Method and system for generating a portrait of a vehicle owner based on dynamic multi-granularity granular cell fusion

Abstract

The application discloses a vehicle owner portrait generation method and system based on dynamic multi-granularity granular ball fusion, and belongs to the technical field of data processing. The method comprises the following steps: collecting multi-source heterogeneous driving data of a target vehicle, converting different dimension data to a unified dimensionless space through multi-core manifold mapping; constructing a bottom layer information granular ball in a unit of a homogenous and semantic feature group in a same time window, each granular ball comprising a center vector and a radius; calculating the spatial convolution response between the granular balls, and determining a fusion weight in combination with an offline information gain; and generating a vehicle owner portrait hypersphere representing the behavior characteristics of the vehicle owner by weightedly fusing the bottom layer information granular balls according to the fusion weight. The application realizes interpretable feature fusion by granular ball modeling, and the generated vehicle owner portrait hypersphere can stably and timely reflect the behavior mode of the vehicle owner, thereby solving the problems of data heterogeneity conflict, unstable static portrait and poor interpretability in the prior art.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for generating vehicle owner profiles based on dynamic multi-granularity particle fusion. Background Technology

[0002] With the rapid development of intelligent connected vehicles, in-vehicle sensing systems, and vehicle networking platforms, vehicles continuously generate a large amount of data related to owner behavior during operation, including vehicle trajectory data, CAN bus data, driving operation data, usage time data, and environmental context data. Utilizing this data to build owner profiles helps achieve application goals such as driving risk identification, user segmentation operations, insurance pricing optimization, personalized service recommendations, and fleet safety management.

[0003] Existing methods for constructing vehicle owner profiles mainly fall into three categories. The first is to directly concatenate multi-source features and input them into a classification or clustering model; the second is to weight and summarize statistical features to form a fixed-dimensional vector; and the third is to utilize deep learning to learn end-to-end representations of time-series data. However, these methods have significant shortcomings in the context of connected vehicle data.

[0004] First, the sources of driving data are significantly heterogeneous. For example, trajectory coordinates, speed, braking pressure, and timestamps have different physical dimensions and statistical distributions. Direct splicing or simple normalization will lead to a certain type of high-variance feature dominating the profile result, causing profile deviation.

[0005] Secondly, most existing methods lack interpretable intermediate structures and cannot clearly explain how behavioral features from different sources play a role in the profile formation process, which is not conducive to model credibility analysis and subsequent iterative optimization.

[0006] Therefore, it is necessary to propose a new method for generating car owner profiles to solve the above problems. Summary of the Invention

[0007] This invention provides a method and system for generating vehicle owner profiles based on dynamic multi-granularity particle fusion, in order to solve the problems in the prior art where the profile results are biased due to the inconsistent dimensions of multi-source heterogeneous driving data, and the lack of interpretability in the profile process.

[0008] This invention is achieved through the following technical solution:

[0009] A first aspect of the present invention provides a method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion, comprising:

[0010] The vehicle's driving data is collected, including spatiotemporal trajectory data, CAN bus behavior data, vehicle usage time data, and environmental context data.

[0011] The driving data is synchronized according to a preset time window, and the driving data with different physical dimensions is converted into a unified dimensionless manifold vector through multi-core manifold mapping.

[0012] The underlying information spheres are constructed using feature groups from the dimensionless manifold vector that belong to the same data source and the same semantic category within the same time window; each of the underlying information spheres contains its center vector and radius.

[0013] Calculate the spatial convolution response between any two of the underlying information particles within the same time window, and determine the fusion weight by combining the offline information gain of each underlying information particle relative to the preset standard label;

[0014] The underlying information particles are weighted and fused according to the fusion weights to generate the owner profile hypersphere of the target vehicle.

[0015] Furthermore, the method also includes:

[0016] Collect incremental driving data of the target vehicle;

[0017] Calculate the deviation level of the incremental driving data relative to the hypersphere of the vehicle owner profile;

[0018] When the deviation levels of N consecutive incremental driving data exceed a preset threshold, the vehicle owner profile hypersphere is updated using the N incremental driving data.

[0019] Furthermore, updating the vehicle owner profile hypersphere using N incremental driving data includes:

[0020] A time decay factor is introduced for each of the incremental driving data; wherein, the more recent the incremental driving data is, the larger its time decay factor.

[0021] The time decay factor is used as a weighting coefficient to calculate the weighted mean of the center vectors of the N incremental driving data.

[0022] The center vector of the hypersphere of the car owner's portrait is corrected by using the weighted mean to obtain the corrected center vector.

[0023] The radius of the hypersphere of the vehicle owner profile is expanded and updated using N incremental driving data to obtain the corrected radius;

[0024] The updated hypersphere of the car owner profile is obtained based on the corrected center vector and the corrected radius.

[0025] Furthermore, the step of constructing the underlying information sphere using the set of data points belonging to the same data source and the same semantic category within the same time window from the dimensionless manifold vector includes:

[0026] The set of data points in the dimensionless manifold vector that are within the same time window, belong to the same data source, and have the same semantic category is taken as a feature group.

[0027] For each feature group, the center vector and radius of the corresponding underlying information sphere are calculated based on its set of data points.

[0028] The center vector of the underlying information sphere is the mean of all data points in the feature group, and the radius is calculated as follows:

[0029] Calculate the covariance matrix of all data points in the feature group, and calculate the first radius component based on the trace of the covariance matrix;

[0030] The mean distance from all data points in the feature group to the center vector is calculated to obtain the second radius component, and the second radius component is corrected by a robust correction coefficient.

[0031] The radius of the underlying information sphere is obtained by adding the first radius component to the robustly corrected second radius component.

[0032] Furthermore, the spatial convolution response between the two underlying information particles is calculated as follows:

[0033] ;

[0034] in, This represents the spatial convolution response of the underlying information spheres s and u. , These represent the center vectors of the underlying information particles s and u, respectively. , Let radii be s and u, representing the bottom-level information spheres, respectively. To prevent extremely small constants with a denominator of zero.

[0035] Furthermore, the determination of fusion weights by combining the offline information gain of each underlying information particle relative to a preset standard label includes:

[0036] Calculate the spatial convolution sum of each bottom-level information sphere:

[0037] ;

[0038] in, This represents the spatial convolution sum of the underlying information particle s with all other underlying information particles;

[0039] Calculate the information gain of the feature group corresponding to each bottom-level information particle on the label set:

[0040] ;

[0041] in, This represents the information gain corresponding to the underlying information particle s. Represents a set of tags Information entropy This indicates that the underlying information spheres The corresponding feature group after binning the state variables Conditional label set Conditional information entropy, The state variables of the feature group corresponding to the bottom information particle s after binning;

[0042] Calculate the fusion weights of each underlying information particle:

[0043] ;

[0044] in, This represents the fusion weight of the underlying information particle s. This represents the spatial convolution sum of the underlying information particle u with all other underlying information particles. This represents the information gain corresponding to the underlying information particle u.

[0045] Further, the step of weightedly fusing the underlying information particles according to the weights to generate the owner profile hypersphere of the target vehicle includes:

[0046] Calculate the center vector of the hypersphere containing the car owner's profile:

[0047] ;

[0048] in, The fusion weights for the underlying information particles s The center vector of the underlying information particle s;

[0049] Calculate the radius of the hypersphere containing the car owner's portrait:

[0050] ;

[0051] in, Let be the radius of the underlying information particle s;

[0052] Construct the hypersphere of the vehicle owner profile based on the center vector and radius of the hypersphere of the vehicle owner profile: .

[0053] Furthermore, the method also includes:

[0054] The vehicle owner portrait hypersphere is compared with the pre-generated standard label particles, and the topological overlap between the vehicle owner portrait hypersphere and each of the standard label particles is calculated.

[0055] Calculate the membership degree of the hypersphere of the car owner profile to each standard label based on the topological overlap, and output a car owner profile report containing the membership degree.

[0056] Furthermore, the topological overlap between the vehicle owner's portrait hypersphere and each of the standard tag particles is calculated using an improved Jaccard topological overlap method; the improved Jaccard topological overlap method is:

[0057] Calculate the equivalent intersection hypervolume between the vehicle owner's portrait hypersphere and the standard label particle sphere. Union supervolume ;

[0058] Calculate the improved Jaccard topology overlap:

[0059] ;

[0060] in, Represents the relationship between the hypersphere center vector of the car owner's portrait and the first... The Euclidean distance between the center vectors of the standard label particles. This indicates that the car owner's portrait exceeds the radius of the sphere. This indicates the radius of the standard label pellet. To prevent extremely small constants with a denominator of zero.

[0061] A second aspect of the present invention provides a vehicle owner profile generation system based on dynamic multi-granularity particle fusion, used to execute the vehicle owner profile generation method based on dynamic multi-granularity particle fusion as described in any one of the first aspects of the present invention, the system comprising:

[0062] The data acquisition module is used to collect driving data of the target vehicle, including spatiotemporal trajectory data, CAN bus behavior data, vehicle usage time data, and environmental context data.

[0063] The preprocessing and mapping module is used to perform time synchronization processing on each of the driving data according to a preset time window, and to convert driving data with different physical dimensions into a unified dimensionless manifold vector through multi-core manifold mapping.

[0064] The particle sphere construction module is used to construct a low-level information particle sphere using feature groups from the dimensionless manifold vector that belong to the same data source and the same semantic category within the same time window; each low-level information particle sphere contains its center vector and radius;

[0065] The fusion weight generation module is used to calculate the spatial convolution response between any two of the underlying information particles within the same time window, and to determine the fusion weight by combining the offline information gain of each underlying information particle relative to the preset standard label.

[0066] The profile construction module is used to perform weighted fusion of the underlying information particles according to the fusion weights to generate a hypersphere of the owner profile of the target vehicle. Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0067] The method of this invention transforms data with different physical dimensions into a unified dimensionless space through multi-kernel manifold mapping, fundamentally eliminating the impact of heterogeneous data dimension differences on the profiling results. By constructing a low-level information particle sphere with a central vector and radius, an interpretable intermediate structure is provided for subsequent fusion. The fusion weights are determined by spatial convolution response and information gain, ensuring that the fusion process considers both the structural synergy within the data and utilizes prior knowledge of label discrimination. The final generated hypersphere of the car owner profile has a clear geometric boundary in the unified space, stably representing the car owner's behavioral patterns. These technical means collectively solve the problems of profile offset caused by data heterogeneity and the lack of interpretability in the profiling process in existing technologies. Attached Figure Description

[0068] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0069] Figure 1 This is a flowchart of a method for generating a vehicle owner profile based on dynamic multi-granularity particle fusion according to an embodiment of the present invention;

[0070] Figure 2 This is a flowchart of a method for calculating the radius of a bottom-level information particle according to an embodiment of the present invention;

[0071] Figure 3 This is a flowchart of a method for dynamically evolving and updating a hypersphere of a vehicle owner's profile, according to an embodiment of the present invention.

[0072] Figure 4 This is a block diagram of a vehicle owner profile generation system based on dynamic multi-granularity particle fusion, according to an embodiment of the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0074] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims, and accompanying drawings of this invention are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to other steps or units inherent in the device.

[0075] The terminology used in the various embodiments of the invention is for the purpose of describing particular embodiments only and is not intended to limit the various embodiments of the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. The terms (such as those defined in commonly used dictionaries) are to be interpreted as having the same meaning as in the context of the relevant technical field and are not to be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.

[0076] To address the problems of existing technologies, this invention provides an interpretable granular profiling modeling mechanism. By performing unified manifold mapping on multi-source heterogeneous vehicle network data, data from different sources, with different dimensions and statistical distributions, is transformed into a unified dimensionless feature space. Within this unified space, underlying information granular spheres are constructed at the granularity of time windows and feature groups. Then, spatial convolution and information gain are used to jointly determine the fusion weights, generating a hypersphere representing the current behavioral state of the vehicle owner. This mechanism is applicable to vehicle network cloud platforms, edge computing nodes, in-vehicle terminals, or their combined deployment architectures.

[0077] like Figure 1 As shown, the present invention proposes a method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion, which includes the following steps:

[0078] S101, Multi-source heterogeneous data acquisition: Collects driving data of the target vehicle, including: spatiotemporal trajectory data, CAN bus behavior data, vehicle usage time data, and environmental context data.

[0079] The spatiotemporal trajectory data includes latitude and longitude, speed, acceleration, driving direction, road grade, and mileage increment; the CAN bus behavior data includes braking pressure, throttle opening, steering wheel angle, engine speed, gear position, and light status; the vehicle usage time data includes departure time, end time, day / night markers, and weekday / non-working day markers; and the environmental context data includes weather type, temperature range, road congestion level, and typical road segment attributes.

[0080] S102, Data preprocessing: Time synchronization processing of each vehicle data according to the preset time window, and conversion of vehicle data with different physical dimensions into a unified dimensionless manifold vector through multi-core manifold mapping.

[0081] The raw data is segmented according to a preset time window. Starting from a unified time point, the raw data is segmented, and each driving data point is processed into a segment with a length of [length missing]. Multiple data segments.

[0082] To eliminate the influence of differences in physical dimensions and distributions, multiple kernel functions are used to perform manifold mapping on different types of data, resulting in dimensionless manifold vectors. Kernel functions suitable for the statistical characteristics of different data sources are selected.

[0083] In one specific implementation, the spatiotemporal trajectory data is mapped using a Gaussian radial basis kernel function, as shown below:

[0084]

[0085] in, Represents a data segment (length is...) The i-th sampling point in ) This represents the j-th sampling point. This represents the scaling parameter of the Gaussian radial basis kernel function corresponding to the spatiotemporal trajectory data.

[0086] The Laplace kernel function is used to map CAN bus behavioral data, as shown below:

[0087]

[0088] in, Represents a data segment (length is...) The i-th sampling point in ) This represents the j-th sampling point. This represents the scale parameter of the Laplace kernel function corresponding to the CAN bus behavior data.

[0089] The vehicle usage time data is mapped using a periodic kernel function, as shown below:

[0090]

[0091] in, Indicates the first Individual vehicle usage time data sample, Indicates the first Individual vehicle usage time data sample, Indicates the first The timestamp corresponding to each vehicle usage time data sample Indicates the first The timestamp corresponding to each vehicle usage time data sample This is a periodic parameter, expressed in hours or days, preferably 24 hours or 7 days. The scaling parameter represents the periodic kernel function corresponding to the vehicle usage time data.

[0092] For environmental context data, a Gaussian radial basis function (or a mixture kernel function) is used for mapping to handle its mixed categorical (e.g., weather type, road segment attributes) and numerical (e.g., temperature, congestion level) features, as shown below:

[0093]

[0094] in, , Represents a data segment (length is...) The first in ) , One sampling point, This refers to the scaling parameter of the corresponding kernel function. For discrete categorical features such as weather type and typical road segment attributes, one-hot encoding or embedding representation can be performed first, and then the above kernel function can be applied; ordered numerical features such as temperature range and congestion level can be directly used in distance calculation.

[0095] Then, by normalizing each kernel matrix and summing them according to their weights, the combined kernel matrix can be obtained:

[0096]

[0097] in, , , and These are the scaling parameters for the corresponding kernel functions. Preferably, It can be determined based on the classification contribution or mutual information contribution on the offline validation set, usually taking [value missing]. In order to balance the impact of environmental factors on car owners' behavior patterns.

[0098] Preferably, the scaling parameters corresponding to each kernel function can be determined based on the classification contribution, cluster silhouette coefficient, or cross-validation accuracy on the offline validation set. The parameters of the environmental context kernel function can also be adjusted prior to incorporate domain knowledge (such as the weight of the influence of weather on driving behavior).

[0099] By performing kernel principal component analysis, local linear embedding, Laplacian eigenmaps, or their equivalent manifold embedding methods on the combined kernel matrix, a unified dimensionless manifold vector is obtained.

[0100] The raw data is segmented according to a preset time window. This time window can be a fixed sliding time window, an event-triggered time window, or a hybrid time window. Preferably, a time window with a length of [missing information] is used. A sliding time window, and set the sliding step size to... There are phases, and adjacent data segments exist. The overlapping time periods.

[0101] Furthermore, within each time window, the following preprocessing operations are performed on the raw data:

[0102] 1. Time synchronization processing: Align data from different sampling frequencies according to a unified timestamp;

[0103] 2. Missing value repair: Short-term missing segments are filled in using linear interpolation, nearest neighbor imputation, or model estimation.

[0104] 3. Outlier suppression: Remove, truncate, or replace data that exceeds the physically reasonable range;

[0105] 4. Standard Feature Group Division: Based on data source and semantic consistency, features are divided into several feature groups with the same origin and semantics;

[0106] 5. Multi-kernel manifold mapping processing: Select a kernel function that matches the statistical properties of each feature group corresponding to the same time window and convert it into a unified dimensionless manifold vector.

[0107] S103, Construction of underlying information spheres: Underlying information spheres are constructed using feature groups from dimensionless manifold vectors that belong to the same data source and the same semantic category within the same time window; each underlying information sphere contains its center vector and radius.

[0108] In this step, the underlying information granules are not constructed individually for each raw feature, but rather as groups within the same time window, belonging to the same data source, and of the same semantic category. For example, for various types of driving data, they can be divided into trajectory feature groups, CAN behavior feature groups, and vehicle usage time feature groups. The trajectory feature group consists of data of different semantic categories, such as average vehicle speed, speed fluctuation, turning frequency, and road grade proportion; the CAN behavior feature group consists of data of different semantic categories, such as braking intensity, rapid acceleration frequency, and steering wheel angle change rate; and the vehicle usage time feature group consists of data of different semantic categories, such as nighttime travel proportion, weekday vehicle usage proportion, and single continuous vehicle usage duration.

[0109] For the s-th feature group within the w-th time window, its mapped sample set is denoted as:

[0110]

[0111] in, Indicates the first One sample point, The total number of sample points for the s-th feature group within the w-th time window.

[0112] Define the underlying information sphere as:

[0113]

[0114] in, This represents the underlying information sphere constructed from the s-th feature group within the w-th time window. The center vector of the underlying information sphere. Let be the radius of the underlying information sphere. The center vector reflects the typical position of this feature group in the manifold space, i.e., the central tendency of the data group. The radius reflects the dispersion of the data within the feature group; a larger radius indicates a more unstable or diverse behavioral pattern within the feature group.

[0115] The information of the underlying information spheres may also include the number of sample points that construct them. .

[0116] In one specific implementation, the center vector is defined as the mean of all data points in the feature set, and it is expressed as:

[0117]

[0118] The radius is defined as the statistical characteristic of the covariance matrix of a feature group and the degree of data dispersion. Specifically, for example... Figure 2 As shown, the radius is calculated as follows:

[0119] S201, calculate the covariance matrix of all data points in the feature group, and calculate the first radius component based on the trace of the covariance matrix;

[0120] S202, calculate the mean distance of all data points in the feature group to the center vector to obtain the second radius component, and correct the second radius component by the robust correction coefficient;

[0121] S203, add the first radius component to the robustly corrected second radius component to obtain the radius of the bottom information particle.

[0122] Underlying information spheres The radius is expressed as:

[0123]

[0124] in, Represents the covariance matrix. The trace of the covariance matrix reflects the statistical properties of the second moment. The dimension and radius of the covariance matrix are... The first term represents the average dispersion of each dimension, and the second term is the robustness correction term. This is a correction factor.

[0125] Among them, the correction coefficient Used to balance the statistical properties of covariance (global second-order dispersion) and robust first-order dispersion (more robust to outliers), its value range is typically [range missing]. In one specific implementation, It can be determined in the following ways:

[0126] 1. Offline Cross-Validation Optimization: On the labeled offline training dataset, the optimal cross-validation method is determined using the topological overlap (or final classification / clustering accuracy, F1 score) between the hypersphere of the vehicle owner profile and the standard label particle sphere as the optimization objective. This method employs grid search, random search, or Bayesian optimization to determine the optimal cross-validation. value.

[0127] 2. Data-driven adaptive: For each feature group, dynamically calculate based on its sample distribution characteristics. ,For example:

[0128]

[0129] Here, MAD stands for Median Absolute Deviation, used to measure the discrete robustness of the distribution. This method enables... Automatically reduce the weights in feature groups with a large number of outliers to improve robustness.

[0130] 3. Empirical default value: In actual deployment, it can be initially set to... And make periodic fine-tuning based on online feedback (such as the stability of profile updates and business metrics).

[0131] The covariance matrix is:

[0132]

[0133] S104, Multi-sphere fusion: Calculate the spatial convolution response between any two bottom-level information spheres within the same time window, and determine the fusion weights by combining the offline information gain of each bottom-level information sphere relative to the preset standard label.

[0134] Within the same time window, multiple underlying information particles can be obtained, denoted as:

[0135]

[0136] To reflect the structural synergy between the underlying information particles, this invention applies the following to any two underlying information particles: and Calculate the spatial convolution response term:

[0137]

[0138] in, Represents the underlying information sphere and Spatial convolution response, To prevent extremely small constants with a denominator of zero.

[0139] Calculate the underlying information sphere Spatial convolution sum:

[0140]

[0141] in, Represents the underlying information sphere and all other underlying information particles within the same time window ( The sum of spatial convolution responses.

[0142] At the same time, in order to reflect different feature groups To assess the importance of label recognition, the information gain corresponding to the feature set is calculated based on offline training samples.

[0143]

[0144] in, This indicates the construction of the underlying information sphere. Feature groups Information gain Represents a set of tags. Representing feature groups The state variables after discretization Represents a set of tags Information entropy This indicates that the underlying information spheres The corresponding feature group after binning the state variables Conditional label set The conditional information entropy.

[0145] Based on this, the underlying information spheres are calculated. The fusion weights are:

[0146]

[0147] in, , Representing the underlying information particles respectively Spatial convolution and information gain, This represents all underlying information particles within the same time window. ( The sum of the spatial convolution and the product of information gain.

[0148] S105, based on the fusion weight, weighted fusion of each underlying information particle to generate a supersphere of the target vehicle's owner profile.

[0149] After obtaining the fusion weights, construct a hypersphere of car owner profiles: ,like Figure 4 The image shown is a schematic diagram of a hypersphere representing a car owner's profile. Its center vector is calculated as follows:

[0150]

[0151] The radius is:

[0152]

[0153] in, This indicates the typical behavioral position of the target vehicle within the current time window. It comprehensively reflects the dispersion of the grain center and the dispersion within the grain, and is used to characterize the fluctuation range and uncertainty boundary of the behavior pattern.

[0154] The portrait construction method of this invention first unifies heterogeneous data into a dimensionless space through multi-kernel manifold mapping, resolving the issue of dimension conflict. Then, it constructs underlying information spheres based on homologous and semantically similar feature groups. The center of each sphere represents a typical behavioral pattern, and its radius represents the range of behavioral fluctuations, providing an interpretable intermediate structure. Next, it measures the structural synergy between spheres through spatial convolution response and measures the contribution of each feature group to label recognition through offline information gain; the two are combined to determine the fusion weight. Finally, weighted fusion yields a hypersphere for the vehicle owner's portrait. The geometric boundary of this hypersphere has clear statistical significance in the unified space, and can stably and intuitively represent the vehicle owner's behavioral patterns.

[0155] In one specific implementation, the method further includes a step of dynamically evolving and updating the hypersphere of the vehicle owner's profile, such as... Figure 3 As shown, it includes:

[0156] S301, collect incremental driving data of the target vehicle;

[0157] S302, calculate the deviation level of incremental driving data relative to the hypersphere of the vehicle owner profile;

[0158] S303: When the deviation levels of N consecutive incremental driving data exceed the preset threshold, the hypersphere of the car owner profile is updated using the N incremental driving data.

[0159] Driver profiles are dynamic over time; driving styles may drift due to changes in road conditions, vehicle status, commuting habits, or driving experience. Existing static profiling methods often build profiles based on fixed historical intervals, making it difficult to reflect behavioral evolution trends in a timely manner. This can lead to profile lag or severe fluctuations caused by short-term abnormal behavior. To avoid drastic fluctuations in profiles caused by short-term abnormal samples, while ensuring timely updates as behavior continues to change, this embodiment introduces a continuous threshold triggering mechanism to dynamically evolve and update the hypersphere of the driver profile.

[0160] Specifically, for the new sample set arriving at the current moment (Incremental driving data), defined as its hypersphere relative to the driver profile at the previous moment. The deviation energy level is:

[0161]

[0162] in, To prevent extremely small constants with a denominator of zero.

[0163] When arriving in chronological order All incremental driving data meet the requirements. This triggers the evolution and update of the hypersphere. Among them, This represents the evolution threshold.

[0164] In one specific implementation, the method for updating the vehicle owner profile hypersphere using N incremental driving data includes:

[0165] S401 introduces a time decay factor for each incremental driving data; the closer the incremental driving data is to the current time, the larger its time decay factor.

[0166] S402, using the time decay factor as a weighting coefficient to calculate the weighted mean of the center vector of N incremental driving data;

[0167] S403, use the weighted mean to correct the drift of the center vector of the hypersphere of the car owner's portrait, and obtain the corrected center vector;

[0168] S404, using N incremental driving data to expand and update the radius of the hypersphere of the car owner profile, and obtain the corrected radius;

[0169] S405, the updated hypersphere of the car owner profile is obtained based on the corrected center vector and the corrected radius.

[0170] Specifically, for N consecutive incremental driving data points that trigger an update, it is represented as follows:

[0171]

[0172] Definition of the first Time decay factor for incremental driving data:

[0173]

[0174] in, For the first The time interval between each incremental driving data point and the current moment. This is the time decay factor; the closer the incremental driving data is to the current moment, the larger the time decay factor.

[0175] Calculate the weighted mean of all samples within the trigger window:

[0176]

[0177] Correct the center for drift:

[0178]

[0179] in, This is the evolutionary step size.

[0180] Expand and update the radius:

[0181]

[0182] In this implementation, the update mechanism adopts an incremental correction method of "historical profile + new data triggered by the window", which does not require backtracking all historical samples, thus balancing update efficiency and profile continuity.

[0183] In one specific implementation, the standard tagged particle library construction step is also included, comprising:

[0184] S501, offline collection of multi-source heterogeneous driving data of multiple sample vehicles / sample owners, and execution of steps S102 and S103 to generate a standard granular representation of each sample vehicle / sample owner in a unified manifold space;

[0185] S502, cluster the center vectors of all standard particles of sample vehicles / sample owners with the same preset behavior label to form multiple label clusters;

[0186] S503, For each tag cluster, calculate its geometric center as the center vector of the standard tag particle sphere corresponding to that tag;

[0187] S504, calculate the Euclidean distance from all sample points in the tag cluster to their center vector, and take the preset high quantile as the radius of the standard tag sphere, thereby forming the standard tag sphere corresponding to the tag cluster.

[0188] The standard tag pellets corresponding to all tag clusters constitute the standard tag pellet library.

[0189] Specifically, sample vehicles / owners with the same preset behavioral labels are clustered in a unified manifold space, and each label cluster is further clustered. The center vector of its standard tag pellet is calculated as follows:

[0190]

[0191] in, Represents a tag cluster The i-th sample in the series, For tag clusters The number of sample points.

[0192] With tag clusters The quantiles of the distances from each sample point to the cluster center are used as the radius of the standard tag sphere, expressed as:

[0193]

[0194] in, The optimal value is 80% to 95%.

[0195] Tag clusters The corresponding standard label pellet representation is as follows: .

[0196] In one specific implementation, the method further includes the step of matching the vehicle owner profile hypersphere with a standard tag particle library to output a vehicle owner profile report, including:

[0197] S601, compare the hypersphere of the vehicle owner's portrait with the pre-generated standard label particles, and calculate the topological overlap between the hypersphere of the vehicle owner's portrait and each standard label particle.

[0198] S602 calculates the membership degree of the hypersphere of the car owner profile to each standard label based on the topological overlap, and outputs a car owner profile report containing the membership degree.

[0199] Specifically, let the hypersphere of the current car owner's profile be:

[0200]

[0201] No. The standard labeled pellets are:

[0202]

[0203] The distance between their centers is:

[0204]

[0205] Calculate the equivalent intersection radius:

[0206]

[0207] in, The hypersphere center vector of the car owner profile and the first The Euclidean distance between the center vectors of the standard label particles.

[0208] calculate Volume constant of a 3D hypersphere:

[0209]

[0210] in, Let be the dimension of the manifold space. This is the Gamma function.

[0211] The equivalent intersecting hypervolume is defined as:

[0212]

[0213] Define the union hypervolume:

[0214]

[0215] The Jaccard topological overlap is calculated as follows:

[0216]

[0217] in, To prevent extremely small constants with a denominator of zero.

[0218] Furthermore, the improved Jaccard topological overlap is used to calculate the topological overlap between the vehicle owner's portrait hypersphere and each standard label particle; the improved Jaccard topological overlap is:

[0219]

[0220] The improvement lies in introducing a radius matching and center distance decay mechanism into the traditional Jaccard similarity. Traditional Jaccard similarity (or Jaccard index) only considers the volume ratio of the intersection to the union of two sets, i.e. It only reflects the degree of volume overlap between hyperspheres and is not sensitive to the distance between the centers or the difference in radius.

[0221] The improved Jaccard topological overlap of this invention has two key extensions based on the traditional method:

[0222] Radius matching item ( This implicitly considers the influence of radius differences on the equivalent intersection / union volume. It involves calculating the equivalent intersection hypervolume of two hyperspheres in high-dimensional space. Union supervolume This makes the overlap more sensitive to differences in radius size.

[0223] Center distance attenuation term ( ): Introducing the Euclidean distance between the center of the hypersphere and the center of the standard label particle sphere in the car owner profile The overlap is normalized to the sum of the radii of the two hyperspheres. When the centers of the two hyperspheres are far apart, the overlap decreases rapidly, thus better reflecting the geometric positional relationship.

[0224] The above improvements enable the overlap calculation to reflect not only the volume overlap of the two hyperspheres, but also the relative positional relationship of the centers, thus enhancing the stability and geometric rationality of the label determination.

[0225] Traditional vehicle owner profiling typically employs single-label or hard classification methods, compressing complex behavioral patterns into a single label. This makes it difficult to represent the actual overlapping labels, transitional states, and ambiguous boundary features. For example, a vehicle owner may simultaneously possess both the attributes of steady driving and frequent nighttime travel, which existing technologies struggle to interpretably represent within a unified framework. To address this issue, this invention adopts the following classification strategy:

[0226] S6021, normalize the topological overlap between the vehicle owner profile hypersphere and each standard label particle sphere, and use the normalized value as the dynamic membership degree.

[0227] S6022, when the dynamic membership degree of a tag is greater than a preset first threshold. When this happens, the tag is designated as the main tag and output.

[0228] S6023, when at least two labels have a dynamic membership degree that is simultaneously greater than a preset second threshold. Furthermore, the difference in membership degrees between the two labels is less than a preset third threshold. When this happens, the two tags are output together as a composite tag;

[0229] S6024, when the dynamic membership degree of all tags is low, the first threshold... When the target vehicle is identified as a transitional user, its fuzzy boundary attributes are output.

[0230] Among them, transitional users refer to car owners whose behavior patterns are in a transitional or mixed state between multiple standard labels. Their driving habits and car use characteristics have not yet stabilized into a typical profile (such as conservative, high-frequency night driving, aggressive driving, etc.), or they simultaneously possess local characteristics of multiple labels.

[0231] The criterion is that the topological overlap between the target vehicle's owner profile hypersphere and all preset standard label particles is low, resulting in low calculated dynamic membership degrees. All less than the first threshold ( It can be configured according to business needs, for example, taking At this point, the system does not force a single label, but marks it as a transitional state and outputs the membership vector of each label as a fuzzy boundary attribute for subsequent personalized services, risk warnings, or manual review.

[0232] Among them, the target vehicle is the first The dynamic membership degree of a label is represented as follows:

[0233]

[0234] in, This indicates that the car owner's portrait is a hypersphere and the standard label sphere is a particle. Topological overlap.

[0235] This invention effectively solves the problem of ambiguous scenarios such as overlapping multiple labels and transitional behavior evolution that are difficult to handle by traditional hard classification methods by introducing transitional user categories, thereby improving the robustness and interpretability of car owner profiles in real and complex car use environments.

[0236] Furthermore, this embodiment of the invention provides a method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion, the specific steps of which are as follows.

[0237] Step 1: Multi-source heterogeneous data acquisition, synchronization and feature grouping.

[0238] Collect multi-source heterogeneous data generated during the continuous operation of the target vehicle. Among them:

[0239] The sampling period for spatiotemporal trajectory data is 1 to 5 seconds;

[0240] The CAN bus behavior data sampling period is 10 milliseconds to 100 milliseconds;

[0241] Vehicle usage time data is collected in an event logging manner;

[0242] Environmental context data is updated based on road segment switching, weather switching, or a fixed period.

[0243] Data is sliced ​​using a sliding time window with a length of 30 minutes and a step size of 10 minutes. For each time window:

[0244] 1. Align all raw data to a unified timeline;

[0245] 2. Perform linear interpolation on short-term missing data with a length of no more than 5 sampling points;

[0246] 3. Truncate data that exceeds the physically reasonable range;

[0247] 4. Group features according to semantics, for example:

[0248] The trajectory feature set includes: average vehicle speed, speed standard deviation, number of turns, and road grade ratio;

[0249] The CAN behavioral characteristic group includes: braking peak, frequency of rapid acceleration, and steering wheel angle change rate;

[0250] The time-related characteristic groups include: the proportion of nighttime travel, the proportion of weekday travel, and the duration of a single trip.

[0251] Step 2: Multi-kernel manifold mapping.

[0252] For each of the aforementioned feature groups, a corresponding kernel matrix is ​​constructed, and then combined and fused, as shown in the formula above. In this step, the kernel parameters are determined using an offline validation set grid search method, as follows:

[0253] The value range is 0.1 to 3.0;

[0254] The value range is 0.1 to 2.0;

[0255] The value range is 0.1 to 1.5;

[0256] This was determined through cross-validation.

[0257] Performing manifold embedding on the combined kernel matrix yields a unified dimensionless vector space with dimension Preferably, it has 4 to 16 dimensions.

[0258] Step 3: Construction of the underlying information spheres.

[0259] For each time window and each feature group The mapped sample set is represented as:

[0260]

[0261] Calculate the center, covariance matrix, and radius of the underlying information particles. An example is provided below:

[0262] Suppose that within a certain time window, a feature set is mapped to obtain 20 samples, with a dimension of... The calculation yielded the following:

[0263]

[0264] And the mean of the robustness term is:

[0265]

[0266] Pick ,but:

[0267]

[0268] Therefore, the underlying information sphere corresponding to this feature group can be represented as:

[0269]

[0270] Step 4: The underlying information particles are fused to generate a hypersphere of the car owner profile.

[0271] The spatial convolutional response term and convolution sum are calculated for the underlying information particles within the same time window, and then the fusion weights are calculated in conjunction with the information gain. Assume there are three underlying information particles within the current time window, and their convolution sums are as follows:

[0272]

[0273] Their information gains are as follows:

[0274]

[0275] The corresponding weights are:

[0276]

[0277] The calculation yielded the following result:

[0278]

[0279] The weighted summation of the centers of each sphere yields the supersphere center of the car owner's profile. Then, the radius of the supersphere is obtained by calculating the car owner's portrait using the radius calculation formula. Therefore, the hypersphere representing the car owner's profile in the current time window is obtained:

[0280]

[0281] Step 5: Dynamic evolution of the portrait.

[0282] For new samples arriving in real time Calculate the deviation energy level:

[0283]

[0284] When continuous All samples satisfy:

[0285]

[0286] This triggers an evolutionary update.

[0287] Preferably, the following is taken: ; ; ; .

[0288] Perform time-decay weighted sampling on samples within the trigger window, update the center and radius, and obtain a new hypersphere for the image. .

[0289] Step 6: Tag matching and profile output.

[0290] The pre-built library of standard tag particles is traversed, and the improved Jaccard topological overlap between the current image hypersphere and each standard tag particle is calculated.

[0291] Let the radius of the current image hypersphere be:

[0292]

[0293] The radius of a certain standard label pellet is:

[0294]

[0295] The center distance is:

[0296]

[0297] The equivalent intersection radius is:

[0298]

[0299] If the dimension of the manifold space is ,but:

[0300]

[0301] Further, we can obtain:

[0302]

[0303]

[0304] Therefore, we get:

[0305]

[0306] Then for all tags Normalization is performed to obtain the dynamic membership degree. When the membership degree corresponding to a certain label is at its maximum and exceeds the threshold... When the membership degree of multiple tags is close, output the main tag; if multiple tags have similar membership degrees, output the composite tag or the transition tag.

[0307] Compared with the prior art, the present invention has at least the following beneficial effects:

[0308] 1. The multi-core manifold mapping mechanism can effectively solve the problems of inconsistent dimensions and distribution differences in multi-source heterogeneous vehicle network data;

[0309] 2. Construct low-level information spheres based on groups of features with the same origin and semantic meaning to avoid representation instability caused by excessively fine granularity of a single feature;

[0310] 3. By combining spatial convolution and information gain for weighting, the underlying behavioral structure and label discrimination capabilities are synergistically integrated;

[0311] 4. The calculation method of particle radius and image hypersphere radius driven by second moment is adopted to make the image boundary have clear statistical significance;

[0312] 5. Through time decay and continuous over-threshold triggering update mechanism, it can both suppress short-term abnormal disturbances and respond promptly to continuous changes in behavior;

[0313] 6. By improving the Jaccard topological overlap and dynamic membership output, we can realize the fuzzy boundary expression and transition state characterization of vehicle owner tags, thereby improving the authenticity and interpretability of the profile results.

[0314] Embodiments of the present invention also provide a vehicle owner profile generation system based on dynamic multi-granularity particle fusion, used to execute the vehicle owner profile generation method based on dynamic multi-granularity particle fusion in any of the above embodiments, such as... Figure 4 As shown, the system includes:

[0315] The data acquisition module 401 is used to collect driving data of the target vehicle, including spatiotemporal trajectory data, CAN bus behavior data, vehicle usage time data and environmental context data.

[0316] The preprocessing and mapping module 402 is used to perform time synchronization processing on each vehicle data according to a preset time window, and to convert vehicle data with different physical dimensions into a unified dimensionless manifold vector through multi-core manifold mapping.

[0317] The particle building module 403 is used to construct the underlying information particle spheres using feature groups in the dimensionless manifold vector that belong to the same data source and the same semantic category within the same time window; each underlying information particle sphere contains its center vector and radius.

[0318] The fusion weight generation module 404 is used to calculate the spatial convolution response between any two bottom information particles within the time window, and to determine the fusion weight by combining the offline information gain of each bottom information particle relative to the preset standard label.

[0319] The profile building module 405 is used to perform weighted fusion of the underlying information particles according to the fusion weight to generate a supersphere of the owner profile of the target vehicle.

[0320] Furthermore, the system also includes a dynamic evolution module, which monitors the deviation level of new samples from the current owner profile hypersphere and incrementally updates the current owner profile hypersphere when the continuous threshold condition is met.

[0321] Furthermore, the system also includes a tag matching module, which loads a standard tag particle library and calculates the improved Jaccard topological overlap between the current vehicle owner profile hypersphere and each tag particle.

[0322] Furthermore, the system also includes a profile output module, which is used to generate a vehicle owner profile report based on dynamic membership and output the main label, composite label, transition label and their corresponding confidence information.

[0323] Embodiments of the present invention also provide an electronic device, which includes a processor and a memory, wherein the number of processors may be one or more. The memory, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules. The processor executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory, thereby implementing the vehicle owner profile generation method based on dynamic multi-granularity particle fusion according to any of the above embodiments of the present invention.

[0324] The memory may primarily comprise a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory may further include memory remotely located relative to the processor, which can be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks (LANs), mobile communication networks, and combinations thereof.

[0325] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the vehicle owner profile generation method based on dynamic multi-granularity particle fusion according to any embodiment of the present invention.

[0326] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0327] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0328] Embodiments of the present invention also provide a computer program product that, when run on a computer, causes the computer to execute the vehicle owner profile generation method based on dynamic multi-granularity particle fusion of any of the above embodiments of the present invention.

[0329] The above embodiments are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the appended claims.

Claims

1. A method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion, characterized in that, include: The vehicle's driving data is collected, including spatiotemporal trajectory data, CAN bus behavior data, vehicle usage time data, and environmental context data. The driving data is synchronized according to a preset time window, and the driving data with different physical dimensions is converted into a unified dimensionless manifold vector through multi-core manifold mapping. The underlying information spheres are constructed using feature groups from the dimensionless manifold vector that belong to the same data source and the same semantic category within the same time window; each of the underlying information spheres contains its center vector and radius. Calculate the spatial convolution response between any two of the underlying information particles within the same time window, and determine the fusion weight by combining the offline information gain of each underlying information particle relative to the preset standard label; The underlying information particles are weighted and fused according to the fusion weights to generate the owner profile hypersphere of the target vehicle.

2. The method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion according to claim 1, characterized in that, The method further includes: Collect incremental driving data of the target vehicle; Calculate the deviation level of the incremental driving data relative to the hypersphere of the vehicle owner profile; When the deviation levels of N consecutive incremental driving data exceed a preset threshold, the vehicle owner profile hypersphere is updated using the N incremental driving data.

3. The method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion according to claim 2, characterized in that, The process of updating the vehicle owner profile hypersphere using N incremental driving data includes: A time decay factor is introduced for each of the incremental driving data; wherein, the time decay factor is larger for incremental driving data that is closer to the current time. The time decay factor is used as a weighting coefficient to calculate the weighted mean of the center vectors of the N incremental driving data. The center vector of the hypersphere of the car owner's portrait is corrected by using the weighted mean to obtain the corrected center vector. The radius of the hypersphere of the vehicle owner profile is expanded and updated using N incremental driving data to obtain the corrected radius; The updated hypersphere of the car owner profile is obtained based on the corrected center vector and the corrected radius.

4. The method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion according to claim 1, characterized in that, The construction of the underlying information sphere using the set of data points belonging to the same data source and the same semantic category within the same time window from the dimensionless manifold vector includes: The set of data points in the dimensionless manifold vector that are within the same time window, belong to the same data source, and have the same semantic category is taken as a feature group. For each feature group, the center vector and radius of the corresponding underlying information sphere are calculated based on its set of data points. The center vector of the underlying information sphere is the mean of all data points in the feature group, and the radius is calculated as follows: Calculate the covariance matrix of all data points in the feature group, and calculate the first radius component based on the trace of the covariance matrix; The mean distance from all data points in the feature group to the center vector is calculated to obtain the second radius component, and the second radius component is corrected by a robust correction coefficient. The radius of the underlying information sphere is obtained by adding the first radius component to the robustly corrected second radius component.

5. The method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion according to claim 1, characterized in that, The spatial convolution response between the two underlying information spheres is calculated as follows: ； in, This represents the spatial convolution response of the underlying information spheres s and u. , These represent the center vectors of the underlying information particles s and u, respectively. , Let radii be s and u, representing the bottom-level information spheres, respectively. To prevent extremely small constants with a denominator of zero.

6. The method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion according to claim 5, characterized in that, The determination of fusion weights by combining the offline information gain of each underlying information particle relative to a preset standard label includes: Calculate the spatial convolution sum of each bottom-level information sphere: ； in, This represents the spatial convolution sum of the underlying information particle s with all other underlying information particles; Calculate the information gain of the feature group corresponding to each bottom-level information particle on the label set: ； in, This represents the information gain corresponding to the underlying information particle s. Represents a set of tags Information entropy This indicates that the underlying information spheres The corresponding feature group after binning the state variables Conditional label set Conditional information entropy, The state variables of the feature group corresponding to the bottom information particle s after binning; Calculate the fusion weights of each underlying information particle: ； in, This represents the fusion weight of the underlying information particle s. This represents the spatial convolution sum of the underlying information particle u with all other underlying information particles. This represents the information gain corresponding to the underlying information particle u.

7. The method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion according to claim 5, characterized in that, The step of weighted fusing the underlying information particles according to the weights to generate the owner profile hypersphere of the target vehicle includes: Calculate the center vector of the hypersphere containing the car owner's profile: ； in, The fusion weights for the underlying information particles s The center vector of the underlying information particle s; Calculate the radius of the hypersphere containing the car owner's portrait: ； in, Let be the radius of the underlying information particle s; Construct the hypersphere of the vehicle owner profile based on the center vector and radius of the hypersphere of the vehicle owner profile: .

8. The method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion according to any one of claims 1-7, characterized in that, The method further includes: The vehicle owner portrait hypersphere is compared with the pre-generated standard label particles, and the topological overlap between the vehicle owner portrait hypersphere and each of the standard label particles is calculated. Calculate the membership degree of the hypersphere of the car owner profile to each standard label based on the topological overlap, and output a car owner profile report containing the membership degree.

9. The method for generating vehicle owner profiles based on dynamic multi-granularity particle fusion according to claim 8, characterized in that, The topological overlap between the vehicle owner's portrait hypersphere and each of the standard tag particles is calculated using an improved Jaccard topological overlap method; the improved Jaccard topological overlap method is: Calculate the equivalent intersection hypervolume between the vehicle owner's portrait hypersphere and the standard label particle sphere. Union supervolume ; Calculate the improved Jaccard topology overlap: ； in, Indicates the relationship between the hypersphere center vector of the car owner's portrait and the first... The Euclidean distance between the center vectors of the standard label particles. This indicates that the car owner's portrait exceeds the radius of the sphere. This indicates the radius of the standard label pellet. To prevent extremely small constants with a denominator of zero.

10. A vehicle owner profile generation system based on dynamic multi-granularity particle fusion, characterized in that, The system is used to execute the vehicle owner profile generation method based on dynamic multi-granularity particle fusion as described in any one of claims 1-9, the system comprising: The data acquisition module is used to collect driving data of the target vehicle, including spatiotemporal trajectory data, CAN bus behavior data, vehicle usage time data, and environmental context data. The preprocessing and mapping module is used to perform time synchronization processing on each of the driving data according to a preset time window, and to convert driving data with different physical dimensions into a unified dimensionless manifold vector through multi-core manifold mapping. The particle sphere construction module is used to construct a low-level information particle sphere using feature groups from the dimensionless manifold vector that belong to the same data source and the same semantic category within the same time window; each low-level information particle sphere contains its center vector and radius; The fusion weight generation module is used to calculate the spatial convolution response between any two of the underlying information particles within the same time window, and to determine the fusion weight by combining the offline information gain of each underlying information particle relative to the preset standard label. The profile building module is used to perform weighted fusion of the underlying information particles according to the fusion weight to generate the owner profile hypersphere of the target vehicle.

Images