, the system reduces the original maximum time synchronization deviation tolerance value from 2 seconds to 1 second, and reduces the time correction step from 1 second to 0.5 second, generating "optimized synchronization parameters". The parameters are used to correct the timestamp of the behavior signal in the next cycle to enhance the behavior and navigation synchronization accuracy, improve the accuracy of subsequent intention recognition and the overall response sensitivity of the system.

[0059] The mode recognition module analyzes the behavior pattern by comparing the intention category and the historical behavior data in path compliance rate and turning frequency using a decision tree algorithm to determine the habit feature, including: comparing the intention category and the historical behavior data in path compliance rate and turning frequency, using a decision tree algorithm to classify the behavior pattern, and obtaining the habit feature; according to the habit feature, obtaining the behavior sequence feature in the time series data, using a clustering algorithm to group the behavior sequence feature, and obtaining the behavior pattern grouping; if the path compliance rate of the behavior pattern grouping is lower than a preset threshold, calculating the trend of the turning frequency by the time series data, and obtaining the turning frequency change feature; according to the turning frequency change feature, using a feature extraction method to reduce the dimension of the behavior sequence feature, and obtaining the reduced dimension behavior feature; if the correlation coefficient of the reduced dimension behavior feature and the habit feature is greater than a preset threshold, updating the behavior pattern grouping by the mode classification result, and obtaining the optimized behavior pattern; according to the optimized behavior pattern and the historical behavior data, calculating the dynamic score of the path compliance rate, and obtaining the path optimization feature; using a decision tree algorithm to reclassify the behavior sequence feature by the path optimization feature and the intention category, and obtaining the final habit feature.

[0060] In this embodiment, after receiving the "intention category" in the previous step, the pattern recognition module first calls the "historical behavior data" saved in the local database, extracts the historical behavior samples with the same intention category as the current one, and extracts two key indicators: path following rate and turning frequency. The calculation method of path following rate is as follows: in each 30-second analysis window, record the spatial distance between the vehicle trajectory point and the navigation path, if the distance is less than 10 meters, it is considered as a "path inside" point; the number of path inside points divided by the total number of trajectory points, that is, the path following rate of the window, the result is expressed in percentage. The tolerance value of 10 meters is set based on the average lane width of urban roads. The calculation method of turning frequency is as follows: in the same 30-second window, count the number of events with a change in direction angle greater than 10 degrees, divide the number by 30 seconds and multiply by 60, the unit is times per minute. 10 degrees is the angle threshold of effective turning, which comes from the average angle change standard of turning behavior in real vehicle test.

[0061] The system compares the current behavior data with the path following rate and turning frequency of the corresponding intention category in the historical sample, forms a training data set containing current features and historical features, and inputs it into the decision tree classification model. The model is constructed using the information gain algorithm, the goal is to maximize the purity after partitioning, and each split node of the tree is determined by the path following rate or the turning frequency, and the leaf node outputs the habit feature category, such as "prefer stable" and "easy to change lane". After the system performs reasoning on the current data, it outputs the preliminary "habit feature" of this period.

[0062] After completing the preliminary habit feature determination, the system further extracts the original high-frequency behavior sequence data within the current time window, called "behavior sequence feature". This feature includes 10 dimensions such as direction angle, direction angle speed, vehicle speed, vehicle acceleration, path offset, speed offset, in-lane position, navigation deviation, turning trigger flag, and stop time, recorded 10 times per second, with a total of 300 sample points in 30 seconds. The system inputs these 300 10-dimensional samples into the clustering module and performs K-means clustering algorithm, with the initial K value set to 4, and the system automatically evaluates the optimal K value through the silhouette coefficient analysis. The clustering result forms multiple "behavior pattern groups", each group containing several continuous or non-continuous behavior segments.

[0063] For each behavior pattern group, the system calculates its path compliance rate separately, in the same way as described above. If the average path compliance rate of a group is lower than 80%, it is a low path consistency group. The 80% threshold is set based on the median of the city driving trajectory coincidence rate, which can effectively separate deviant behaviors. For all the turning frequency data in this group, arrange them in chronological order, and divide them into 6 time segments, each of which is 5 seconds long. Perform linear regression on the turning frequency of each segment, and take the regression slope as the trend value. If the trend slope is greater than 0.1 or less than negative 0.1, it means that there is a significant upward or downward trend in the turning frequency, and this value is the "turning frequency change feature". The slope threshold of 0.1 is derived from the statistical standard deviation of 300 typical turning behavior trend changes.

[0064] After identifying that the turning frequency has a trend of change, the system performs principal component analysis dimension reduction operation on all original 10-dimensional features in this behavior pattern group. The system first constructs a 10-dimensional feature covariance matrix, performs eigenvalue decomposition on it, and selects the first several eigenvectors whose cumulative contribution rate exceeds 95% as principal components, usually 2 to 3 principal components can be retained, to obtain the "dimension-reduced behavior features". This dimension-reduced feature is used to further judge the consistency with the historical habits.

[0065] The system performs correlation analysis on the dimension-reduced behavior features and the "habit features" determined initially. The specific operation is: calculate the Pearson correlation coefficient between each principal component vector and the reference feature vector formed by the habit feature code. First, the system has reduced the original high-dimensional behavior sequence features (a total of 10 dimensions) to 2 to 3 principal components through principal component analysis PCA method. Each principal component vector is a one-dimensional numerical sequence composed of multiple time points, for example, principal component 1 contains 300 numerical values, corresponding to the number of samples generated by sampling 10 times per second within a 30-second window.

[0066] At the same time, the system converts the "habit features" obtained by decision tree analysis into reference feature vectors. This reference vector is generated by the system based on the statistical mean of historical habit behaviors in the dimension-reduced feature space, that is, for a large number of historical behavior data belonging to a certain habit feature label, the dimension-reduced principal component vectors are averaged according to the time sequence to generate a vector with the same length as the principal component vector, called "reference feature vector".

[0067] Subsequently, the system compares each principal component vector in the current behavior sequence with the corresponding reference feature vector one by one. The comparison method is as follows: sort all sample values in the principal component vector in the sampling order to form sequence A; form sequence B by the sample values at the corresponding positions in the reference feature vector; and calculate the linear correlation of sequences A and B. First, the average values of A and B are calculated respectively; then the difference between each sample value and the average value of the sequence in which it is located is calculated point by point; each difference value in A is multiplied by the difference value at the corresponding position in B, and then the sum of all product values is calculated to obtain the amount of cooperative change; the sum of squares of each difference value in A and the sum of squares of each difference value in B are calculated, and then the square roots of the two sums are calculated to obtain two standard deviations; the amount of cooperative change is divided by the product of the two standard deviations to obtain the correlation coefficient of the principal component and the reference feature. The above calculation process is performed once for each pair of principal component and reference feature in the system in a programmatic manner. Usually, there are 2 to 3 principal components after dimensionality reduction, and the system finally calculates the average of all correlation coefficient values to obtain a global correlation coefficient, which is used to judge the overall consistency of the current behavior feature and the historical habit. The average of the correlation coefficients of all principal components and reference vectors is taken as the overall correlation coefficient. If the value is greater than 0.7, it is considered to be highly correlated. This value is an industry experience boundary value used to determine behavior consistency in driving behavior style recognition.

[0068] If the correlation coefficient is greater than 0.7, the system determines that the current behavior sequence is consistent with the historical habit feature, and classifies the behavior sequence into the "optimized behavior mode", and updates the original behavior mode grouping structure, i.e., reclassifies the sequences originally belonging to other categories into the grouping corresponding to the habit feature, to complete the optimization of the mode clustering result.

[0069] The system then calculates the path score of the optimized behavior mode. In each optimized behavior mode grouping, it is detected whether each point is within the navigation path tolerance interval, the number of "on-path" points is counted, divided by the total number of points, and multiplied by 100 to obtain the "path following rate dynamic score", which has a full score of 100. This score reflects the enhancement effect of the consistency of the behavior path under the optimized grouping.

[0070] Finally, the system inputs the above path following rate dynamic score and the initial "intention category" into the decision tree for reclassification. The reconstructed decision tree takes the path score as the root node, the intention category as the first layer branch, and combines speed, steering, acceleration, and other original features to construct deep nodes, so as to rejudge the behavior sequence features, and finally output the "final habit feature", which represents the stable behavior expression label of the current driving behavior under the cross conditions of multi-dimensional features, time sequence structure, historical style, and navigation purpose, to provide a basis for personalized matching of advertising content.

[0071] The modeling module fuses the user interest weight and the real-time behavior preference according to the habit feature and the intention category to generate a consumption tendency probability distribution, including: calculating the user interest weight by using a weighted fusion method through the habit feature and the intention category to obtain a user preference vector; generating an initial consumption tendency probability distribution by using a probability distribution model according to the user preference vector and the real-time behavior preference; if the correlation coefficient of the initial consumption tendency probability distribution and a preset threshold is lower than a standard value, extracting a behavior mode grouping through the behavior sequence data to obtain the behavior mode grouping; performing dimension reduction processing on the behavior sequence data by using a feature extraction method according to the behavior mode grouping to obtain a reduced behavior feature; if the matching degree of the reduced behavior feature and the user preference vector is higher than a preset threshold, updating the consumption tendency probability distribution by using the probability distribution model to obtain an optimized tendency probability distribution; calculating a dynamic score of the consumption tendency probability according to the optimized tendency probability distribution and the consumption behavior feature to obtain the consumption behavior feature; and classifying the behavior sequence data by using a decision tree algorithm through the consumption behavior feature and the user preference vector to obtain a final consumption tendency feature.

[0072] In the embodiment, after receiving the habit feature and the intention category, the modeling module first performs encoding conversion on the two. The habit feature is obtained by the previous decision tree classification, for example, “stable driving” is the encoding label “category 0”, and the intention category such as “intention deviation” is “category 1”. The system predefines a behavior feature weight set corresponding to each label to form an encoding vector in multiple dimensions such as direction deviation angle, path offset, speed stability, etc. The system sets the habit feature weight coefficient to 0.6 and the intention category weight coefficient to 0.4, multiplies the two encoding vectors by the respective coefficients in proportion, and adds the elements to obtain a fusion vector, which is the “user preference vector”.

[0073] The system then obtains the driving behavior data in the last five minutes, with a sampling frequency of ten times per second, a total of three thousand records, containing multi-dimensional information such as direction angle, instantaneous speed, acceleration, pause state, path offset, etc., extracts statistical features such as mean, standard deviation, maximum and minimum, and combines the extraction results into a “real-time behavior preference vector”. The system takes the “user preference vector” and the “real-time behavior preference vector” as inputs and puts them into a Gaussian mixture model for training. In the training process, the expectation maximization algorithm is used to calculate the mean, covariance and mixing weight of each advertisement category in the input feature space, and finally the model outputs the probability estimate of each category, and the result is standardized to obtain the “initial consumption tendency probability distribution”, with a value between zero and one, and a total of one.

[0074] The system then correlates the initial distribution with a preset "standard consumption propensity template". The standard template is obtained by the system through training on a large number of user behavior and advertisement acceptance rate data, forming a set of typical probability values for each advertisement category. After comparing the two, the system calculates the Pearson correlation coefficient, and if the coefficient is lower than the set threshold of 0.5, it is considered that the current distribution deviates significantly from the user's historical interest structure, and the behavior pattern needs to be extracted again. The threshold of 0.5 is set by the system according to the correlation statistics of five thousand groups of user historical behavior and recommendation feedback, and a value above the threshold indicates good structural consistency.

[0075] If there is a deviation, the system clusters the current five-minute behavior time series according to the data sampled every ten seconds, initially sets the K-means algorithm to 4 categories, and then adjusts the best cluster number (usually 3 to 5 categories) by measuring the cluster quality through the silhouette coefficient. The clustering features include ten dimensions such as direction angle, speed, path deviation, and pause time. After clustering, each category of time period is taken as a "behavior pattern group". The system performs principal component analysis on each group to reduce the dimension: constructs a feature matrix from the included samples, calculates the covariance matrix, performs eigenvalue decomposition, and selects the feature vectors that achieve a cumulative contribution rate of 95% as principal components in turn. Usually, 2 to 3 principal components are retained to form "reduced dimension behavior features".

[0076] The system calculates the matching degree of the reduced dimension features in each behavior pattern group and the "user preference vector" obtained by fusion. The matching degree is calculated by calculating the Euclidean distance between the two vectors, then normalizing the distance to 0 to 1, and multiplying the obtained value by 1 to get the matching degree score; the matching degree threshold is set to 0.75, indicating a high match behavior pattern. This threshold is determined by the system after statistical standard deviation of a large number of user behavior error distribution data to filter unmatched behavior patterns.

[0077] For behavior pattern groups with a matching degree of 0.75 or higher, the system inputs the pattern group samples and the user preference vector as training samples into the Gaussian mixture model again to retrain and obtain an "optimized propensity probability distribution". This distribution takes into account both real-time behavior patterns and historical preference structures, so it can better reflect the current user state than the initial distribution.

[0078] Next, the system takes the advertisement category with the highest probability in the optimized distribution as the target, and counts the number of trigger behaviors and the duration of the category from the behavior records. The system divides the number of triggers by 300 seconds (five minutes) to obtain the trigger rate, multiplies it by the probability value of the category, and finally multiplies it by the standardization coefficient 100 to obtain the "consumption propensity probability dynamic score", with a score range of 0 to 100. The score value reflects the attractiveness intensity of the advertisement category under the current behavior.

[0079] Finally, the system takes the "consumption tendency probability dynamic score" and "user preference vector" as inputs, and uses a decision tree classifier for reclassification judgment. The first layer node of the decision tree is the dynamic score. If the score is higher than the set threshold (such as 80 points), it will enter the judgment of the secondary node of the feature, such as speed stability, steering regularity, etc. until the leaf node outputs the "final consumption tendency feature". This feature is used to identify the high confidence interest of the current user for a certain type of advertisement, which is used for accurate pushing by the subsequent recommendation module.

[0080] The content recommendation module uses a recommendation algorithm to match the relevance of the advertisement content and the regional targeting from the pre-established advertisement library according to the consumption tendency probability distribution, to generate a personalized advertisement push list, including: obtaining a user preference vector from the consumption tendency probability distribution, clustering the user preferences using a collaborative filtering algorithm to obtain user interest groups; extracting advertisement content features and regional targeting features from the pre-set advertisement library according to the user interest groups, calculating the matching degree of the advertisement content and the user interest groups using cosine similarity to obtain an advertisement matching score; if the advertisement matching score is higher than the pre-set threshold, selecting the highest score advertisement content from the advertisement matching score to generate a candidate advertisement set; using a logistic regression model to predict the click probability of the user on the advertisement according to the candidate advertisement set and the user behavior data to obtain an advertisement click prediction value; selecting the advertisements higher than the pre-set threshold from the advertisement click prediction value, combining the regional targeting features to generate a preliminary push list; using a sorting algorithm to prioritize the advertisements for the preliminary push list to obtain an optimized push list; dynamically adjusting the push order according to the optimized push list and the real-time behavior data of the user to obtain the final personalized advertisement push list.

[0081] In the present application, the content recommendation module first extracts the user preference vector from the consumption tendency probability distribution data output by the modeling module. The vector is a probability vector containing multiple dimensions, each dimension corresponding to an advertisement category, such as food, shopping, travel, and refueling. The values of each dimension in the user preference vector are between 0 and 1, and the sum of all dimensions is 1. The system obtains a historical user preference vector dataset from the database and combines the current user preference vector to perform clustering using a collaborative filtering algorithm. The clustering algorithm optimally selects the K-means method, with an initial clustering number of 10. The elbow method is used to analyze the change rate of the within-group sum of squares under each clustering number. When the error reduction brought by adding a new clustering group is less than 5%, the clustering number is stopped increasing, and the clustering number at this point is the final K value. The final output is the center vector of each interest group, which is the average preference of the user interest features in each group.

[0082] The system first receives the consumption tendency probability distribution output by the modeling module, which is a multi-dimensional probability vector, each dimension representing an advertising category, such as food, fuel, shopping, travel, etc. The value of each dimension in the vector is between 0 and 1, and the sum of all dimensions is equal to 1. The probability vector is taken as the user preference vector of the current user, and the consumption tendency probability distribution of all historical users is retrieved from the user behavior database in the same format to form a user preference vector set. Then, the system uses a collaborative filtering algorithm to analyze the similarity of the user preference vector set of all users. Collaborative filtering uses a user-user similarity model to cluster users by calculating the cosine similarity between the current user preference vector and other user preference vectors. The cosine similarity is calculated by dividing the dot product of the two vectors by the product of their respective magnitudes, and the final output is a similarity score between 0 and 1. If the preference vectors of two users are in the same direction, the similarity is 1, indicating that the users have consistent interests; if the angle between them is 90 degrees, the similarity is 0, indicating no correlation. After calculating the similarity matrix between all users, the system uses the K-means clustering algorithm to group users by interest. The initial number of clusters in K-means clustering is set to 10, and the elbow method is used for optimization, that is, the change trend of the sum of squared errors is calculated when K is from 1 to 20, and the K value at the inflection point of the error reduction curve is selected as the final cluster number. The system then assigns users to each cluster center based on the Euclidean distance between them in each iteration, and continuously updates the cluster center until the clustering converges, that is, the grouping result of users does not change or the change rate is less than 1% in two consecutive iterations. Finally, each clustering result is a user interest group, and users in the same group have similar consumption tendency characteristics. The system saves the center vector of each group as a representative of the interest group for subsequent ad matching and recommendation strategy formulation.

[0083] Next, the system scans all ad records in the ad library according to the above user interest grouping results. Each ad is encoded into a content feature vector and a regional feature vector, where the content feature vector includes 12 fields such as ad type label, product category number, promotion activity label, visual element style label, etc., all of which are standardized to values between 0 and 1; the regional feature vector includes administrative district code, geographic center point latitude and longitude, coverable radius (in kilometers), and cover type such as service area, commercial district, highway exit, etc. The regional features of each ad are combined into a unified dimension vector. The system calculates the similarity between the user interest group center vector and the ad content feature vector using cosine similarity, and the closer the cosine value is to 1, the more consistent the ad content is with the interest group. The system sets the ad matching score threshold to 0.7, which is obtained by statistical analysis of large-scale user click feedback data, that is, the click rate of ads with a cosine similarity higher than 0.7 is much higher than the average, so it is used as the screening threshold.

[0084] After obtaining the user interest group, the system reads the central preference vector of the group, which is a set of normalized dimension values representing the collective preference of the interest group users on multiple advertisement types. For example, the vector may include consumption tendency weights for categories such as food, gas, supermarket, scenic spot, repair, etc., and the sum of the weights is 1. The system then extracts the advertisement content features and regional targeting features from the preset advertisement library one by one. Advertisement content features are a fixed dimension vector, the dimension content is consistent with the user preference vector, including advertisement type, product subcategory, promotion label, service period label, etc., which have been standardized to values between 0 and 1 to ensure consistency in numerical structure with the user preference vector. Regional targeting features include target city code, geographic center point coordinates, and service radius of the advertisement service, which are used for subsequent area matching and do not participate in cosine similarity calculation at this time. The system calculates the cosine similarity between the central preference vector of the user interest group and the content feature vector of each advertisement. The specific operation is as follows: the system first performs point multiplication on the two vectors to obtain the numerator; then calculates the Euclidean norm of each vector as the denominator; finally, the numerator is divided by the denominator to obtain a similarity score between 0 and 1. The closer the similarity is to 1, the more consistent the advertisement content is with the user interest group, and the more suitable it is for pushing. The system generates an advertisement matching score for each advertisement, which is the cosine similarity value between the advertisement and the central vector of a certain interest group. If an advertisement needs to be matched with multiple interest groups, the system can record the similarity results of each advertisement with all interest groups, and take the highest value as the final advertisement matching score of the advertisement.

[0085] If the matching score of an advertisement content with any interest group is higher than 0.7, the advertisement is included in the candidate advertisement set. Each candidate advertisement needs to be further analyzed for click probability prediction. The system collects all click behavior data of users in the past 7 days, including click frequency, dwell time, bounce rate, display position, etc., and concatenates user behavior features and advertisement content features as training samples, and uses a logistic regression model for training. The training target is to predict the click probability value of a user on a certain advertisement, and the model output is a probability value between 0 and 1, representing the likelihood of the user clicking the advertisement under the current conditions. The system sets the click probability threshold to 0.3, which is obtained by comparing the click probability at the 75th percentile in the click rate distribution. Only when the click probability prediction value of a certain advertisement is higher than the threshold, it is included in the preliminary push list.

[0086] The system sorts all the advertisements in the preliminary push list by priority using a gradient boosting tree-based sorting learning algorithm. During the sorting process, the system calculates a comprehensive score for each advertisement, which is based on factors such as the advertisement's click prediction value, the remaining time-to-live for the advertisement, the current user's geographic distance from the advertisement's target location, and the advertisement's historical click success rate. A feature matrix is constructed with each feature as a column, which is input into the sorting model for training, and the sorting score is output. The final output is the "optimized push list".

[0087] The system adjusts the push link dynamically by combining the user's current real-time behavior data. If the vehicle's current driving direction points to the advertisement target location, and the straight-line distance between the vehicle and the advertisement target location is less than or equal to 2 kilometers, and the advertisement's ranking in the optimized push list is not in the top three, it will be promoted to the top. If the current user is in a high-speed state and is expected to pass through the advertisement coverage area within 30 seconds, the system will use an early trigger mechanism to display the advertisement in advance. The advertisement display control part sets the minimum display interval time to 5 minutes to avoid repeated display. The final adjusted advertisement set is the "personalized advertisement push list", which is pushed to the vehicle terminal for display or voice broadcast.

[0088] The behavior update module updates the behavior signal sequence by real-time monitoring of the direction deviation angle and speed matching degree, and obtains new dynamic features of steering frequency and pause time difference, including: obtaining direction deviation angle and speed matching data through sensors, using a sliding window method to process real-time collected data to generate a behavior signal sequence; according to the behavior signal sequence, using a fast Fourier transform algorithm to extract the steering frequency feature, obtaining the frequency change trend; if the frequency change trend exceeds the preset threshold, calculating the pause time difference by weighted average method to generate time difference distribution; according to the time difference distribution, using K-means clustering algorithm to divide the steering behavior mode, determining the behavior mode classification; if the behavior mode classification matches the preset mode, updating the dynamic feature vector by sequence comparison method to obtain dynamic behavior signal; according to the dynamic behavior signal, using a linear regression model to predict the frequency change trend, generating a behavior prediction value; according to the behavior prediction value, adjusting the priority of the dynamic feature vector by sorting algorithm to obtain an optimized behavior signal sequence.

[0089] In the behavior updating module, first, the real-time direction deviation angle and speed matching degree data of the vehicle are obtained by the vehicle-mounted sensor at a sampling frequency of 5 times per second. The direction deviation angle refers to the included angle between the driving direction of the vehicle and the recommended direction of the navigation, and is calculated by differentiating the heading angle of the vehicle and the navigation path angle; the speed matching degree is a proportional value calculated by dividing the current vehicle speed by the recommended speed of the navigation, and is limited to the range of 0 to 1, wherein 1 represents complete matching and 0 represents complete mismatch. The system arranges the obtained real-time data in chronological order and adopts a sliding window processing method for structured processing. The length of the sliding window is set to 30 seconds, and the sliding step is 10 seconds, that is, every 10 seconds, the data in the last 30 seconds is reprocessed once. In each window, the direction deviation angle and speed matching degree sequence are composed into a time series to form a behavior signal sequence for subsequent feature analysis.

[0090] The system performs fast Fourier transform processing on the behavior signal sequence to convert the time domain signal into a frequency domain signal, extracts the main frequency component, and records the frequency amplitude. Between two consecutive time windows, the system calculates the difference value of the main frequency amplitude as the current frequency change trend. If the difference value exceeds 0.15 hertz (this threshold value is obtained by statistically analyzing the 95th percentile data of the frequency change in a large number of driving samples), it is considered that the vehicle steering behavior has changed significantly.

[0091] First, the system collects the behavior signal sequence in the sliding window, which contains the direction deviation angle data indexed by time. The system sets the window length to 30 seconds and the sampling frequency to 5 times per second, so there are a total of 150 data points of the direction deviation angle in each window. These direction deviation angles are arranged in chronological order to form a time series. To ensure the accuracy of frequency feature extraction, the sequence is first de-meaned, that is, all data points are subtracted by the average value of the sequence, so as to eliminate the interference of direct current offset on frequency domain analysis. Then, the system calls the fast Fourier transform algorithm to convert the de-meaned direction deviation angle sequence into a frequency domain. This algorithm converts the original time series into a set of frequency components, each frequency component corresponds to an amplitude representing the strength of the frequency component in the original signal. The system records the amplitudes of all frequency components and finds the frequency with the largest amplitude, called the main frequency, which is the main steering frequency feature in the current time window. Subsequently, the system saves the main frequency value and compares it with the main frequency of the previous time window. The specific method is to calculate the absolute value of the difference between the current main frequency and the previous main frequency, which is the frequency change trend. If the trend value is large, it indicates that the vehicle has changed significantly in the steering behavior mode in this time period; if the trend value is small, it indicates that the vehicle steering behavior is stable. Finally, the frequency change trend is recorded as a time series indicator and used to determine whether to trigger the subsequent pause time difference calculation and behavior mode recognition process.

[0092] Subsequently, the system identifies all time points in the signal sequence where the speed matching degree is below 0.2 to determine the stopping behavior of the vehicle. Each time a period below the value is identified, the starting time and the ending time are recorded, and the stopping time difference is calculated. The system collects all stopping time difference samples in the last 90 seconds and divides them into three segments: the data in the last 30 seconds is given a weight of 0.6, the data in the middle 30 seconds is given a weight of 0.3, and the data in the first 30 seconds is given a weight of 0.1, and then the statistical distribution value of the current stopping time difference is calculated in a weighted average manner. The value is used for subsequent clustering analysis.

[0093] After obtaining the steering frequency feature and the stopping time difference distribution, the system applies the K-means clustering algorithm to classify the behavior patterns of the samples. The K value is set to 3, indicating that the system divides three types of steering behavior: high-frequency steering, medium-frequency steering, and low-frequency steering. The system initially randomly selects three centroid points from the data set, corresponding to the three types of behavior patterns, and then calculates the Euclidean distance of each data point to the three centroids, and assigns it to the nearest centroid category. After completing the initial classification, the system recalculates the center value of the sample points in each category, updates the centroid, and repeats the iteration process until the centroid no longer changes significantly or the maximum iteration count reaches 100 times.

[0094] After clustering is completed, the system counts the label proportion of the data in the current window. If the label with the highest proportion is consistent with the preset standard driving behavior pattern, and the proportion of samples of this type in the total number of window samples is greater than 70%, the system considers that the current behavior pattern matches the expected pattern. At this time, the system calls the sequence comparison method to compare the current behavior signal sequence with the standard behavior signal template item by item. The two sequences are aligned through the dynamic time warping algorithm, and if there are missing time points, they are filled by linear interpolation, and if there are redundant time points, they are deleted, so that the current signal and the template remain consistent in length and rhythm, thereby generating a time-standardized dynamic feature vector.

[0095] The dynamic feature vector is then input into a linear regression model, which uses time steps as the independent variable and steering frequency as the dependent variable, and uses the least squares method to calculate the regression coefficients. After training is completed, the system predicts the frequency values 5 seconds and 10 seconds in the future based on the model to form behavior prediction values, which represent the possible behavior trends in the future. The system then sorts all behavior dimensions in the dynamic feature vector according to the absolute gradient of the prediction values, with the gradient of the behavior feature representing a more important behavior change trend. The system uses the quicksort algorithm to sort and obtain an optimized behavior signal sequence arranged according to behavior significance, which maintains the original dimensions but reflects the importance of behavior changes, providing a priority processing order for the subsequent recommendation system.

[0096] After the dynamic behavior signal generation and behavior prediction value calculation are completed, the system will enter the ranking optimization phase. At this time, the system has obtained the prediction value of each dynamic feature component (such as steering frequency, pause time difference, etc.) in the next 5 seconds, 10 seconds through the linear regression model. The system calculates the change gradient of each dynamic feature in the next 5 seconds, that is, the prediction value is subtracted from the current value and then divided by the prediction time interval to obtain the change rate per unit time. The larger the gradient value, the greater the change of the feature in a short time, and the higher the importance.

[0097] The system ranks all dynamic features according to their respective gradient values. The sorting process uses the quicksort algorithm. The specific operation is: first select an intermediate value from the feature list as a reference, move the features with gradient less than the reference to one side, and move the features with gradient greater than the reference to the other side, recursively execute the process until all features are arranged in descending order. After sorting, the system arranges the order as the priority of the dynamic feature, and the high-priority feature will be arranged at the front of the optimized behavior signal sequence, and the low-priority feature will be arranged at the back.

[0098] Finally, the optimized behavior signal sequence is output as a feature list rearranged according to the importance of feature change. This list not only retains all the information of the original behavior signal, but also clearly indicates which features need to be prioritized and processed.

[0099] The push strategy adjustment module adjusts the weights of real-time behavior preferences and interaction feedback factors in the consumption tendency probability distribution according to new dynamic features, and regenerates a personalized advertisement push list including advertisement display frequency and push time sensitivity, including: obtaining user behavior data and interaction records, extracting real-time behavior preferences and interaction feedback factors through data cleaning processing, and obtaining user behavior feature set; calculating the consumption tendency probability by using the weighted average algorithm according to the user behavior feature set, updating the dynamic probability distribution, and obtaining the probability distribution vector; if the change amplitude of the probability distribution vector exceeds the preset threshold, predicting the advertisement display frequency trend by using the linear regression model to obtain the frequency prediction value; adjusting the weight adjustment strategy according to the frequency prediction value, updating the weight of the advertisement display frequency, and obtaining the optimized weight parameter; generating an advertisement push sequence by combining the push time sensitivity through the optimized weight parameter, and obtaining a preliminary push list; dividing user preference groups by using the K-means clustering algorithm according to the preliminary push list, updating the personalized push list, and obtaining the final push list; matching user interaction records through the final push list, adjusting the weight of the push time sensitivity, and obtaining the optimized push sequence.

[0100] In one possible implementation, the push strategy adjustment module adjusts the weights of real-time behavior preferences and interaction feedback factors in the consumption tendency probability distribution according to new dynamic features, and regenerates a personalized advertisement push list containing advertisement display frequency and push time sensitivity. The module first acquires the behavior data and log information of interaction with advertisements of the user in real time through the data interface of the vehicle-mounted system, including but not limited to advertisement clicking, sliding, ignoring, closing operation, etc., all of which are attached with accurate time stamps, advertisement identifiers, display positions, and geographical positions of the vehicle. The data cleaning link performs null deletion, outlier rejection, and time alignment processing on the behavior data, and collects the behavior sequence of each user according to the user identifier. Then, the system extracts real-time behavior preference features from the cleaned behavior data, including advertisement click frequency per unit time, display frequency, dwell time, and bounce rate, etc., and extracts interaction feedback factors from the interaction log, mainly including user click response rate, click depth, and time period activity, etc. All features are standardized to form a user behavior feature set.

[0101] Subsequently, the system uses a weighted average algorithm to combine the real-time behavior preferences and interaction feedback factors into a set of consumption tendency probability values. The behavior preference weight is set to 0.7, and the interaction feedback factor weight is set to 0.3. The two sets of feature vectors are multiplied by the corresponding weights and summed, and then normalized, so that the sum of the tendency probabilities of all advertisement categories is 1, thereby constructing a probability distribution vector. The system compares the vector with the probability distribution vector generated in the historical period element by element. If the change amplitude of any dimension exceeds the set threshold value 0.1 (this threshold value is determined by calculating the standard deviation in 10,000 user samples and taking twice the value), it is considered that the tendency has changed significantly, and the system starts the prediction process.

[0102] In the prediction process, the system takes the current probability distribution vector as the independent variable and the advertisement display frequency in the historical unit time as the dependent variable to construct a linear regression model. The model training uses the least squares method, and after obtaining the regression coefficient, the future display frequency trend value is calculated by substituting the current probability vector, and the frequency prediction value is output. If the prediction value exceeds the upper threshold value of 60 times per hour, it means that the push is too frequent, and the system will lower the basic weight of each advertisement display by 20%; if it is lower than the lower limit of 20 times per hour, it will be increased by 10%. The optimized weight parameters are normalized after the weight adjustment and used as the basis for advertisement sorting and display control.

[0103] The system generates an advertisement push sequence in combination with the optimized weight parameters and the push time sensitivity rules of the current time period and geographical location. The sensitivity rules include, but are not limited to, a minimum display interval of 3 minutes in a high-speed driving state, 2 minutes on urban trunk roads, and allowing instant push once in a user waiting for a red light or parking state. The system inserts the advertisement content that meets the rules into the timeline to form a preliminary push list, and sorts the content according to the weight to form an optimized push list.

[0104] Next, the system performs user interest clustering processing on the optimized push list. The K-means clustering algorithm is used to group user preference vectors, and the number of cluster centers is set to 5. For each group of user interest labels, similar topics are filtered from the advertisement content, and after uniform sorting, the corresponding list is replaced to update the personalized push list. Finally, the system compares the user's latest interaction feedback record to determine their current activity and advertisement acceptance. If multiple ignore or skip are detected within a short period of time, the system dynamically increases the push time interval, extends the interval to a minimum of 4 minutes, and prioritizes high-click-rate advertisement content to form the final optimized push sequence and push it to the vehicle-mounted terminal display system.

[0105] The system first extracts the user preference vector corresponding to each advertisement from the preliminary push list. This vector usually contains multiple dimensions, such as advertisement type preference, content style preference, time period preference, and interaction tendency, each dimension being a normalized probability value ranging from 0 to 1. All user preference vectors are combined into a sample set and input into the K-means clustering algorithm. The system sets the number of clusters K to 5 (this value is determined during sample training through the evaluation of the silhouette coefficient), and uses the Euclidean distance as the similarity measurement index. The clustering steps include: first, randomly select 5 initial center points from the sample set, then iteratively perform the following process: in each round, divide all user samples into the cluster with the nearest center point, then update the center point position by averaging the user preference vectors in each cluster. Repeat this process until the cluster center point changes less than the set convergence threshold of 0.01 or reaches the maximum iteration number of 100 times.

[0106] After clustering, the system selects the advertisement content with the highest similarity to the cluster center from the advertisement library according to the average preference features of each cluster center. The advertisement content features include content theme, industry, image style, target location, and other fields, which are also encoded into vectors with the same structure as the user preference vector. The similarity is calculated by the cosine similarity, and the advertisements with a similarity value higher than 0.85 are selected as candidate content to replace the advertisements with a matching degree lower than 0.7 in the original preliminary push list, forming personalized push content for each user group. This replacement operation is performed independently for all user groups to ensure that the final push list contains advertisement content that is highly matched with the target user group features, ultimately forming a personalized and optimized final push list.

[0107] The system obtains the history record of user interaction with the advertisement, including click time, display time, whether to skip, stay time after clicking, bounce rate and other data, and matches the user identifier with the advertisement item in the final push list one by one, counts the click rate and bounce rate of each advertisement in different time periods, and calculates the average response time. The advertisement with a click rate higher than 0.5, a bounce rate lower than 0.2 and a response time shorter than 5 seconds is marked as a high response advertisement, and the advertisement below the standard is marked as a low response advertisement. Based on the response level of the advertisement, the system adjusts the weight of time sensitivity. In the time sensitivity strategy, the time sensitivity benchmark weight of the morning peak period, lunch break period, evening peak period and evening leisure period of weekdays is set to 0.6, 0.7, 0.5 and 0.8 respectively. If an advertisement is marked as a high response advertisement in the corresponding time period, the time sensitivity weight of the period is increased by 20%, that is, the original weight is multiplied by 1.2; if it is a low response advertisement, it is reduced by 10%, that is, the original weight is multiplied by 0.9. After adjustment, all advertisements are sorted according to the weighted score, combined with real-time vehicle state and driving scene to judge whether to push, and finally generate an optimized advertisement push sequence and push to the user interface for display. The whole process strictly bases on user behavior and interaction data to dynamically adapt the advertisement push strategy, ensuring that the push content and time window can maximize user acceptance rate and response effect.

[0108] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1.A scenario-based vehicle advertisement recommendation system for intelligent transportation, characterized in that, The method comprises the following steps: A data acquisition module acquires direction deviation angle, speed matching degree and path compliance rate in real time through vehicle sensors and a navigation system to obtain original behavior data; A feature extraction module processes the original behavior data using time series analysis to extract dynamic features of steering frequency, pause time difference and signal acquisition frequency, and determines a behavior signal sequence; An association analysis module calculates a correlation coefficient according to the matching degree of the behavior signal sequence and navigation setting information in terms of time synchronization and environmental interference factors to obtain an association feature vector; An intention recognition module classifies the behavior signal sequence according to direction deviation angle and speed matching degree through a clustering algorithm to determine an intention category if the correlation coefficient of the association feature vector is greater than a preset threshold; A pattern recognition module analyzes a behavior pattern using a decision tree algorithm by comparing the intention category and historical behavior data in terms of path compliance rate and steering frequency to determine habit features; A modeling module generates a consumption tendency probability distribution according to the habit features and the intention category, fuses user interest weights and real-time behavior preferences, and generates a personalized advertisement push list; A content recommendation module generates the personalized advertisement push list by matching advertisement content relevance and regional targeting from a pre-established advertisement library according to the consumption tendency probability distribution; A behavior updating module updates the behavior signal sequence by monitoring direction deviation angle and speed matching degree in real time to obtain new dynamic features of steering frequency and pause time difference; A push strategy adjustment module adjusts the weights of real-time behavior preferences and interaction feedback factors in the consumption tendency probability distribution according to the new dynamic features to regenerate the personalized advertisement push list containing advertisement display frequency and push time sensitivity. 2.The smart traffic oriented scenario-based in-vehicle advertisement recommendation system of claim 1, wherein: The data acquisition module acquires direction deviation angle, speed matching degree and path compliance rate in real time through vehicle sensors and a navigation system to obtain original behavior data, which comprises the following steps: Real-time acquisition of direction deviation angle, speed matching degree and path compliance rate through vehicle sensors and a navigation system to obtain original driving behavior data; If the direction deviation angle exceeds a preset threshold, the direction deviation angle is smoothed using a Kalman filtering algorithm to obtain smoothed direction data; According to the smoothed direction data and navigation system data, path planning consistency is calculated to determine a path deviation vector; By using a support vector machine algorithm to classify driving behavior features through the path deviation vector and the speed matching degree, the behavior abnormal state is determined; If the behavior abnormal state is true, real-time control correction parameters are calculated according to the speed adjustment amplitude and the path compliance rate to obtain corrected driving instructions; The navigation system data is updated through the corrected driving instructions and real-time acquisition data to obtain optimized path planning; According to the optimized path planning and data acquisition frequency, the sensor acquisition parameters are adjusted to obtain updated original driving behavior data. 3.The smart traffic oriented scenario-based in-vehicle advertisement recommendation system according to claim 1, characterized in that: The feature extraction module processes the original behavior data using time series analysis to extract dynamic features of steering frequency, pause time difference and signal acquisition frequency, and determines a behavior signal sequence, which comprises the following steps: Steering frequency, pause time difference and signal acquisition frequency are extracted from the original driving behavior data through time series analysis to generate a behavior signal sequence; According to the behavior signal sequence, a sliding time window method is used to divide data segments, and the feature change trend in each segment is determined; If the feature change trend exceeds the preset threshold, the periodicity of the steering frequency is analyzed using Fourier transform to obtain a periodic feature parameter; By the difference between the periodic feature parameter and the pause time, a stability index of the driving behavior is calculated to generate a stability evaluation sequence; According to the stability evaluation sequence, a K-means clustering algorithm is used to classify the driving behavior to determine an abnormal behavior pattern; If the abnormal behavior pattern is true, the data sampling rate is adjusted according to the signal collection frequency to obtain an optimized behavior signal sequence; Through the optimized behavior signal sequence and the time window division result, the feature analysis method is updated to generate a new behavior signal sequence. 4.The smart traffic oriented scenario-based in-vehicle advertisement recommendation system of claim 1, wherein: The correlation analysis module calculates a correlation coefficient according to the matching degree of the behavior signal sequence and the navigation setting information in time synchronization and environmental interference factors to obtain a correlation feature vector, including: By aligning the time stamps of the behavior signal sequence and the navigation setting information, the time synchronization deviation is calculated to obtain a synchronization adjustment parameter; According to the synchronization adjustment parameter, the time axis of the behavior signal sequence is corrected to generate a corrected signal sequence; Using the corrected signal sequence and the navigation setting information, the weighted influence of the environmental interference factor is calculated to determine the interference factor weight; Through the interference factor weight and the corrected signal sequence, the weighted correlation coefficient is calculated to generate a correlation feature vector; If the dimension of the correlation feature vector exceeds the preset threshold, the principal component analysis algorithm is used to compress the feature vector to obtain a reduced feature vector; According to the matching degree of the reduced feature vector and the navigation setting information, a dynamic matching score is calculated to determine the behavior signal classification result; Through the behavior signal classification result and the environmental interference factor weight, the time synchronization adjustment strategy is updated to generate an optimized synchronization parameter. 5.The smart traffic oriented scenario-based in-vehicle advertisement recommendation system according to claim 1, characterized in that: The intention recognition module classifies the behavior signal sequence according to the direction deviation angle and the speed matching degree by a clustering algorithm if the correlation coefficient of the correlation feature vector is greater than the preset threshold to determine the intention category, including: The intention category is obtained by classifying the behavior signal sequence using the K-means clustering algorithm; According to the intention category, the time axis of the behavior signal sequence is corrected using the time synchronization parameter to generate a corrected signal sequence; Through the corrected signal sequence and the navigation setting information, the weighted value of the environmental interference factor is calculated to determine the interference factor weight; If the interference factor weight exceeds the preset threshold, the principal component analysis algorithm is used to reduce the dimension of the corrected signal sequence to obtain a reduced signal sequence; According to the reduced signal sequence and the navigation setting information, a dynamic score of the direction deviation angle and the speed matching degree is calculated to obtain a matching feature vector; By using the support vector machine algorithm, the intention category of the behavior signal is judged through the matching feature vector and the intention category; According to the intention category and the interference factor weight, the time synchronization parameter is updated to generate an optimized synchronization parameter. 6.The smart traffic oriented scenario-based in-vehicle advertisement recommendation system according to claim 1, characterized in that: The pattern recognition module analyzes the behavior pattern by comparing the intention category and the historical behavior data in the path following rate and the steering frequency using the decision tree algorithm to judge the habit feature, including: The habit features are obtained by comparing the path compliance rate and the turning frequency of the historical behavior data according to the intent category, and the behavior patterns are classified by using a decision tree algorithm to obtain the habit features; According to the habit features, the behavior sequence features in the time series data are obtained, and the behavior sequence features are grouped by using a clustering algorithm to obtain the behavior pattern grouping; If the path compliance rate of the behavior pattern grouping is lower than a preset threshold, the change trend of the turning frequency is calculated through the time series data to obtain the turning frequency change feature; According to the turning frequency change feature, the behavior sequence features are processed by dimension reduction by using a feature extraction method to obtain the reduced behavior features; If the correlation coefficient of the reduced behavior features and the habit features is greater than a preset threshold, the behavior pattern grouping is updated by the pattern classification result to obtain the optimized behavior pattern; According to the optimized behavior pattern and the historical behavior data, the dynamic score of the path compliance rate is calculated to obtain the path optimization feature; The final habit features are obtained by reclassifying the behavior sequence features by using a decision tree algorithm according to the path optimization feature and the intent category. 7.The smart traffic oriented scenario-based in-vehicle advertisement recommendation system according to claim 1, characterized in that: The modeling module generates a consumption tendency probability distribution according to the habit features and the intent category, and fuses a user interest weight and a real-time behavior preference, including: The user interest weight is calculated by using a weighted fusion method according to the habit features and the intent category to obtain a user preference vector; An initial consumption tendency probability distribution is generated by using a probability distribution model according to the user preference vector and the real-time behavior preference; If the correlation coefficient of the initial consumption tendency probability distribution and a preset threshold is lower than a standard value, the behavior pattern grouping is obtained by extracting the behavior pattern grouping from the behavior sequence data; The behavior sequence data is processed by dimension reduction by using a feature extraction method according to the behavior pattern grouping to obtain reduced behavior features; If the matching degree of the reduced behavior features and the user preference vector is higher than a preset threshold, the consumption tendency probability distribution is updated by using a probability distribution model to obtain an optimized tendency probability distribution; According to the optimized tendency probability distribution and the consumption behavior feature, the dynamic score of the consumption tendency probability is calculated to obtain the consumption behavior feature; The final consumption tendency feature is obtained by classifying the behavior sequence data by using a decision tree algorithm according to the consumption behavior feature and the user preference vector. 8.The smart traffic oriented scenario-based in-vehicle advertisement recommendation system according to claim 1, characterized in that: The content recommendation module generates a personalized advertisement pushing list by using a recommendation algorithm according to the consumption tendency probability distribution, and matches the advertisement content relevance and the regional targeting from a pre-established advertisement library, including: The user interest grouping is obtained by clustering the user interest by using a collaborative filtering algorithm according to the user preference vector obtained from the consumption tendency probability distribution; According to the user interest grouping, the advertisement content features and the regional targeting features are extracted from the preset advertisement library, and the matching degree of the advertisement content and the user interest grouping is calculated by using a cosine similarity to obtain an advertisement matching score; If the advertisement matching score is higher than a preset threshold, the advertisement content with the highest score is selected from the advertisement matching score to generate a candidate advertisement set; According to the candidate advertisement set and the user behavior data, the click probability of the user to the advertisement is predicted by using a logistic regression model to obtain an advertisement click prediction value; Filtering out the advertisements higher than a preset threshold from the advertisement click prediction values, generating a preliminary push list in combination with regional targeted features; Prioritizing the advertisements in the preliminary push list by using a sorting algorithm to obtain an optimized push list; According to the optimized push list and real-time behavior data of the user, dynamically adjusting the push order to obtain a final personalized advertisement push list. 9.The smart traffic oriented scenario-based in-vehicle advertisement recommendation system according to claim 1, characterized in that: The behavior updating module updates the behavior signal sequence by real-time monitoring of the direction deviation angle and the speed matching degree, and obtains new dynamic features of the steering frequency and the pause time difference, including: Obtaining the direction deviation angle and the speed matching data through a sensor, processing the real-time collected data by using a sliding window method to generate a behavior signal sequence; According to the behavior signal sequence, extracting the steering frequency feature by using a fast Fourier transform algorithm to obtain a frequency change trend; If the frequency change trend exceeds a preset threshold, calculating the pause time difference by using a weighted average method to generate a time difference distribution; According to the time difference distribution, dividing the steering behavior mode by using a K-means clustering algorithm to determine a behavior mode classification; If the behavior mode classification matches a preset mode, updating a dynamic feature vector by using a sequence comparison method to obtain a dynamic behavior signal; According to the dynamic behavior signal, predicting the frequency change trend by using a linear regression model to generate a behavior prediction value; According to the behavior prediction value, adjusting the priority of the dynamic feature vector by using a sorting algorithm to obtain an optimized behavior signal sequence. 10.The smart traffic oriented scenario-based in-vehicle advertisement recommendation system according to claim 1, characterized in that: The push strategy adjusting module adjusts the weights of the real-time behavior preference and the interaction feedback factor in the consumption tendency probability distribution according to the new dynamic features, and regenerates a personalized advertisement push list including the advertisement display frequency and the push time sensitivity, including: Obtaining user behavior data and interaction records, extracting the real-time behavior preference and the interaction feedback factor by data cleaning processing to obtain a user behavior feature set; According to the user behavior feature set, calculating the consumption tendency probability by using a weighted average algorithm to update a dynamic probability distribution to obtain a probability distribution vector; If the change amplitude of the probability distribution vector exceeds a preset threshold, predicting the advertisement display frequency trend by using a linear regression model to obtain a frequency prediction value; According to the frequency prediction value, adjusting the weight adjustment strategy to update the weight of the advertisement display frequency to obtain an optimized weight parameter; By using the optimized weight parameter in combination with the push time sensitivity, generating an advertisement push sequence to obtain a preliminary push list; According to the preliminary push list, dividing the user preference groups by using a K-means clustering algorithm to update the personalized push list to obtain a final push list; By matching the user interaction records through the final push list, adjusting the weight of the push time sensitivity to obtain an optimized push sequence.

Images