A non-contact vehicle overload dynamic detection method and system

Abstract

The application discloses a kind of non-contact vehicle overload dynamic detection method and system, it is related to intelligent transportation technical field.First, the trajectory information of target detection vehicle is continuously collected, and vehicle trajectory sequence is constructed;Vehicle trajectory sequence is encoded, and high-dimensional space-time feature based on vehicle dynamics mapping relationship is obtained;High-dimensional space-time feature is decoupled and separated, and decoupled feature is obtained, actual observation and theoretical deduction level feature reconstruction and difference quantization are carried out to decoupled feature, fusion motion trajectory offset degree, quality attribute offset degree and global reconstruction fidelity, obtain final abnormal score;Vehicle overload state is identified according to final abnormal score and physical quality inference result.The application realizes dynamic identification and online early warning to overload risk by mining the influence of overload behavior on vehicle dynamics characteristics and space-time trajectory mode, so as to improve the efficiency of overload control and intelligent supervision level.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and more specifically to a non-contact dynamic detection method and system for vehicle overload. Background Technology

[0002] Current methods for detecting overloaded freight vehicles primarily rely on fixed weighing equipment, weighbridges, or dynamic weighing systems. These methods typically require vehicles to slow down, be diverted, or even briefly stop before detection, resulting in limited detection efficiency, high deployment costs, limited coverage, and insufficient continuous monitoring capabilities. Furthermore, existing technologies mostly focus on directly measuring vehicle weight, with insufficient utilization of vehicle behavior, dynamic response, and trajectory evolution. Even some studies attempting to incorporate vehicle motion trajectories for overload detection largely remain at a rudimentary kinematic logic level, exhibiting bottlenecks such as low detection accuracy and weak scenario generalization ability, making it difficult to achieve continuous identification and risk warning of suspected overloaded vehicles without stopping inspections. Therefore, a new method for overload identification that requires no contact with the vehicle and no stopping is urgently needed.

[0003] Therefore, in view of the shortcomings of the existing technology, how to provide a non-contact dynamic detection method and system for vehicle overload is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a non-contact vehicle overload dynamic detection method and system. By exploring the influence of overload status on vehicle dynamic behavior and spatiotemporal trajectory patterns, it realizes dynamic identification and online early warning of overload risk, thereby improving the efficiency of overload control and the level of intelligent supervision. Overload identification can be achieved without contacting the vehicle or stopping the vehicle.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a non-contact dynamic detection method for vehicle overload, comprising: Collect trajectory data of the target vehicle running continuously, and construct a vehicle trajectory input sequence based on the trajectory data; Collect trajectory data of the target vehicle running continuously, and construct a vehicle trajectory input sequence based on the trajectory data; An intervention-sensing spatiotemporal Transformer model is constructed, which encodes the vehicle trajectory sequence to obtain high-dimensional spatiotemporal features based on vehicle dynamics mapping relationships. A dual-branch motion-mass co-decoupling structure is used to decouple and separate the high-dimensional spatiotemporal features to obtain decoupled features, which include baseline motion features, baseline mass state features, observed motion features, and observed mass inference features. The decoupling features are reconstructed and quantized at the level of actual observation and theoretical deduction to calculate the motion trajectory offset, mass attribute offset and global reconstruction fidelity. The motion trajectory offset, mass attribute offset, and global reconstruction fidelity are combined, and a motion-mass co-consistency penalty factor is introduced to obtain the final anomaly score. The vehicle overload status is identified based on the final anomaly score and the physical quality inference results.

[0006] Preferably, the intervention-sensing spatiotemporal Transformer model encodes the vehicle trajectory sequence to obtain high-dimensional spatiotemporal features based on vehicle dynamics mapping relationships, including: The vehicle trajectory input sequence is input into the embedding layer of the intervention perception spatiotemporal Transformer model, and the initial representation is obtained by superimposing position encoding: ; in, Input a sequence for the vehicle trajectory. For position encoding; The initial representation The quality intervention state command m, inferred from within the model, is input to the intervention perception coding layer. The quality intervention state command m is used to characterize the implicit physical load attributes of the target vehicle. Feature encoding is achieved through a self-attention mechanism modulated by quality causal intervention do(m): ; in, These are the query matrix, key matrix, and value matrix, respectively. Scaling factor For the causal attention mask modulated by the implicit quality state, do(m) represents the computation of self-attention under the causal premise of the quality intervention state instruction m; After the above feature encoding, a high-dimensional spatiotemporal feature representation is obtained: .

[0007] Preferably, the dual-branch motion-mass co-decoupling structure includes: a reference branch and a response offset branch; The reference branch uses theoretical deduction to simulate the ideal dynamic response of the target vehicle under the reference standard load mass; The response offset branch uses actual observations to capture the true response of the target vehicle under actual driving conditions.

[0008] Preferably, the reference branch decouples the high-dimensional spatiotemporal features into reference motion features and reference mass state features through a mapping network; The reference motion characteristics Used to characterize the standard motion characteristics of a standard-load vehicle under current road conditions: ; The reference quality state characteristics Characterized mass state features under standard load conditions: ; in, , For mapping functions with different parameters.

[0009] Preferably, the response offset branch decouples the high-dimensional spatiotemporal features into observation motion features and observation quality inference features through a mapping network; The observed motion characteristics Used to capture motion features in actual trajectory sequences: ; The observation quality inference feature The target vehicle mass state features used to infer backward from the current trajectory: ; in, , For mapping functions with different parameters.

[0010] Preferably, the decoupled features are reconstructed and quantized using both actual observation and theoretical deduction methods to calculate the trajectory offset, mass attribute offset, and global reconstruction fidelity, including: Reconstructing the normal baseline trajectory using baseline motion characteristics and observed motion characteristics respectively Reconstructing the trajectory of facts The motion trajectory offset is obtained: ; By calculating the distance between the observed quality inferred features and the baseline quality state features in the causal latent space, the degree of intrinsic quality attribute offset caused by overload is quantified, and the quality attribute offset is obtained: ; in, Indicates the characteristics for inferring observation quality. Indicates the characteristics of the reference quality state; Calculate the original vehicle trajectory input sequence Reconstructing the trajectory of facts The residuals between them yield the global reconstruction fidelity: ; in, Indicates the global reconstruction fidelity. Indicates the deviation of the motion trajectory. This indicates the offset of the quality attribute.

[0011] Preferably, the motion trajectory offset, mass attribute offset, and global reconstruction fidelity are integrated, and a motion-mass co-consistency penalty factor is introduced to obtain the final anomaly score, including: By integrating the motion trajectory offset, mass attribute offset, and global reconstruction fidelity, a hybrid anomaly scoring function with physical perception capabilities is constructed: ; in, and This is the global balance coefficient. and For the dynamic adjustment coefficients of the motion and mass items, Indicates the global reconstruction fidelity. Indicates the deviation of the motion trajectory. Indicates the offset of the quality attribute; A motion-quality consistency penalty factor is introduced. When kinematic abnormalities and quality abnormalities show a positive correlation, nonlinear abnormality enhancement is performed to obtain the final abnormality score. ; in, To characterize and Consistency function of unidirectional and coordinated change trends.

[0012] Preferably, identifying the vehicle overload status based on the final anomaly score and physical quality inference results includes: If and only if the final anomaly score Exceeding the threshold Furthermore, the relative deviation of its observed implicit quality from the reference quality exceeds the allowable threshold. At that time, it was determined that the target vehicle was overloaded. Otherwise, it is considered to be in normal driving condition. The expression is as follows: .

[0013] Preferably, a non-contact vehicle overload dynamic detection system includes: The data acquisition and enhancement module is used to acquire the trajectory data of the target vehicle running continuously, and to construct a vehicle trajectory input sequence based on the trajectory data; The feature encoding module is used to construct an intervention-sensing spatiotemporal Transformer model, which encodes the vehicle trajectory sequence to obtain high-dimensional spatiotemporal features based on vehicle dynamics mapping relationships. The decoupling module is used to decouple and separate the high-dimensional spatiotemporal features using a dual-branch motion-mass collaborative decoupling structure to obtain decoupling features, which include reference motion features, reference mass state features, observation motion features, and observation mass inference features. The feature reconstruction and difference quantization module is used to perform feature reconstruction and difference quantization at the level of actual observation and theoretical deduction on the decoupled features, and calculate the motion trajectory offset, mass attribute offset and global reconstruction fidelity. The scoring module is used to fuse the motion trajectory offset, quality attribute offset, and global reconstruction fidelity, and introduce a motion-quality co-consistency penalty factor to obtain the final anomaly score. The identification module is used to identify the vehicle overload status based on the final anomaly score and the physical quality inference result.

[0014] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a non-contact vehicle overload dynamic detection method and system, which has at least the following beneficial effects: 1. Reduce hardware construction costs and improve traffic efficiency: There is no need for vehicles to stop, slow down, or install expensive weighbridges and dynamic weighing (WIM) equipment on the road surface. Overload risk identification can be completed solely based on existing continuous trajectory data of highway vehicles, which significantly reduces construction costs and improves road traffic efficiency. 2. Filtering environmental and human interference significantly reduces false alarm rate: Traditional trajectory analysis easily misjudges "slow driving" or "congestion ahead" as overloading. This invention introduces a causal intervention mechanism, which can forcibly remove road noise such as driving style and traffic congestion, and extract the inertial drag performance caused purely by overloading, greatly improving the detection accuracy under complex road conditions; 3. Based on a vehicle dynamics-based construction model, it possesses strong interpretability: overcoming the "black box" deficiency of purely data-driven models. By utilizing the intervention-perception Transformer forced model, attention is focused on the starting acceleration and long downhill braking phases where overload characteristics are most obvious. The detection logic fully conforms to the physical common sense of vehicle dynamics, making the results more convincing. 4. Motion-mass cross-validation enhances detection sensitivity: The dual-branch structure not only compares whether the vehicle's trajectory is "correct" (motion trajectory deviation), but also infers the vehicle's "inertia as many tons" (dynamic property shift). This cross-validation of appearance and underlying physical characteristics significantly improves the model's sensitivity to slight overloading or complex cheating behaviors; 5. Possesses dynamic adaptive capability for different road conditions: Supports online updates and dynamic adaptation. The system can automatically calibrate the "load dynamic benchmark" according to different road sections (such as long uphill slopes and plains) and different seasons (changes in road surface adhesion due to rain and snow), meeting the needs of long-term, maintenance-free intelligent monitoring. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of a non-contact vehicle overload dynamic detection method provided by the present invention.

[0017] Figure 2 This is a schematic diagram of the high-dimensional spatiotemporal feature acquisition process provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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] This invention discloses a non-contact dynamic detection method for vehicle overload, such as... Figure 1 As shown, it includes: Collect trajectory data of the target vehicle running continuously, and construct a vehicle trajectory input sequence based on the trajectory data; An intervention-sensing spatiotemporal Transformer model is constructed, which encodes the vehicle trajectory sequence to obtain high-dimensional spatiotemporal features based on vehicle dynamics mapping relationships. A dual-branch motion-mass co-decoupling structure is used to decouple and separate the high-dimensional spatiotemporal features to obtain decoupled features, which include baseline motion features, baseline mass state features, observed motion features, and observed mass inference features. The decoupling features are reconstructed and quantized at the level of actual observation and theoretical deduction to calculate the motion trajectory offset, mass attribute offset and global reconstruction fidelity. The motion trajectory offset, mass attribute offset, and global reconstruction fidelity are combined, and a motion-mass co-consistency penalty factor is introduced to obtain the final anomaly score. The vehicle overload status is identified based on the final anomaly score and the physical quality inference results.

[0020] A dual-branch motion-mass collaborative decoupling structure is used to decouple and separate the high-dimensional spatiotemporal features, resulting in decoupled features. Specifically, the motion trajectory and mass attributes of the high-dimensional spatiotemporal features are decoupled. Since the high-dimensional spatiotemporal feature representation output from the previous stage satisfies… It contains the initial representation The motion trajectory information carried, and the intervention instructions The injected quality attribute constraints form a coupled spatiotemporal feature; the decoupling and separation is to orthogonally decouple and purify the high-dimensional spatiotemporal feature based on the reference branch and the influence offset branch, and decompose the mixed high-dimensional spatiotemporal feature into mutually independent and pure decoupled features. The decoupled features include motion trajectory features and quality attribute features. Finally, the decoupled features include reference motion features, reference quality state features, observation motion features, and observation quality inference features.

[0021] This invention employs a technical approach of "physical feature enhancement—intervention-aware modeling—motion-mass hierarchy decoupling—dual-criteria joint discrimination": First, continuous vehicle trajectory data is collected and virtual dynamics proxy features are introduced to enhance physical properties; second, an intervention-aware Transformer is used to extract high-dimensional features containing underlying dynamic causal relationships; third, motion features and mass attributes are deeply decoupled and separated through a dual-main-branch structure and causal operators; finally, overload status is identified based on multi-dimensional intervention deviation and consistency verification. This approach can be used in scenarios such as highway overload control for freight vehicles, traffic operation monitoring, intelligent vehicle supervision, and abnormal behavior identification.

[0022] Specifically, the vehicle trajectory input sequence constructed based on the trajectory data includes: Collect continuous trajectory data of the target vehicle, extract basic kinematic features such as position, velocity, acceleration and jerk, and derive virtual dynamic proxy features that integrate kinetic energy and traction / braking power changes to construct a physically enhanced vehicle trajectory input sequence.

[0023] Specifically, continuous trajectory data of the target freight vehicle during its road operation is collected, and a single-moment feature vector is constructed as follows: ; in, This indicates the spatial position of the vehicle at time t. Indicates speed, Indicates acceleration. It indicates jerk.

[0024] Based on the single-moment feature vector, a time-series vehicle trajectory input sequence of length T is constructed: .

[0025] Specifically, the intervention perception spatiotemporal Transformer model uses the vehicle's implicit mass as a potential causal confounding variable, and utilizes a dynamic attention mask modulated by the causal state of mass to encode the input sequence, extracting high-dimensional spatiotemporal features that filter out environmental noise and contain underlying physical causal relationships. Specifically, the intervention-perception spatiotemporal Transformer model encodes the vehicle trajectory sequence to obtain high-dimensional spatiotemporal features based on vehicle dynamics mapping relationships, including: The vehicle trajectory input sequence is input into the embedding layer of the intervention perception spatiotemporal Transformer model, and the initial representation is obtained by superimposing position encoding: ; in, Input a sequence for the vehicle trajectory. For position encoding; The initial representation The quality intervention state command m, inferred from within the model, is input to the intervention perception coding layer. The quality intervention state command m is used to characterize the implicit physical load attributes of the target vehicle. The quality intervention state command m is based on a large amount of historical vehicle trajectory data, under the premise of the lack of direct weighing equipment. After pre-training, when facing a target vehicle, latent variables are inferred in real time based on its continuous trajectory data to characterize the target vehicle's implicit physical load properties: ; in, The network feature encoder is used for inferring latent variables, with the following parameters: , For MLP inference head, the parameters are: , Indicates parameters , Through historical data Training yielded results.

[0026] Feature encoding is achieved through a self-attention mechanism modulated by quality causality intervention: ; in, These are the query matrix, key matrix, and value matrix, respectively, and all are initialized with an initial representation. Mapped to obtain, Scaling factor For the causal attention mask modulated by the implicit quality state, do(m) represents the computation of self-attention under the causal premise of the quality intervention state instruction m; When the trajectory sequence is in dynamically sensitive sections such as initial acceleration or long downhill braking, the sudden changes in the vehicle's motion state are constrained by its own load during the process of overcoming huge physical inertia and changing speed. This results in the trajectory features of this stage containing high-density mass causal perturbation information. After capturing this strong physical signal and high causal information manifold, the conditional feature mapping network reaches its maximum activation value, and the output causal attention mask... The corresponding matrix element values ​​are approximated to 1 by the causal attention mask. The element-wise multiplication with the original attention score matrix amplifies the attention weights of overload-sensitive segments; conversely, during the constant-speed cruise phase, the causal attention mask... The corresponding matrix element values ​​are close to 0 to shield environmental noise during the cruise phase; in this process, in order to minimize the causal inference loss function, the model spontaneously focuses attention on the key segments with the most obvious overload characteristics; like Figure 2 As shown, after the above feature encoding, a high-dimensional spatiotemporal feature representation containing the underlying dynamic causal relationships is obtained: .

[0027] Specifically, the dual-branch motion-mass collaborative decoupling structure introduces theoretical deduction and actual observation quality intervention commands, implements hierarchical feature stripping, and learns the motion characteristics and mass inference characteristics of the target vehicle under the baseline load state and the actual observation state, respectively. This embodiment of the invention obtains high-dimensional causal features. Based on this, a dual main branch is constructed, and hierarchical feature stripping is implemented within it to achieve deep decoupling between apparent motion performance and underlying physical properties.

[0028] Specifically, the dual-branch motion-mass co-decoupling structure includes: a NormalReference Branch and an Anomaly Response Branch. The reference branch is derived theoretically. Simulated target vehicle under baseline load mass Ideal dynamic response under the given conditions; The response offset branch is based on actual observations. Capture the real-world response of the target vehicle under actual driving conditions.

[0029] Specifically, the reference branch decouples the high-dimensional spatiotemporal features into reference motion features and reference mass state features through a mapping network; The reference motion characteristics Used to characterize the standard motion characteristics of a standard-load vehicle under current road conditions: ; The reference quality state characteristics Characterized mass state features under standard load conditions: ; in, , For mapping functions with different parameters.

[0030] Specifically, the response offset branch decouples the high-dimensional spatiotemporal features into observation motion features and observation quality inference features through a mapping network; The observed motion characteristics Used to capture motion features in actual trajectory sequences: ; The observation quality inference feature The target vehicle mass state features used to infer backward from the current trajectory: ; in, , For mapping functions with different parameters.

[0031] This invention constructs a network architecture that decouples intervention perception from motion-mass hierarchy, introducing theoretical derivation into trajectory anomaly detection for the first time. Even without knowing the vehicle's true weight, it successfully separates and independently models the mixed motion trajectory and mass attributes.

[0032] At the same time, a comprehensive anomaly index is constructed by combining actual observations and theoretical deductions in multiple dimensions: the simple reconstruction error judgment is abandoned, and the motion deviation and consistency verification mechanism is combined, so that overload identification no longer depends on a single threshold, and the system is more rigorous.

[0033] Specifically, a hybrid intervention perception and evaluation system is constructed based on cross-dimensional motion trajectory offset, dynamic attribute offset, and global reconstruction fidelity. A dual-criteria joint verification is introduced in conjunction with the motion-mass synergistic consistency mechanism. Based on the comprehensive anomaly score and physical quality inference results, the accurate identification and judgment of vehicle overload status is completed.

[0034] Specifically, the decoupling features are reconstructed and quantized using both actual observation and theoretical deduction methods to calculate the trajectory offset, mass attribute offset, and global reconstruction fidelity, including: Reconstructing the normal baseline trajectory using baseline motion characteristics and observed motion characteristics respectively Reconstructing the trajectory of facts The motion trajectory offset is obtained: ; The normal baseline trajectory The reference motion characteristics are decoupled from the reference reference branch. The input is fed into the spatiotemporal trajectory decoding network, and the feature recovery mapping is obtained as follows: ; The trajectory of fact reconstruction The observed motion characteristics are decoupled from the response offset branch output. The inputs are fed into the same spatiotemporal trajectory decoding network, and the mapping is obtained through the same feature recovery: ; in, For trajectory reconstruction mapping function.

[0035] By calculating the distance between the observed quality inferred features and the baseline quality state features in the causal latent space, the degree of intrinsic quality attribute offset caused by overload is quantified, and the quality attribute offset is obtained: ; in, Indicates the characteristics for inferring observation quality. Indicates the characteristics of the reference quality state; Calculate the original vehicle trajectory input sequence Reconstructing the trajectory of facts The residuals between them are used to evaluate the model's ability to reconstruct spatiotemporal sequences, and the global reconstruction fidelity is obtained. ; in, Indicates global reconstruction fidelity, used to characterize the degree of difference between the original input trajectory and the actual reconstructed trajectory. This represents the trajectory offset, used to characterize the degree of feature deviation between the reference trajectory features and the observed trajectory features. It represents the quality attribute offset, used to characterize the degree of feature deviation between the baseline quality state characteristics and the observed quality state characteristics.

[0036] Specifically, the motion trajectory offset, mass attribute offset, and global reconstruction fidelity are integrated, and a motion-mass co-consistency penalty factor is introduced to obtain the final anomaly score, including: By integrating the motion trajectory offset, mass attribute offset, and global reconstruction fidelity, a hybrid anomaly scoring function with physical perception capabilities is constructed: ; in, and This is the global balance coefficient. and For the dynamic adjustment coefficients of the motion and mass items, Indicates the global reconstruction fidelity. Indicates the deviation of the motion trajectory. Indicates the offset of the quality attribute; To further eliminate false anomalies caused by sudden changes in driving style (such as sudden braking or sharp turns), a motion-mass consistency penalty factor is introduced. Nonlinear anomaly enhancement is performed only when kinematic anomalies and mass anomalies show a high degree of co-correlation, resulting in a final anomaly score. When the two are not co-correlated, the exponential term... The system directly uses the basic calculation results of the hybrid anomaly scoring function as the final anomaly score (i.e., This avoids amplifying the scores of operational false anomalies, thus ensuring that they are not misjudged as overload states in subsequent dual-criteria checks. ; in, To characterize and Consistency function of unidirectional and coordinated change trends.

[0037] Specifically, set a threshold for overall system anomaly assessment. and the implicit quality relative deviation threshold To reduce the false alarm rate, a dual-criteria verification logic coupled with the final anomaly score and physical quality inference results is used to identify vehicle overload status, specifically including: If and only if the final anomaly score Exceeding the threshold Furthermore, the relative deviation of its observed implicit quality from the reference quality exceeds the allowable threshold. At that time, it was determined that the target vehicle was overloaded. Otherwise, it is considered to be in normal driving condition. The final state determination rule expression for the target vehicle is as follows: .

[0038] Specifically, this also includes: Introducing an online update mechanism: To improve the model's adaptability to different road environments, traffic conditions, and seasonal conditions, this invention further introduces a sliding window online update mechanism. A data window is constructed for historical trajectory samples at time t: ; The model parameters are incrementally updated using newly arrived samples, and the parameter update can be expressed as follows: ; in, For learning rate, This is the model loss function.

[0039] This invention, through online updates, improves the model's generalization ability to new scenarios and operating conditions. It constructs a highly interference-resistant, purely data-driven intelligent monitoring closed loop, supporting contactless, continuous, and online intelligent monitoring modes, and possesses promising engineering application prospects.

[0040] In one specific embodiment of the present invention, a non-contact vehicle overload dynamic detection system includes: The data acquisition and enhancement module is used to acquire the trajectory data of the target vehicle running continuously, and to construct a vehicle trajectory input sequence based on the trajectory data; The feature encoding module is used to construct an intervention-sensing spatiotemporal Transformer model, which encodes the vehicle trajectory sequence to obtain high-dimensional spatiotemporal features based on vehicle dynamics mapping relationships. The decoupling module is used to decouple and separate the high-dimensional spatiotemporal features using a dual-branch motion-mass collaborative decoupling structure to obtain decoupling features, which include reference motion features, reference mass state features, observation motion features, and observation mass inference features. The feature reconstruction and difference quantization module is used to perform feature reconstruction and difference quantization at the level of actual observation and theoretical deduction on the decoupled features, and calculate the motion trajectory offset, mass attribute offset and global reconstruction fidelity. The scoring module is used to fuse the motion trajectory offset, quality attribute offset, and global reconstruction fidelity, and introduce a motion-quality co-consistency penalty factor to obtain the final anomaly score. The identification module is used to identify the vehicle overload status based on the final anomaly score and the physical quality inference result.

[0041] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0042] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A non-contact dynamic detection method for vehicle overload, characterized in that, include: Collect trajectory data of the target vehicle running continuously, and construct a vehicle trajectory input sequence based on the trajectory data; An intervention-sensing spatiotemporal Transformer model is constructed, which encodes the vehicle trajectory sequence to obtain high-dimensional spatiotemporal features based on vehicle dynamics mapping relationships. A dual-branch motion-mass co-decoupling structure is used to decouple and separate the high-dimensional spatiotemporal features to obtain decoupled features, which include baseline motion features, baseline mass state features, observation motion features, and observation mass inference features. The decoupling features are reconstructed and quantized at the level of actual observation and theoretical deduction to calculate the motion trajectory offset, mass attribute offset and global reconstruction fidelity. The motion trajectory offset, mass attribute offset, and global reconstruction fidelity are combined, and a motion-mass co-consistency penalty factor is introduced to obtain the final anomaly score. The vehicle overload status is identified based on the final anomaly score and the physical quality inference results.

2. The non-contact dynamic detection method for vehicle overload according to claim 1, characterized in that, The intervention-sensing spatiotemporal Transformer model encodes the vehicle trajectory sequence to obtain high-dimensional spatiotemporal features based on vehicle dynamics mapping relationships, including: The vehicle trajectory input sequence is input into the embedding layer of the intervention perception spatiotemporal Transformer model, and the initial representation is obtained by superimposing position encoding: ； in, Input a sequence for the vehicle trajectory. For position encoding; The initial representation The quality intervention state command m, inferred from within the model, is input to the intervention perception coding layer. The quality intervention state command m is used to characterize the implicit physical load attributes of the target vehicle. Feature encoding is achieved through a self-attention mechanism modulated by quality causality intervention: ； in, These are the query matrix, key matrix, and value matrix, respectively. Scaling factor For the causal attention mask modulated by the implicit quality state, do(m) represents the computation of self-attention under the causal premise of the quality intervention state instruction m; After the above feature encoding, a high-dimensional spatiotemporal feature representation is obtained: 。 3. The non-contact dynamic detection method for vehicle overload according to claim 1, characterized in that, The dual-branch motion-mass co-decoupling structure includes: a reference branch and a response offset branch; The reference branch uses theoretical deduction to simulate the ideal dynamic response of the target vehicle under the reference standard load mass; The response offset branch uses actual observations to capture the true response of the target vehicle under actual driving conditions.

4. The non-contact dynamic detection method for vehicle overload according to claim 3, characterized in that, The reference branch decouples the high-dimensional spatiotemporal features into reference motion features and reference mass state features through a mapping network; The reference motion characteristics Used to characterize the standard motion characteristics of a standard-load vehicle under current road conditions: ； The reference quality state characteristics Characterized mass state features under standard load conditions: ； in, , For mapping functions with different parameters.

5. The non-contact dynamic detection method for vehicle overload according to claim 3, characterized in that, The response offset branch decouples the high-dimensional spatiotemporal features into observation motion features and observation quality inference features through a mapping network; The observed motion characteristics Used to capture motion features in actual trajectory sequences: ； The observation quality inference feature Features of the target vehicle's mass state inferred backward from the current trajectory: ； in, , For mapping functions with different parameters.

6. The non-contact dynamic detection method for vehicle overload according to claim 1, characterized in that, The decoupled features are reconstructed and quantized using both actual observation and theoretical deduction methods to calculate the trajectory offset, mass attribute offset, and global reconstruction fidelity, including: Reconstructing the normal baseline trajectory using baseline motion characteristics and observed motion characteristics respectively Reconstructing the trajectory of facts The motion trajectory offset is obtained: ； The quality attribute offset is obtained by calculating the distance between the observed quality inferred features and the baseline quality state features in the causal latent space: ； in, Indicates the characteristics for inferring observation quality. Indicates the characteristics of the reference quality state; Calculate the original vehicle trajectory input sequence Reconstructing the trajectory of facts The residuals between them yield the global reconstruction fidelity: ； in, Indicates the global reconstruction fidelity. Indicates the deviation of the motion trajectory. This indicates the offset of the quality attribute.

7. The non-contact dynamic detection method for vehicle overload according to claim 1, characterized in that, By integrating the motion trajectory offset, mass attribute offset, and global reconstruction fidelity, and introducing a motion-mass co-consistency penalty factor, a final anomaly score is obtained, including: By integrating the motion trajectory offset, mass attribute offset, and global reconstruction fidelity, a hybrid anomaly scoring function with physical perception capabilities is constructed: ； in, and This is the global balance coefficient. and For the dynamic adjustment coefficients of the motion and mass items, Indicates the global reconstruction fidelity. Indicates the deviation of the motion trajectory. Indicates the offset of the quality attribute; A motion-quality consistency penalty factor is introduced. When kinematic abnormalities and quality abnormalities show a positive correlation, nonlinear abnormality enhancement is performed to obtain the final abnormality score. ； in, To characterize and Consistency function of unidirectional and coordinated change trends.

8. The non-contact dynamic detection method for vehicle overload according to claim 7, characterized in that, The vehicle overload status is identified based on the final anomaly score and physical quality inference results, including: If and only if the final anomaly score Exceeding the threshold Furthermore, the relative deviation of its observed implicit quality from the reference quality exceeds the allowable threshold. At that time, it was determined that the target vehicle was overloaded. Otherwise, it is considered to be in normal driving condition. The expression is as follows: 。 9. A non-contact vehicle overload dynamic detection system based on trajectory analysis, employing the non-contact vehicle overload dynamic detection method according to any one of claims 1-8, characterized in that, include: The data acquisition and enhancement module is used to acquire the trajectory data of the target vehicle running continuously, and to construct a vehicle trajectory input sequence based on the trajectory data; The feature encoding module is used to construct an intervention-sensing spatiotemporal Transformer model, which encodes the vehicle trajectory sequence to obtain high-dimensional spatiotemporal features based on vehicle dynamics mapping relationships. The decoupling module is used to decouple and separate the high-dimensional spatiotemporal features using a dual-branch motion-mass collaborative decoupling structure to obtain decoupling features, which include reference motion features, reference mass state features, observation motion features, and observation mass inference features. The feature reconstruction and difference quantization module is used to perform feature reconstruction and difference quantization at the level of actual observation and theoretical deduction on the decoupled features, and calculate the motion trajectory offset, mass attribute offset and global reconstruction fidelity. The scoring module is used to fuse the motion trajectory offset, quality attribute offset, and global reconstruction fidelity, and introduce a motion-quality co-consistency penalty factor to obtain the final anomaly score. The identification module is used to identify the vehicle overload status based on the final anomaly score and the physical quality inference result.

Images