Big model-based vehicle networking data security enhancement method and system

Abstract

This invention relates to the field of vehicle-to-everything (V2X) data security technology, and particularly to a method and system for enhancing V2X data security based on a large model. The method includes: collecting and splitting the original time-series service data sequence from the vehicle; separating the macroscopic trajectory node sequence from the spatial displacement parameters and the onboard high-frequency inertial sensor sequence within adjacent timestamp intervals; fusing the spatial displacement parameters and the onboard high-frequency inertial sensor sequence to extract deviation feature values ​​as kinematic feature parameters; associating time-series mutation digital tags when the deviation exceeds a dynamic tolerance threshold; importing the macroscopic trajectory node sequence into a global semantic feature matching network to output the comparison results; and inversely mapping the time-series mutation digital tags to the corresponding time nodes. When the comparison results are compliant and the tag frequency exceeds a preset service fault tolerance threshold, the original time-series service data sequence from the vehicle is blocked from entering the commercial value assessment stage.

Description

Technical Field

[0001] This invention relates to the field of vehicle network data security technology, and in particular to a method and system for enhancing vehicle network data security based on a large model. Background Technology

[0002] Commercial intelligent connected vehicle fleet operation monitoring technology uploads driving trajectory and vehicle status data to a cloud server in real time through onboard terminals. This data is then analyzed using big data analytics algorithms to assess fleet operational efficiency and commercial value. To prevent malicious data tampering for subsidy fraud, the industry has introduced large-scale models as data security review engines. These engines use algorithms to extract global semantic features from long-term data series. Global semantic features represent the logical coherence of a data segment at a macro level. The security review engine compares the matching degree of these global semantic features with normal business operation logic to determine whether the business flow data uploaded by the onboard terminals has been forged.

[0003] Existing methods for enhancing vehicle network data security have significant limitations. A core technical challenge lies in the overemphasis on global semantic feature extraction by large-scale models, neglecting temporal mutations at the terminal's micro-sampling layer. This allows high-frequency commercial forgery of data to successfully bypass security checks. When processing massive fleet operation data, the algorithms within the large-scale model's computational framework allocate substantial computational resources to the macro-level business process rationality assessment steps. Attackers intentionally create micro-temporal signal mutations in the normal data stream by frequently injecting forged state values ​​with extremely short durations. These minute signal mutations do not disrupt the overall logical coherence of the fleet operation data at the macro level. Using the fleet cargo transportation route monitoring process as a business scenario, unauthorized users continuously insert forged location coordinates deviating from the normal route into the trajectory data packet every few seconds. During the review and evaluation, the large-scale model only determines that the overall starting and ending points and main routes conform to the global semantic features of conventional logistics transportation, causing the security review engine to fail to recognize the spatial location anomalies occurring between adjacent second-level timestamps. Sudden shifts in coordinates on the high-frequency time axis are incorrectly identified by data analysis algorithms as legitimate fluctuations in sensor environmental noise. This allows illegal operators to bypass regulatory mechanisms and obtain commercial logistics transportation subsidies by using falsely accumulated mileage data generated at high frequencies. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for enhancing the data security of the Internet of Vehicles (IoV) based on a large model. This invention solves the technical problem that the large model's excessive focus on global semantics and neglect of temporal mutations in the terminal's micro-sampling layer leads to the successful bypassing of security reviews by high-frequency commercial forged data in the IoV.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: In a first aspect, the vehicle-to-everything (V2X) data security enhancement method based on a large model provided by the present invention includes: Collect the vehicle-side raw time-series business data sequence continuously uploaded by the vehicle-mounted node; Data splitting is performed on the original time-series service data sequence of the vehicle to separate the macroscopic trajectory node sequence, as well as the spatial displacement parameters and vehicle high-frequency inertial sensing sequence within the adjacent timestamp intervals corresponding to the original time-series service data sequence of the vehicle. By integrating spatial displacement parameters and vehicle-mounted high-frequency inertial sensing sequences to extract deviation feature values, these deviation feature values ​​are set as kinematic feature parameters. The kinematic feature parameters are compared with dynamic tolerance thresholds. When the kinematic feature parameters are greater than the dynamic tolerance threshold, temporal abrupt change digital tags are associated with adjacent timestamp intervals. The macroscopic trajectory node sequence is imported into the global semantic feature matching network of the large model, and the comparison results between the macroscopic trajectory node sequence and the preset logistics transportation business specifications are output. Establish a baseline time sliding alignment window, reverse map the time-series mutation digital tags to the corresponding time nodes included in the macro trajectory node sequence, and count the frequency of occurrence of the time-series mutation digital tags within the baseline time sliding alignment window. Under the condition that the comparison result is compliant and the frequency of occurrence is greater than the preset business fault tolerance threshold, block the original time-series business data sequence of the vehicle end from entering the business value assessment stage.

[0006] Furthermore, the vehicle-to-everything (V2X) data security enhancement method based on a large model of the present invention performs data splitting on the original time-series service data sequence of the vehicle, separating the macroscopic trajectory node sequence, as well as the spatial displacement parameters and onboard high-frequency inertial sensing sequence within adjacent timestamp intervals corresponding to the original time-series service data sequence of the vehicle, including: Extract the time-series synchronization message headers from the original time-series service data sequence on the vehicle side; Compare the absolute timestamp recorded in the timing synchronization message header with the relative time offset parameter of the vehicle bus; The data splitting instruction is triggered when the absolute timestamp is perfectly aligned with the relative time offset parameter of the vehicle bus.

[0007] Furthermore, the vehicle network data security enhancement method based on a large model of the present invention further includes, after triggering the data splitting instruction: Based on the spatial grid division standard, the original time-series business data sequence of the vehicle end is separated to generate a macroscopic trajectory node sequence; Extract the coordinate difference of the original time-series business data sequence from the vehicle end within adjacent sampling periods to generate spatial displacement parameters; Simultaneously extract the three-dimensional spatial acceleration signal and yaw rate signal within adjacent sampling periods, and merge the three-dimensional spatial acceleration signal and yaw rate signal into an on-board high-frequency inertial sensing sequence.

[0008] Furthermore, the vehicle-to-everything (V2X) data security enhancement method based on a large model of the present invention, after merging the three-dimensional spatial acceleration signal and the yaw rate signal into an onboard high-frequency inertial sensing sequence, further includes: Extract chassis topology location labels from the vehicle-mounted high-frequency inertial sensing sequence; Match the chassis topology location labels with the preset vehicle firmware component assembly mapping table; When the chassis topology location label matches the preset vehicle firmware component assembly mapping table, a calculus calculation trigger signal is output.

[0009] Furthermore, the vehicle-to-everything (V2X) data security enhancement method based on a large model of the present invention integrates spatial displacement parameters and vehicle-mounted high-frequency inertial sensing sequences to extract deviation feature values, including: It receives the calculus operation trigger signal, performs dual time calculus operations on the three-dimensional spatial acceleration signal and the yaw rate signal, and generates a dynamic displacement vector; An interpolation algorithm is used to force the temporal resolution of the spatial displacement parameter to be aligned with the vehicle-mounted high-frequency inertial sensing sequence, and a standardization algorithm is used to eliminate the absolute dimensional differences between the spatial displacement parameter and the dynamically derived displacement vector. Based on the normalized data, the deviation covariance matrix between the spatial displacement parameter and the dynamically derived displacement vector is established. Singular value decomposition is performed on the deviation covariance matrix to extract the principal eigenvalues ​​as kinematic feature parameters.

[0010] Furthermore, the vehicle-to-everything (V2X) data security enhancement method based on a large model of the present invention compares kinematic feature parameters with dynamic tolerance thresholds, including: Obtain the fleet dynamic load status descriptor that is perfectly aligned with the adjacent timestamp interval. The fleet dynamic load status descriptor records the vehicle chassis suspension airbag pressure sensor readings. The vehicle chassis suspension airbag pressure sensor reading is used as an index key value to input into the dynamic tolerance database, and the nonlinear suspension system dynamic error tolerance value corresponding to the vehicle chassis suspension airbag pressure sensor reading is retrieved as the dynamic tolerance threshold. When the principal eigenvalue is greater than the dynamic error tolerance value of the nonlinear suspension system, associate time-series abrupt change digital tags with adjacent timestamp intervals.

[0011] Furthermore, the vehicle-to-everything (V2X) data security enhancement method based on a large model of the present invention imports the macroscopic trajectory node sequence into the global semantic feature matching network of the large model, including: Map the macroscopic trajectory node sequence to discretized business operation status identifiers; Convert discretized business operation status identifiers into high-dimensional semantic vectors; High-dimensional semantic vectors are input into the global semantic feature matching network of a large model.

[0012] Furthermore, the vehicle network data security enhancement method based on a large model of the present invention outputs a comparison result between the macroscopic trajectory node sequence and the preset logistics transportation business specifications, including: Retrieve the vector template of compliant business specifications included in the preset logistics and transportation business specifications; Calculate the spatial cosine similarity between the high-dimensional semantic vector and the compliance business specification vector template; If the spatial cosine similarity meets the preset standard threshold, a comparison result that is in compliance status is generated.

[0013] Furthermore, the vehicle-to-everything (V2X) data security enhancement method based on a large model of the present invention includes statistically analyzing the frequency of occurrence of time-series abrupt change digital tags within a reference time sliding alignment window, including: Accumulate the total number of occurrences of the timing abrupt change digital labels within the reference time sliding alignment window; Calculate the ratio of the total number of occurrences of the time-series abrupt change digital label to the total length of the reference time sliding alignment window, and generate the occurrence frequency; If the comparison results are compliant and the frequency of occurrence exceeds the preset business fault tolerance threshold, the transmission of the original time-series business data sequence from the vehicle to the business value assessment stage is blocked by closing the vehicle data receiving port, and the unprocessed original time-series business data sequence from the vehicle is cleared simultaneously.

[0014] Secondly, the vehicle network data security enhancement system based on a large model provided by the present invention is applied to the vehicle network data security enhancement method based on a large model as described above, including: The data acquisition module is used to collect the original time-series business data sequences continuously uploaded by the vehicle-mounted nodes. The data splitting module is used to split the original time-series business data sequence of the vehicle end, separating the macro trajectory node sequence, as well as the spatial displacement parameters and vehicle high-frequency inertial sensing sequence within the adjacent timestamp intervals corresponding to the original time-series business data sequence of the vehicle end. The feature comparison module is used to extract deviation feature values ​​by fusing spatial displacement parameters and vehicle-mounted high-frequency inertial sensing sequences. The deviation feature values ​​are set as kinematic feature parameters. The kinematic feature parameters are compared with the dynamic tolerance threshold. When the kinematic feature parameters are greater than the dynamic tolerance threshold, the module associates time-series change digital tags with adjacent timestamp intervals. The semantic matching module is used to import the macro trajectory node sequence into the global semantic feature matching network of the large model and output the comparison results between the macro trajectory node sequence and the preset logistics and transportation business specifications. The safety blocking module is used to establish a reference time sliding alignment window, reverse map the time-series mutation digital tags to the corresponding time nodes included in the macro trajectory node sequence, count the frequency of occurrence of the time-series mutation digital tags in the reference time sliding alignment window, and block the original time-series business data sequence on the vehicle side from entering the business value assessment stage if the comparison result is compliant and the frequency of occurrence is greater than the preset business fault tolerance threshold.

[0015] Beneficial effects of this invention: This invention provides a method and system for enhancing vehicle network data security based on a large model, overcoming the technical limitations of existing large models that overemphasize global semantic feature extraction while neglecting temporal mutations at the terminal's micro-sampling layer. It performs dual-channel data splitting on the collected original vehicle-side time-series business data sequence, separating the macro-trajectory node sequence and the spatial displacement parameters and onboard high-frequency inertial sensing sequences within adjacent timestamp intervals. The deviation feature values ​​extracted from the spatial displacement parameters and onboard high-frequency inertial sensing sequences are fused and set as kinematic feature parameters. When these deviations exceed a dynamic tolerance threshold, they are associated with digital tags representing temporal mutations, accurately capturing abnormal instantaneous coordinate shifts in the micro-temporal dimension. The digital tags representing temporal mutations are then inversely mapped to the corresponding time nodes in the macro-trajectory node sequence. Cross-dimensional security conflict verification is performed by combining the compliance status output by the large model with the frequency of tag occurrence. This effectively identifies and blocks high-frequency injected forged status values, solving the technical challenge of successfully bypassing security reviews with high-frequency commercial forged data in the vehicle network, and preventing false mileage data from participating in commercial logistics transportation subsidy calculations. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the method for enhancing vehicle network data security based on a large model according to the present invention. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Firstly, please refer to Figure 1 The present invention provides a method for enhancing vehicle network data security based on a large model, comprising: Step 1: Collect the raw time-series service data sequence continuously uploaded by the vehicle-mounted node; Step 2: Perform data splitting on the original time-series service data sequence of the vehicle to separate the macroscopic trajectory node sequence, as well as the spatial displacement parameters and vehicle high-frequency inertial sensing sequence within the adjacent timestamp intervals corresponding to the original time-series service data sequence of the vehicle. Step 3: Extract deviation feature values ​​by fusing spatial displacement parameters and vehicle-mounted high-frequency inertial sensing sequences, set the deviation feature values ​​as kinematic feature parameters, compare the kinematic feature parameters with the dynamic tolerance threshold, and associate time-series abrupt change digital tags with adjacent timestamp intervals when the kinematic feature parameters are greater than the dynamic tolerance threshold. Step 4: Import the macro trajectory node sequence into the global semantic feature matching network of the large model, and output the comparison results between the macro trajectory node sequence and the preset logistics transportation business specifications. Step 5: Establish a baseline time sliding alignment window, reverse map the time-series mutation digital tags to the corresponding time nodes included in the macro trajectory node sequence, count the frequency of occurrence of the time-series mutation digital tags in the baseline time sliding alignment window, and block the original time-series business data sequence from entering the business value assessment stage if the comparison result is compliant and the frequency of occurrence is greater than the preset business fault tolerance threshold.

[0020] After initiating the data reception service of the vehicle-mounted nodes, the system begins collecting the raw time-series service data sequences continuously uploaded by the vehicle-mounted nodes. The vehicle-mounted nodes are specifically intelligent connected communication terminals installed on commercial transport vehicles. The raw time-series service data sequences include global satellite positioning coordinate streams, inertial sensor reading streams, and vehicle control bus status streams.

[0021] The data splitting module performs data splitting on the original time-series business data sequence from the vehicle, separating the macroscopic trajectory node sequence. Simultaneously, the data splitting module separates the spatial displacement parameters and the vehicle-mounted high-frequency inertial sensor sequence within adjacent timestamp intervals.

[0022] The system extracts the timing synchronization message header from the original timing service data sequence on the vehicle side, and then compares the absolute timestamp recorded in the timing synchronization message header with the relative time offset parameter of the vehicle bus. The absolute timestamp originates from the timing signal of the satellite navigation system, while the relative time offset parameter of the vehicle bus originates from the internal crystal oscillator timer of the vehicle controller. When the absolute timestamp and the relative time offset parameter of the vehicle bus are perfectly aligned, it is determined that the current data packet has not been subjected to a time replay attack, thus triggering a data splitting command.

[0023] In a specific embodiment where the absolute timestamp is determined to be perfectly aligned with the relative time offset parameter of the vehicle bus, the system computing core pre-obtains the initial satellite timing timestamp when the commercial transport vehicle's engine ignites. The system computing core adds the initial satellite timing timestamp to the relative time offset parameter of the vehicle bus to generate a theoretically derived timestamp. Subsequently, the system computing core calculates the absolute time difference between the absolute timestamp and the theoretically derived timestamp. When the absolute time difference is less than or equal to a preset network transmission delay error range (e.g., a set error range of 50 milliseconds), the system determines that the absolute timestamp and the relative time offset parameter of the vehicle bus are perfectly aligned, thereby generating an authentication signal confirming that the data has not been tampered with and triggering a data splitting instruction.

[0024] After triggering the data splitting command, the original time-series service data sequence from the vehicle end is separated according to the spatial grid partitioning standard. The spatial grid partitioning standard adopts a hexagonal geospatial discretization algorithm to convert continuous latitude and longitude coordinates into a fixed-resolution grid code, generating a macroscopic trajectory node sequence. The coordinate differences between adjacent sampling periods of the original time-series service data sequence from the vehicle end are extracted, and spatial displacement parameters are calculated using the coordinate differences.

[0025] In a specific embodiment of generating spatial displacement parameters through coordinate difference calculation, the original time-series service data sequence at the vehicle end includes continuous longitude and latitude values. The system data processing center extracts the longitude and latitude coordinates of the previous and subsequent times within adjacent sampling periods. The system data processing center uses a spherical distance calculation algorithm to calculate the difference in spherical distance values ​​between the previous and subsequent longitude and latitude coordinates on the meridional and zonal projection axes, respectively. The system data processing center performs vector merging operations on the meridional and zonal projection axis distance differences to generate spatial displacement parameters with absolute direction and absolute distance scale attributes.

[0026] The data splitting module synchronously extracts the three-dimensional spatial acceleration signal and yaw rate signal within adjacent sampling periods, merges the three-dimensional spatial acceleration signal and yaw rate signal to form an on-board high-frequency inertial sensing sequence, and finally inputs it into the kinematic verification stage.

[0027] After entering the kinematic verification phase, the system performs the operation of fusing spatial displacement parameters and vehicle-mounted high-frequency inertial sensing sequences to extract deviation feature values. Specifically, the fusing of spatial displacement parameters and vehicle-mounted high-frequency inertial sensing sequences includes: establishing a deviation covariance matrix between spatial displacement parameters and dynamically derived displacement vectors based on normalized data, and performing singular value decomposition on the deviation covariance matrix to extract principal eigenvalues; the system extracts the principal eigenvalues ​​as deviation feature values, sets the deviation feature values ​​as kinematic feature parameters, and compares the kinematic feature parameters with the dynamic tolerance threshold.

[0028] After merging to form the vehicle-mounted high-frequency inertial sensing sequence, the chassis topology location tags included in the vehicle-mounted high-frequency inertial sensing sequence are extracted. The chassis topology location tags are then matched with a preset vehicle firmware component assembly mapping table.

[0029] If the chassis topology location label matches the preset vehicle firmware component assembly mapping table without error, confirming that the sensor has not been illegally replaced, a calculus calculation trigger signal is then output. The preset vehicle firmware component assembly mapping table is dynamically and freshly maintained through the online upgrade link of the vehicle-side security gateway. When a commercial transport vehicle undergoes a legitimate component replacement or firmware online upgrade, the cloud management platform sends a firmware update log message with a digital signature to the vehicle-side security gateway. The feature comparison module receives the firmware update log message and performs key verification on it. After successful verification, the latest sensor unique device code and installation space three-dimensional coordinates are overwritten into the preset vehicle firmware component assembly mapping table, thereby ensuring that the verification benchmark is updated dynamically in real time.

[0030] In a specific embodiment, a pre-defined vehicle firmware component assembly mapping table pre-registers the unique device factory code and corresponding three-dimensional coordinates of each sensor component when the vehicle leaves the factory. The chassis topology location label records the device factory code of the sensor currently transmitting data. The system compares the device factory code in the chassis topology location label with the information registered in the pre-defined vehicle firmware component assembly mapping table. If the coding information is completely consistent, it confirms that the sensor component has not been illegally replaced, and then outputs a calculus calculation trigger signal. By verifying the device factory code, it prevents attackers from connecting external fake sensor devices to send forged data.

[0031] Upon receiving the calculus operation trigger signal, the computation core performs dual time-integral calculus operations on the three-dimensional spatial acceleration signal and the yaw rate signal. In a specific embodiment of performing dual time-integral calculus operations on the three-dimensional spatial acceleration signal and the yaw rate signal, the computation core first performs a time integration on the yaw rate signal to obtain the vehicle heading angle change value. The computation core constructs a coordinate system rotation transformation matrix based on the vehicle heading angle change value. Subsequently, the computation core uses the coordinate system rotation transformation matrix to convert the three-dimensional spatial acceleration signal from the vehicle relative coordinate system to a spatial acceleration vector in the world absolute coordinate system. After completing the coordinate system transformation, the computation core performs a first time integration on the spatial acceleration vector to obtain the three-dimensional velocity vector, and then performs a second time integration on the three-dimensional velocity vector to generate the dynamic displacement vector. By introducing the yaw rate signal for coordinate system compensation, the spatial pointing accuracy of the dynamic displacement vector can be effectively improved.

[0032] The acceleration is first integrated over time to obtain a three-dimensional velocity vector, and then the three-dimensional velocity vector is integrated over time a second time to generate a dynamic displacement vector. To perform the step of fusing spatial displacement parameters and extracting deviation feature values ​​from the vehicle-mounted high-frequency inertial sensing sequence, the system data processing center prioritizes performing timestamp interpolation alignment and spatial dimension normalization of the multi-source heterogeneous data before establishing the deviation covariance matrix between the spatial displacement parameters and the dynamic displacement vector. The subsequently established deviation covariance matrix will be used to quantify the degree of dispersion of the two different sources of displacement data.

[0033] Because of the frequency gap between the low-frequency sampling rate of the global satellite positioning coordinate stream and the high-frequency sampling rate of the inertial sensor reading stream in the original time-series service data sequence of the vehicle, the system data processing center uses a spline interpolation algorithm to upsample the low-frequency spatial displacement parameters to a time resolution equal to that of the vehicle-mounted high-frequency inertial sensing sequence, performing a forced alignment operation. Simultaneously, addressing the absolute dimensional difference between the spatial displacement parameters and the dynamically derived displacement vector, the system data processing center uses a Z-score normalization algorithm to map the interpolated spatial displacement parameters and the dynamically derived displacement vector to a dimensionless standard normal distribution interval with a mean of 0 and a standard deviation of 1, thereby eliminating mathematical weight interference caused by the difference in absolute numerical magnitudes between different sensor components in the extraction of the principal eigenvalues ​​of the covariance matrix.

[0034] The system data processing center establishes a deviation covariance matrix between spatial displacement parameters and dynamically derived displacement vectors based on normalized data. This established deviation covariance matrix is ​​used to quantify the degree of dispersion of displacement data from two different sources.

[0035] In a specific embodiment for establishing the deviation covariance matrix, the system computation unit calculates the independent numerical differences between the spatial displacement parameters and the dynamically derived displacement vector on three independent spatial coordinate axes. The computation unit uses these independent numerical differences to calculate the variance and covariance, thereby constructing a 3x3 deviation covariance matrix. A singular value decomposition algorithm is then applied to the deviation covariance matrix to extract the principal eigenvalues ​​representing the direction of maximum error divergence as kinematic characteristic parameters. The larger the principal eigenvalue, the more severe the spatial geometric discrepancy between the satellite positioning trajectory and the internally calculated inertial trajectory, indicating that the externally input spatial displacement parameters are highly likely to have been artificially altered.

[0036] Considering the significant difference in mechanical inertia between fully loaded and unloaded commercial trucks, a fleet dynamic load status descriptor is obtained that is perfectly aligned with adjacent timestamp intervals. This descriptor records vehicle chassis suspension airbag pressure sensor readings, which directly reflect the actual total weight of the vehicles.

[0037] The vehicle chassis suspension airbag pressure sensor readings are used as index keys to input into the dynamic tolerance database, and the corresponding nonlinear suspension system dynamic error tolerance values ​​are retrieved as dynamic tolerance thresholds. Specifically, the dynamic tolerance threshold is determined by mapping the vehicle chassis suspension airbag pressure sensor readings to the nonlinear suspension system dynamic error tolerance values ​​extracted in a two-column correlation matrix, thereby achieving dynamic configuration of the threshold.

[0038] The mechanism for establishing the dynamic tolerance database is as follows: The R&D system pre-collects historical vibration test samples of the test vehicle under no-load, half-load, and full-load conditions. A quadratic continuous function mapping curve between the suspension airbag pressure and the standard deviation of the sensor zero-point drift is established through a nonlinear least squares fitting algorithm. The continuous function mapping curve is discretized into a two-column correlation matrix with multiple pressure intervals and tolerance boundary values. The two-column correlation matrix is ​​persistently stored in the dynamic tolerance database so that the feature comparison module can accurately retrieve the dynamic filtering boundary of the kinematic feature parameters based on the real-time total vehicle weight status.

[0039] In a specific embodiment, when a commercial transport vehicle is unloaded, the airbag pressure sensor readings of the vehicle chassis suspension are in the low-pressure range. During vehicle operation, the vertical vibration amplitude of the cargo compartment is large, resulting in significant natural fluctuations in the calculated principal characteristic values. Therefore, the system retrieves a larger nonlinear suspension system dynamic error tolerance value based on the low-pressure range to prevent normal mechanical vibrations from being incorrectly identified. When a commercial transport vehicle is fully loaded, the airbag pressure sensor readings of the vehicle chassis suspension are in the high-pressure range. The vehicle's driving state is stable and has high inertia. The system retrieves a smaller nonlinear suspension system dynamic error tolerance value based on the high-pressure range. By accurately matching the dynamic tolerance threshold with the airbag pressure, legitimate mechanical vibrations and illegal coordinate transitions can be effectively distinguished.

[0040] When the principal eigenvalue is greater than the dynamic error tolerance value of the nonlinear suspension system, it is determined that the human-caused instantaneous coordinates of satellite positioning lack corresponding mechanical energy support inside the vehicle body, and the time-series abrupt change digital tags are associated with adjacent timestamp intervals.

[0041] After completing the micro-level labeling, the macro-track node sequence is imported into the global semantic feature matching network of the large model, outputting the comparison results between the macro-track node sequence and the preset logistics transportation business specifications. The macro-track node sequence is mapped to discrete business operation status identifiers, for example, mapping continuous spatial grid codes to text description labels such as leaving the logistics park, entering the highway, and passing through service areas. Using natural language embedding algorithms, the discrete business operation status identifiers are converted into high-dimensional semantic vectors, and the high-dimensional semantic vectors are input into the global semantic feature matching network of the large model for deep feature extraction. The global semantic feature matching network of the large model adopts a bidirectional transformer topology architecture with a multi-layer self-attention mechanism. The bidirectional transformer topology architecture includes twelve core coding units cascaded in series. Each core coding unit has sixteen parallel self-attention probes and a feedforward fully connected network layer with two thousand and forty-eight neurons. The input tensor dimension of the high-dimensional semantic vector is set to seven hundred and sixty-eight dimensions. The global semantic feature matching network of the large model extracts the macro-context business logic chain of the entire transportation process by calculating the attention association weight matrix between different sequence positions in the input tensor. It outputs a one-dimensional normalized scalar value representing the overall route compliance probability from the end of the core coding unit, thus enabling the global semantic feature matching network of the large model to have technical determinism that can be completely reproducibly constructed.

[0042] In a specific embodiment of converting discretized business operation status identifiers into high-dimensional semantic vectors, the system invokes a pre-trained natural language word vector encoding network. The word vector encoding network receives a sequence of discretized business operation status identifiers arranged chronologically, such as the text sequence "exiting the logistics park," "entering the highway," and "passing through a service area" arranged sequentially. The word vector encoding network extracts the contextual temporal dependency features of the text sequence and outputs a floating-point high-dimensional semantic vector with 768 dimensions. This high-dimensional semantic vector, in the form of a digital matrix, accurately represents the logical continuity of the entire transportation journey, facilitating in-depth mathematical comparisons by large models.

[0043] The global semantic feature matching network of the large model is constructed based on a multi-layer self-attention mechanism framework. For model pre-training, the network model training server pre-collects historical fleet operation status label data to construct a training corpus, and uses a one-hot encoding algorithm to transform the discrete status labels contained in the training corpus into a basic word vector matrix. The network model training server then superimposes positional encoding vectors reflecting temporal order onto the basic word vector matrix, thereby generating an input tensor with a fused temporal dimension using action sequence logic.

[0044] Following the completion of data preprocessing, the network model training server feeds the input tensor into the core encoder assembly for processing. This core encoder assembly is specifically formed by alternating cascades of multi-head self-attention layers and feedforward neural network layers. The multi-head self-attention layers within the core encoder assembly capture the contextual dependencies of different state labels within the macro-level business process by calculating attention weight scores between state labels at different positions in the input tensor. Subsequently, the network model training server projects the extracted data features with contextual dependencies into multiple independent and parallel sub-representation spaces to perform feature concatenation. After feature concatenation, the feedforward neural network layer performs nonlinear spatial mapping on the concatenated data features. After outputting the mapped features, the network model training server uses the backpropagation algorithm to continuously minimize the cross-entropy loss function value between the mapped features and the business logic labels. By continuously reducing the cross-entropy loss function value, iterative updates of the model network weights are achieved. Finally, the updated model network weights are solidified to obtain the global semantic feature matching network of the trained large model.

[0045] The core encoder assembly configures residual connection paths and layer normalization units between the multi-head self-attention layer and the feedforward neural network layer. The attention feature matrix output from the multi-head self-attention layer is added to the initial input tensor to generate a residual feature matrix. The layer normalization unit then performs normalization mathematical calculations on the residual feature matrix along the last feature dimension, calculating the mean and standard deviation. The normalized matrix is ​​then used as the input data for the feedforward neural network layer.

[0046] In the iterative update step of the network weights, the network model training server loads an adaptive moment estimation optimizer as the core of the backpropagation calculation. The network model training server presets an initial learning rate of 0.0001 and simultaneously introduces a cosine annealing learning rate decay function. The network model training server sets the batch processing sample data size to a scalar value of 512 and the total number of training iterations to 100.

[0047] The cross-entropy loss function takes the predicted probability distribution and the one-hot encoded distribution of the business logic labels as input data. The network model training server calculates the log-likelihood product of the corresponding dimensions of the predicted probability distribution matrix and the one-hot encoded distribution matrix, performs a summation and negation operation, and outputs a scalar loss value. The network model training server calculates the gradient partial derivatives of the network weight matrix based on the scalar loss value.

[0048] During the real-time inference phase, the global semantic feature matching network of the large model receives discretized business operation status identifiers to initiate the data flow path. The embedding layer within the global semantic feature matching network maps the received discretized business operation status identifiers into high-dimensional dense feature representations. Based on these high-dimensional dense feature representations, the self-attention layer within the global semantic feature matching network calculates the global correlation matrix between the high-dimensional dense feature representations. Subsequently, the non-linear activation function within the global semantic feature matching network performs feature extraction on the global correlation matrix. After completing the feature extraction step, the global semantic feature matching network outputs a high-dimensional semantic vector from the end of the network topology. This output high-dimensional semantic vector is specifically used to represent the consistency state of the overall route logic.

[0049] Retrieve the compliant business specification vector template included in the preset logistics and transportation business specifications. The preset logistics and transportation business specifications are composed of the compliant business specification vector template and the spatial cosine similarity calculation logic. The compliant business specification vector template represents the regular execution process of standard transportation business.

[0050] In a specific embodiment for retrieving the compliance business specification vector template, a large amount of historical freight vehicle travel data without abnormal alarm states is pre-collected for the pre-defined logistics transportation business specifications. The system model training unit performs cluster analysis on the historical freight vehicle travel data to extract high-frequency path node sequences and standard dwell time spans representing the regular point-to-point transportation process. The system model training unit inputs the high-frequency path node sequences and standard dwell time spans into a natural language word vector encoding network, outputting a fixed-dimensional standard reference vector sequence. The system model training unit persistently stores the standard reference vector sequence as a compliance business specification vector template in the business logic database for real-time business verification retrieval.

[0051] Calculate the spatial cosine similarity between the high-dimensional semantic vector and the compliance business specification vector template. If the spatial cosine similarity meets the preset specification threshold, it indicates that the data flow is logically coherent at the macro-business level, generating a comparison result that is in a compliant state.

[0052] In a specific embodiment, a preset standard threshold of 85% is set. When the calculated spatial cosine similarity is greater than or equal to 85%, the system determines that the macroscopic driving route fully conforms to the normal logistics transportation business process, such as passing through the loading area, highway entrance and unloading area in sequence.

[0053] The numerical labels of temporal abrupt changes are inversely mapped to the corresponding time nodes included in the macroscopic trajectory node sequence, and the frequency of occurrence of the numerical labels of temporal abrupt changes within the reference time sliding alignment window is statistically analyzed. A sliding observation interval of fixed time length is set as the reference time sliding alignment window.

[0054] In a specific embodiment, the total length of the reference time sliding alignment window is set to 60 seconds. The reverse mapping is implemented as follows: extract the absolute timestamp recorded inside the time-series mutation digital tag, accurately find the business trajectory point in the macro trajectory node sequence that matches the absolute timestamp time, and accurately associate the micro-generated time-series mutation digital tag with the corresponding macro business trajectory point. In order to eliminate the time span gap between high-frequency micro signals and low-frequency macro states, the security blocking module retrieves the entry absolute time and departure absolute time corresponding to each grid code in the macro trajectory node sequence. If the absolute timestamp recorded inside the time-series mutation digital tag is within the closed interval formed by the entry absolute time and departure absolute time of the corresponding grid code, the security blocking module determines that the time-series mutation digital tag and the grid code have a time alignment correlation. The security blocking module binds the spatial coordinates of the time-series mutation digital tag to the grid code, thereby establishing a mapping correlation between clock synchronization and spatial location binding.

[0055] The total number of occurrences of the time-series mutation number labels within the reference time sliding alignment window is accumulated, and the ratio of the total number of occurrences of the time-series mutation number labels to the total time length of the reference time sliding alignment window is calculated to generate the occurrence frequency.

[0056] If the comparison results are compliant and the frequency of occurrence exceeds the preset business fault tolerance threshold, a highly concealed commercial data forgery anomaly is determined to have occurred. A compliant comparison result indicates that the macro-level forgery route conforms to the regulatory logic, and a frequency exceeding the preset business fault tolerance threshold indicates that the timeline contains dense, high-frequency coordinate tampering.

[0057] In a specific embodiment, a preset business fault tolerance threshold is set to 10%. When the number of trajectory points with mutation markers in the reference time sliding alignment window exceeds 10% of the total number of nodes in the window, the system determines that the commercial transport vehicle is suffering from a high-frequency commercial forgery attack that injects fake coordinates intensively in a short period of time.

[0058] Upon confirming the abnormal behavior, the uplink communication link of the terminal is disconnected by shutting down the vehicle data receiving port, preventing the original time-series business data sequence from entering the commercial value assessment stage and preventing false data from participating in logistics subsidy calculation. The memory management module simultaneously clears any unprocessed original time-series business data sequences from memory, ensuring the data reliability of the entire vehicle-to-everything (V2X) commercial operation accounting system.

[0059] In a specific embodiment of synchronously clearing unprocessed vehicle-side original time-series service data sequences, the system memory management unit issues a forced interrupt reset command to the circular buffer queue responsible for temporarily storing data. Upon receiving the forced interrupt reset command, the circular buffer queue directly forces its read / write pointers to zero and overwrites and releases the storage address space occupied by all vehicle-side original time-series service data sequences that have not yet entered the security verification stage. By overwriting the storage address space, the storage records of high-frequency forged data in the receive buffer are eliminated, preventing abnormal data transmission to the commercial value assessment stage.

[0060] The system pre-writes the specific numerical range of the dynamic tolerance threshold, setting it to between 0.5 meters and 2.5 meters. The system acquires the dynamic error tolerance value of the nonlinear suspension system under full load conditions for commercial transport vehicles, using this value as the dynamic tolerance threshold, which is set to 0.5 meters. Subsequently, the system acquires the dynamic error tolerance value of the nonlinear suspension system under empty conditions for commercial transport vehicles, using this value as the dynamic tolerance threshold, which is set to 2.5 meters.

[0061] In the system configuration file, a preset service fault tolerance threshold is set, and the numerical range of the preset service fault tolerance threshold is defined as 10% to 15%. During system operation, the preset service fault tolerance threshold recorded in the configuration file is read, and then the preset service fault tolerance threshold is specifically set to 10%. Specifically, the preset service fault tolerance threshold is the ratio limit of the number of trajectory points with abrupt change markers in the reference time sliding alignment window to the total number of nodes in the window.

[0062] Based on the network environment, a preset network transmission delay error range is established, with the value set between 30 milliseconds and 80 milliseconds. When verifying the absolute timestamp, the system reads the preset network transmission delay error range and ultimately sets its specific value to 50 milliseconds.

[0063] To standardize the calculation process, a preset range of standardization thresholds is defined, ranging from 80% to 95%. When calculating spatial cosine similarity, the system retrieves the preset standardization threshold and sets its specific value to 85%.

[0064] Formula for calculating the absolute difference in time: In the formula: symbols Represents the absolute difference over time. Symbol Represents an absolute timestamp. (Symbol) Represents the initial satellite timing timestamp at the time of engine ignition. (Symbol) This represents the relative time offset parameter of the vehicle bus. (Symbol) Represents the absolute value symbol. Substitute the numerical value into the absolute timestamp. The initial satellite timing timestamp is 10050 milliseconds. The relative time offset parameter of the vehicle bus is 10000 milliseconds. The absolute time difference is calculated as 30 milliseconds. The absolute time difference of 20 milliseconds is compared with the preset network transmission delay error range of 50 milliseconds. Since the absolute time difference of 20 milliseconds is less than the preset network transmission delay error range of 50 milliseconds, the conclusion is that the absolute timestamp is completely aligned with the relative time offset parameter of the vehicle bus.

[0065] Formula for calculating spatial displacement parameters: In the formula: symbols Represents spatial displacement parameters. Symbol Represents the Earth's average radius. (Symbol) Represents the longitude coordinates at the previous moment. (Symbol) Represents the longitude coordinates at the next moment. Symbol Represents the latitude coordinates at the previous moment. (Symbol) Represents the latitude coordinates at the next moment. Symbol This represents the cosine function. Substituting the values, the Earth's average radius R is taken as 6371 km, the longitude coordinate λ1 at the previous moment is taken as 110 degrees, the longitude coordinate λ2 at the next moment is taken as 110.01 degrees, the latitude coordinate φ1 at the previous moment is taken as 30 degrees, and the latitude coordinate φ2 at the next moment is taken as 30.01 degrees. The longitude difference and latitude difference are multiplied by pi and divided by 180 to convert them into radians. Substituting these values ​​into the formula, the geographical span of the meridional axis is calculated to be 0.96 km, and the geographical span of the zonal axis is calculated to be 1.11 km. Performing the vector merging operation of the square root of the sum of squares, the spatial displacement parameter S is calculated to be 1.5 km.

[0066] The formula for transforming spatial acceleration vectors and the dual calculus formula for dynamic derivation of displacement vectors: In the formula: symbols Represents the spatial acceleration vector. Symbol This represents the coordinate system rotation transformation matrix constructed based on the yaw rate signal. (Symbol) Represents a three-dimensional spatial acceleration signal. Symbol Represents the displacement vector derived from dynamics. Symbol Represents the time integral variable. Symbol This represents the double-time calculus operator. Substitute the numerical three-dimensional spatial acceleration signal. The horizontal rotation transformation matrix is ​​2 meters per second squared. The angle is 45 degrees, and the spatial acceleration vector is calculated. With an axial component of 1.41 meters per second squared in the absolute coordinate system, the spatial acceleration vector... Integrating the variable over a time interval of 3 seconds Performing dual-time calculus operations within the interval yields the dynamic displacement vector. It is 6.34 meters.

[0067] Formulas for calculating the deviation covariance matrix and extracting principal eigenvalues ​​from singular value decomposition: In the formula: symbols This represents the deviation covariance matrix. (Symbol) This represents the total number of data sampling points within adjacent timestamp intervals. (Symbol) Represents elements of the spatial displacement parameter matrix. Symbols These represent elements of the displacement vector matrix used in dynamic derivation. (Symbols) This represents the matrix transpose operation. (Symbol) This represents the summation operator. (Symbol) Represents a left singular vector matrix. (Symbol) Represents a singular value diagonal matrix. Symbol Represents a right singular vector matrix. (Symbol) Represents the principal eigenvalue. (Symbol) This represents the function for finding the maximum value. (Symbol) This represents the function for extracting diagonal elements. Substitute elements from the spatial displacement parameter matrix. Elements of the displacement vector matrix in dynamic derivation From the coordinate difference sequence, construct a third-order deviation covariance matrix. For the bias covariance matrix Performing singular value decomposition yields a singular value diagonal matrix. The elements within the range are 0.8, 0.2, and 0.1. The maximum value on the diagonal is extracted to obtain the principal eigenvalue. The value is 0.8, which is the principal eigenvalue. The kinematic feature parameter is set as 0.8. The kinematic feature parameter 0.8 is compared with the dynamic tolerance threshold 0.5. The kinematic feature parameter 0.8 is greater than the dynamic tolerance threshold 0.5. Therefore, the conclusion is that the kinematic feature parameter exceeds the dynamic tolerance threshold.

[0068] Formula for calculating spatial cosine similarity: In the formula: symbols Represents spatial cosine similarity. Symbol Represents the vector dimension index. Symbol Dimensional elements representing high-dimensional semantic vectors. (Symbol) Dimensional elements representing the compliance business specification vector template. Symbols This represents the summation operator. (Symbol) Represents the square root operation. Substitute the elements of the high-dimensional semantic vector. Compliance Business Specification Vector Template Dimension Elements The spatial cosine similarity was calculated. The spatial cosine similarity is 92%. The spatial cosine similarity of 92% is compared with the preset standard threshold of 85%. Since the spatial cosine similarity of 92% is greater than the preset standard threshold of 85%, the conclusion is that the spatial cosine similarity meets the preset standard threshold.

[0069] Formula for calculating the frequency of occurrence of time-series abrupt change digital tags: In the formula: symbols Represents frequency of occurrence. Symbol This represents the total number of occurrences of the timing abrupt change numerical label within the reference time sliding alignment window. (Symbol) This represents the total length of the sliding alignment window at the reference time. Substitute this with the total number of occurrences of the timing abrupt change labels. The total time length of the sliding alignment window is fifteen times, substituted into the reference time. The frequency of occurrence is calculated over a period of 60 seconds. The frequency of occurrence is set at 25%. The frequency of occurrence at 25% is compared with the preset business fault tolerance threshold of 10%. Since the frequency of occurrence at 25% is greater than the preset business fault tolerance threshold of 10%, the conclusion is that the frequency of occurrence is greater than the preset business fault tolerance threshold.

[0070] Numerical integral formula for vehicle heading angle change: In the formula: symbols Represents the numerical value of the vehicle's heading angle change. Symbol This represents the yaw rate signal. (Symbol) Represents the time integral variable. Symbol This represents the time integration operator.

[0071] Formula for constructing the coordinate system rotation transformation matrix: In the formula: symbols Represents the coordinate system rotation transformation matrix. Symbol Represents the numerical value of the vehicle's heading angle change. Symbol Represents the cosine function. Symbol It represents the sine function.

[0072] Z-score normalization algorithm formula: In the formula: symbols Represents the normalized data. Symbol Represents spatial displacement parameters or displacement vectors derived from dynamics. Symbols This represents the sample mean of spatial displacement parameters or dynamically derived displacement vectors. (Symbol) The standard deviation of the sample represents the spatial displacement parameter or the displacement vector derived from dynamics.

[0073] Formula for calculating the mapping curve of a quadratic continuous function: In the formula: symbols This represents the numerical value of the dynamic error tolerance of the nonlinear suspension system. (Symbol) This represents the pressure sensor reading of the vehicle's chassis suspension airbags. (Symbol) Represents the fitting coefficient for the quadratic term. Symbol Represents the coefficient of fit for the linear term. Symbol The constant term fit coefficients.

[0074] The formula for calculating the attention association weight matrix is ​​as follows: In the formula: symbols This represents the attention-related weight matrix. (Symbol) Represents the query feature tensor matrix. (Symbol) Represents the key feature tensor matrix. Symbols Representative value characteristic tensor matrix. Symbols The transpose of the key feature tensor matrix. (Symbol) The numerical value representing the scaled feature dimension of the key feature tensor matrix. (Symbol) This represents the normalized exponential function.

[0075] The formula for calculating the cross-entropy loss function is as follows: In the formula: symbols Represents the scalar loss value. Symbol Represents the total dimension value for the category. (Symbol) Represents the category index number. (Symbol) This represents the numerical value of the corresponding dimension of the one-hot encoded distribution matrix. (Symbol) This represents the predicted probability value for the corresponding dimension of the predicted probability distribution matrix. (Symbol) This represents the summation operator. (Symbol) This represents the natural logarithm function.

[0076] Secondly, the vehicle network data security enhancement system based on a large model provided by the present invention is applied to the vehicle network data security enhancement method based on a large model as described above, including: The data acquisition module is used to collect the original time-series business data sequences continuously uploaded by the vehicle-mounted nodes. The data splitting module is used to split the original time-series business data sequence of the vehicle end, separating the macro trajectory node sequence, as well as the spatial displacement parameters and vehicle high-frequency inertial sensing sequence within the adjacent timestamp intervals corresponding to the original time-series business data sequence of the vehicle end. The feature comparison module is used to extract deviation feature values ​​by fusing spatial displacement parameters and vehicle-mounted high-frequency inertial sensing sequences. The deviation feature values ​​are set as kinematic feature parameters. The kinematic feature parameters are compared with the dynamic tolerance threshold. When the kinematic feature parameters are greater than the dynamic tolerance threshold, the module associates time-series change digital tags with adjacent timestamp intervals. The semantic matching module is used to import the macro trajectory node sequence into the global semantic feature matching network of the large model and output the comparison results between the macro trajectory node sequence and the preset logistics and transportation business specifications. The safety blocking module is used to establish a reference time sliding alignment window, reverse map the time-series mutation digital tags to the corresponding time nodes included in the macro trajectory node sequence, count the frequency of occurrence of the time-series mutation digital tags in the reference time sliding alignment window, and block the original time-series business data sequence on the vehicle side from entering the business value assessment stage if the comparison result is compliant and the frequency of occurrence is greater than the preset business fault tolerance threshold.

[0077] In this embodiment of the invention, in a commercial inter-provincial cold chain logistics transportation application scenario, an on-board gateway is deployed on a commercial cold chain transport tractor to continuously collect the original time-series business data sequence from the vehicle. The original time-series business data sequence from the vehicle includes real-time dynamic differential positioning coordinates, controller area network bus pedal opening, and voltage values ​​of the six-degree-of-freedom inertial measurement unit.

[0078] The data processing module receives the original time-series business data sequence from the vehicle and performs data splitting to extract the data portion used for macro-path evaluation. The macro-path evaluation module, based on a hexagonal geospatial discretization algorithm, processes the extracted data portion and generates a macro-trajectory node sequence. The state transition module converts the macro-trajectory node sequence into discretized business operation status identifiers, which have specific forms such as leaving a fresh food cold storage, cruising on an inter-provincial highway, and entering a loading and unloading distribution center. The global semantic feature matching network is pre-constructed using historical trajectory label data of millions of compliant freight vehicles, trained through masked language modeling. The semantic analysis module inputs the discretized business operation status identifiers into the global semantic feature matching network; the global semantic feature matching network maps the input discretized business operation status identifiers into 768-dimensional high-dimensional semantic vectors and calculates the contextual association weights between long-term routing node sequences through a multi-layer self-attention mechanism configured within the global semantic feature matching network architecture. Finally, the global semantic feature matching network outputs a scoring result from the terminal network layer of the network structure. The scoring result represents the logical coherence of the global trajectory and serves as the comparison result.

[0079] When the simulator injects false differential positioning coordinates that deviate 30 meters from the main road into the vehicle gateway every few seconds during long-distance trunk line driving, the comparison results output by the global semantic feature matching network are still compliant because the false differential positioning coordinates do not disrupt the overall macro logic of the starting point, the ending point, and the provinces traversed.

[0080] The microscopic verification module is responsible for running a microscopic temporal mutation verification mechanism, extracting three-dimensional spatial acceleration signals including longitudinal, lateral, and vertical acceleration, and merging these signals with the yaw rate signal into an onboard high-frequency inertial sensing sequence. Simultaneously, the load monitoring module acquires the vehicle chassis suspension airbag pressure sensor readings from the 0-10 bar range on the corresponding trailer axle in the fleet dynamic load state descriptor. The threshold calculation module, based on the vehicle chassis suspension airbag pressure sensor readings under full load conditions in the high-pressure range, searches the dynamic tolerance database, retrieves the nonlinear suspension system dynamic error tolerance value, and sets it as the dynamic tolerance threshold. The nonlinear suspension system dynamic error tolerance value corresponds to a relatively small tolerance range.

[0081] Because the spatial displacement parameters generated by the injected spurious differential positioning coordinates lack the chassis's mechanical inertial calculus response, the kinematic feature parameter calculation module extracts and fuses the kinematic feature parameters. These kinematic feature parameters represent the degree of geometric divergence between the spatial trajectory and the trajectory obtained through mechanical calculation in the data processing module. When the value of the kinematic feature parameter exceeds the dynamic tolerance threshold, the tag association module associates temporally abrupt digital tags with adjacent timestamp intervals.

[0082] The frequency statistics module counts the frequency of occurrence of time-series abrupt change digital tags within a set 60-second reference time sliding alignment window. When the frequency of occurrence of these tags exceeds a preset 10% service fault tolerance threshold, the security blocking module generates a communication blocking control signal and sends it to the commercial cold chain transport tractor. The control unit executes the communication blocking control signal, disconnecting the uplink transmission link of the network transmission control protocol between the vehicle gateway and the cloud monitoring platform, and simultaneously clearing the local communication cache queue in the vehicle gateway's memory area. The disconnection of the communication link cuts off the terminal data reporting channel, enabling direct intervention and control over the communication status of the commercial cold chain transport tractor. This prevents the original time-series service data sequence containing high-frequency abnormal transition coordinates from entering the commercial value assessment stage, thus completing the identification and blocking of microscopic high-frequency mileage spoofing vulnerabilities.

Claims

1. A method for enhancing vehicle-to-everything (V2X) data security based on a large model, characterized in that: include: Collect the vehicle-side raw time-series business data sequence continuously uploaded by the vehicle-mounted node; Data splitting is performed on the original time-series service data sequence of the vehicle to separate the macroscopic trajectory node sequence, as well as the spatial displacement parameters and vehicle high-frequency inertial sensing sequence within the adjacent timestamp intervals corresponding to the original time-series service data sequence of the vehicle. By integrating spatial displacement parameters and vehicle-mounted high-frequency inertial sensing sequences to extract deviation feature values, these deviation feature values ​​are set as kinematic feature parameters. The kinematic feature parameters are compared with dynamic tolerance thresholds. When the kinematic feature parameters are greater than the dynamic tolerance threshold, temporal abrupt change digital tags are associated with adjacent timestamp intervals. The macroscopic trajectory node sequence is imported into the global semantic feature matching network of the large model, and the comparison results between the macroscopic trajectory node sequence and the preset logistics transportation business specifications are output. Establish a baseline time sliding alignment window, reverse map the time-series mutation digital tags to the corresponding time nodes included in the macro trajectory node sequence, and count the frequency of occurrence of the time-series mutation digital tags within the baseline time sliding alignment window. Under the condition that the comparison result is compliant and the frequency of occurrence is greater than the preset business fault tolerance threshold, block the original time-series business data sequence of the vehicle end from entering the business value assessment stage.

2. The method for enhancing vehicle network data security based on a large model as described in claim 1, characterized in that, Data splitting is performed on the original time-series service data sequence from the vehicle to separate the macroscopic trajectory node sequence, as well as the spatial displacement parameters and onboard high-frequency inertial sensing sequences within adjacent timestamp intervals corresponding to the original time-series service data sequence from the vehicle, including: Extract the time-series synchronization message headers from the original time-series service data sequence on the vehicle side; Compare the absolute timestamp recorded in the timing synchronization message header with the relative time offset parameter of the vehicle bus; The data splitting instruction is triggered when the absolute timestamp is perfectly aligned with the relative time offset parameter of the vehicle bus.

3. The method for enhancing vehicle network data security based on a large model as described in claim 2, characterized in that, After the data splitting instruction is triggered, it also includes: Based on the spatial grid division standard, the original time-series business data sequence of the vehicle end is separated to generate a macroscopic trajectory node sequence; Extract the coordinate difference of the original time-series business data sequence from the vehicle end within adjacent sampling periods to generate spatial displacement parameters; Simultaneously extract the three-dimensional spatial acceleration signal and yaw rate signal within adjacent sampling periods, and merge the three-dimensional spatial acceleration signal and yaw rate signal into an on-board high-frequency inertial sensing sequence.

4. The method for enhancing vehicle network data security based on a large model as described in claim 3, characterized in that, After merging the three-dimensional spatial acceleration signal and the yaw rate signal into an onboard high-frequency inertial sensing sequence, it also includes: Extract chassis topology location labels from the vehicle-mounted high-frequency inertial sensing sequence; Match the chassis topology location labels with the preset vehicle firmware component assembly mapping table; When the chassis topology location label matches the preset vehicle firmware component assembly mapping table, a calculus calculation trigger signal is output.

5. The method for enhancing vehicle network data security based on a large model as described in claim 4, characterized in that, The deviation feature values ​​are extracted by fusing spatial displacement parameters with vehicle-mounted high-frequency inertial sensing sequences, including: It receives the calculus operation trigger signal, performs dual time calculus operations on the three-dimensional spatial acceleration signal and the yaw rate signal, and generates a dynamic displacement vector; An interpolation algorithm is used to force the temporal resolution of the spatial displacement parameter to be aligned with the vehicle-mounted high-frequency inertial sensing sequence, and a standardization algorithm is used to eliminate the absolute dimensional differences between the spatial displacement parameter and the dynamically derived displacement vector. Based on the normalized data, the deviation covariance matrix between the spatial displacement parameter and the dynamically derived displacement vector is established. Singular value decomposition is performed on the deviation covariance matrix to extract the principal eigenvalues ​​as kinematic feature parameters.

6. The method for enhancing vehicle network data security based on a large model as described in claim 5, characterized in that, Comparison of kinematic characteristic parameters with dynamic tolerance thresholds, including: Obtain the fleet dynamic load status descriptor that is perfectly aligned with the adjacent timestamp interval. The fleet dynamic load status descriptor records the vehicle chassis suspension airbag pressure sensor readings. The vehicle chassis suspension airbag pressure sensor reading is used as an index key value to input into the dynamic tolerance database, and the nonlinear suspension system dynamic error tolerance value corresponding to the vehicle chassis suspension airbag pressure sensor reading is retrieved as the dynamic tolerance threshold. When the principal eigenvalue is greater than the dynamic error tolerance value of the nonlinear suspension system, associate time-series abrupt change digital tags with adjacent timestamp intervals.

7. The method for enhancing vehicle network data security based on a large model as described in claim 6, characterized in that, Importing macroscopic trajectory node sequences into the global semantic feature matching network of a large model includes: Map the macroscopic trajectory node sequence to discretized business operation status identifiers; Convert discretized business operation status identifiers into high-dimensional semantic vectors; High-dimensional semantic vectors are input into the global semantic feature matching network of a large model.

8. The method for enhancing vehicle network data security based on a large model as described in claim 7, characterized in that, Output the comparison results between the macroscopic trajectory node sequence and the preset logistics transportation business specifications, including: Retrieve the vector template of compliant business specifications included in the preset logistics and transportation business specifications; Calculate the spatial cosine similarity between the high-dimensional semantic vector and the compliance business specification vector template; If the spatial cosine similarity meets the preset standard threshold, a comparison result that is in compliance status is generated.

9. The method for enhancing vehicle network data security based on a large model as described in claim 8, characterized in that, The frequency of occurrence of statistical time-series abrupt change numerical labels within a sliding alignment window at the baseline time includes: Accumulate the total number of occurrences of the timing abrupt change digital labels within the reference time sliding alignment window; Calculate the ratio of the total number of occurrences of the time-series abrupt change digital label to the total length of the reference time sliding alignment window, and generate the occurrence frequency; If the comparison results are compliant and the frequency of occurrence exceeds the preset business fault tolerance threshold, the transmission of the original time-series business data sequence from the vehicle to the business value assessment stage is blocked by closing the vehicle data receiving port, and the unprocessed original time-series business data sequence from the vehicle is cleared simultaneously.

10. A vehicle-to-everything (V2X) data security enhancement system based on a large model, applied to the V2X data security enhancement method based on a large model as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to collect the original time-series business data sequences continuously uploaded by the vehicle-mounted nodes. The data splitting module is used to split the original time-series business data sequence of the vehicle end, separating the macro trajectory node sequence, as well as the spatial displacement parameters and vehicle high-frequency inertial sensing sequence within the adjacent timestamp intervals corresponding to the original time-series business data sequence of the vehicle end. The feature comparison module is used to extract deviation feature values ​​by fusing spatial displacement parameters and vehicle-mounted high-frequency inertial sensing sequences. The deviation feature values ​​are set as kinematic feature parameters. The kinematic feature parameters are compared with the dynamic tolerance threshold. When the kinematic feature parameters are greater than the dynamic tolerance threshold, the module associates time-series change digital tags with adjacent timestamp intervals. The semantic matching module is used to import the macro trajectory node sequence into the global semantic feature matching network of the large model and output the comparison results between the macro trajectory node sequence and the preset logistics and transportation business specifications. The safety blocking module is used to establish a reference time sliding alignment window, reverse map the time-series mutation digital tags to the corresponding time nodes included in the macro trajectory node sequence, count the frequency of occurrence of the time-series mutation digital tags in the reference time sliding alignment window, and block the original time-series business data sequence on the vehicle side from entering the business value assessment stage if the comparison result is compliant and the frequency of occurrence is greater than the preset business fault tolerance threshold.

Images