Driving behavior data analysis method and system for premium management

Abstract

The present application provides a driving behavior data analysis method and system for premium management. By identifying each driving behavior event included in the user driving behavior data set to be analyzed, the driving behavior event identification result is obtained, and an event feature description array of each driving behavior event is generated. The behavior sequence relationship between each driving behavior event included in the user driving behavior data set to be analyzed is extracted based on the event feature description array of each driving behavior event to obtain a driving behavior event relationship description array. Based on two features, whether the user driving behavior data set matches the reference driving behavior data is identified, thereby obtaining a driving behavior pattern and formulating a premium strategy. The present application expresses the driving behavior data more accurately based on the event feature description array of the driving behavior event and the driving behavior event relationship description array, thereby improving the accuracy of matching and identification, helping to obtain a more accurate driving behavior pattern, and formulating a more accurate premium strategy.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a driving behavior data analysis method and system for premium management. Background Art

[0002] In the insurance industry, premium management strategies targeting driving behavior are gaining increasing attention. With the rapid development of intelligent transportation systems and in-vehicle sensor technology, insurance companies are able to collect a large amount of data on users' driving behavior. This data makes it possible to personalize premium pricing, but it also brings challenges to data processing and analysis.

[0003] Traditional premium management strategies often rely on static factors, such as vehicle type, driver age, and driving experience. Although these factors reflect driving risks to a certain extent, they cannot fully capture the dynamic behavioral characteristics of users in actual driving. In recent years, with the rise of big data and machine learning technologies, insurance companies have begun to try to use driving behavior data to formulate more accurate premium strategies. However, how to effectively extract and utilize key information from these data to accurately assess users' driving risks remains an urgent problem to be solved.

[0004] In the existing technology, the analysis of driving behavior data mainly focuses on the identification of single driving behavior events, such as acceleration, braking, turning, etc. Although these technologies can identify which driving behaviors the user has performed, they lack in-depth analysis of the relationship between behaviors. Driving behavior is a complex process, and there are often temporal correlations and logical dependencies between different behaviors. Analysis based solely on a single behavior event cannot fully reflect the user's driving habits and risk level.

[0005] In addition, when matching driving behavior patterns, existing technologies usually only consider the event features themselves, while ignoring the behavioral sequence relationship between events. This matching method often leads to inaccurate recognition results and cannot effectively distinguish the subtle differences between different users. Summary of the invention

[0006] In view of this, the embodiment of the present application at least provides a driving behavior data analysis method and system for premium management. The technical solution of the present application is implemented as follows:

[0007] On the one hand, the present application provides a driving behavior data analysis method for premium management, comprising: obtaining multiple groups of driving behavior collection data sets collected during a user's driving process during an evaluation period, and using one group of driving behavior collection data sets as a user driving behavior data set to be analyzed; identifying each driving behavior event included in the user driving behavior data set to be analyzed, and obtaining a driving behavior event recognition result; generating an event feature description array of each driving behavior event based on the driving behavior event recognition result; extracting a behavior sequence relationship between each driving behavior event included in the user driving behavior data set to be analyzed based on the event feature description array of each driving behavior event, and obtaining a driving behavior event relationship. description array; identifying whether the user driving behavior data set to be analyzed matches the reference driving behavior data in the reference behavior pattern data set based on the event feature description array and the driving behavior event relationship description array; if the reference driving behavior data is matched, taking the driving behavior pattern corresponding to the reference driving behavior data as the driving behavior pattern for the group of driving behavior collection data sets; analyzing the remaining groups of driving behavior collection data sets to obtain the driving behavior pattern corresponding to each group of driving behavior collection data sets, and weightedly fusing the driving behavior patterns corresponding to the multiple groups of driving behavior collection data sets to obtain a fused driving behavior pattern; generating a corresponding premium management strategy based on the fused driving behavior pattern.

[0008] On the other hand, the present application provides a computer system, including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the above method when executing the program.

[0009] In an embodiment of the present application, an event feature description array of each driving behavior event is generated based on the driving behavior event recognition result, and the behavior sequence relationship between each driving behavior event included in the user driving behavior data set to be analyzed is extracted based on the event feature description array of each driving behavior event to obtain a driving behavior event relationship description array, and then based on the event feature description array of each driving behavior event and the driving behavior event relationship description array, it is identified whether the user driving behavior data set to be analyzed matches the reference driving behavior data in the reference behavior pattern data set. In this way, when mining data features to complete matching and identification, not only the event feature description array of each driving behavior event is considered, but also the behavior sequence relationship between each driving behavior event is considered. In this way, the driving behavior data can be more accurately expressed based on the event feature description array of the driving behavior event and the driving behavior event relationship description array, thereby improving the accuracy of matching and identification, helping to obtain a more accurate driving behavior pattern, and formulating a more accurate premium strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A schematic diagram of the implementation flow of a driving behavior data analysis method for premium management provided in an embodiment of the present application.

[0011] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0012] The embodiment of the present application provides a driving behavior data analysis method for premium management, which can be executed by a processor of a computer system, wherein the computer system can refer to a device with data processing capabilities such as a server, a laptop, a tablet computer, and a desktop computer.

[0013] Figure 1 A schematic diagram of the implementation flow of a driving behavior data analysis method for premium management provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method includes the following operations 110 to 150:

[0014] Step S100: obtaining multiple sets of driving behavior data sets collected during the user's driving process during the evaluation period, and taking one set of driving behavior data sets as the user's driving behavior data set to be analyzed.

[0015] In step S100, the evaluation period is a time range used to define the period of time for collecting driving behavior data to be analyzed. For example, this period can be one month, one quarter, or one year, depending on the needs of the premium management strategy and the purpose of data analysis. Clarifying the evaluation period helps ensure the timeliness and relevance of the data, so that the analysis results can accurately reflect the user's driving behavior characteristics in a specific time period.

[0016] After determining the evaluation period, the computer system begins to collect multiple sets of driving behavior data sets collected from the user's driving process during the period. These data sets are usually generated by various sensors and devices installed on the vehicle, including but not limited to GPS locators, accelerometers, gyroscopes, cameras, etc. These sensors and devices can monitor the vehicle's driving speed, acceleration, steering angle, braking frequency, lane departure and other driving behavior parameters in real time, and record these data in the form of a time series. It should be noted that before collecting this data, it is necessary to request permission from the user and do so in compliance with laws and regulations.

[0017] The multiple sets of driving behavior data collected by the computer system form the basis for subsequent analysis. These data sets may contain driving behavior data of different users, different vehicles, and different time periods, and are highly diverse and complex. In order to extract useful information from these massive data, the computer system needs to organize and manage them effectively.

[0018] When selecting a user driving behavior dataset to be analyzed, the computer system may consider a variety of factors. For example, if the purpose of the analysis is to evaluate the driving risk level of a specific user, the system may select a subset of all driving behavior datasets collected from the user during the evaluation period for analysis; if the purpose of the analysis is to compare driving behavior differences between different users under the same conditions, the system may select driving behavior datasets collected from multiple users during the same time period during the evaluation period for analysis.

[0019] Taking the evaluation of a specific user's driving risk level as an example, suppose the computer system selects a set of data from all the driving behavior data collected during the evaluation period for analysis. This set of data may contain all the driving records of the user during the evaluation period, including the start time, end time, mileage, average speed, number of brakes, number of accelerations, number of lane deviations, and other parameters of each drive. By conducting in-depth analysis of this set of data, the computer system can identify the user's driving behavior pattern, such as whether he often speeds, whether he brakes frequently, whether he is prone to lane deviation, etc., and then evaluate his driving risk level.

[0020] Step S200: identifying each driving behavior event included in the user driving behavior data set to be analyzed, and obtaining a driving behavior event identification result.

[0021] In step S200, driving behavior events refer to actions or state changes with specific meanings and characteristics that occur during the user's driving process. These events can cover multiple aspects, including but not limited to the driving state of the vehicle, the driver's operating behavior, changes in the external environment, etc. For example, vehicle acceleration, braking, turning, changing lanes, parking, starting, etc. can all be regarded as driving behavior events. In addition, unsafe behaviors such as driver distraction, fatigue driving, using mobile phones, and external environmental factors such as weather changes and road conditions may also trigger specific driving behavior events.

[0022] When identifying driving behavior events, computer systems usually rely on predefined event models and rules. These models and rules define the characteristics and identification conditions of different driving behavior events based on a large amount of historical data and expert knowledge. For example, for the event "vehicle acceleration", the computer system can define its characteristics as a significant increase in the vehicle's speed in a short period of time, and set a speed increase threshold as the identification condition. When a data segment that meets this characteristic appears in the data set to be analyzed, the system will identify it as a "vehicle acceleration" event.

[0023] In addition to rule- and model-based event recognition methods, computer systems can also use machine learning techniques to improve the accuracy and efficiency of event recognition. For example, supervised learning algorithms (such as support vector machines, random forests, etc.) can be used to train an event classification model. During the training process, the system uses annotated historical driving behavior data sets as training samples to allow the model to learn the characteristics and classification boundaries of different driving behavior events. After the training is completed, the model can perform real-time event recognition and classification on newly collected driving behavior data.

[0024] When identifying driving behavior events, the computer system also needs to consider the timing and continuity of the data. Since driving behavior is a continuous process, there are often temporal correlations and dependencies between different events. Therefore, the system needs to pay attention to the temporal relationship and contextual information between events while identifying individual events. For example, when identifying the "vehicle braking" event, the system must not only pay attention to the data characteristics of the braking action itself (such as braking force, braking time, etc.), but also consider the vehicle's speed, acceleration and other state information before braking, as well as subsequent data such as vehicle response after braking.

[0025] Taking the identification of "fatigue driving" events as an example, the computer system comprehensively considers data and information from multiple aspects. First, the system can analyze the driver's operating behavior data, such as the frequency and amplitude of steering wheel operations, braking and acceleration reaction time, etc., to determine whether the driver has signs of distraction or slow response. Secondly, the system can analyze the vehicle's driving status data, such as driving speed, acceleration, lane departure, etc., to evaluate the stability and safety of the vehicle's driving. Finally, the system can also combine external environmental data (such as weather, road conditions, etc.) and the driver's personal information (such as age, gender, driving experience, etc.) to further determine the possibility of fatigue driving.

[0026] After obtaining the driving behavior event recognition results, the computer system will represent each event in a structured form, including the event type, occurrence time, duration, related parameters, etc. This event information will be used in subsequent steps such as event feature description, behavior sequence relationship extraction, and pattern matching recognition. For example, for the "vehicle acceleration" event, the system can record its type as "acceleration", occurrence time as "2023-05-15 14:30:00", duration as "5 seconds", and related parameters including "speed before acceleration 60km / h", "speed after acceleration 80km / h", etc.

[0027] Step S300: generating an event feature description array of each driving behavior event according to the driving behavior event recognition result.

[0028] In step S300, the computer system specifies the characteristic dimensions of each driving behavior event. Characteristic dimensions refer to various parameters or attributes required to describe an event. These dimensions can cover multiple aspects, including but not limited to the time attributes, spatial attributes, dynamic attributes, driver operation attributes, etc. of the event. For example, for the "vehicle acceleration" event, its characteristic dimensions may include the acceleration start time, acceleration end time, acceleration duration, speed before acceleration, speed after acceleration, acceleration change rate, etc. These characteristic dimensions together constitute a multidimensional feature vector that describes the event.

[0029] After determining the feature dimensions, the computer system begins to generate an event feature description array for each driving behavior event. This array is a multidimensional array, each dimension of which corresponds to a feature dimension, and the elements in the array are specific values ​​on that dimension. For example, for the above-mentioned "vehicle acceleration" event, its event feature description array may be as follows:

[0030] [acceleration start time, acceleration end time, acceleration duration, speed before acceleration, speed after acceleration, acceleration change rate] = [2023-05-15 14:30:00, 2023-05-15 14:30:05, 5 seconds, 60km / h, 80km / h, 4m / s²].

[0031] In this array, each element represents the specific performance of the event in the corresponding feature dimension. These values ​​not only reflect the objective attributes of the event, but also contain rich driving behavior information, providing an important basis for subsequent data analysis and pattern recognition.

[0032] In order to generate accurate and effective event feature description arrays, the computer system uses a variety of data processing and analysis technologies. For example, for time attribute features (such as acceleration start time, acceleration end time, etc.), the system can directly extract relevant information from the original data; for dynamic attribute features (such as acceleration change rate, etc.), the system uses physical formulas or mathematical models for calculation; for driver operation attribute features (such as steering wheel angle, braking force, etc.), the system needs to obtain them through sensor data or image processing technology.

[0033] In addition, in order to improve the accuracy and robustness of event feature description, the computer system can also use machine learning technology to further process and optimize the features. For example, feature selection algorithms can be used to screen out the feature dimensions that have the greatest impact on classification or regression tasks; feature dimensionality reduction algorithms can be used to reduce the number of feature dimensions and reduce computational complexity; feature transformation algorithms can also be used to enhance the expressiveness of features and improve the generalization ability of the model.

[0034] In the process of generating event feature description arrays, the computer system also needs to consider the feature differences and associations between different events. Since driving behavior is a complex process, there are often temporal associations and dependencies between different events. Therefore, the system needs to pay attention to the feature relationships between events while generating a single event feature description array. For example, when describing the "vehicle acceleration" event, the system considers the vehicle's driving state before acceleration (such as speed, acceleration, etc.) and the vehicle's response after acceleration (such as speed changes, driving trajectory, etc.) at the same time, so as to more comprehensively reflect the impact of the event on driving behavior.

[0035] Taking the identification of "fatigue driving" events as an example, the computer system comprehensively considers the feature description arrays of multiple events to construct a complete feature representation of the event. These events may include the driver's blinking frequency, head posture changes, vehicle driving trajectory deviation, etc. The system can generate a feature vector that comprehensively reflects the driver's fatigue state by weighted fusion or feature splicing of the feature description arrays of these events. This vector not only contains the specific feature information of each event, but also reflects the association and dependency between events, providing more comprehensive and accurate data support for subsequent pattern matching recognition.

[0036] Step S300 converts the identified driving behavior events into event feature description arrays with clear mathematical representations. By reasonably defining feature dimensions, using a variety of data processing and analysis techniques, and considering feature differences and associations between events, the computer system can generate accurate and effective event feature description arrays, providing strong support for subsequent data analysis and pattern recognition.

[0037] Step S400: extracting the behavior sequence relationship between each driving behavior event included in the user driving behavior data set to be analyzed according to the event feature description array of each driving behavior event, and obtaining a driving behavior event relationship description array.

[0038] In step S400, the behavior sequence relationship refers to the internal connection and regularity of different driving behavior events when they occur in chronological order. These relationships may be manifested as direct causal relationships between events (such as braking immediately after acceleration), or indirect correlations between events (such as frequent lane changes and high-speed driving). By extracting these relationships, the computer system can have a deeper understanding of the user's driving habits, risk preferences, and potential safety hazards.

[0039] In order to extract the behavior sequence relationship, the computer system encodes the event feature description array of each driving behavior event. The purpose of encoding is to simplify the complex event features into a form that is easy to process and compare, while retaining the key information between events. The encoding method can be varied, such as discrete encoding based on event type, continuous encoding based on event features, or mixed encoding. Taking the "vehicle acceleration" event as an example, its encoding may include the quantized values ​​of feature dimensions such as event type (such as "acceleration"), acceleration start speed, acceleration end speed, and acceleration duration.

[0040] After the encoding is completed, the computer system generates a coding array based on the temporal distribution of each driving behavior event in the data set and their coding results. This array not only records the specific coding of each event, but also retains the order in which they occur. For example, a coding array containing three events of "acceleration-constant speed driving-braking" may be as follows:

[0041] [Event type 1 (acceleration), acceleration start speed, acceleration end speed, acceleration duration, ..., Event type 2 (constant speed), constant speed, duration, ..., Event type 3 (braking), speed before braking, braking time]

[0042] Next, the computer system extracts the driving behavior event relationship description array from the encoding array. This step involves in-depth analysis of the encoding array to identify features such as the time interval between events, frequency distribution, pattern repetition, and the behavioral sequence relationships implied by these features. For example, the system may find that the pattern of "acceleration followed by braking" appears frequently, which suggests that the user may tend to respond quickly to changes in road conditions, but there may also be a risk of overdriving.

[0043] In order to more accurately extract the behavior sequence relationship, the computer system can use machine learning techniques such as recurrent neural networks (RNN), long short-term memory networks (LSTM) or graph neural networks (GNN). These models can process sequence data and capture the temporal dependencies and potential patterns between events. Taking LSTM as an example, the system can take the encoding array as the input sequence, and output a vector or matrix containing the description of the event relationship through the processing of the LSTM layer. This output not only reflects the direct connection between events, but also takes into account the long-term dependencies between events.

[0044] In the process of extracting behavioral sequence relationships, the computer system can also use feature extraction and dimensionality reduction techniques, such as principal component analysis (PCA), t-SNE, etc., to further simplify data representation and highlight key information. For example, by reducing the dimensionality of the encoded array through PCA, the system can obtain an event distribution map in a low-dimensional space, so as to intuitively observe the clustering phenomenon and potential relationship between events. In addition, in order to verify the accuracy and effectiveness of the extracted behavioral sequence relationship, the computer system can perform cross-validation or model evaluation. For example, the system can divide the data set into a training set and a test set, use the training set to train the behavioral sequence relationship extraction model, and then use the test set to evaluate the performance of the model. Evaluation indicators can include accuracy, recall rate, F1 score, etc., which can reflect the performance of the model in identifying behavioral sequence relationships.

[0045] Step S500: According to the event feature description array and the driving behavior event relationship description array, it is identified whether the user driving behavior data set to be analyzed matches the reference driving behavior data in the reference behavior pattern data set.

[0046] In step S500, the reference behavior pattern data set is a database containing multiple preset driving behavior patterns, which are constructed based on a large amount of historical driving data, expert knowledge or industry standards, and are intended to reflect ideal or standard driving behaviors in different driving scenarios. For example, the reference behavior pattern data set may include multiple types such as safe driving mode, economic driving mode, and aggressive driving mode, and each mode corresponds to a set of specific event feature descriptions and driving behavior event relationships.

[0047] In the process of identifying the matching, the computer system first compares the event feature description array of the user driving behavior data set to be analyzed with the event feature description array of each reference driving behavior data in the reference behavior pattern data set. This comparison process involves a one-by-one comparison of multiple feature dimensions, including but not limited to speed, acceleration, braking frequency, number of lane departures, etc. The computer system calculates the similarity or distance measurement on these feature dimensions, such as Euclidean distance, cosine similarity, etc., to quantify the difference between the user driving behavior and the reference behavior pattern. Taking the speed feature as an example, suppose that a safe driving mode in the reference behavior pattern data set stipulates that the average driving speed should be between 60-80km / h, and the average speed of a certain section of the user driving behavior data set to be analyzed is 75km / h. In this case, the computer system can calculate the deviation between the speed feature and the speed range specified by the reference mode as an indicator for evaluating the matching.

[0048] In addition to comparing the event feature description arrays, the computer system also needs to consider the matching of the driving behavior event relationship description arrays. This comparison process is more complicated because it involves the time sequence, logical associations, and potential pattern recognition between events. The computer system will use machine learning techniques, such as sequence matching algorithms and graph matching algorithms, to compare the similarity between the user's driving behavior event relationship and the reference behavior pattern event relationship.

[0049] Taking the sequence matching algorithm as an example, suppose an economic driving mode in the reference behavior pattern data set specifies the event sequence of "constant speed driving-slow acceleration-smooth braking", and the user driving behavior data set to be analyzed also contains a similar event sequence. The computer system can calculate the similarity score between the user event sequence and the reference event sequence through the dynamic time warping (DTW) algorithm or the longest common subsequence (LCS) algorithm. If the score exceeds the preset threshold, it is considered that the user's driving behavior matches the reference pattern in terms of event relationship.

[0050] During the comparison process, the computer system also needs to consider the weights of different feature dimensions and event relationships in the matching assessment. Since different driving behavior characteristics may have different impacts on safe, economical or aggressive driving modes, the computer system needs to assign appropriate weights to each feature dimension and event relationship. These weights can be determined through expert scoring, historical data analysis or machine learning model training.

[0051] For example, when evaluating the compatibility of a safe driving mode, the computer system can give higher weights to feature dimensions directly related to safety, such as speed and braking frequency, and give lower weights to feature dimensions indirectly related to safety, such as the number of lane departures. Similarly, when evaluating the compatibility of an economic driving mode, the computer system can pay more attention to feature dimensions directly related to economic benefits, such as fuel consumption and driving distance.

[0052] After obtaining the similarity or distance measurement of each feature dimension and event relationship, the computer system needs to integrate this information to make the final matching judgment. This judgment process may involve a variety of methods such as threshold comparison, weighted summation, cluster analysis, etc. For example, the computer system can set a comprehensive similarity threshold, and only when the comprehensive similarity between the user's driving behavior and the reference behavior pattern in all feature dimensions and event relationships exceeds the threshold, the two are considered to match.

[0053] It is worth noting that matching judgment is not a simple binary classification problem, but a multi-level evaluation process. In some cases, the user's driving behavior may partially match multiple reference behavior patterns. At this time, the computer system needs to comprehensively consider factors such as matching degree and pattern priority to make the final decision.

[0054] Step S600: If the reference driving behavior data is matched, the driving behavior pattern corresponding to the reference driving behavior data is used as the driving behavior pattern for the set of driving behavior collection data sets.

[0055] In step S500, the computer system compares the event feature description array and the driving behavior event relationship description array of the user driving behavior data set with the reference driving behavior data in the reference behavior pattern data set to perform a matching evaluation. When the evaluation result shows that the user driving behavior and a certain reference driving behavior data show a high degree of similarity in multiple key feature dimensions and event relationships, the match is considered successful. This matching not only takes into account the characteristics of a single driving behavior event, but also comprehensively considers the sequence relationship and overall pattern between events, thereby ensuring the accuracy and reliability of the match. Once the match is successful, the computer system assigns the driving behavior pattern in the corresponding reference driving behavior data to the driving behavior collection data set currently being analyzed. The driving behavior pattern is an abstract concept that represents the overall characteristics and laws exhibited by a specific driving behavior data set. These patterns may include safe driving modes, economic driving modes, aggressive driving modes, and other types, each of which corresponds to a set of specific driving behavior characteristics and event relationships.

[0056] Taking safe driving mode as an example, it may be manifested as smooth acceleration and deceleration, moderate driving speed, frequent sight checking, compliance with traffic rules and other behavioral characteristics. When the computer system matches the safe driving mode, it means that the characteristics of the driving behavior collection data set currently being analyzed are highly consistent with these safe driving behaviors. Therefore, the computer system will use these characteristics as the driving behavior mode of the data set for subsequent premium management strategy formulation.

[0057] It is worth noting that the determination of driving behavior patterns is not a simple labeling process, but a result based on in-depth analysis and precise matching. In the matching process, the computer system needs to comprehensively consider multiple factors, such as the similarity of feature dimensions, the degree of consistency of event relationships, the setting of matching thresholds, etc. These factors jointly determine the accuracy of matching and the effectiveness of the pattern.

[0058] In actual applications, the computer system can also further refine and classify driving behavior patterns. For example, in the safe driving mode, it can be further divided into different levels such as "very safe", "safe" and "relatively safe" to more accurately reflect the differences in user driving behaviors. These refined driving behavior patterns can provide insurance companies with richer pricing options to better meet the needs and preferences of different users. At the same time, the computer system can also combine driving behavior patterns with factors such as user personal information and vehicle information to conduct more comprehensive risk assessment and premium management. For example, for older users with more experienced driving, even if their driving behavior patterns tend to be aggressive, they may enjoy lower premiums due to their higher driving skills and risk awareness; while for young users with less driving experience, even if their driving behavior patterns tend to be safe, they may face higher premiums due to potential risks.

[0059] Step S700: Analyze the remaining groups of driving behavior collection data sets to obtain the driving behavior patterns corresponding to each group of driving behavior collection data sets, and perform weighted fusion on the driving behavior patterns corresponding to the multiple groups of driving behavior collection data sets to obtain a fused driving behavior pattern.

[0060] In step S700, first, for the remaining multiple sets of driving behavior collection data sets, the computer system repeats the analysis process of steps S200 to S600. This means that for each set of data sets, the system will identify the driving behavior events therein, generate an event feature description array and a driving behavior event relationship description array, and then match them with the reference driving behavior data in the reference behavior pattern data set to determine the driving behavior pattern of the set of data sets.

[0061] Taking one of the datasets as an example, suppose the system identifies that the dataset contains frequent acceleration and braking events, and the time intervals between these events are short, and the vehicle's driving speed is relatively high. After matching with the reference behavior pattern dataset, the system can classify the driving behavior pattern of this dataset as "aggressive driving mode". This mode reflects that the user tends to adopt a more aggressive driving style, which may be accompanied by higher risks.

[0062] After completing the analysis of all remaining data sets, the computer system obtains a set of driving behavior patterns, each of which corresponds to a specific set of driving behavior collection data sets. However, since the user's driving behavior may show differences in different time periods or different driving scenarios, these individual patterns may not fully reflect the user's overall driving behavior characteristics.

[0063] To solve this problem, the computer system uses a weighted fusion method. Weighted fusion is a technique that integrates multiple patterns or data points into a unified pattern, taking into account the importance and reliability of different patterns or data points. In driving behavior data analysis, weighted fusion can assign different weights to each set of data according to their importance (such as collection time, collection scene, etc.) or the similarity between patterns.

[0064] Taking the "aggressive driving mode" mentioned earlier as an example, suppose that the data set corresponding to this mode is collected when the user drives long distances on weekends, while the data set corresponding to the other "safe driving mode" is collected when the user commutes on weekdays. In this case, the system can give a higher weight to the "safe driving mode" because it is more likely to reflect the user's regular behavior in daily driving. On the contrary, although the "aggressive driving mode" is also important, it may only represent the user's special behavior in a specific situation.

[0065] After determining the weight of each mode, the computer system integrates these modes into a fusion driving behavior model through weighted averaging or other fusion algorithms. This fusion model not only contains the characteristics of each set of data sets, but also reflects their importance and relative importance through weight adjustment. For example, if the weight of "safe driving mode" is higher, the fusion model can reflect more safe and stable driving characteristics; if the weight of "aggressive driving mode" is higher, the fusion model can include more aggressive and high-risk driving characteristics.

[0066] It is worth noting that the weighted fusion process is not a simple arithmetic average. Instead, it needs to consider the complex relationship between different modes and their impact on the user's overall driving behavior. Therefore, in practical applications, computer systems use advanced machine learning algorithms or statistical models to optimize the fusion process to ensure the accuracy and reliability of the fusion results.

[0067] In addition, the result of weighted fusion - the fused driving behavior pattern, provides an important reference for the subsequent formulation of premium management strategies. Insurance companies can assess the risk level of users based on their fused driving behavior patterns and formulate personalized premium strategies accordingly. For example, for users who exhibit safe driving behavior, insurance companies can provide more favorable premiums; while for users who exhibit aggressive driving behavior, additional premiums are charged to cover potential risks.

[0068] Step S800: Generate a corresponding premium management strategy based on the fused driving behavior pattern.

[0069] In step S800, the computer system needs to conduct an in-depth analysis of the fused driving behavior pattern to extract key information related to premium management. This information may include multiple aspects such as the user's driving risk level, driving habit stability, and driving scenario diversity. For example, if the fused driving behavior pattern shows that the user exhibits a low driving risk in different time periods and driving scenarios, then it can be considered that the user's driving habits are relatively stable and safe, and the corresponding premium management strategy may tend to provide more favorable premiums. On the contrary, if the fused driving behavior pattern shows that the user exhibits a high driving risk in a specific time period or driving scenario, then a more stringent premium management strategy is adopted, such as increasing the premium or requiring the user to participate in driving safety training. The mapping relationship between the premium management strategy and the fused driving behavior pattern can be pre-set.

[0070] On the basis of extracting key information, the computer system combines this information with the insurance company's premium management policy to generate a specific premium management strategy. This process may involve multiple considerations, such as the calculation method of the premium, the frequency of premium adjustment, and the preferential policies of the premium. Taking the calculation method of the premium as an example, the computer system can set different premium coefficients for different users based on the risk level indicators in the integrated driving behavior pattern. The higher the risk level of the user, the higher the premium coefficient may be, and vice versa. In this way, the insurance company can customize the premium according to the user's actual driving behavior to achieve a reasonable match between risk and premium. In addition to the premium calculation method, the computer system can also consider the premium adjustment frequency and preferential policies. For example, for users with stable driving habits and low risk levels, the computer system can recommend that the insurance company reduce the premium adjustment frequency to reduce the user's financial burden. At the same time, in order to encourage users to maintain good driving habits, the computer system can also recommend that the insurance company provide preferential policies such as premium discounts or point rewards for users who have shown low-risk behavior for multiple consecutive evaluation cycles.

[0071] As an implementation mode, step S400 extracts the behavior sequence relationship between each driving behavior event included in the user driving behavior data set to be analyzed according to the event feature description array of each driving behavior event, and obtains the driving behavior event relationship description array, including:

[0072] Step S410: performing encoding operation on each driving behavior event according to the event feature description array of each driving behavior event to obtain the encoding result of each driving behavior event;

[0073] Step S420: generating a coding array corresponding to the user driving behavior data set to be analyzed according to the time series distribution of each driving behavior event in the user driving behavior data set to be analyzed and the coding result of each driving behavior event;

[0074] Step S430: extracting the driving behavior event relationship description array from the coding array.

[0075] In step S410, the computer system performs encoding operations on these events based on the event feature description array of each driving behavior event to generate corresponding encoding results. The purpose of encoding is to simplify complex event features into a form that is easy to process and compare, while retaining key information between events. First, a set of encoding rules needs to be defined. This set of rules should be able to comprehensively cover all possible driving behavior events and assign a unique code to each event. The encoding can be determined based on multiple factors such as the type of event, feature dimensions, and time of occurrence. For example, for a "vehicle acceleration" event, its encoding may include quantified values ​​of feature dimensions such as event type (such as "acceleration"), acceleration start speed, acceleration end speed, and acceleration duration.

[0076] After defining the coding rules, the computer system begins to encode each driving behavior event. This process involves quantifying and normalizing each feature dimension in the event feature description array to ensure consistency and comparability of the coding. For example, for the speed feature, the system can convert it into a relative speed value (such as the current speed divided by the road speed limit) for comparison under different road conditions. After the coding is completed, each driving behavior event will obtain a unique coding result. This coding result is a multidimensional vector or array, each dimension of which corresponds to a feature dimension, and the elements in the array are the specific values ​​of the dimension. For example, the coding result of the "vehicle acceleration" event may be an array containing multiple elements such as event type code, acceleration start speed code, acceleration end speed code, acceleration duration code, etc.

[0077] In step S420, the computer system generates a coding array corresponding to the data set according to the temporal distribution of each driving behavior event in the user driving behavior data set to be analyzed and their coding results. The coding array is a two-dimensional array or matrix, whose rows represent different events and columns represent different feature dimensions or coding elements.

[0078] First, the computer system determines the number of rows and columns of the encoding array. The number of rows is equal to the total number of driving behavior events in the data set to be analyzed, and the number of columns is equal to the number of elements in the encoding result of each event. For example, if the data set contains 10 driving behavior events and the encoding result of each event contains 5 elements, the encoding array will be a matrix with 10 rows and 5 columns.

[0079] Next, the computer system fills the coding results of each event into the corresponding rows of the coding array in sequence according to the temporal distribution of the events in the data set. In this way, the first row of the coding array will contain the coding results of the first event, the second row will contain the coding results of the second event, and so on. In this way, the coding array not only records the specific coding of each event, but also preserves the temporal relationship between them.

[0080] In step S430, the computer system extracts the driving behavior event relationship description array from the coding array. This step aims to reveal the time sequence, logical association and potential patterns between events, and provide key information for subsequent pattern matching and premium management strategy formulation.

[0081] First, the computer system selects a suitable algorithm or model to process the encoded array to extract the relationship between events. These algorithms or models may include sequence matching algorithms, graph matching algorithms, recurrent neural networks (RNNs), long short-term memory networks (LSTMs), etc. These models can process sequence data and capture the temporal dependencies and potential patterns between events.

[0082] Taking LSTM as an example, the computer system can use the encoding array as the input sequence of the LSTM model. Through its internal memory unit and gating mechanism, the LSTM model can learn the long-term dependencies in the sequence data and output a vector or matrix containing the event relationship description. This output vector or matrix not only reflects the direct connection between events, but also takes into account the long-term dependencies between events. In the process of extracting the event relationship description array, the computer system can also use feature extraction and dimensionality reduction techniques to further simplify the data representation and highlight key information. For example, the output of the LSTM model can be reduced in dimension using principal component analysis (PCA) to obtain an event relationship description array in a low-dimensional space. This array not only retains the key information of the event relationship, but also reduces the dimension and complexity of the data. Finally, the driving behavior event relationship description array generated by the computer system is a multidimensional array or matrix, each dimension or column of which represents an event relationship feature. These features may include the time interval between events, frequency distribution, pattern repetition, etc., which together constitute a comprehensive description of the event relationship. For example, if the "acceleration-braking" pattern frequently appears between two events, the corresponding time interval feature and pattern repetition feature in the event relationship description array will have a higher value.

[0083] As an implementation mode, step S410 performs encoding operation on each driving behavior event according to the event feature description array of each driving behavior event to obtain the encoding result of each driving behavior event, including:

[0084] Step S411: obtaining a plurality of group centroids, where the group centroids are obtained by grouping and clustering the event feature description arrays of the driving behavior events contained in the reference driving behavior data;

[0085] Step S412: determining the group centroid having the smallest distance from the event feature description array of each driving behavior event according to the vector space distance between the event feature description array of each driving behavior event and a plurality of group centroids;

[0086] Step S413: Generate encoding results for each driving behavior event according to the group code corresponding to the centroid of the group with the smallest distance in the event feature description array of each driving behavior event.

[0087] In step S411, the computer system obtains a plurality of group centroids. These group centroids are obtained by grouping and clustering the event feature description arrays of the driving behavior events contained in the reference driving behavior data. Group clustering is an unsupervised learning method that aims to cluster similar data points together to form different clusters or groups.

[0088] In order to obtain the centroid of the group, the computer system collects a large amount of reference driving behavior data. This data may come from multiple users, different time periods and driving scenarios, and contains a rich array of driving behavior events and event feature descriptions. Then, the system uses clustering algorithms (such as K-means, DBSCAN, etc.) to group and cluster these data.

[0089] Taking the K-means algorithm as an example, the system first randomly selects K data points as the initial centroids, then calculates the Euclidean distance between each data point and these centroids, and assigns the data point to the cluster represented by the nearest centroid. Next, the system recalculates the centroid of each cluster (that is, the mean of all data points in the cluster) and repeats the above process until the position of the centroid no longer changes significantly or reaches the preset number of iterations. In the end, the system will obtain K group centroids, each of which represents the center position of a cluster.

[0090] These grouping centroids not only reflect the characteristic distribution of different driving behavior events in the reference driving behavior data, but also provide a benchmark for subsequent encoding operations. By comparing the distance between the feature description array of the event to be encoded and these centroids, the system can assign the event to the most similar cluster and generate the encoding result accordingly.

[0091] In step S412, the computer system determines the group centroid with the smallest distance from each event feature description array based on the vector space distance between the event feature description array of each driving behavior event and multiple group centroids. This step is the core of the encoding operation, which determines which cluster the event will be assigned to and generates the corresponding code accordingly.

[0092] Vector space distance is a method to measure the similarity between data points. In multidimensional space, the Euclidean distance, cosine similarity or Manhattan distance between two data points can be used as a metric for vector space distance. In computer systems, Euclidean distance is usually chosen as a metric because it is simple, intuitive and easy to calculate.

[0093] For each driving behavior event to be coded, the computer system first calculates the Euclidean distance between its event feature description array and all group centroids. These distances reflect the similarity between the event and different clusters. Then, the system selects the group centroid with the smallest distance as the representative centroid of the event. This means that the event is most similar in features to the cluster represented by the selected centroid.

[0094] For example, suppose there is a driving behavior event whose event feature description array is [speed, acceleration, braking frequency], and its value is [80km / h, 2m / s², 0.1 times / minute]. At the same time, the system has obtained two grouping centroids through the K-means algorithm, namely centroid A ([60km / h, 1m / s², 0.05 times / minute]) and centroid B ([100km / h, 3m / s², 0.2 times / minute]). The system calculates the Euclidean distance between the event and the two centroids, and obtains distance A as 20 (assuming it is calculated in some way) and distance B as 30. Since distance A is smaller than distance B, the system selects centroid A as the representative centroid of the event.

[0095] In step S413, the computer system generates the coding results of each driving behavior event according to the group code corresponding to the centroid of the group with the smallest distance in the event feature description array of each driving behavior event. This step is the process of assigning a unique code or label to the event after assigning the event to a specific cluster.

[0096] In the clustering algorithm, each cluster is usually assigned a unique grouping code or label. This code can be a number, a letter, or a combination thereof, which is used to distinguish different clusters. In step S411, when the system obtains multiple grouping centroids through the clustering algorithm, a grouping code is also assigned to each centroid.

[0097] For each driving behavior event to be coded, the computer system has determined the most similar group centroid in step S412. In step S413, the system directly uses the group code corresponding to the centroid as the coding result of the event. In this way, each event will obtain a unique code, which reflects which cluster the event is most similar to in terms of characteristics.

[0098] Continuing with the above example, assume that the group code of centroid A is "A" and the group code of centroid B is "B". Since the system has determined that the event to be coded is most similar to centroid A, the coding result of the event will be set to "A". This means that the event is closest to the cluster represented by centroid A in terms of features and may have similar driving behavior characteristics.

[0099] Through the implementation of steps S411, S412 and S413, the computer system successfully encodes each driving behavior event and generates corresponding encoding results. These encoding results not only retain the key information between events, but also provide a basis for subsequent encoding array generation and event relationship extraction. By comparing the encoding results of different events, the computer system can more easily identify the similarities and differences between events, and then extract valuable behavior sequence relationships and patterns.

[0100] As an embodiment, the method further includes:

[0101] Step S414: obtaining the number of known driving behavior events and the number of driving behavior events associated with the known driving behavior events, and obtaining an event feature description array of the driving behavior events in the reference driving behavior data;

[0102] Step S415: determining the number of group centroids according to the number of known driving behavior events and the number of associated driving behavior events;

[0103] Step S416: grouping and clustering the event feature description arrays of the driving behavior events included in the reference driving behavior data according to the number of group centroids to obtain a plurality of group centroids.

[0104] In step S414, the computer system obtains a series of key data and information to provide a basis for subsequent grouping, clustering and coding operations. These data and information include the number of known driving behavior events, the number of associated driving behavior events of the known driving behavior events, and an event feature description array of the driving behavior events in the reference driving behavior data.

[0105] The number of known driving behavior events refers to the total number of different driving behavior events that have been clearly identified in the reference driving behavior data. These events may include acceleration, braking, turning, lane change, etc. The number of associated driving behavior events refers to the number of other events associated with each known driving behavior event. For example, an acceleration event may be followed by a braking event, so the braking event can be considered an associated event of the acceleration event.

[0106] The event feature description array is a multidimensional array or vector obtained after feature extraction of each driving behavior event, which contains multiple parameters or indicators describing the event characteristics. These features may include physical quantities such as speed, acceleration, braking force, steering angle, etc., and may also include environmental information such as the time, location, and weather of the event. In step S414, the computer system needs to extract these event feature description arrays from the reference driving behavior data for subsequent grouping and clustering.

[0107] In step S415, the computer system determines the number of grouping centroids based on the number of known driving behavior events and the number of associated driving behavior events. The number of grouping centroids is an important parameter in the clustering algorithm, which determines the number of clusters finally formed. Selecting the appropriate number of grouping centroids is crucial to ensure clustering effect and coding accuracy.

[0108] There are many ways to determine the number of group centroids, one of which is the elbow method. The elbow method calculates clustering effect indicators (such as the total internal sum of squares SSE) under different numbers of group centroids, and then plots a curve chart showing the indicators as the number of group centroids changes. In the curve chart, when the number of group centroids increases, the clustering effect usually improves significantly first, and then gradually stabilizes. The elbow method selects the number of group centroids corresponding to the "elbow" in the curve chart as the optimal value, because at this time, increasing the number of group centroids will not significantly improve the clustering effect, but the calculation complexity will increase significantly.

[0109] In addition to the elbow rule, the computer system can also use other methods to determine the number of group centroids, such as the Silhouette Coefficient method, the Calinski-Harabasz index method, etc. These methods have their own advantages and disadvantages, and the choice of which method depends on the specific application scenario and data characteristics.

[0110] When determining the number of group centroids, the computer system also needs to consider the relationship between known driving behavior events and related driving behavior events. If there is a strong correlation between certain events, they are assigned to the same cluster so that this correlation can be preserved during encoding. Therefore, when determining the number of group centroids, the computer system needs to comprehensively consider multiple factors such as the number of events, correlation, and clustering effect.

[0111] In step S416, the computer system groups and clusters the event feature description arrays of the driving behavior events contained in the reference driving behavior data according to the determined number of grouping centroids, thereby obtaining a plurality of grouping centroids. Grouping and clustering is an unsupervised learning method that can cluster similar data points together to form clusters, each of which has a representative point or centroid.

[0112] There are many algorithms for grouping and clustering, such as K-means, DBSCAN, hierarchical clustering, etc. In step S416, the computer system can select one of the algorithms to perform the grouping and clustering task. Taking the K-means algorithm as an example, the algorithm first randomly selects K data points as the initial centroids (K is the number of grouping centroids determined in step S415), then calculates the distance between each data point and these centroids, and assigns the data point to the cluster represented by the nearest centroid. Next, the algorithm recalculates the centroid of each cluster (that is, the mean of all data points in the cluster), and repeats the above process until the position of the centroid no longer changes significantly or reaches a preset number of iterations. Finally, the algorithm will output K grouping centroids and the cluster label to which each data point belongs.

[0113] During the grouping and clustering process, the computer system needs to ensure that the dimension and metric of the event feature description array match the requirements of the clustering algorithm. If the dimension of the event feature description array is too high or there are redundant features, it may lead to a decrease in clustering effect or an increase in computational complexity. Therefore, before grouping and clustering, the computer system performs preprocessing operations such as dimensionality reduction or feature selection on the event feature description array. In addition, since grouping and clustering is an iterative process, its results can be affected by the initial centroid selection. In order to reduce this effect and improve the stability and accuracy of the clustering results, the computer system can use a method of running the K-means algorithm multiple times and selecting the optimal result, or use more advanced clustering algorithms such as K-means++, MiniBatch K-means, etc.

[0114] As an implementation mode, step S420 generates a coding array corresponding to the user driving behavior data set to be analyzed according to the temporal distribution of each driving behavior event in the user driving behavior data set to be analyzed and the coding result of each driving behavior event, including:

[0115] Step S421: replacing the time series data corresponding to each driving behavior event based on the encoding result of each driving behavior event according to the time series distribution of each driving behavior event in the user driving behavior data set to be analyzed;

[0116] Step S422: changing the data other than the time series corresponding to each driving behavior event in the user driving behavior data set to be analyzed into preset data, and obtaining a coding array corresponding to the user driving behavior data set to be analyzed.

[0117] In step S421, the computer system first focuses on the temporal distribution of each driving behavior event in the user driving behavior data set to be analyzed. Temporal distribution refers to the order and interval of events on the time axis, which reflects the dynamic change process of the user's driving behavior. In order to capture this dynamic change, the system needs to ensure that the encoding array can accurately record the specific time point of each event.

[0118] However, directly recording the time point can introduce a lot of redundant information and is not conducive to subsequent data processing and analysis. Therefore, in step S421, the computer system adopts a more efficient method: replacing the corresponding time series data based on the encoding results of each driving behavior event. This means that the system no longer directly records the specific time point when the event occurs, but uses the encoding results of the event to represent the driving behavior at that time point.

[0119] The specific operation is as follows: the system traverses the user driving behavior data set to be analyzed. For each driving behavior event, the system first searches for its timestamp in the data set (i.e., the specific time point when the event occurred). Then, based on the encoding result of the event obtained in step S410, the system replaces the data corresponding to the timestamp with the encoding result. In this way, each element in the encoding array is no longer a specific time point, but a coding value representing a specific driving behavior.

[0120] For example, suppose the user driving behavior data set to be analyzed contains three driving behavior events: acceleration, constant speed driving, and braking. The system has generated coding results for these events: acceleration corresponds to code A, constant speed driving corresponds to code B, and braking corresponds to code C. If the timestamps of these three events in the data set are T1, T2, and T3, respectively, then after step S421, the data corresponding to these three time points in the coding array will be replaced by A, B, and C.

[0121] Through this replacement operation, the encoding array not only retains the time series characteristics of the events (i.e. the relative order of the events), but also represents the specific content of each event in a more concise and efficient way, which facilitates subsequent data processing and analysis.

[0122] In step S422, the computer system further processes the user driving behavior data set to be analyzed, and changes the data other than the time series corresponding to the event into the preset data. The purpose of this step is to ensure that the encoding array only contains information directly related to the driving behavior event, and excludes other redundant data that may interfere with subsequent analysis.

[0123] Data other than time series usually refers to data at time points other than driving behavior events in the dataset. These data may contain various non-behavioral information of the user during driving, such as vehicle status, environmental noise, etc. Although this information may be useful for some analysis tasks, it is not required in the current encoding array generation process. Therefore, the system chooses to replace them with preset data to simplify data representation and reduce computational complexity.

[0124] The preset data can be any information that is not directly related to the driving behavior event, but is usually selected as a simple and easy-to-process numerical value or symbol. For example, the system can choose to replace all data other than the time series with 0 or a specific placeholder symbol. In this way, in the encoding array, except for the encoding value representing the driving behavior event, all other elements will have the same preset value.

[0125] Continuing with the previous example, assume that the user driving behavior data set to be analyzed also contains data at other time points (such as T0 and T4) between acceleration, constant speed driving and braking events. In step S422, the system replaces the data at these time points with preset data (such as 0). Therefore, if T0 and T4 originally contain information such as vehicle speed and acceleration, then after step S422, this information will be ignored, and the elements corresponding to T0 and T4 in the encoding array will be set to 0.

[0126] Through the processing of step S422, the coding array becomes more concise and clear. It only contains information directly related to the driving behavior event (i.e., the coding result of the event), and excludes other redundant data that may interfere with the analysis. This provides a more accurate and effective data basis for subsequent tasks such as event relationship extraction and pattern recognition.

[0127] In summary, step S420 successfully converts the user driving behavior data set to be analyzed into a coding array through the implementation of steps S421 and S422. This conversion process not only retains the time series characteristics and coding information of the event, but also excludes redundant data, providing strong support for subsequent data processing and analysis. Through the coding array, the computer system can more efficiently extract the relationship between driving behavior events, identify the user's driving behavior pattern, and provide a scientific basis for the formulation of premium management strategies.

[0128] As an implementation mode, step S430, extracting a driving behavior event relationship description array from a coding array, includes:

[0129] Step S431: loading the encoding array into the driving behavior event relationship mining network, the driving behavior event relationship mining network includes a plurality of cascaded feature extraction components, and the encoding array is loaded into the first feature extraction component among the plurality of feature extraction components;

[0130] Step S432: fusing the outputs of several feature extraction components among the multiple feature extraction components based on the dimensionality reduction operation to obtain a driving behavior event relationship description array.

[0131] In step S431, the computer system loads the encoding array into the driving behavior event relationship mining network. This network is a deep learning model specially designed for mining the relationship between driving behavior events, which contains multiple cascaded feature extraction components. The function of these components is to gradually extract high-level feature representations from the encoding array to reveal the internal connections and laws between events.

[0132] The specific structure of the driving behavior event relationship mining network may vary depending on the application scenario and data characteristics, but generally includes an input layer, multiple hidden layers (each hidden layer contains multiple feature extraction components), and an output layer. In step S431, the encoding array is loaded into the input layer as input data and then passed to the first feature extraction component for processing.

[0133] The feature extraction component can be various types of neural network layers, such as convolutional layers, loop layers, fully connected layers, etc. These layers transform the input data through different calculation methods (such as convolution operations, loop operations, linear transformations, etc.) to extract useful feature information. In the first feature extraction component, each element in the encoding array (that is, the encoding result of each driving behavior event) is converted into a higher-dimensional feature vector. This higher-dimensional vector contains richer information than the original encoding and can better describe the characteristics and attributes of the event.

[0134] As data is passed through the network, each subsequent feature extraction component will further process and transform the output of the previous component. These components may have different structures and parameters to meet the requirements of feature extraction at different levels. Through the cascade processing of multiple feature extraction components, the driving behavior event relationship mining network can gradually extract more abstract and advanced feature representations, which can better reflect the complex relationships between events.

[0135] In step S432, the computer system fuses the outputs of multiple feature extraction components to generate a driving behavior event relationship description array. The purpose of this step is to integrate feature representations at different levels to form a comprehensive description that fully reflects the event relationship.

[0136] Since the output of each feature extraction component represents feature information at different levels, there may be some redundancy and complementarity between them. In order to make full use of this information and reduce computational complexity, the system needs to perform dimensionality reduction operations on these outputs. Dimensionality reduction operations can be achieved through algorithms such as principal component analysis (PCA), linear discriminant analysis (LDA), and t-SNE, which can map high-dimensional data into low-dimensional space while retaining the main structure and features of the data.

[0137] After the dimensionality reduction operation, the system fuses the outputs of multiple feature extraction components. The fusion can be simple concatenation, weighted averaging, or more complex nonlinear combination. The concatenation operation directly connects the outputs of different components together to form a longer vector; the weighted average assigns different weights to each component according to its importance and calculates the weighted sum; the nonlinear combination may involve more complex calculations, such as using a neural network for further processing.

[0138] Regardless of the fusion method used, the final result is a vector or matrix containing comprehensive feature information, that is, an array describing the relationship between driving behavior events. This array not only contains the feature representation of a single event, but also reflects the time sequence, logical associations, and potential patterns between events. By analyzing and interpreting this array, the computer system can have a deeper understanding of the user's driving behavior characteristics and behavior patterns, providing strong support for the subsequent formulation of premium management strategies.

[0139] Taking the actual scenario as an example, assume that the encoding array contains the encoding results of a series of driving behavior events, such as acceleration, constant speed driving, braking, etc. In step S431, these encoding results are loaded into the driving behavior event relationship mining network, and high-level feature representations are gradually extracted through the processing of multiple feature extraction components. In step S432, the system performs dimensionality reduction and fusion processing on these feature representations, and finally generates a driving behavior event relationship description array. This array may reveal useful information such as frequent co-occurrence patterns between certain events (such as acceleration followed by braking), time interval distribution patterns between events, etc. By analyzing and utilizing this information, insurance companies can more accurately assess the user's driving risk level and formulate personalized premium management strategies accordingly.

[0140] As an embodiment, the method further includes:

[0141] Step S430A: obtaining a network training sample for calibrating a driving behavior event relationship mining network, wherein the network training sample includes basic driving behavior data, positive driving behavior data, and negative driving behavior data; wherein the basic driving behavior data, the positive driving behavior data, and the negative driving behavior data are all coded data;

[0142] Step S430B: loading the network training sample into the driving behavior event relationship mining network to obtain a first feature mapping array for basic driving behavior data, a second feature mapping array for negative driving behavior data, and a third feature mapping array for positive driving behavior data output by the driving behavior event relationship mining network;

[0143] Step S430C: determining a spacing error between the first spatial spacing and the second spatial spacing based on a first spatial spacing between the first feature mapping array and the second feature mapping array, and a second spatial spacing between the first feature mapping array and the third feature mapping array;

[0144] Step S430D: Determine the distance error that is not less than the preset data as the adjustment direction, and optimize the network parameters of the driving behavior event relationship mining network.

[0145] In step S430A, the computer system obtains network training samples for adjusting the driving behavior event relationship mining network. These samples include basic driving behavior data (as anchor points), positive driving behavior data, and negative driving behavior data. All of these data have been encoded into a form suitable for network processing.

[0146] Baseline driving behavior data represents a standard or neutral driving behavior pattern, which serves as a benchmark or anchor point during the tuning process. Positive driving behavior data represents driving behavior patterns that demonstrate higher driving skills, lower risk, or comply with safety regulations. Conversely, negative driving behavior data represents driving behavior patterns that demonstrate lower driving skills, higher risk, or violate safety regulations.

[0147] In order to obtain these training samples, the computer system can screen representative samples from historical driving behavior data sets, or generate new samples through simulated driving experiments. When screening or generating samples, the system needs to ensure the diversity and representativeness of the samples to cover different driving scenarios, driving habits, and risk levels.

[0148] For example, basic driving behavior data may include driving records of driving at a steady speed on a highway in clear weather; active driving behavior data may include driving records of quickly and safely avoiding obstacles in emergency situations; and passive driving behavior data may include driving records of frequently changing lanes and speeding while drunk.

[0149] In step S430B, the computer system loads the acquired network training samples into the driving behavior event relationship mining network. The network processes these samples and outputs a feature mapping array for each sample.

[0150] The feature map array is the result of the network extracting and transforming the input samples. It contains the representation of the samples in a specific feature space, and these features can reflect the similarities and differences between the samples. For basic driving behavior data, positive driving behavior data, and negative driving behavior data, the network will output the first feature map array, the second feature map array, and the third feature map array, respectively.

[0151] In the process of loading samples into the network for feature mapping, the computer system needs to ensure that the network's input layer matches the encoding format of the samples. At the same time, the system also needs to set the network's hyperparameters (such as learning rate, batch size, etc.) and training strategies (such as training rounds, optimization algorithms, etc.) to control the network's training process.

[0152] In step S430C, the computer system calculates a first spatial distance between the first feature mapping array and the second feature mapping array, and a second spatial distance between the first feature mapping array and the third feature mapping array. Spatial distance is an indicator that measures the distance or similarity between two arrays in feature space, which can help the system evaluate the network's ability to distinguish different driving behavior patterns.

[0153] Commonly used spatial distance measurement methods include Euclidean distance, cosine similarity, Manhattan distance, etc. Among these methods, Euclidean distance is the most intuitive and commonly used one. It calculates the square root of the sum of the squares of the differences between the corresponding elements of two arrays in the feature space. Cosine similarity evaluates the similarity of two arrays by calculating the cosine value of the angle between them. The smaller the angle (the larger the cosine value), the more similar they are.

[0154] After calculating the first spatial spacing and the second spatial spacing, the system needs to determine the spacing error between the two spacings. The spacing error reflects the network's inadequacy in distinguishing basic driving behavior patterns from positive and negative driving behavior patterns. If the spacing error is small, it means that the network can distinguish these patterns well; if the spacing error is large, it means that the network needs to be further optimized to improve its distinguishing ability. The calculation formula for the spacing error may vary depending on the specific measurement method and application scenario. For example, if the Euclidean distance is used as the measurement method, the spacing error can be defined as the absolute value of the difference (or ratio) between the first spatial spacing and the second spatial spacing. The larger this absolute value, the weaker the network's ability to distinguish basic driving behavior from positive or negative driving behavior.

[0155] In step S430D, the computer system determines the situation where the distance error is not less than the preset data as the adjustment direction, and optimizes the network parameters of the driving behavior event relationship mining network. Network parameters refer to adjustable parameters in the network, such as weights and bias items, which determine the output and performance of the network.

[0156] The purpose of optimization is to reduce the spacing error by adjusting the network parameters and improve the network's ability to distinguish driving behavior patterns. Common optimization algorithms include gradient descent, stochastic gradient descent, Adam optimizer, etc. These algorithms calculate the gradient of the loss function (i.e., spacing error) to the network parameters and update the parameter values ​​in the opposite direction of the gradient, thereby gradually reducing the loss function and optimizing network performance.

[0157] During the optimization process, the computer system needs to set appropriate optimization algorithms, learning rates, batch sizes and other hyperparameters, and determine the number of training rounds and termination conditions based on the size and complexity of the training samples. At the same time, the system also needs to monitor the loss function value during training and the performance indicators on the validation set to ensure that the network does not overfit or underfit.

[0158] Taking the Adam optimizer as an example, it combines the advantages of the momentum method and the RMSprop algorithm, and dynamically adjusts the learning rate of each parameter by calculating the first-order moment estimate and the second-order moment estimate of the gradient. During the optimization process, the system updates the value of each parameter according to the gradient information of the current batch of samples and gradually reduces the spacing error. When the training reaches the preset round or the performance index on the validation set no longer improves, the training process will terminate and the optimized driving behavior event relationship mining network will be obtained.

[0159] As an implementation mode, step S430D, determining the distance error not less than the preset data as the adjustment direction, and optimizing the network parameter variables of the driving behavior event relationship mining network, includes:

[0160] Step S430D1: Generate a calibration cost corresponding to the network training sample according to the spacing error and preset data;

[0161] Step S430D2: summing up the calibration costs corresponding to a preset number of network training samples, as a comprehensive cost in a calibration process;

[0162] Step S430D3: Optimize the network parameters of the driving behavior event relationship mining network according to the comprehensive cost.

[0163] In step S430D1, the computer system first generates a calibration cost corresponding to the network training sample according to the spacing error and the preset data. The calibration cost is an indicator to measure the gap between the current performance of the network and the target performance, which reflects the amount of error that needs to be reduced during the training process of the network.

[0164] The spacing error is calculated in step S430C, which represents the spatial spacing difference between the feature mapping array of the basic driving behavior data and the feature mapping array of the positive and negative driving behavior data. The preset data is a threshold set in advance by the system to determine whether the spacing error is acceptable. If the spacing error is greater than the preset data, it means that the network is insufficient in distinguishing different driving behavior patterns and needs to be optimized.

[0165] In order to generate the calibration cost, the computer system can use a variety of methods. A common method is to use the difference between the spacing error and the preset data as the calibration cost. This method is simple and intuitive, and can directly reflect the gap between network performance and target performance. Another method is to use more complex cost functions, such as mean square error (MSE), cross entropy loss, etc., which can more comprehensively measure the difference between the network output and the true label.

[0166] Taking the mean square error as an example, assuming that the spacing error is E and the preset data is T, the calibration cost C can be expressed as: C = (ET)2 .

[0167] This formula calculates the square of the difference between the spacing error and the preset data as the network's tuning cost on the current training sample. If E is greater than T, C is a positive number, indicating that the network needs further optimization; if E is less than or equal to T, C is zero or negative (but in practical applications, C is usually limited to non-negative numbers), indicating that the network has met the performance requirements or performed better than expected.

[0168] In step S430D2, the computer system sums the calibration costs corresponding to the preset number of network training samples as the comprehensive cost in a calibration process. The comprehensive cost is an overall measure of the network performance on the entire training set, which reflects the average performance of the network on multiple samples.

[0169] The preset number refers to a value set in advance by the system to determine the number of training samples that need to be processed during a calibration process. This value can be determined based on factors such as the size of the training set, the complexity of the network, and computing resources. Generally, the larger the preset number, the better the network training effect, but the higher the computing cost.

[0170] When calculating the comprehensive cost, the computer system traverses each sample in the training set, calculates the corresponding adjustment cost of each sample according to the method in step S430D1, and adds these costs to obtain the comprehensive cost. This process can be expressed as: total = ΣC i Among them, C t otal represents the comprehensive cost, C i represents the calibration cost corresponding to the i-th training sample, and Σ represents the summation operation.

[0171] By calculating the comprehensive cost, the computer system can fully understand the performance of the network on the entire training set. If the comprehensive cost is high, it means that the network has a large performance gap on multiple samples and needs further optimization; if the comprehensive cost is low, it means that the network has good performance and you can consider ending the tuning process or fine-tuning.

[0172] In step S430D3, the computer system optimizes the network parameters of the driving behavior event relationship mining network according to the comprehensive cost. The network parameters refer to the adjustable parameters in the network, such as weights and bias terms, which determine the output and performance of the network.

[0173] The purpose of optimization is to reduce the comprehensive cost by adjusting the value of network parameters and improve the network's ability to distinguish different driving behavior patterns. Common optimization algorithms include gradient descent, stochastic gradient descent, Adam optimizer, etc. These algorithms calculate the gradient of the loss function (i.e., comprehensive cost) to the network parameters and update the parameter values ​​in the opposite direction of the gradient, thereby gradually reducing the loss function and optimizing network performance.

[0174] Taking the Adam optimizer as an example, it combines the advantages of the momentum method and the RMSprop algorithm, and dynamically adjusts the learning rate of each parameter by calculating the first-order moment estimate and the second-order moment estimate of the gradient. During the optimization process, the system first calculates the gradient information on the current batch of training samples, and then updates the value of each parameter based on this information. The specific update formula is as follows:

[0175] m t = β1 * m (t-1) + (1-β1) * g t

[0176] v t = β2 * v (t-1) + (1-β2) * g t 2

[0177] M t ' = m t / (1-β1 t )

[0178] V t ' = v t / (1-β2 t )

[0179] θ t +1 = θ t -α * m t ' / (√v t '+ ε)

[0180] Among them, m t and v t They represent the first-order moment estimate and the second-order moment estimate of the gradient, respectively, and g t represents the gradient information on the current batch of training samples, β1 and β2 are momentum coefficients, α is the learning rate, ε is a very small number (used to prevent division by zero errors), and θ t and θ t +1 indicates the current and updated parameter values ​​respectively.

[0181] By updating the parameter values ​​for multiple iterations, the Adam optimizer can gradually reduce the comprehensive cost and improve network performance. When the comprehensive cost is reduced to a certain level or no longer decreases significantly, the optimization process will terminate and the optimized driving behavior event relationship mining network will be obtained.

[0182] As an embodiment, the method further includes:

[0183] Step S430E: obtaining a plurality of coding arrays and a driving behavior event description sequence corresponding to each coding array, wherein the driving behavior event description sequence is obtained by arranging the coding results of the driving behavior events contained in the coding array according to a preset sequence;

[0184] Step S430F: determining a target coding array from among the plurality of coding arrays, and searching, according to the target driving behavior event description sequence corresponding to the target coding array, for a first coding array whose corresponding driving behavior event description sequence is the same as or covers the target driving behavior event description sequence from among the plurality of coding arrays;

[0185] Step S430G: Use the target coding array as basic driving behavior data, use the second coding array as positive driving behavior data corresponding to the target coding array, and use the remaining coding arrays other than the target coding array and the first coding array among the multiple coding arrays as negative driving behavior data corresponding to the target coding array to generate network training samples.

[0186] In step S430E, the computer system first obtains multiple coding arrays and driving behavior event description sequences corresponding to each coding array. The coding array is obtained by encoding the driving behavior event in the previous step, and it contains the coding results of each event, which are arranged in a certain order in the array. The driving behavior event description sequence is obtained by arranging the coding results in the coding array according to a preset sequence, which reflects the time series relationship between the events.

[0187] For example, suppose there is a coding array [A, B, C], where A, B, and C represent the coding results of three driving behavior events: acceleration, constant speed driving, and braking. According to the preset sequence (i.e. the order in which the events occur), the corresponding driving behavior event description sequence "acceleration-constant speed driving-braking" can be generated. This sequence clearly describes the entire process of the vehicle from acceleration to constant speed driving to braking.

[0188] The computer system provides a rich data source for subsequent sample generation by collecting multiple such encoding arrays and corresponding driving behavior event description sequences. These data sources can come from historical driving behavior data, simulated driving experiments or other reliable sources.

[0189] In step S430F, the computer system determines a target coding array from multiple coding arrays, and based on the driving behavior event description sequence corresponding to the target coding array, retrieves the first coding array that is identical to or covers the sequence from the remaining coding arrays. The purpose of this step is to find a coding array that is similar or related to the target coding array so as to subsequently generate positive and negative driving behavior data.

[0190] The target encoding array can be determined randomly or based on a specific selection criterion (such as the diversity of event types, the complexity of the sequence, etc.). Once the target encoding array is determined, the system will traverse the remaining encoding arrays and compare their driving behavior event description sequences with the target sequence for similarity.

[0191] Similarity comparison can be achieved through a variety of methods, such as string matching, edit distance calculation, sequence alignment, etc. Taking string matching as an example, the system can directly check whether the driving behavior event description sequence of each remaining encoding array is exactly the same as the target sequence. If a completely matching sequence is found, the encoding array is used as the first encoding array; if there is no completely matching sequence, but there is a longer sequence that covers the target sequence (that is, the target sequence is a subsequence of the longer sequence), the encoding array corresponding to the longer sequence can also be selected as the first encoding array.

[0192] In step S430G, the computer system generates network training samples using the target coding array, the first coding array, and the remaining coding arrays. Specifically, the system uses the target coding array as basic driving behavior data, the first coding array as positive driving behavior data corresponding to the target coding array, and the remaining coding arrays (except the target coding array and the first coding array) as negative driving behavior data.

[0193] The basic driving behavior data represents a standard or neutral driving behavior pattern, which serves as a benchmark for network training. The positive driving behavior data represents a driving behavior pattern that is better or more positive than the basic data, which may include a more efficient, safer or more environmentally friendly driving strategy. On the contrary, the negative driving behavior data represents a driving behavior pattern that is worse or more negative than the basic data, which may include unsafe, uneconomical or non-compliant driving behavior.

[0194] The training samples generated in this way can cover a variety of driving behavior patterns, which helps the network learn the differences and connections between different patterns. During the training process, the network tries to distinguish the feature differences between basic data, positive data, and negative data, and adjusts its internal parameters accordingly to improve performance. Of course, in practical applications, the generation of positive driving behavior data and negative driving behavior data may not always directly depend on the similarity comparison of the encoding array. Sometimes, the system manually annotates or synthesizes this data based on specific business rules or expert knowledge. For example, in some cases, positive data may come from the driving records of professional drivers, while negative data may come from records of traffic accidents or illegal driving.

[0195] As an implementation manner, step S430, obtaining a network training sample for calibrating a driving behavior event relationship mining network, includes:

[0196] Step S431: obtaining selected basic driving behavior data;

[0197] Step S432: When the driving behavior event relationship in the basic driving behavior data remains unchanged, the basic driving behavior data is data-enhanced to obtain active driving behavior data corresponding to the basic driving behavior data; wherein, when the driving behavior event relationship in the basic driving behavior data remains unchanged, the basic driving behavior data is data-enhanced to obtain active driving behavior data corresponding to the basic driving behavior data in a manner including one or more of the following manners:

[0198] Method 1: changing the order of the basic driving behavior data or all driving behavior events included in the basic driving behavior data;

[0199] Method 2: interpolating or deleting basic driving behavior data;

[0200] Method 3: Under the premise of maintaining the driving behavior events included in the basic driving behavior data, the basic driving behavior data is changed.

[0201] In step S431, the computer system selects representative basic driving behavior data from the massive amount of driving behavior data. These data should cover a variety of driving scenarios, road conditions and driving habits to ensure that the network can learn a wide range of driving behavior characteristics. Basic driving behavior data usually includes a series of driving behavior events arranged in chronological order, such as acceleration, braking, steering, lane change, etc. Each event contains a detailed event feature description, such as timestamp, speed, acceleration, steering angle, etc.

[0202] For example, a basic driving behavior dataset may contain a driving record on a highway, which includes a series of driving behavior events such as the vehicle accelerating from a stationary state, maintaining a constant speed, slowing down and braking when encountering a vehicle in front, and finally stopping smoothly. These data provide a solid foundation for subsequent data reinforcement and network training.

[0203] In step S432, the computer system performs data enhancement processing on the selected basic driving behavior data to generate active driving behavior data. The purpose of data enhancement is to increase the diversity and complexity of the data without changing the relationship between driving behavior events, so that the network can learn more subtle driving behavior characteristics. The data enhancement process needs to follow certain principles, that is, the enhanced data should still maintain the original driving logic and safety.

[0204] In the first method, the computer system rearranges the order of events in the basic driving behavior data to generate a new driving behavior sequence. The core idea of ​​this method is to keep the relative relationship between events unchanged, but change the order in which they occur. For example, in the original data, the vehicle may accelerate first and then brake, while in the enhanced data, this order may be adjusted to brake first and then accelerate (of course, this adjustment needs to conform to the actual driving logic and safety regulations). In this way, the network can learn the impact of different event sequences on driving behavior patterns.

[0205] In the second method, the computer system performs interpolation or data deletion operations on the basic driving behavior data. Interpolation is to insert new data points between existing data points to increase the density and continuity of the data. For example, in the original data, the speed change of the vehicle may be discrete. Through interpolation, a smoother speed change curve can be generated. Data deletion is to randomly delete some data points to simulate the situation of missing data or insufficient sampling. Both operations can increase the diversity and complexity of the data and improve the network's ability to handle incomplete or noisy data.

[0206] In method three, the computer system fine-tunes or changes the event features in the basic driving behavior data to generate new driving behavior patterns. The core idea of ​​this method is to change the specific features of the event (such as speed, acceleration, steering angle, etc.) while keeping the event type and relationship unchanged. For example, in the original data, the vehicle may be traveling at a constant speed of 80km / h, and in the enhanced data, this speed may be adjusted to 90km / h or 70km / h. In this way, the network can learn the impact of different feature values ​​on driving behavior patterns and improve its sensitivity to subtle feature changes.

[0207] Through the implementation of steps S431 to S432 and the application of three methods of data enhancement, the computer system successfully generated a series of high-quality network training samples. These samples not only cover a wide range of driving behavior characteristics, but also increase the diversity and complexity of the data through data enhancement technology. These samples provide strong support for the subsequent adjustment of the driving behavior event relationship mining network, which helps to improve the performance and generalization ability of the network.

[0208] As an implementation manner, step S430, obtaining a network training sample for calibrating the driving behavior event relationship mining network, further includes:

[0209] Step S433: obtaining selected basic driving behavior data;

[0210] Step S434: changing one or more of the driving behavior events and the driving behavior event relationships in the basic driving behavior data to obtain negative driving behavior data corresponding to the basic driving behavior data; wherein changing one or more of the driving behavior events and the driving behavior event relationships in the basic driving behavior data to obtain negative driving behavior data corresponding to the basic driving behavior data includes one or more of the following methods:

[0211] Method A: changing the order of some driving behavior events in the basic driving behavior data;

[0212] Method B: randomly changing one or more driving behavior events in the basic driving behavior data;

[0213] Method C: replacing the remaining driving behavior events in the basic driving behavior data that are different from the designated driving behavior event based on the designated driving behavior event in the similar driving behavior data of the basic driving behavior data;

[0214] Method D: changing the sequence of several driving behavior events in the basic driving behavior data;

[0215] Method E: changing the timing of the specified driving behavior events in the basic driving behavior data to a preset timing.

[0216] In step S433, the computer system selects representative basic driving behavior data from the driving behavior database. These data usually cover a variety of driving scenarios, road conditions and driving habits, and are the basis for the subsequent generation of positive and negative driving behavior data. The basic driving behavior data should contain a series of driving behavior events arranged in chronological order, and each event records in detail the event type, occurrence time, event characteristics (such as speed, acceleration, steering angle, etc.) and the relationship between events. These data provide the network with rich driving behavior samples, which helps the network learn the intrinsic connection between different driving behavior patterns.

[0217] For example, a basic driving behavior dataset may contain a driving record on a city road, including a series of driving behavior events such as the vehicle starting from a stationary state, accelerating, maintaining a constant speed, stopping at a red light, continuing to drive after the green light comes on, and finally stopping at the destination. These data not only record the changes in physical quantities such as the vehicle's speed and acceleration, but also imply information such as the driver's judgment of road conditions, reaction speed, and driving habits.

[0218] In step S434, the computer system processes the selected basic driving behavior data to generate corresponding negative driving behavior data. Negative driving behavior data refers to data that shows unsafe, uneconomical or non-compliant driving behavior, which is crucial for training the network to identify negative driving behavior.

[0219] In method A, the computer system partially adjusts the sequence of driving behavior events in the basic data to simulate the driver's decision-making errors or improper operations during driving. For example, the original sequence may be "start-accelerate-drive at a constant speed-stop", and the adjusted sequence may become "start-accelerate-stop-accelerate" (assuming that the car suddenly accelerates after stopping, which may be an unsafe driving behavior). In this way, the network can learn the impact of changes in the sequence of events on driving behavior patterns and improve the ability to identify improper operating behaviors.

[0220] In method B, the computer system randomly selects one or more driving behavior events in the basic data to change, such as changing the event type, event characteristics, or the relationship between events. This randomness helps simulate uncertainty and unexpected situations in the driving process. For example, an original "constant speed driving" event may be changed to a "sudden acceleration" event, or a "braking" event may be inserted between two consecutive "acceleration" events. These changes may cause significant changes in driving behavior patterns, allowing the network to learn more about negative driving behaviors.

[0221] In method C, the computer system first finds other driving behavior data (called similar data) that are similar to the basic driving behavior data, and then selects one or more specified driving behavior events from the similar data and replaces these events with the remaining positions in the basic data that are different from the specified events. This method aims to enrich the diversity of the basic data by introducing negative driving behavior features from external data. For example, if the similar data contains a "running a red light" event, but there is no such event in the basic data, then the system can replace this "running a red light" event with a suitable position in the basic data to simulate the driver's behavior of violating traffic rules.

[0222] In method D, the computer system makes an overall adjustment to the order of multiple consecutive driving behavior events in the basic data. Compared with method A, method D involves more events and a wider range of changes, so it may produce more significant changes in driving behavior patterns. For example, the original order may be "start-accelerate-drive at a constant speed-decelerate-stop", and the adjusted order may become "start-decelerate-accelerate-drive at a constant speed-stop" (assuming that an unnecessary deceleration operation is performed before acceleration). This method helps the network learn the impact of changes in the order of consecutive events on the overall pattern of driving behavior.

[0223] In method E, the computer system selects one or more specified driving behavior events in the basic data and adjusts their occurrence time to a preset timing position. This adjustment may cause the time interval between events to change, thereby affecting the consistency and rationality of driving behavior. For example, an original "brake" event may occur a short time after the "acceleration" event, but after adjusting the timing, it may be moved to a position far before the "acceleration" event (assuming that the driver performed unnecessary braking operations long before accelerating). This method helps the network learn the impact of time factors on driving behavior patterns and improve the ability to identify time-related negative driving behaviors.

[0224] Through the application of the above five methods, the computer system successfully generated diversified negative driving behavior data from basic driving behavior data. These data not only enriched the diversity of network training samples, but also improved the network's recognition ability and robustness for negative driving behaviors. In the subsequent network calibration process, these negative driving behavior data will be used together with positive driving behavior data to train the network and jointly improve the performance and accuracy of the network.

[0225] As an implementation manner, step S500, based on the event feature description array and the driving behavior event relationship description array, identifies whether the user driving behavior data set to be analyzed matches the reference driving behavior data in the reference behavior pattern data set, including:

[0226] Step S510: when the event feature description array is similar to the event feature description array corresponding to the specified reference driving behavior data, and the driving behavior event relationship description array is similar to the driving behavior event relationship description array corresponding to the specified reference driving behavior data, it is determined that the user driving behavior data set to be analyzed matches the specified reference driving behavior data;

[0227] Step S520: When the event feature description array is not similar to the event feature description array corresponding to the specified reference driving behavior data, or the driving behavior event relationship description array is not similar to the driving behavior event relationship description array corresponding to the specified reference driving behavior data, it is determined that the user driving behavior data set to be analyzed does not match the specified reference driving behavior data.

[0228] In step S510, the computer system first checks whether the event feature description array of the user driving behavior data set to be analyzed is similar to the event feature description array corresponding to the specified reference driving behavior data, and at the same time checks whether the driving behavior event relationship description array is similar to the driving behavior event relationship description array corresponding to the specified reference driving behavior data. Only when these two conditions are met at the same time, the system will determine that the user driving behavior matches the reference driving behavior data.

[0229] The similarity of event feature description arrays can be measured by calculating the similarity score between the two arrays. The similarity score can be calculated using a variety of methods, such as Euclidean distance, cosine similarity, Pearson correlation coefficient, etc. Taking Euclidean distance as an example, the system first calculates the sum of the squares of the differences between the corresponding elements of the two arrays, and then takes the square root to get the distance value. The smaller the distance value, the more similar the two arrays are. In order to determine whether they are similar, the system usually sets a similarity threshold. Only when the calculated similarity score is less than or equal to the threshold, the two arrays are considered similar.

[0230] The similarity judgment of the driving behavior event relationship description array is more complicated because it involves the time series relationship and logical association between events. The system can use methods such as sequence matching algorithms, graph matching algorithms, or deep learning models to evaluate the similarity between two arrays. Taking the sequence matching algorithm as an example, the system can regard the two arrays as time series data, and then calculate the dynamic time warping (DTW) distance or the longest common subsequence (LCS) length between them to measure the similarity. Similarly, the system will set a similarity threshold to determine whether the two arrays are similar.

[0231] The system determines that the user's driving behavior matches the reference driving behavior data only if and only if both the event feature description array and the driving behavior event relationship description array are similar to the arrays corresponding to the specified reference driving behavior data. This means that the user's driving behavior is highly consistent with the reference behavior pattern in both event features and time series relationships.

[0232] In step S520, if the event feature description array is not similar to the event feature description array corresponding to the specified reference driving behavior data, or the driving behavior event relationship description array is not similar to the driving behavior event relationship description array corresponding to the specified reference driving behavior data (or both are not similar), the system will determine that the user's driving behavior does not match the reference driving behavior data.

[0233] The judgment basis of dissimilarity is opposite to the similarity judgment in step S510. If the calculated similarity score between the event feature description arrays is greater than the set similarity threshold, or the similarity index between the driving behavior event relationship description arrays does not meet the matching conditions (such as the DTW distance is too large or the LCS length is too short), the system will consider the two arrays to be dissimilar.

[0234] Mismatches may be caused by a variety of factors. For example, the user's driving speed, acceleration and other characteristics may be significantly different from the reference behavior pattern; or the user's event sequence, time interval and other relationships during driving are inconsistent with the reference behavior pattern; or the user adopts a driving strategy that is different from the reference behavior pattern in certain specific situations (such as emergency avoidance, sudden road conditions, etc.).

[0235] It is worth noting that a mismatch does not necessarily mean that the user's driving behavior is unsafe or non-compliant. In some cases, the mismatch may be caused by the user's personalized driving habits, changes in road conditions, or other external factors. Therefore, when a mismatch is identified, the system needs to further analyze the cause and make a reasonable judgment based on the actual situation.

[0236] As an embodiment, the method further includes:

[0237] Step S900: when it is identified that there is reference driving behavior data that needs to be saved in the reference behavior pattern data set, identifying each driving behavior event included in the reference driving behavior data, and obtaining a driving behavior event identification result corresponding to the reference driving behavior data;

[0238] Step S1000: generating an event feature description array of each driving behavior event in the reference driving behavior data according to the driving behavior event recognition result corresponding to the reference driving behavior data, and obtaining the event feature description array corresponding to the reference driving behavior data;

[0239] Step S1100: extracting the behavior sequence relationship between each driving behavior event included in the reference driving behavior data according to the event feature description array corresponding to each driving behavior event in the reference driving behavior data, and obtaining a driving behavior event relationship description array corresponding to the reference driving behavior data;

[0240] Step S1200: The reference driving behavior data and the event feature description array and the driving behavior event relationship description array corresponding to the reference driving behavior data are associated and saved in the reference behavior pattern data set.

[0241] In step S900, when the computer system recognizes that there is new reference driving behavior data that needs to be saved in the reference behavior pattern data set, it first analyzes the data to identify the various driving behavior events contained therein. This step is the basis for subsequent processing, and it requires the system to accurately distinguish different types of driving behavior events, such as acceleration, braking, turning, lane change, etc.

[0242] To achieve this goal, the system can use event recognition algorithms or models. These algorithms or models usually distinguish different types of driving behavior events based on event characteristics (such as speed changes, acceleration, steering angle, etc.). For example, the system can set a series of thresholds and rules. When a sudden increase in speed is detected and the acceleration exceeds a certain threshold, it is identified as an "acceleration" event; when a rapid decrease in speed is detected and the braking force is large, it is identified as a "braking" event.

[0243] Through this step, the system can obtain the driving behavior event recognition results corresponding to the reference driving behavior data, that is, a list or array containing all recognized driving behavior events and their related information. This information may include the type of event, the time of occurrence, the duration, and related parameters (such as speed, acceleration, etc.).

[0244] In step S1000, the computer system generates an event feature description array for each driving behavior event in the reference driving behavior data based on the driving behavior event recognition result obtained in step S900. This array is a multi-dimensional vector or matrix, which contains multiple parameters or indicators describing the event characteristics.

[0245] The generation of event feature description arrays usually involves the extraction, quantification, and normalization of event-related data. For example, for an "acceleration" event, the system can extract characteristic parameters such as its starting speed, ending speed, acceleration, and acceleration duration, and quantify these parameters into numerical form. Then, in order to eliminate the dimensional differences and order of magnitude differences between different parameters, the system can normalize these values ​​and convert them to the same scale.

[0246] Through this step, the system can obtain a set of event feature description arrays corresponding to the reference driving behavior data, that is, a list or matrix containing all event feature description arrays. These arrays provide basic data for subsequent behavior sequence relationship extraction and pattern recognition.

[0247] In step S1100, the computer system further extracts the behavior sequence relationship between each driving behavior event in the reference driving behavior data based on the event feature description array set obtained in step S1000, and generates a driving behavior event relationship description array. This step is the key to understanding the driving behavior pattern, which requires the system to capture the time sequence, logical association and potential pattern between events. To achieve this goal, the system can use sequence analysis algorithms or models. These algorithms or models can process time series data and identify the time dependency and potential pattern between events. For example, the system can use a dynamic time warping (DTW) algorithm to compare the similarity between different event sequences, or use a hidden Markov model (HMM) to learn the generation process of event sequences.

[0248] Through this step, the system can obtain the driving behavior event relationship description array corresponding to the reference driving behavior data. This array may be a multi-dimensional vector or matrix, which contains multiple features or indicators describing the relationship between events. These features or indicators may include the time interval between events, frequency distribution, pattern repetition, etc.

[0249] In step S1200, the computer system associates and saves the reference driving behavior data, event feature description array, and driving behavior event relationship description array obtained in steps S900 to S1100, and adds them together to the reference behavior pattern data set. This step is an important part of data management and maintenance, which ensures the integrity and consistency of the reference data.

[0250] When performing associated storage, the system needs to assign a unique identifier or label to each reference driving behavior data so that it can be easily identified and used in subsequent data retrieval and analysis. At the same time, the system also needs to ensure that there is a clear association between the event feature description array and the driving behavior event relationship description array and the corresponding reference driving behavior data so that relevant data can be quickly and accurately obtained when needed.

[0251] Through this step, the system not only completes the processing and storage of new reference driving behavior data, but also provides reliable data support for subsequent driving behavior pattern recognition, premium management strategy formulation, etc. As the reference behavior pattern data set continues to be enriched and improved, the system's analysis capabilities and decision-making level will also be continuously improved.

[0252] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application is shown in FIG. Figure 2As shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

[0253] The above is only an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A driving behavior data analysis method for premium management, characterized in that: include: Acquire multiple sets of driving behavior data sets collected during the user's driving process during the evaluation period, and use one set of the driving behavior data sets as the user driving behavior data set to be analyzed; Identify each driving behavior event included in the user driving behavior data set to be analyzed, and obtain a driving behavior event identification result; Generating an event feature description array of each driving behavior event according to the driving behavior event recognition result; Acquire a plurality of group centroids, wherein the group centroids are obtained by grouping and clustering an event feature description array of a driving behavior event included in the reference driving behavior data; Determine the group centroid having the smallest distance from the event feature description array of each driving behavior event according to the vector space distance between the event feature description array of each driving behavior event and the plurality of group centroids; Generate encoding results for each driving behavior event according to a group code corresponding to a group centroid with the smallest distance in an event feature description array of each driving behavior event; Generate a coding array corresponding to the user driving behavior data set to be analyzed according to the time series distribution of each driving behavior event in the user driving behavior data set to be analyzed and the coding result of each driving behavior event; Extracting a driving behavior event relationship description array from the encoding array; According to the event feature description array and the driving behavior event relationship description array, identifying whether the user driving behavior data set to be analyzed matches the reference driving behavior data in the reference behavior pattern data set; If the reference driving behavior data is matched, the driving behavior pattern corresponding to the reference driving behavior data is used as the driving behavior pattern for the set of driving behavior collection data sets; Analyze the remaining groups of driving behavior collection data sets to obtain the driving behavior patterns corresponding to each group of driving behavior collection data sets, and perform weighted fusion on the driving behavior patterns corresponding to the multiple groups of driving behavior collection data sets to obtain a fused driving behavior pattern; A corresponding premium management strategy is generated based on the fused driving behavior pattern.

2. The driving behavior data analysis method for premium management according to claim 1, characterized in that: The method further comprises: Acquire the number of known driving behavior events and the number of driving behavior events associated with the known driving behavior events, and acquire an event feature description array of the driving behavior events in the reference driving behavior data; Determining the number of group centroids according to the number of the known driving behavior events and the number of the associated driving behavior events; According to the number of the grouping centroids, the event feature description arrays of the driving behavior events contained in the reference driving behavior data are grouped and clustered to obtain the plurality of grouping centroids.

3. The driving behavior data analysis method for premium management according to claim 1, characterized in that: According to the time series distribution of each driving behavior event in the user driving behavior data set to be analyzed, and the encoding result of each driving behavior event, a coding array corresponding to the user driving behavior data set to be analyzed is generated, including: According to the time series distribution of each driving behavior event in the user driving behavior data set to be analyzed, replacing the time series data corresponding to each driving behavior event based on the encoding result of each driving behavior event; The data other than the time series corresponding to each driving behavior event in the user driving behavior data set to be analyzed is changed into preset data to obtain a coding array corresponding to the user driving behavior data set to be analyzed; Extracting the driving behavior event relationship description array from the encoding array includes: Loading the encoding array into a driving behavior event relationship mining network, wherein the driving behavior event relationship mining network includes a plurality of cascaded feature extraction components, and the encoding array is loaded into a first feature extraction component among the plurality of feature extraction components; The outputs of several feature extraction components among the multiple feature extraction components are respectively fused based on the dimension reduction operation to obtain the driving behavior event relationship description array.

4. The driving behavior data analysis method for premium management according to claim 3, characterized in that: The method further comprises: Acquire a network training sample for calibrating the driving behavior event relationship mining network, wherein the network training sample includes basic driving behavior data, positive driving behavior data, and negative driving behavior data; wherein the basic driving behavior data, positive driving behavior data, and negative driving behavior data are all coded data; Loading the network training sample into the driving behavior event relationship mining network to obtain a first feature mapping array for the basic driving behavior data, a second feature mapping array for the negative driving behavior data, and a third feature mapping array for the positive driving behavior data output by the driving behavior event relationship mining network; Determining a spacing error between the first spatial spacing and the second spatial spacing based on a first spatial spacing between the first feature mapping array and the second feature mapping array, and a second spatial spacing between the first feature mapping array and the third feature mapping array; The distance error is not less than the preset data and is determined as the adjustment direction, and the network parameter variables of the driving behavior event relationship mining network are optimized.

5. The driving behavior data analysis method for premium management according to claim 4, characterized in that: The spacing error is not less than the preset data and is determined as the adjustment direction, and the network parameter variables of the driving behavior event relationship mining network are optimized, including: Generating a calibration cost corresponding to the network training sample according to the spacing error and the preset data; The sum of the tuning costs corresponding to the preset number of network training samples is taken as the comprehensive cost in a tuning process; Optimizing network parameters of the driving behavior event relationship mining network according to the comprehensive cost; The method further comprises: Acquire a plurality of coding arrays and a driving behavior event description sequence corresponding to each coding array, wherein the driving behavior event description sequence is obtained by arranging coding results of driving behavior events contained in the coding array according to a preset sequence; Determine a target coding array among the multiple coding arrays, and according to a target driving behavior event description sequence corresponding to the target coding array, retrieve a first coding array among the multiple coding arrays whose corresponding driving behavior event description sequence is the same as or covers the target driving behavior event description sequence; The target coding array is used as basic driving behavior data, the first coding array is used as positive driving behavior data corresponding to the target coding array, and the remaining coding arrays among the multiple coding arrays, except the target coding array and the first coding array, are used as negative driving behavior data corresponding to the target coding array to generate the network training samples.

6. The driving behavior data analysis method for premium management according to claim 4, characterized in that: Obtaining a network training sample for calibrating the driving behavior event relationship mining network includes: Obtain selected basic driving behavior data; When the driving behavior event relationship in the basic driving behavior data remains unchanged, performing data enhancement on the basic driving behavior data to obtain active driving behavior data corresponding to the basic driving behavior data; Wherein, when the driving behavior event relationship in the basic driving behavior data remains unchanged, the basic driving behavior data is enhanced to obtain the active driving behavior data corresponding to the basic driving behavior data, including one or more of the following methods: Method 1: changing the order of the basic driving behavior data or all driving behavior events included in the basic driving behavior data; Method 2: interpolating or deleting the basic driving behavior data; Method 3: changing the basic driving behavior data while maintaining the driving behavior events included in the basic driving behavior data; Acquiring a network training sample for calibrating the driving behavior event relationship mining network also includes: Obtain selected basic driving behavior data; Changing one or more of the driving behavior events and driving behavior event relationships in the basic driving behavior data to obtain negative driving behavior data corresponding to the basic driving behavior data; Among them, changing one or more of the driving behavior events and driving behavior event relationships in the basic driving behavior data to obtain the negative driving behavior data corresponding to the basic driving behavior data includes one or more of the following methods: Mode A: changing the order of some driving behavior events in the basic driving behavior data; Method B: randomly changing one or more driving behavior events in the basic driving behavior data; Mode C: replacing the remaining driving behavior events in the basic driving behavior data that are different from the designated driving behavior event based on the designated driving behavior event in the similar driving behavior data of the basic driving behavior data; Mode D: changing the sequence of several driving behavior events in the basic driving behavior data; Mode E: changing the timing of the specified driving behavior events in the basic driving behavior data to a preset timing.

7. The driving behavior data analysis method for premium management according to any one of claims 1 to 6, characterized in that: According to the event feature description array and the driving behavior event relationship description array, identifying whether the user driving behavior data set to be analyzed matches the reference driving behavior data in the reference behavior pattern data set, including: When the event feature description array is similar to the event feature description array corresponding to the specified reference driving behavior data, and the driving behavior event relationship description array is similar to the driving behavior event relationship description array corresponding to the specified reference driving behavior data, determining that the user driving behavior data set to be analyzed matches the specified reference driving behavior data; When the event feature description array is not similar to the event feature description array corresponding to the specified reference driving behavior data, or the driving behavior event relationship description array is not similar to the driving behavior event relationship description array corresponding to the specified reference driving behavior data, determining that the user driving behavior data set to be analyzed does not match the specified reference driving behavior data; The method further comprises: When it is identified that there is reference driving behavior data that needs to be saved in the reference behavior pattern data set, identifying each driving behavior event included in the reference driving behavior data, and obtaining a driving behavior event identification result corresponding to the reference driving behavior data; generating an event feature description array of each driving behavior event in the reference driving behavior data according to a driving behavior event recognition result corresponding to the reference driving behavior data, to obtain an event feature description array corresponding to the reference driving behavior data; Extracting the behavior sequence relationship between each driving behavior event included in the reference driving behavior data according to the event feature description array corresponding to each driving behavior event in the reference driving behavior data, and obtaining the driving behavior event relationship description array corresponding to the reference driving behavior data; The reference driving behavior data is associated with an event feature description array and a driving behavior event relationship description array corresponding to the reference driving behavior data and saved in the reference behavior pattern data set.

8. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 7 are implemented.

Images