A driving behavior-based safety education recommendation method and device

Abstract

The application discloses a kind of based on driving behavior's safety education recommendation method and device, and the method is executed by vehicle terminal.The method includes: collecting vehicle driving data and detecting bad driving behavior;According to single trip and week report granularity respectively, carry out statistics, extract the high-frequency behavior of top list, mapping generation ordered risk label set;According to the label set and driving level, recall the safety education content matched;To candidate content, calculate multidimensional comprehensive score and sort, combine behavior severity and occurrence frequency distribution display material quantity weight;When vehicle is in low interference state, content arrangement is pushed to " for you recommended " area display;Finally, collect user learning feedback index to dynamically update the recommendation strategy parameter of next cycle.The application will recent driving weak item priority rectification push, constructs self-adapting closed loop, effectively improves the improvement efficiency of bad behavior.

Description

Technical Field

[0001] This invention belongs to the field of vehicle information systems and data processing technology, specifically relating to a method and device for recommending safety education based on driving behavior. Background Technology

[0002] With the increasing intelligence of in-vehicle terminals and the improvement of driving data collection capabilities, driving behavior recognition based on vehicle operation data has become a common technical approach in this field. Existing systems can typically collect vehicle operation data such as vehicle speed, acceleration, steering, following distance, and the status of lights or turn signals, and identify undesirable driving behaviors such as sudden braking, sudden acceleration, speeding, and following too closely through rule thresholds, and then provide voice or interface reminders. On the other hand, although in-vehicle safety education content (such as text, images, and videos) is widely available, it is mostly in the form of static knowledge base displays or fixed column pushes.

[0003] Currently, the closest technical solution to this invention is the "behavioral alert + general educational content push" model. Typical steps of this model include: collecting driving data; identifying risky behaviors according to preset rules; triggering real-time alerts; and pushing general safety content after the trip. In terms of device structure, this solution typically includes a data acquisition module, a behavior recognition module, an alert output module, and a content display module executed in series.

[0004] However, this traditional, universal push notification model, due to its lack of deep integration of individual driver risk vulnerabilities into the content recommendation process and the absence of personalized strategies and feedback loops, typically reveals the following specific shortcomings in practical implementation: First, the content relevance is insufficient. Because the recommendation mechanism mainly relies on fixed sections or general hot topics, it fails to effectively map users' individual high-frequency risk behaviors to the content tags of safety education. As a result, the educational content received by users is inconsistent with their actual driving weaknesses, and the corrective effect on their behavior after learning is very limited.

[0005] Second, the difficulty matching is insufficient. The existing content recommendation process does not integrate the driver's driving level and risk stratification information. This often results in novice drivers receiving professional content that is too difficult to understand, while experienced drivers receive general content that is too simple, leading to a significant decrease in overall learning efficiency.

[0006] Third, the system lacks a closed-loop optimization mechanism. After recommending content, the system fails to continuously track the driver's subsequent learning behavior and the trend of improvement in response to poor driving behavior. This prevents the system from adaptively adjusting and iterating its recommendation strategy, resulting in unstable long-term behavior correction effects.

[0007] Fourth, the integration of multi-source information is insufficient. The system's behavioral alerts, historical driving records, scene tags, and content tags fail to form a unified model and joint ranking, which limits the accuracy, timeliness, and overall scalability of the final pushed content.

[0008] In summary, existing vehicle driving safety education delivery mechanisms lack in-depth behavioral statistical feature mapping, refined difficulty stratification, and effective learning feedback mechanisms, thus failing to achieve accurate on-demand content recommendation and long-term behavioral correction. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides a method and apparatus for recommending safety education based on driving behavior. Instead of employing traditional, generic, static push notifications, this invention extracts ordered risk tags based on multi-granular statistical features of a driver's single trip and weekly reporting cycle. This is combined with on-demand recall of driving level-based content and multi-dimensional comprehensive scoring and ranking. Furthermore, by collecting user feedback indicators after learning, the recall, ranking, and material allocation strategies are dynamically adjusted. Ultimately, an adaptive technical closed loop of "monitoring—recommendation—feedback—update" is constructed in the in-vehicle environment, thereby effectively ensuring and improving the accuracy of safety education recommendations and the efficiency of driving behavior correction.

[0010] To achieve the above objectives, the first aspect of the present invention provides a safety education recommendation method based on driving behavior, executed by a computer in an in-vehicle terminal, the method comprising the following steps: Step 1: Collect the current vehicle's driving data; Step 2: Perform undesirable driving behavior detection processing on the collected driving data, identify risky behavior events based on a preset rule engine, and obtain a set of undesirable driving behavior events; Step 3: Perform statistical processing on the set of bad driving behavior events at the single trip statistical granularity and the weekly report statistical granularity to obtain the occurrence frequency and severity information of each behavior type, and construct a behavior feature vector accordingly; Step 4: Based on the behavioral feature vector, extract recommended risk features from the driving report data to obtain an ordered risk label set. The extraction process includes: sorting each behavior type under the single trip statistical granularity in descending order of occurrence frequency, selecting the top 3 behavior types, and mapping them to generate the risk label group for the most recent trip, denoted as the TOP3 label group; sorting each behavior type under the weekly report statistical granularity in descending order of occurrence frequency, selecting the top 5 behavior types, and mapping them to generate the weekly report risk label group, denoted as the TOP5 label group; merging the TOP3 label group and the TOP5 label group to remove duplicates, and sorting them according to a preset priority rule to obtain an ordered risk label set, wherein the priority of each label in the TOP3 label group is higher than that of the labels in the TOP5 label group, and when the same behavior type appears in both groups, the ranking corresponding to the TOP3 label group is retained in the ordered risk label set. Step 5: Based on the ordered risk tag set, perform candidate content recall processing sequentially from the preset safety education content library according to the tag order to obtain a candidate safety education content set. Each safety education content matches the corresponding risk tag in the ordered risk tag set, and the content difficulty matches the stored driving level information. The content recalled corresponding to the TOP3 tag group is marked as high priority content in the candidate set. Step 6: Calculate the comprehensive score for each item in the candidate safety education content set, and sort them from high to low according to the comprehensive score to obtain the sorted recommended content list; among them, for high-priority content from the TOP3 tag group, the source priority bonus is added to the comprehensive score calculation; Step 7: Based on the severity and frequency of dangerous driving behaviors corresponding to each risk label in the ordered risk label set, determine the weight of the recommended materials corresponding to each risk label, and select recommended materials from the sorted recommended content list accordingly to generate a recommended list to be displayed. Step 8: Arrange the safety education content in the recommended list to be displayed in the "Recommended for You" display area of ​​the user terminal according to the order of the ordered risk tag set, so that the content corresponding to the TOP3 tag group is placed first; and complete the push or update display when the vehicle is in a low interference state. Step 9: Collect learning feedback metrics after users have studied the safety education content in the recommended list to be displayed, and obtain a set of learning feedback metrics; Step 10: Based on the set of learning feedback indicators, and combined with the latest trip statistics and weekly report statistics obtained in the next statistical period, dynamically adjust the recommendation strategy parameters to update the recommendation strategy for subsequent periods, and return to Step 1 to continue execution.

[0011] Furthermore, the driving data includes at least vehicle speed, longitudinal acceleration, lateral acceleration, steering wheel angle change rate, following distance, turn signal status, and gear status.

[0012] Furthermore, the risky behavioral events include at least one or more of the following: sudden braking, sudden acceleration, sudden steering, speeding, following too closely, failure to use turn signals, continuous lane changes, incorrect gear selection, and failure to decelerate in areas where deceleration is required.

[0013] Furthermore, the formula for calculating the comprehensive score in step six is: Comprehensive score = Risk matching degree × w1 + Difficulty matching degree × w2 + Content freshness × w3 + Correction gain × w4 + Source priority × w5, where w1, w2, w3, w4, and w5 are weight parameters.

[0014] Furthermore, step seven, which involves determining the weight of the recommended material quantity corresponding to each risk label and selecting recommended materials from the sorted list of recommended content accordingly to generate a list of recommendations to be displayed, includes: the higher the severity and frequency of occurrence, the greater the weight of the recommended material quantity assigned to the corresponding risk label; the mapping relationship between the weight of the recommended material quantity and the specific number of items to be displayed is a configurable strategy parameter.

[0015] Furthermore, the configuration strategy for selecting recommended materials from the sorted list of recommended content to generate the list of recommended items to be displayed includes one or more of the following methods: Method 1: Let the weight of the recommended material quantity for the i-th risk tag be Qi = α × Si + β × Ci, where Si is the severity score of the behavior corresponding to the tag, Ci is the number of occurrences, and α and β are configuration coefficients; allocate the total number of recommended items N according to the proportion of Qi to the sum of the weights of all tags to obtain the number of items displayed under each tag ni; Method 2: Divide the recommended materials into three weighted levels: high, medium, and low. High-level materials are allocated more recommended materials, medium-level materials are allocated a medium number of recommended materials, and low-level materials are allocated fewer recommended materials. The threshold for the number of materials in each level is preset by the operation strategy. Method 3: Allocate at least one recommended material to each tag in the TOP3 tag group, and distribute the remaining materials among the other tags in the TOP5 tag group.

[0016] Furthermore, the low-interference state includes a parked state or a trip-end state, and the form of the safety education content includes text, images, or videos.

[0017] Furthermore, the driving report data is presented in the user interface as a “recent” trip view and a “weekly report” view; the “recent” view displays the statistical results of dangerous driving behavior at the statistical granularity of a single trip and the TOP3 tag group, and the “weekly report” view displays the statistical results of dangerous driving behavior at the statistical granularity of a weekly report and the TOP5 tag group. The "Recommended for You" display area on the user terminal is used to present the list of recommendations to be displayed output in step eight.

[0018] Furthermore, the dynamic adjustment of the recommendation strategy parameters in step ten includes: dynamically adjusting the weight parameters w1 to w5 in step six, the weight mapping relationship of the recommended material quantity in step seven, and the recall strategy in step five.

[0019] A second aspect of the present invention provides a safety education recommendation device based on driving behavior, deployed in a computer within an in-vehicle terminal, the device comprising: The data acquisition module is configured to execute step one, which is used to collect the current vehicle's driving data; The detection module is configured to perform step two, which is used to perform bad driving behavior detection processing on the collected driving data, identify risky behavior events based on a preset rule engine, and obtain a set of bad driving behavior events. The profiling module is configured to execute steps three to four, which are used to construct behavioral feature vectors and extract recommended risk features from driving report data to obtain an ordered set of risk labels. The recall module is configured to execute step five, which is used to sequentially recall candidate content from the preset safety education content library according to the ordered risk label set; The sorting module is configured to execute step six, which is used to calculate the comprehensive score of each item in the candidate safety education content set and sort them to obtain a sorted list of recommended content. The allocation module is configured to execute step seven, which is used to determine the weight of the number of recommended materials corresponding to each risk label and select recommended materials to generate a list of recommendations to be displayed. The push module is configured to execute step eight, which is used to arrange and display the safety education content in the recommended list to be displayed in the "Recommended for You" display area and complete the push when the vehicle is in a low-interference state. The feedback module is configured to execute step nine, which is used to collect learning feedback metrics after the user's learning and obtain a set of learning feedback metrics. The update module is configured to execute step ten, which is used to dynamically update and adjust the strategy parameters in the recall module, sorting module and allocation module based on the learning feedback indicator set and the data of the next statistical period. All of the above modules are deployed on the vehicle terminal, or in a way that combines the vehicle terminal with the cloud server.

[0020] Compared with the prior art, the present invention has the following significant advantages: First, it improves the correlation and matching accuracy between selected safety education content and drivers' recent actual driving weaknesses. This invention combines single-trip features (TOP3 tag group) reflecting high-frequency undesirable behaviors during a driver's most recent trip with weekly report features (TOP5 tag group) reflecting long-term behavioral trends, and assigns higher selection priority to single-trip features during set deduplication and merging. This ensures that the education content presented on the terminal is no longer general static knowledge, but precisely corresponds to the urgent risks exposed by the driver in recent trips, thereby more effectively guiding the correction of undesirable behaviors in subsequent driving and improving the efficiency of behavior improvement.

[0021] Secondly, by integrating multiple attribute features for joint ranking, the system improves the rationality of content recommendations and the quality of information distribution. When generating the recommendation list, the system comprehensively calculates scores from multiple dimensions, including risk matching degree, difficulty matching degree, content freshness, correction gain, and source priority, and dynamically adjusts the amount of material based on the severity and frequency of dangerous driving behaviors. This not only avoids excessive duplication of similar safety content, ensuring the timeliness and diversity of the content, but also ensures that the quantity and type of distributed content match the actual needs of drivers.

[0022] Third, it achieves an adaptive match between the difficulty of safety education content and the individual driver's skill level. By introducing stored driving level information as a constraint during the content selection stage, it ensures that the difficulty of the recalled content is appropriate to the driver's individual experience. This effectively avoids the problem of novice drivers struggling to understand overly complex professional content, or experienced drivers becoming fatigued by overly simplistic general content, thereby significantly improving the user's learning completion rate and knowledge absorption efficiency.

[0023] Fourth, by introducing parameter updates linked to user feedback and subsequent behavioral trends, the long-term stability of system strategy adjustments is ensured. This invention collects user feedback metrics after learning and, upon entering the next statistical cycle, combines this feedback with updated driving behavior statistics, dynamically influencing parameters at each level of recall, ranking, and quantity allocation. This mechanism of adjusting input strategies based on execution results overcomes the shortcomings of traditional one-way push solutions that cannot self-adjust based on usage effects, achieving periodic parameter iteration and guaranteeing behavioral improvement effects over long-term application. Attached Figure Description

[0024] Figure 1This is a flowchart of the safety education recommendation method in Embodiment 1 of the present invention; Figure 2 This is a diagram illustrating the overall architecture of the safety education recommendation in Embodiment 2 of the present invention. Figure 3 This is a flowchart of the method in Embodiment 2 of the present invention; Figure 4 This is a timing diagram of the execution process in Embodiment 2 of the present invention; Figure 5 This is a schematic diagram of the device structure in Embodiment 3 of the present invention. Detailed Implementation Example

[0025] like Figure 1 As shown, this embodiment provides a safety education recommendation method based on driving behavior, executed by a computer in an in-vehicle terminal. The method includes the following steps: Step 1: Collect current vehicle driving data. In this step, by acquiring the underlying multi-dimensional physical parameters of vehicle driving in real time, a high-fidelity and continuous basic data stream can be provided to support the subsequent refined and quantitative assessment of the user's individual driving status.

[0026] Step 2: Perform undesirable driving behavior detection processing on the collected driving data, identify risky behavior events based on a preset rule engine, and obtain a set of undesirable driving behavior events. In this step, the rule engine discretizes the multidimensional continuous physical parameter data stream and converts it into well-defined risky operation events, which can capture the abnormal weaknesses of the driver in the driving operation process in a timely and accurate manner.

[0027] Step 3: Perform statistical processing on the set of adverse driving behavior events at both the single-trip statistical granularity and the weekly report statistical granularity to obtain the frequency and severity information of each behavior type, and construct a behavior feature vector accordingly. In this step, by introducing a dual statistical perspective with different time spans, the sudden and occasional risks of a single trip can be combined with the periodic and regular risks, so that the final constructed behavior feature vector can comprehensively depict the overall risk profile of an individual driver from both breadth and depth.

[0028] Step 4: Based on the behavioral feature vector, extract recommended risk features from the driving report data to obtain an ordered risk label set. The extraction process includes: sorting each behavior type at the single trip statistical granularity in descending order of frequency of occurrence, selecting the top 3 behavior types, and mapping them to generate a risk label group for the most recent trip, denoted as the TOP3 label group; sorting each behavior type at the weekly report statistical granularity in descending order of frequency of occurrence, selecting the top 5 behavior types, and mapping them to generate a weekly report risk label group, denoted as the TOP5 label group; combining the TOP3 label group and the TOP5 label group, removing duplicates, and... The risk labels are sorted according to a preset priority rule to obtain an ordered risk label set. The priority of each label in the TOP3 label group is higher than that of the labels in the TOP5 label group. When the same behavior type appears in two groups at the same time, the ranking corresponding to the TOP3 label group is retained in the ordered risk label set. In this step, by determining the deduplication and recombination rule that the risk of the most recent trip is superior to the risk of the weekly report, the technical effect of prioritizing the correction of recent weaknesses is achieved. This ensures that the final risk label set can be arranged in order of risk urgency from high to low, thereby greatly improving the targeting and correction guidance of subsequent information recommendations.

[0029] Step 5: Based on the ordered risk label set, candidate content is recalled sequentially from the preset safety education content library according to the label order to obtain a candidate safety education content set. Each safety education content matches the corresponding risk label in the ordered risk label set, and the content difficulty matches the stored driving level information. Content recalled corresponding to the TOP3 label group is marked as high-priority content in the candidate set. In this step, by automatically establishing a precise match between risk labels and content labels, and introducing the stored driving level information as a hard difficulty constraint for content recall, it is possible to adaptively implement a refined stratification of content comprehension difficulty for drivers at different experience levels. This effectively avoids the problem of decreased learning efficiency caused by novice drivers receiving overly difficult content or experienced drivers receiving overly simple content. At the same time, by marking high-priority content, an urgency feature foundation is laid for subsequent joint ranking.

[0030] Step Six: Calculate the comprehensive score for each item in the candidate safety education content set, and sort them from high to low according to the comprehensive score to obtain the sorted recommended content list; wherein, for high-priority content from the TOP3 tag group, a source priority bonus is added to the comprehensive score calculation; in this step, by adding a specific source priority bonus to high-priority content in the multi-dimensional scoring, the final calculated recommended sequence can form a strong correlation with the driver's recent urgent safety defects in the ranking result, ensuring that the most dangerous and urgent recent weakness correction materials can be stably ranked at the front of the final list.

[0031] Step 7: Based on the severity and frequency of dangerous driving behaviors corresponding to each risk label in the ordered risk label set, determine the weight of the recommended materials corresponding to each risk label, and select recommended materials from the sorted recommended content list accordingly to generate a list of recommendations to be displayed. In this step, by combining the severity and frequency of dangerous driving behaviors to determine the weight of the material allocation, fine-grained control of on-demand and quantity-based distribution is achieved, ensuring that high-risk and high-frequency weak points can obtain a larger proportion of content presentation quota, avoiding the inefficient push of blind and equal distribution in traditional solutions.

[0032] Step 8: Display the safety education content from the recommended list in the "Recommended for You" display area of ​​the user terminal according to the order of the ordered risk tag set, so that the content corresponding to the TOP3 tag group is placed first; and complete the push or update display when the vehicle is in a low-interference state; in this step, by forcing the correction content corresponding to the most recent trip to be visually absolutely first in the spatial arrangement, the consistency between the interface display order and the urgency of risks is maintained, which significantly improves the hit rate of learning content; at the same time, limiting the push to a low-interference state can prevent the distraction of driving caused by multimedia information pop-ups, and ensure the active safety of vehicle interaction.

[0033] Step 9: Collect learning feedback metrics after users learn the safety education content in the recommended list to be displayed, and obtain a set of learning feedback metrics. In this step, by quantitatively capturing users' real learning performance and behavior, the system can be provided with objective data support to judge the quality of recommended content and user acceptance and compliance, laying a data foundation for subsequent parameter calibration iterations.

[0034] Step 10: Based on the set of learning feedback indicators, and combined with the latest trip statistics and weekly report statistics obtained in the next statistical period, dynamically adjust the recommendation strategy parameters to update the recommendation strategy for subsequent periods, and return to Step 1 to continue execution. In this step, by obtaining actual learning feedback and combining it with the behavioral evolution trend of the next period to dynamically adjust the algorithm parameters in reverse, a feedback loop structure for adaptive dynamic adjustment of parameters is constructed. This overcomes the limitation of traditional push strategies that cannot adjust strategies based on long-term behavioral improvement effects, ensuring that safe recommendations in subsequent periods can achieve self-evolution.

[0035] As one implementation method, the driving data in this embodiment includes at least vehicle speed, longitudinal acceleration, lateral acceleration, steering wheel angle change rate, following distance, turn signal status, and gear status. The above-mentioned diverse underlying sensor physical parameters can seamlessly cover multiple core control dimensions such as vehicle dynamics, spatial distance, and human control intentions, and can provide sufficient feature source data for the upper layer to finely identify various complex bad driving actions.

[0036] As one implementation method, the risky behavior events described in this embodiment include at least one or more of the following: sudden braking, sudden acceleration, sudden steering, speeding, following too closely, failure to use turn signals, continuous lane changes, incorrect gear selection, and failure to decelerate in areas requiring deceleration. The above-mentioned subdivided types of poor driving comprehensively cover typical high-risk driving hazards in urban and complex driving conditions, thereby assisting the system in establishing a comprehensive diagnostic and correction database for individual drivers' poor driving habits.

[0037] As one implementation method, the formula for calculating the comprehensive score in step six of this embodiment is: Comprehensive Score = Risk Matching Degree × w1 + Difficulty Matching Degree × w2 + Content Freshness × w3 + Correction Gain × w4 + Source Priority × w5, where w1, w2, w3, w4, and w5 are weight parameters. The calculation logic of the comprehensive score is as follows: multiply the risk matching degree, difficulty matching degree, content freshness, correction gain, and source priority by their respective preset weight parameters and then sum them up. This joint ranking formula deeply integrates information from multiple sources, such as matching accuracy, individual experience, freshness to prevent duplicate pushes, expected improvement benefits from subsequent behavior, and time urgency, into a single indicator. By adjusting the weight parameters, recommendation preferences can be flexibly switched, significantly improving the scientific nature, timeliness, and scalability of multi-information flow joint recommendation.

[0038] As one implementation method, step seven in this embodiment, which involves determining the recommended material quantity weight corresponding to each risk label and selecting recommended materials from the sorted recommended content list accordingly to generate a list of recommendations to be displayed, includes: the higher the severity and frequency of occurrence, the greater the recommended material quantity weight allocated to the corresponding risk label; the mapping relationship between the recommended material quantity weight and the specific number of displayed items is a configurable strategy parameter; this setting organically combines the severity of the hazard source with the high frequency of occurrence, which can prevent single-dimensional, occasional data from interfering with the system's quota judgment, and make the recommended content quota accurately and scientifically tilted towards high-risk, high-frequency key weaknesses.

[0039] As one implementation method, the configuration strategy for selecting recommended materials from the sorted list of recommended content and generating the list of recommended materials to be displayed in this embodiment includes one or more of the following methods: Method 1: Let the weight of the recommended material quantity for the i-th risk tag be Qi = α × Si + β × Ci, where Si is the severity score of the behavior corresponding to the tag, Ci is the number of occurrences, and α and β are configuration coefficients; allocate the total number of recommended items N according to the proportion of Qi to the sum of the weights of all tags to obtain the number of items displayed under each tag ni; For specific risk tags, the weighting of recommended content quantity is calculated as follows: the severity score and frequency of occurrence of the behavior corresponding to the tag are multiplied by the corresponding configuration coefficient and then summed; the total number of recommended items N is allocated according to the proportion of this weight to the sum of the weights of all tags, resulting in the number of items displayed under each tag, ni. This rigorous linear coupling formula determines the quantity allocation weight, allowing the number of displayed content items to smoothly adapt to the specific weight ratio in a continuously adjustable manner, achieving precise mathematical allocation of recommendation capacity.

[0040] Method 2: The recommended material quantity weight is divided into three levels: high, medium, and low. The high level is allocated more recommended items, the medium level is allocated a medium number of recommended items, and the low level is allocated fewer recommended items. The corresponding number of items for each level is preset by the operation strategy. The hierarchical mapping scheme with tiered dimensionality reduction effectively simplifies the online calculation logic, making it easier for the vehicle terminal to quickly complete the threshold matching and lightweight update of the number of displayed items in a low computing power environment.

[0041] Method 3: Allocate at least one recommended material to each tag in the TOP3 tag group, and distribute the remaining materials among the other tags in the TOP5 tag group; a strong priority guarantee line is established to ensure that the three most urgent core security vulnerabilities in the current process can all get exposure opportunities with at least one material, completely eliminating the technical risk that recent pain points will be diluted and submerged due to the large base of weekly report tag data.

[0042] As one implementation method, the low-interference state mentioned in this embodiment includes a parked state or a trip-end state, and the form of the safety education content includes text, images, or videos. By limiting the push and update timing to a static state when not driving, not only are the driving distraction hazards caused by information pop-ups during driving eliminated from a safety perspective, but the driver's brief review time after the trip can also be fully utilized to significantly improve the complete reading rate and compliance of the text, images, or videos of safety education content.

[0043] As one implementation method, the driving report data in this embodiment includes a "Recent Trip" view and a "Weekly Report" view when presented on the user interface. The "Recent Trip" view displays the statistical results of dangerous driving behaviors at the single trip statistical granularity and the TOP3 tag group, while the "Weekly Report" view displays the statistical results of dangerous driving behaviors at the weekly report statistical granularity and the TOP5 tag group. The "Recommended for You" display area of ​​the user terminal is used to present the list of recommendations to be displayed output in step eight. By separating and presenting two statistical granularity sub-views with differences in time span and spatial dimension on the human-computer interaction user interface, and simultaneously coordinating them with an orderly list of recommendations to be displayed in the golden display area, not only is the readability and information hierarchy of the in-vehicle interaction interface improved, but the driver is also guided to establish a clear cognitive logic from reviewing recent short pain points to comparing the evolution of long-term habits in terms of visual cognition.

[0044] As one implementation method, the dynamic adjustment of recommendation strategy parameters in step ten of this embodiment includes: dynamically adjusting the weight parameters w1 to w5 in step six, the weight mapping relationship of the recommended material quantity in step seven, and the recall strategy in step five. This multi-objective joint tuning setting accurately decomposes the parameter update instructions and links them in reverse to three independent technical modules such as content recall, multi-dimensional ranking, and quantity allocation, ensuring that the system can achieve linkage coordination and efficient self-adaptation of various control nodes after receiving external feedback. Example

[0045] This embodiment is a further supplementary description of Embodiment 1, providing a safety education recommendation method based on driving behavior, executed by a computer in an in-vehicle terminal, and applicable to in-vehicle terminal software systems; the computer executes the following steps sequentially according to a preset program.

[0046] like Figure 2As shown, the computer's internal structure is functionally divided into a data acquisition module, a detection module, a profiling module, a recall module, a sorting module, an allocation module, a push module, a feedback module, and an update module. The main data flow is as follows: data flows from the acquisition module through the detection module, profiling module, recall module, sorting module, and allocation module, finally reaching the push module. The feedback flow is as follows: data flows from the push module through the feedback module to the update module, and is ultimately fed back to the recall module, sorting module, and allocation module by the update module. The update module adjusts the recall, sorting, and allocation strategy parameters based on the learning feedback indicators output by the feedback module to achieve closed-loop optimization. Based on the data flow characteristics shown in the attached diagram, in the main data link, the acquisition module executes step one and transmits the data to the detection module to execute step two. After the detection module outputs the risk events, it uses the top 3 most recent trips in the driving report and the top 5 in the weekly report as feature inputs to the profiling module for executing steps three and four. Subsequently, the output of the profiling module is sequentially sent to the recall module to execute step five, the sorting module to execute step six, and the allocation module to execute step seven. Finally, the push module executes step eight, displaying the safety education content on the "Recommended for You" interface. In the feedback link, after the push module completes the display, the feedback module executes step nine to collect data and outputs the learning feedback indicators to the update module. The update module executes step ten based on the learning feedback indicators output by the feedback module, radially writing back the strategies to the recall module, sorting module, and allocation module along the dotted line in the attached diagram to adjust the recall, sorting, and allocation strategy parameters, thereby achieving closed-loop optimization.

[0047] like Figure 3 As shown, when the computer executes the recommended method of this invention, each step is executed sequentially, and the input data of the subsequent step comes from the output result of the previous step. A cyclical mechanism oriented towards the next statistical period is formed through the path shown by the dashed line in the attached figure, specifically including: Step 1: The computer collects the current vehicle's driving data, which includes at least: vehicle speed, longitudinal acceleration, lateral acceleration, steering wheel angle change rate, following distance, turn signal status, and gear status. Step 2: The computer performs undesirable driving behavior detection processing on the driving data collected in Step 1, identifies risky behavior events based on a preset rule engine, and obtains a set of undesirable driving behavior events; the risky behavior events include at least one or more of the following: sudden braking, sudden acceleration, sudden steering, speeding, following too closely, not using turn signals, continuous lane changes, incorrect gear selection, and failure to decelerate in areas where deceleration is required. Step 3: The computer performs statistical processing on the set of bad driving behavior events obtained in Step 2 according to different statistical granularities to obtain the occurrence frequency and severity information of each behavior type, and constructs a behavior feature vector accordingly; the statistical granularity includes at least: the statistical granularity of a single trip for the driver's most recent completed trip, and the statistical granularity of a weekly report for a preset weekly reporting period.

[0048] Step 4: The computer extracts recommended risk features from the driving report data based on the behavioral feature vector obtained in Step 3, and obtains an ordered risk label set. The extraction process includes: (a) sorting each behavior type under the single trip statistical granularity in descending order of the number of occurrences, selecting the top 3 behavior types, and mapping them to generate the risk label group for the most recent trip, denoted as the TOP3 label group; (b) sorting each behavior type under the weekly report statistical granularity in descending order of the number of occurrences, selecting the top 5 behavior types, and mapping them to generate the weekly report risk label group, denoted as the TOP5 label group; (c) combining the TOP3 label group and the TOP5 label group, removing duplicates, and sorting them according to a preset priority rule to obtain an ordered risk label set. The priority rule is: the priority of each label in the TOP3 label group is higher than that of the labels in the TOP5 label group, and when the same behavior type appears in both groups, the ranking corresponding to the TOP3 label group is retained in the ordered risk label set.

[0049] Step 5: The computer, based on the ordered risk tag set obtained in Step 4, sequentially performs candidate content recall processing from the preset safety education content library according to the tag order, to obtain a candidate safety education content set; the recall processing shall at least satisfy the following: each safety education content matches the corresponding risk tag in the ordered risk tag set, and the content difficulty matches the stored driving level information; wherein, the content recalled corresponding to the TOP3 tag group is marked as high-priority content in the candidate set.

[0050] Step Six: The computer calculates a comprehensive score for each item in the candidate safety education content set obtained in Step Five, and sorts them from highest to lowest comprehensive score to obtain a sorted list of recommended content. For high-priority content from the TOP3 tag group, a source priority bonus is added to the comprehensive score calculation. The formula for calculating the comprehensive score is: Comprehensive Score = Risk Matching Degree × w1 + Difficulty Matching Degree × w2 + Content Freshness × w3 + Correction Gain × w4 + Source Priority × w5, where w1, w2, w3, w4, and w5 are weighting parameters.

[0051] Step 7: The computer determines the recommended material quantity weight corresponding to each risk label based on the severity and frequency of dangerous driving behavior corresponding to each risk label in the ordered risk label set in Step 4, and selects recommended materials from the sorted recommended content list obtained in Step 6 accordingly to generate a list of recommendations to be displayed; wherein, the higher the severity and frequency of occurrence, the greater the recommended material quantity weight allocated to the corresponding risk label; the mapping relationship between the recommended material quantity weight and the specific number of displayed items is a configurable strategy parameter. Regarding the allocation of the number of recommended materials in Step 7, the present invention provides the following configurable embodiments. The above embodiments can be used individually or in combination, and the specific details can be adjusted and optimized according to product operation needs, and are not limited to the following methods: Example A (Proportional Allocation): Let the weight of the recommended material quantity for the i-th risk label be Qi = α × Si + β × Ci, where Si is the severity score of the behavior corresponding to the label, Ci is the number of occurrences, and α and β are configuration coefficients; allocate the total number of recommended items N according to the proportion of Qi to the sum of the weights of all labels, and obtain the number of items displayed under each label ni.

[0052] Example B (Tiered Allocation): Qi is divided into three tiers: high, medium, and low. High-tier items are allocated more recommendations, medium-tier items are allocated a medium number of recommendations, and low-tier items are allocated fewer recommendations. The threshold number of items for each tier is preset by the operation strategy.

[0053] Example C (Priority Guarantee): Allocate at least one recommended material to each tag in the TOP3 tag group, and distribute the remaining materials to the remaining tags in the TOP5 tag group according to Example A or Example B.

[0054] Step 8: The computer displays the safety education content from the recommended list obtained in Step 7 in the "Recommended for You" display area of ​​the user terminal, arranged in the order of the ordered risk tag set. The "Recommended for You" display area of ​​the safety popularization module is used to present the recommended list output in this step, placing the content corresponding to the TOP3 tag group first. Push or update the display when the vehicle is in a low-interference state, including a parked state or a trip-end state. The safety education content can be in the form of text, images, or videos. In conjunction with the product interface embodiment, the driving report module includes a "Recent" trip view and a "Weekly Report" view. The "Recent" view displays the statistical results of dangerous driving behaviors and the TOP3 tag group at the single trip statistical granularity, while the "Weekly Report" view displays the statistical results of dangerous driving behaviors and the TOP5 tag group at the weekly report statistical granularity.

[0055] Step 9: The computer collects learning feedback metrics after the user learns the safety education content in the recommended list to be displayed, and obtains a set of learning feedback metrics.

[0056] Step 10: The computer dynamically adjusts the weight parameters w1 to w5 in Step 6, the weight mapping relationship of the recommended material quantity in Step 7, and the recall strategy in Step 5 based on the learning feedback indicator set obtained in Step 9, and in combination with the latest trip statistics and weekly report statistics obtained from Steps 2 to 4 in the next statistical period, so as to update the recommendation strategy for subsequent periods.

[0057] like Figure 4 As shown, the execution process of this invention involves the temporal interaction of multiple system components. Within a complete statistical period, the data source, detection module, profiling module (i.e., providing driving reports), recommendation engine, content library, and in-vehicle interface (recommended for you) within the system interact sequentially according to the following nodes: First, the data source reports vehicle data to the detection module; second, after processing, the detection module outputs risk events to the profiling module; third, the profiling module extracts and generates TOP3 and TOP5 ordered risk tags and sends them to the recommendation engine; fourth, the recommendation engine requests candidate content from the content library based on the tags; fifth, after matching the difficulty with the tags, the content library returns the candidate list to the recommendation engine; sixth, the recommendation engine internally performs sorting, allocation, and push processing sequentially, and sends the list to be displayed to the in-vehicle interface for presentation; seventh, after the user completes the learning, the in-vehicle interface sends learning feedback back to the recommendation engine; eighth, after receiving the feedback, the recommendation engine performs the operation of updating strategy parameters, thereby forming a closed loop of strategy iteration.

[0058] In combination with the above Figure 4 The timing logic shown is as follows: within a statistical period, the computer first executes steps one through eight to complete feature extraction, content sorting, and updating the "Recommended for You" display; after the user learns, step nine is executed to collect feedback; before the start of the next statistical period, step ten is executed to update the strategy parameters, and the process is repeated from step one to step eight, thus forming a periodic closed loop. Example

[0059] This embodiment provides a safety education recommendation device based on driving behavior, deployed in a computer within an in-vehicle terminal. The device includes: The data acquisition module is configured to execute step one, which is used to collect the current vehicle's driving data; The detection module is configured to perform step two, which is used to perform bad driving behavior detection processing on the collected driving data, identify risky behavior events based on a preset rule engine, and obtain a set of bad driving behavior events. The profiling module is configured to execute steps three to four, which are used to construct behavioral feature vectors and extract recommended risk features from driving report data to obtain an ordered set of risk labels. The recall module is configured to execute step five, which is used to sequentially recall candidate content from the preset safety education content library according to the ordered risk label set; The sorting module is configured to execute step six, which is used to calculate the comprehensive score of each item in the candidate safety education content set and sort them to obtain a sorted list of recommended content. The allocation module is configured to execute step seven, which is used to determine the weight of the number of recommended materials corresponding to each risk label and select recommended materials to generate a list of recommendations to be displayed. The push module is configured to execute step eight, which is used to arrange and display the safety education content in the recommended list to be displayed in the "Recommended for You" display area and complete the push when the vehicle is in a low-interference state. The feedback module is configured to execute step nine, which is used to collect learning feedback metrics after the user's learning and obtain a set of learning feedback metrics. The update module is configured to execute step ten, which is used to dynamically update and adjust the strategy parameters in the recall module, sorting module and allocation module based on the learning feedback indicator set and the data of the next statistical period. All of the above modules are deployed on the vehicle terminal, or in a way that combines the vehicle terminal with the cloud server.

[0060] like Figure 5 As shown, Figure 5 This is a system device diagram for implementing the recommended method described above in Embodiment 3 of the present invention. Referring to the module names, configured execution steps, and data flow identifiers shown in the accompanying drawings, the physical or logical connection relationships and data interaction methods of the internal functional modules of the system device deployed in the vehicle terminal are as follows: The main data link of the system device is connected in series in the direction shown by the solid arrows in the attached figure: the output of the acquisition module (configured to execute step one) is connected to the input of the detection module (configured to execute step two); the output of the detection module is connected to the input of the profiling module (configured to execute steps three and four); the output of the profiling module is connected to the input of the recall module (configured to execute step five); the output of the recall module is connected to the input of the sorting module (configured to execute step six); the output of the sorting module is connected to the input of the allocation module (configured to execute step seven); and the output of the allocation module is connected to the input of the push module (configured to execute step eight), thereby completing the forward data flow from data acquisition to interface display.

[0061] Furthermore, the system device is also configured with a feedback link for implementing strategy iteration: the feedback module (configured to execute step nine) is located after the push module in the logical flow, and its output is connected to the input of the update module (configured to execute step ten). Figure 5As shown by the dashed arrow path in the lower middle, the output of the update module leads out a reverse data transmission path labeled "feedback write-back". This dashed path crosses back to the recall module, which is used to indicate that the dynamically updated and adjusted strategy parameters are written back to the recall module (and the sorting module and allocation module for coordinated adjustment), thereby forming a control closed loop with parameter adaptive adjustment capability in the underlying architecture of the system device.

[0062] Regarding the specific hardware implementation and deployment of the above system devices, all modules can be centrally deployed in the vehicle-side computing device to ensure its real-time performance, or a distributed architecture can be adopted in which vehicle-side computing nodes and cloud server nodes are deployed collaboratively.

[0063] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A safety education recommendation method based on driving behavior, characterized in that, The method, executed by a computer within the vehicle-mounted terminal, includes the following steps: Step 1: Collect the current vehicle's driving data; Step 2: Perform undesirable driving behavior detection processing on the collected driving data, identify risky behavior events based on a preset rule engine, and obtain a set of undesirable driving behavior events; Step 3: Perform statistical processing on the set of bad driving behavior events at the single trip statistical granularity and the weekly report statistical granularity to obtain the occurrence frequency and severity information of each behavior type, and construct a behavior feature vector accordingly; Step 4: Based on the behavioral feature vector, extract recommended risk features from the driving report data to obtain an ordered risk label set. The extraction process includes: sorting each behavior type under the single trip statistical granularity in descending order of occurrence frequency, selecting the top 3 behavior types, and mapping them to generate the risk label group for the most recent trip, denoted as the TOP3 label group; sorting each behavior type under the weekly report statistical granularity in descending order of occurrence frequency, selecting the top 5 behavior types, and mapping them to generate the weekly report risk label group, denoted as the TOP5 label group; merging the TOP3 label group and the TOP5 label group to remove duplicates, and sorting them according to a preset priority rule to obtain an ordered risk label set, wherein the priority of each label in the TOP3 label group is higher than that of the labels in the TOP5 label group, and when the same behavior type appears in both groups, the ranking corresponding to the TOP3 label group is retained in the ordered risk label set. Step 5: Based on the ordered risk tag set, perform candidate content recall processing sequentially from the preset safety education content library according to the tag order to obtain a candidate safety education content set. Each safety education content matches the corresponding risk tag in the ordered risk tag set, and the content difficulty matches the stored driving level information. The content recalled corresponding to the TOP3 tag group is marked as high priority content in the candidate set. Step 6: Calculate the comprehensive score for each item in the candidate safety education content set, and sort them from high to low according to the comprehensive score to obtain the sorted recommended content list; among them, for high-priority content from the TOP3 tag group, the source priority bonus is added to the comprehensive score calculation; Step 7: Based on the severity and frequency of dangerous driving behaviors corresponding to each risk label in the ordered risk label set, determine the weight of the recommended materials corresponding to each risk label, and select recommended materials from the sorted recommended content list accordingly to generate a recommended list to be displayed. Step 8: Arrange the safety education content in the recommended list to be displayed in the "Recommended for You" display area of ​​the user terminal according to the order of the ordered risk tag set, so that the content corresponding to the TOP3 tag group is placed first; and complete the push or update display when the vehicle is in a low interference state. Step 9: Collect learning feedback metrics after users have studied the safety education content in the recommended list to be displayed, and obtain a set of learning feedback metrics; Step 10: Based on the set of learning feedback indicators, and combined with the latest trip statistics and weekly report statistics obtained in the next statistical period, dynamically adjust the recommendation strategy parameters to update the recommendation strategy for subsequent periods, and return to Step 1 to continue execution.

2. The safety education recommendation method based on driving behavior according to claim 1, characterized in that, The driving data includes at least vehicle speed, longitudinal acceleration, lateral acceleration, steering wheel angle change rate, following distance, turn signal status, and gear status.

3. The safety education recommendation method based on driving behavior according to claim 1, characterized in that, The risky behavioral events include at least one or more of the following: sudden braking, sudden acceleration, sudden steering, speeding, following too closely, not using turn signals, continuous lane changes, incorrect gear selection, and failure to decelerate in areas where deceleration is required.

4. The safety education recommendation method based on driving behavior according to claim 1, characterized in that, The formula for calculating the comprehensive score mentioned in step six is: Comprehensive score = Risk matching degree × w1 + Difficulty matching degree × w2 + Content freshness × w3 + Correction gain × w4 + Source priority × w5, where w1, w2, w3, w4, and w5 are weight parameters.

5. The safety education recommendation method based on driving behavior according to claim 1, characterized in that, Step seven involves determining the weight of the recommended material quantity corresponding to each risk label, and selecting recommended materials from the sorted list of recommended content accordingly to generate a list of recommendations to be displayed. This includes: the higher the severity and frequency of occurrence, the greater the weight of the recommended material quantity assigned to the corresponding risk label; the mapping relationship between the weight of the recommended material quantity and the specific number of items to be displayed is a configurable strategy parameter.

6. The safety education recommendation method based on driving behavior according to claim 5, characterized in that, The configuration strategy for selecting recommended materials from the sorted list of recommended content to generate the list of recommended items to be displayed includes one or more of the following methods: Method 1: Let the weight of the recommended material quantity for the i-th risk tag be Qi = α × Si + β × Ci, where Si is the severity score of the behavior corresponding to the tag, Ci is the number of occurrences, and α and β are configuration coefficients; allocate the total number of recommended items N according to the proportion of Qi to the sum of the weights of all tags to obtain the number of items displayed under each tag ni; Method 2: Divide the recommended materials into three weighted levels: high, medium, and low. High-level materials are allocated more recommended materials, medium-level materials are allocated a medium number of recommended materials, and low-level materials are allocated fewer recommended materials. The threshold for the number of materials in each level is preset by the operation strategy. Method 3: Allocate at least one recommended material to each tag in the TOP3 tag group, and distribute the remaining materials among the other tags in the TOP5 tag group.

7. The safety education recommendation method based on driving behavior according to claim 1, characterized in that, The low-interference state includes the parking state or the end of the trip state, and the form of the safety education content includes text, images or videos.

8. The safety education recommendation method based on driving behavior according to claim 1, characterized in that, The driving report data is presented in the user interface as a "Recent Trip" view and a "Weekly Report" view; The "Recent" view displays the statistical results of dangerous driving behavior at the single trip statistical granularity and the TOP3 tag group, while the "Weekly Report" view displays the statistical results of dangerous driving behavior at the weekly report statistical granularity and the TOP5 tag group. The "Recommended for You" display area on the user terminal is used to present the list of recommendations to be displayed output in step eight.

9. A method for recommending safety education based on driving behavior according to claim 4 or 5, characterized in that, The dynamic adjustment of the recommendation strategy parameters in step ten includes: dynamically adjusting the weight parameters w1 to w5 in step six, the weight mapping relationship of the recommended material quantity in step seven, and the recall strategy in step five.

10. A safety education recommendation device based on driving behavior, characterized in that, The device, deployed in a computer within a vehicle-mounted terminal, includes: The data acquisition module is configured to execute step one, which is used to collect the current vehicle's driving data; The detection module is configured to perform step two, which is used to perform bad driving behavior detection processing on the collected driving data, identify risky behavior events based on a preset rule engine, and obtain a set of bad driving behavior events. The profiling module is configured to execute steps three to four, which are used to construct behavioral feature vectors and extract recommended risk features from driving report data to obtain an ordered set of risk labels. The recall module is configured to execute step five, which is used to sequentially recall candidate content from the preset safety education content library according to the ordered risk label set; The sorting module is configured to execute step six, which is used to calculate the comprehensive score of each item in the candidate safety education content set and sort them to obtain a sorted list of recommended content. The allocation module is configured to execute step seven, which is used to determine the weight of the number of recommended materials corresponding to each risk label and select recommended materials to generate a list of recommendations to be displayed. The push module is configured to execute step eight, which is used to arrange and display the safety education content in the recommended list to be displayed in the "Recommended for You" display area and complete the push when the vehicle is in a low-interference state. The feedback module is configured to execute step nine, which is used to collect learning feedback metrics after the user's learning and obtain a set of learning feedback metrics. The update module is configured to execute step ten, which is used to dynamically update and adjust the strategy parameters in the recall module, sorting module and allocation module based on the learning feedback indicator set and the data of the next statistical period. All of the above modules are deployed on the vehicle terminal, or in a way that combines the vehicle terminal with the cloud server.

Images