A multi-source sensing vehicle cheating detection method based on a space-time graph neural network
By constructing a multi-source sensor vehicle cheating detection method based on spatiotemporal graph neural network, the method achieves accurate identification of single-axis and multiple complex cheating behaviors of vehicles on dynamic weighbridges, solving the problem of difficulty in identifying complex cheating behaviors in existing technologies and improving detection accuracy and real-time performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- FUZHOU UNIV
- Filing Date
- 2026-03-10
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies struggle to accurately identify single-axle and multi-component cheating behaviors on dynamic weighbridges, especially in complex scenarios where multiple vehicles are weighed simultaneously.
A multi-source sensor vehicle cheating detection method based on spatiotemporal graph neural network is constructed. Through multi-source sensor data acquisition, data preprocessing, spatiotemporal graph construction, spatiotemporal graph neural network feature learning, and multi-level decision output, the method can accurately identify vehicle cheating behavior.
It improves the accuracy and real-time performance of vehicle cheating detection, ensures the reliability and fairness of weighing data, can identify single-axle and multiple complex cheating behaviors, and adapts to complex scenarios where multiple vehicles are weighed simultaneously.
Smart Images
Figure CN121808649B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent traffic monitoring and dynamic weighing, and particularly relates to a multi-source sensor vehicle cheating detection method based on spatiotemporal graph neural network, which is used to identify abnormal behaviors of vehicles during the weighing process of dynamic truck scales, including edge pressing, S-shaped driving, dragging the scale, jumping the scale and various complex cheating behaviors. Background Technology
[0002] With the continuous development of the national transportation industry, off-site overload control systems have been widely used. Dynamic truck scales, as the core equipment in these systems, primarily detect axle load and total weight under normal driving conditions. However, in actual road environments, some vehicles resort to cheating tactics to evade detection, such as driving in an S-shape to change the force distribution, pressing down on edges to reduce the force area, dragging the scale, or abnormally stopping to alter the weighing process, or installing support devices (such as jacks) to transfer weight. These actions result in distorted weighing data.
[0003] Existing methods largely rely on data from single sensors, making it difficult to achieve multi-dimensional identification of complex cheating behaviors. Traditional vehicle cheating detection methods mainly depend on image or video surveillance, or detection using dynamic truck scales. Image or video surveillance is affected by external factors such as weather, and can only detect violations by the entire vehicle, failing to accurately identify the weighing of a single axle. While dynamic truck scales can acquire vehicle weight information, they have limitations in judging single cheating behaviors, especially in effectively identifying combinations of cheating behaviors. Furthermore, existing methods rely too heavily on idealized trajectory judgments, assuming uniform contact between the tires and the scale platform, and are ill-suited to complex scenarios where multiple vehicles pass through the scale simultaneously.
[0004] Chinese invention patent 202110979340.3 discloses a training method for a multi-temporal feature model. It describes using the vehicle and surrounding vehicles as graph nodes, and vehicle information (e.g., vehicle position, speed, and acceleration) as node features to construct a multi-temporal feature map. This multi-temporal feature map is then input into a graph convolutional neural network to extract corresponding multi-temporal graph feature vectors; thereby training a driving risk prediction model. However, when applied to the field of driving risk prediction, this technical solution cannot identify multiple complex combinations of cheating behaviors and struggles to handle complex scenarios where multiple vehicles simultaneously pass through a weighbridge.
[0005] Therefore, there is an urgent need in this field for a detection method that can integrate multi-source sensor data and accurately identify complex cheating behaviors of single axles and the whole vehicle. Summary of the Invention
[0006] The technical problem to be solved by this invention is to provide a multi-source sensing vehicle cheating detection method based on spatiotemporal graph neural network, which can accurately identify single-axle, whole-vehicle and multiple complex cheating behaviors during vehicle weighing, improve detection accuracy and real-time performance, and improve the reliability and fairness of dynamic truck scale weighing data.
[0007] This invention is implemented as follows:
[0008] A multi-source sensor vehicle cheating detection method based on spatiotemporal graph neural network constructs a five-layer detection framework. The method sequentially achieves vehicle cheating behavior detection through multi-source sensor data acquisition, data preprocessing and feature extraction, spatiotemporal graph construction, spatiotemporal graph neural network feature learning, and multi-level decision output. Specific steps include:
[0009] S1. Multi-source sensor data acquisition:
[0010] When a vehicle enters the dynamic weighbridge inspection area, the weighing process time start is triggered by the ground induction coil. Simultaneously, radar is used to obtain the vehicle's real-time driving trajectory, speed and acceleration information. The camera captures the spatial relationship between the vehicle's tires and the edge of the weighbridge and the lane lines. The dynamic weighbridge collects the axle weight signals of each axle of the vehicle to obtain full-process monitoring data of the vehicle from entering to leaving the inspection area.
[0011] S2. Data Preprocessing and Feature Extraction:
[0012] The multi-source sensor data collected in step S1 is processed for time synchronization. The data from each sensor are unified to the same time axis by timestamp alignment and interpolation mapping.
[0013] Noise suppression and outlier removal were performed on the synchronized data. Single-source features such as tire lateral position, trajectory curvature, wheelbase and axle time difference, and edge pressure distance were extracted from the original data of dynamic weighbridge, radar, inductive loop, and camera, respectively.
[0014] Spatial alignment is performed based on the radar coordinate system, and the data sampling frequency is unified through time interpolation. The data is then fused to obtain a multi-dimensional feature vector that includes the vehicle's spatial position, temporal evolution, and weighing status.
[0015] S3. Spatiotemporal Graph Construction:
[0016] The vehicle weighing process is modeled as a spatiotemporal graph that evolves dynamically over time, with each axle of the vehicle as a graph node and the node features being the multidimensional feature vector output in step S2. Spatial connections are established between adjacent axle nodes at the same time, and temporal connections are established between the same axle nodes in adjacent time slices.
[0017] A global vehicle feature node is introduced, which contains information on vehicle length, total weight, maximum acceleration, environmental coding, and weight confidence. It is fully connected to all axle nodes to form a dynamic spatiotemporal graph that includes spatial and temporal dependencies.
[0018] S4. Spatiotemporal graph neural network feature learning:
[0019] The dynamic spatiotemporal graph constructed in step S3 is input into the spatiotemporal graph neural network. First, the multi-source sensing features of the same axle node are adaptively fused through the attention weighting mechanism to obtain the initial embedding representation of the node.
[0020] In the spatial dimension, spatial anomaly features are characterized by aggregating the features of axle nodes, their adjacent nodes, and global nodes through a graph neural network based on an attention mechanism.
[0021] In the time dimension, temporal variation trends of features of the same axle node are captured through temporal convolution or recurrent units to identify persistent or abrupt temporal anomalous behaviors;
[0022] The final spatiotemporal embedding features of nodes are obtained by fusing spatial and temporal features;
[0023] S5. Multi-level decision output:
[0024] Based on the final spatiotemporal embedding features of nodes, nodes are embedded into the input classification network at the single-axis level to obtain the probability distribution of single-axis under various cheating modes. An environment-adaptive confidence threshold is introduced to filter suspicious anomalies on the single axis.
[0025] At the vehicle level, the embedded features of all axle nodes are weighted and aggregated to obtain a vehicle-level representation, which is then input into a classification network to obtain a vehicle cheating probability vector.
[0026] Spatiotemporal correlation matrix analysis is introduced to analyze the temporal evolution coupling relationship of each axle embedding and identify complex cheating behaviors;
[0027] The final output includes the anomaly type, confidence level, and structured detection results involving axis position and time interval.
[0028] Furthermore, in S2, noise suppression is achieved by combining wavelet transform and sliding window filtering on the multiple voltage signals of the dynamic truck scale. Then, the lateral position of the tires on the scale platform is obtained by weighting the voltage signals based on the spatial distribution of the weighing sensors. The calculation formula is as follows:
[0029]
[0030] Where, x t Let V be the lateral position of the tire on the weighing platform at time t, i.e., the distance from the left edge of the weighing platform. i Let x be the voltage value of the i-th load cell.i It corresponds to its horizontal position coordinates.
[0031] Furthermore, in step S2, the raw point cloud data acquired by the radar is first filtered out and clustered to separate the background point cloud, extract the vehicle target point set, and obtain the vehicle position sequence through target tracking. Then, the trajectory curvature feature is extracted by differential calculation of the vehicle centroid trajectory. The calculation formula is as follows:
[0032] Where x(t) and y(t) are the horizontal and vertical coordinates of the vehicle on the road surface coordinate system. , The first derivative of position with respect to time. , It is the second derivative.
[0033] Furthermore, in S2, the video data acquired by the camera is first subjected to perspective correction to eliminate lens distortion, and then the tire region is extracted based on the tire detection model. The minimum Euclidean distance between the tire and the edge of the weighing platform is calculated to quantify the edge pressing behavior. The calculation formula is as follows:
[0034] Where, d edge denoted as the edge distance, p is the position of the tire in the coordinate system of the weighing platform, and B is the set of points on the edge of the weighing platform.
[0035] Furthermore, the multidimensional feature vector obtained by fusion in step S2 is:
[0036] Among them, f t Here, xt represents the lateral position of the tire, wt represents the axle load feature, vt represents the vehicle speed, and d represents the node feature vector. edge The distance between the pressure plates is Δt, and the time difference between the axes is Δt. is the trajectory curvature feature, and fc is the contact state feature between the tire and the weighing platform.
[0037] Furthermore, the calculation formula for the global vehicle feature nodes in step S3 is as follows:
[0038]
[0039] Where u is the global vehicle feature node, L veh W is the length of the vehicle. total This refers to the total weight of the vehicle. For maximum acceleration, E env For environment coding, C w For weight confidence;
[0040] Constructed vehicle spatiotemporal graph G t Represented as:
[0041] Among them, V t Let E be the set of axle nodes. t U is a set of spatial and temporal edges. t These are global feature nodes.
[0042] Furthermore, the calculation formula for the initial embedding representation of nodes in S4 is as follows:
[0043] Among them, h i (0) Let i be the temporal characteristics of node i at the initial stage. For the k-th type of sensor features at node i, This is the corresponding feature mapping matrix. This is for normalized attention weights.
[0044] Furthermore, the environment-adaptive confidence threshold in S5 The calculation formula is:
[0045] in, The default threshold is used for judgment. Encoding for the environment, This is the environmental weighting coefficient.
[0046] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a multi-source sensing vehicle cheating detection method based on a spatiotemporal graph neural network as described above.
[0047] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of a multi-source sensing vehicle cheating detection method based on a spatiotemporal graph neural network as described above.
[0048] The advantages of this invention are:
[0049] 1. Multi-source sensor fusion enhances comprehensive identification: By integrating multi-dimensional information from dynamic truck scales, radar, cameras, and inductive loops, it breaks through the detection limitations of single sensors. It can not only identify single cheating behaviors such as edge pressing and scale jumping, but also accurately capture complex cheating patterns such as S-shaped driving and jack + edge pressing, solving the technical problem that existing methods cannot identify complex cheating behaviors.
[0050] 2. Spatiotemporal graph modeling for accurate characterization of vehicle behavior: The vehicle weighing process is modeled as a dynamic spatiotemporal graph, which simultaneously expresses the spatial distribution characteristics and temporal evolution of the vehicle. This achieves a unified representation of single-axle behavior, whole-vehicle behavior, and their temporal relationships, enabling it to handle complex vehicle driving conditions and scenarios where multiple vehicles weigh simultaneously, thereby improving detection robustness.
[0051] 3. End-to-end learning of spatiotemporal graph neural networks to improve detection accuracy: Adaptive fusion of multi-source features is achieved through attention mechanism, and abnormal features are mined in spatial and temporal dimensions respectively. The spatiotemporal correlation information of vehicles in the weighing process is fully learned, effectively identifying continuous, abrupt and complex cheating behaviors, and significantly improving the accuracy of cheating behavior identification.
[0052] 4. Real-time detection to meet engineering application needs: The entire detection process is based on end-to-end deep learning inference, which can quickly process and analyze each vehicle passing through the scale and output detection results in real time. This ensures timely feedback in actual road overload control scenarios and has good engineering practicality.
[0053] 5. Improve the reliability of weighing data and ensure the fairness of overloading control: It can effectively identify and eliminate various irregular weighing cheating behaviors, correct distorted weighing data, improve the reliability and fairness of dynamic truck scale weighing results, provide strong technical support for road transportation overloading control, and ensure the fairness and safety of the transportation industry. Attached Figure Description
[0054] The present invention will now be further described with reference to the accompanying drawings and embodiments.
[0055] Figure 1 This is a schematic diagram of the overall method flow of the present invention. Detailed Implementation
[0056] Please refer to Figure 1 A multi-source sensor vehicle cheating detection method based on spatiotemporal graph neural network is proposed. This method constructs a five-layer detection framework, sequentially achieving vehicle cheating behavior detection through multi-source sensor data acquisition, data preprocessing and feature extraction, spatiotemporal graph construction, spatiotemporal graph neural network feature learning, and multi-level decision output. Specific steps include:
[0057] S1. Multi-source sensor data acquisition:
[0058] When a vehicle enters the dynamic weighbridge's detection area, the system first detects its arrival using inductive loop detectors, marking this event as the starting point for the weighing process. Simultaneously, radar equipment acquires the vehicle's real-time trajectory, speed, and acceleration information. Cameras capture the spatial relationship between the vehicle's tires and the edge of the weighbridge and lane lines, transmitting this data to the processing system. As the vehicle's tires gradually enter the weighbridge, the dynamic weighbridge begins outputting axle load signals for each axle.
[0059] These multi-source sensor data work together to provide a complete monitoring system from the moment a vehicle enters the detection area to the moment it leaves the weighbridge, laying the foundation for subsequent abnormal behavior identification. Through the coordinated operation of these sensor data, the system can accurately record and reflect the vehicle's weighing process.
[0060] Target tracking of radar point clouds can be performed using the SORT / DeepSORT algorithm, and tire detection of cameras can be performed using the YOLO / SSD algorithm (this invention is not limited to the above algorithms, and can use conventional target detection and tracking algorithms in this field).
[0061] S2. Data Preprocessing and Feature Extraction:
[0062] During the weighing process of a vehicle on a dynamic truck scale, the system needs to uniformly process multi-source sensor data from the truck scale, radar, cameras, and inductive loops to form a structured feature representation that accurately reflects the vehicle's weighing behavior. Since various sensors differ in sampling frequency, data format, and signal characteristics, this layer first performs time synchronization processing on the multi-source sensor data. Through timestamp alignment and interpolation mapping, the data from each sensor are uniformly mapped to the same time axis, thereby ensuring consistency and comparability in the description of vehicle behavior at the same moment.
[0063] After time synchronization, noise suppression and outlier removal are performed on the multi-source raw data. For the multiple voltage signals output by the dynamic truck scale, a combination of wavelet transform and sliding window filtering is used to suppress high-frequency noise caused by factors such as vehicle vibration and electromagnetic interference. Based on this, the voltage signals are weighted according to the spatial distribution of the weighing sensors to obtain the lateral position of the tires on the scale platform. The calculation can be expressed as:
[0064] in, x t for t The lateral position of the tire on the scale platform at any given time is its distance from the left edge of the platform. V i For the first i Only the voltage value of the load cell ( i =1, 2, 3), x i It corresponds to its horizontal position coordinates.
[0065] For the raw point cloud data acquired by radar, the system first performs background point cloud filtering and clustering to extract the vehicle target point set, and then obtains the continuous position sequence of the vehicle in the road coordinate system through a target tracking algorithm. Based on this, the system extracts trajectory curvature features by performing differential calculations on the vehicle's centroid trajectory to characterize the degree of lateral sway of the vehicle on the weighing platform. The trajectory curvature is defined as:
[0066] in, For trajectory curvature features, (x ( t ), y ( t )) represents the horizontal and vertical coordinates of the vehicle on the road surface coordinate system. , The first derivative of position with respect to time. , As the second derivative, this feature can effectively characterize abnormal trajectory behaviors such as S-shaped driving.
[0067] After the high-speed pulse signal output by the inductive loop is digitally processed, the axle passing interval is calculated by the time difference between adjacent pulses, and the axle distance parameter is inferred by combining the vehicle speed information detected by radar. At the same time, it can be used to identify abnormal axle sequence or abnormal pulse pattern, providing auxiliary features for single axle behavior analysis.
[0068] For video data captured by the camera, the system performs perspective correction on the images to eliminate lens distortion and extracts the tire region based on the tire detection model. Based on this, it calculates the minimum Euclidean distance between the tire and the edge of the weighing platform to quantify whether the vehicle is engaging in edge-pressing behavior. The distance calculation form is as follows:
[0069] Where, d edge This is the edge pressing distance. p Let be the position of the tire in the coordinate system of the weighing platform. B This is the set of points on the edge of the weighing platform.
[0070] After extracting features from each sensor, the system performs unified fusion processing on the multi-source features. Using the radar coordinate system as the reference coordinate system, it spatially aligns the weighing, visual, and coil data, and uses time interpolation to unify data from different sampling frequencies to a fixed time resolution, forming a multi-dimensional feature vector.
[0071] Among them, f t For node feature vectors, x t The lateral position of the tire. w t Axle load characteristics, v t For vehicle speed, d edge Δ is the edge pressing distance. t For axis time difference, For trajectory curvature characteristics, f c This refers to the contact characteristics between the tire and the weighing platform.
[0072] Through the above data preprocessing and feature extraction processes, the system constructs a high-quality feature set that can simultaneously reflect the spatial positional relationship, temporal evolution characteristics, and weighing status changes of vehicles, providing reliable input for subsequent spatiotemporal graph construction and abnormal behavior recognition of spatiotemporal graph neural networks.
[0073] S3. Spatiotemporal Graph Construction:
[0074] To accurately depict the continuous weighing behavior of vehicles on dynamic truck scales, this invention models the entire process of a vehicle from entering to leaving the scale platform as a spatiotemporal graph structure that evolves dynamically over time. This spatiotemporal graph can simultaneously express the spatial distribution characteristics and temporal evolution patterns of vehicles on the scale platform, providing a unified and structured data representation for the subsequent joint identification of abnormal behaviors by the spatiotemporal graph neural network.
[0075] In the spatiotemporal graph, each axle of the vehicle serves as a basic graph node, with each node corresponding to the single-axle state of the vehicle at a given moment. The node features are composed of a multi-source fusion feature vector output from the second layer, used to comprehensively describe the spatial position, axle load state, and motion characteristics of the axle at the corresponding moment. The node feature vector is represented as f. t (For the specific formula, see Eq. ).
[0076] To reflect the structural characteristics of the vehicle and the cooperative relationships between axles, spatial connections are constructed between adjacent axle nodes of the same vehicle at the same time point to characterize the physical wheelbase relationship and force distribution correlation between axles. Simultaneously, temporal connections are established between the same axle nodes within adjacent time slices to describe the continuous evolution of the axle state over time. Through the joint construction of the above spatial and temporal edges, a dynamic graph structure containing both spatial and temporal dependencies is formed.
[0077] Building upon this, a global vehicle feature node is introduced to represent the overall state information at the vehicle level, including the vehicle length detected by radar, the total weight calculated by the dynamic weighbridge, the vehicle's maximum acceleration, and environmental coding information. The global feature node u is defined as:
[0078] in, L veh For vehicle length, W total This refers to the total weight of the vehicle. For maximum acceleration, E env Encode the environment (e.g., weather: sunny = 1, cloudy = 0.8, rainy = 0.5; lighting: strong light = 1, weak light = 0.6). C wFor weight confidence, the global node is associated with all axle nodes through a fully connected manner, which is used to introduce global constraint information at the vehicle level into the graph structure.
[0079] Finally, the vehicle spatiotemporal graph G constructed within each time window t It can be represented as:
[0080] in, V t For the set of axle nodes, E t It is a set of spatial edges and temporal edges. U t These serve as global feature nodes. The spatiotemporal graph structure is dynamically updated over time, reflecting the real-time behavioral changes of vehicles on the weighbridge. Through this spatiotemporal graph construction layer, the single-axle behavior, overall vehicle behavior, and their temporal evolution are uniformly mapped to the same graph model, providing a high-quality, parsable input representation for the subsequent Spatiotemporal Graph Neural Network Feature Learning (STGNN) model to learn and identify complex cheating behaviors.
[0081] S4. Spatiotemporal Graph Neural Network Feature Learning (STGNN):
[0082] The STGNN core model layer is used for deep analysis and feature learning of the spatiotemporal graph of the vehicle weighing process constructed in the previous layer, thereby uncovering potential abnormal behavior patterns of vehicles on dynamic weighbridges. This layer takes the dynamic graph structure output by the spatiotemporal graph construction layer as input, and achieves a unified representation of the vehicle's single-axle behavior and whole-vehicle behavior by jointly modeling the spatial relationships and temporal evolution laws in the graph structure.
[0083] During the model input phase, the system first performs feature-level fusion processing on the multi-source sensor features. For the feature representations of the same axle node from different sensors, an attention weighting mechanism is introduced to adaptively fuse them, resulting in the initial embedding representation of the node, which is calculated as follows:
[0084] Among them, h i (0) Let i be the temporal characteristics of node i at the initial stage. For nodes i The k Sensor-like characteristics, This is the corresponding feature mapping matrix. The attention weights are normalized to reflect the relative importance of different sensor information to the current node state. This mechanism can effectively suppress the interference of noisy sensor information on the model's decision.
[0085] In the spatial dimension, the model employs an attention-based graph neural network to aggregate features of axle nodes and their adjacency relationships. By weighted fusion of information from adjacent axle nodes and global nodes, the model can learn the relative force relationships and cooperative change characteristics between axles. Its spatial feature update process can be represented as follows:
[0086] in, For nodes i The neighborhood group, Spatial attention weights, For learnable parameter matrix, It is a nonlinear activation function. This process can effectively characterize spatial anomalies such as edge compression and uneven stress.
[0087] In the time dimension, the model models the state sequence of the same axle over consecutive time slices, capturing the changing trends of features such as axle load, position, and speed over time through temporal convolution or recurrent units to identify persistent or abrupt abnormal behaviors such as scale dragging and scale jumping. The update form of the time features is represented as:
[0088] in, For nodes i At any moment t The temporal characteristics are represented.
[0089] When a vehicle drags, stops, or suddenly accelerates on the weighbridge... In v t , and w t The relevant dimensions will produce identifiable timing anomalies.
[0090] Finally, the node embedding is obtained through spatiotemporal joint fusion:
[0091] Where u represents the global vehicle node feature. Based on this embedding, the model obtains a unified representation that includes both single-axis force information and overall vehicle behavior constraints, providing a basis for subsequent decision-making.
[0092] Through the aforementioned STGNN core model layer, the spatial structure relationship, temporal evolution characteristics, and multi-source sensor information of the vehicle during the weighing process are uniformly mapped into a discriminable spatiotemporal feature representation, providing a key basis for the subsequent multi-level decision output layer to achieve highly reliable cheating behavior judgment.
[0093] S5. Multi-level decision output layer:
[0094] In the perfect tense After obtaining the spatiotemporal embedding, this layer does not need to be remodeled. Instead, within a unified spatiotemporal feature space, hierarchical joint reasoning is performed on single-axis and vehicle-level cheating behaviors, outputting structured decision results and confidence assessments. At the single-axis level, nodes are embedded into h. i Input the data into the classification network to obtain the probability distribution of this axis under various cheating modes:
[0095] in, W a and b a These are the parameters for a single-axis classification network.
[0096] As the vehicle travels along the weighing platform, the distance of the pressure edge in the spatiotemporal diagram node... d edge The axle load curve continues to decrease, while within the same time period... w t If the relative reference weight is systematically low, then p i The weight corresponding to "weight reduction by pressing the edge" gradually increases; if the effective weighing time is when passing through the center section of the scale, the effective weighing time is... T w Sharply shortened and accompanied by d v / d t A significant increase in the confidence threshold further reinforces the probability of the same node falling under the "scale skipping" category. To avoid false alarms under different environments and road conditions, an environment-adaptive confidence threshold is introduced:
[0097] in, The default threshold is used for judgment. Encode the environment; This is the environmental weighting coefficient. When... In this case, only the axis is marked as suspicious to ensure that the uniaxial anomaly is localized and not excessively propagated.
[0098] At the vehicle level, the embeddings of all axes are weighted and aggregated to form a vehicle-level representation:
[0099] And based on this, the overall vehicle cheating probability vector is obtained:
[0100] in, W v and b v These are the parameters for a single-axis classification network.
[0101] When the vehicle exhibits trajectory curvature within the weighbridge section The periodic amplification, and the multi-axis blanking distance d edge and w t When there is an alternating offset relationship, the component in q corresponding to "S-shaped driving" increases cumulatively over time; at the same time, when the total weight of the vehicle and the force distribution between the axles are inconsistent under dynamic constraints, the system will increase the confidence of the "weight transfer" category in the vehicle-level probability space, so that local anomalies of a single axle and structural anomalies of the whole vehicle can be explained simultaneously in a unified framework.
[0102] To identify complex cheating behaviors with superposition, collaboration, or temporal dependence characteristics, a spatiotemporal correlation matrix is introduced:
[0103] in, H axis This is a temporal embedding sequence for all axles, used to describe the temporal evolution of each axle's embedding. The "scale jump" at different time points leads to... T w Sudden drop, and simultaneous "edge pressing" caused d edge Continuous compression, and A xz When a highly coupled region is formed, the system no longer outputs two separate labels, but instead determines it as a complex cheating, thus avoiding misjudgments caused by simple superposition. This layer ultimately generates a structured output, which includes the anomaly type, confidence level, involved axle position, and time interval, and is consistent with the spatiotemporal graph structure of the previous layer, achieving multi-level consistent decision-making from local single axle to the whole vehicle, and from instantaneous behavior to temporal evolution.
[0104] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a multi-source sensing vehicle cheating detection method based on a spatiotemporal graph neural network as described above.
[0105] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a multi-source sensing vehicle cheating detection method based on a spatiotemporal graph neural network as described above.
[0106] The off-site overload control system is equipped with various sensors, including dynamic weighbridges, radar, cameras, and loop detectors, enabling comprehensive monitoring of key physical parameters such as vehicle speed, number of axles, mass, and dimensions. By integrating this sensor data and employing robust spatiotemporal graph neural networks for end-to-end learning, the accuracy of cheating behavior detection can be significantly improved. This method can effectively handle various complex cheating behaviors (including edge pressing, S-shaped driving, scale dragging, and scale jumping) and arbitrary combinations, and also greatly improves the overall stability and reliability of the system, providing stronger guarantees for fairness and safety in the transportation industry.
[0107] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A multi-source sensing vehicle cheating detection method based on spatiotemporal graph neural networks, characterized in that: A five-layer detection framework is constructed, which sequentially achieves vehicle cheating behavior detection through multi-source sensor data acquisition, data preprocessing and feature extraction, spatiotemporal graph construction, spatiotemporal graph neural network feature learning, and multi-level decision output. The specific steps include: S1. Multi-source sensor data acquisition: When a vehicle enters the dynamic weighbridge inspection area, the weighing process time start is triggered by the ground induction coil. Simultaneously, radar is used to obtain the vehicle's real-time driving trajectory, speed and acceleration information. The camera captures the spatial relationship between the vehicle's tires and the edge of the weighbridge and the lane lines. The dynamic weighbridge collects the axle weight signals of each axle of the vehicle to obtain full-process monitoring data of the vehicle from entering to leaving the inspection area. S2. Data Preprocessing and Feature Extraction: Time synchronization processing is performed on the multi-source sensor data collected by S1, and the data from each sensor are unified to the same time axis through timestamp alignment and interpolation mapping. Noise suppression and outlier removal were performed on the synchronized data. Single-source features such as tire lateral position, trajectory curvature, wheelbase and axle time difference, and edge pressure distance were extracted from the original data of dynamic weighbridge, radar, inductive loop, and camera, respectively. Spatial alignment is performed based on the radar coordinate system, and the data sampling frequency is unified through time interpolation. The data is then fused to obtain a multi-dimensional feature vector that includes the vehicle's spatial position, temporal evolution, and weighing status. The multidimensional feature vector obtained by fusion is: Among them, f t Let x be the node feature vector. t The lateral position of the tire, w t For axle load characteristics, v t For vehicle speed, d edge The distance between the pressure plates is Δt, and the time difference between the axes is Δt. For the trajectory curvature feature, f c The characteristics of the contact state between the tire and the weighing platform; S3. Spatiotemporal Graph Construction: The vehicle weighing process is modeled as a spatiotemporal graph that evolves dynamically over time, with each axle of the vehicle as a graph node and the node features being the multidimensional feature vectors output by S2. Spatial connections are established between adjacent axle nodes at the same time, and temporal connections are established between the same axle nodes in adjacent time slices. A global vehicle feature node is introduced, which contains information on vehicle length, total weight, maximum acceleration, environmental coding, and weight confidence. It is fully connected to all axle nodes to form a dynamic spatiotemporal graph that includes spatial and temporal dependencies. The calculation formula for the global vehicle feature nodes is as follows: Where u is the global vehicle feature node, L veh W is the length of the vehicle. total This refers to the total weight of the vehicle. For maximum acceleration, E env For environment coding, C w For weight confidence; Constructed vehicle spatiotemporal graph G t Represented as: Among them, V t Let E be the set of axle nodes. t U is a set of spatial and temporal edges. t These are global feature nodes; S4. Spatiotemporal graph neural network feature learning: The dynamic spatiotemporal graph constructed by S3 is input into the spatiotemporal graph neural network. First, the multi-source sensing features of the same axle node are adaptively fused through the attention weighting mechanism to obtain the initial embedding representation of the node. In the spatial dimension, spatial anomaly features are characterized by aggregating the features of axle nodes, their adjacent nodes, and global nodes through a graph neural network based on an attention mechanism. In the time dimension, temporal variation trends of features of the same axle node are captured through temporal convolution or recurrent units to identify persistent or abrupt temporal anomalous behaviors; The final spatiotemporal embedding features of nodes are obtained by fusing spatial and temporal features; S5. Multi-level decision output: Based on the final spatiotemporal embedding features of nodes, nodes are embedded into the input classification network at the single-axis level to obtain the probability distribution of a single axis under various cheating modes. An environment-adaptive confidence threshold is then introduced to filter suspicious anomalies on the single axis. The calculation formula is: in, The default threshold is used for judgment. Encoding for the environment, This is the environmental weighting coefficient; At the vehicle level, the embedded features of all axle nodes are weighted and aggregated to obtain a vehicle-level representation, which is then input into a classification network to obtain a vehicle cheating probability vector. Spatiotemporal correlation matrix analysis is introduced to analyze the temporal evolution coupling relationship of each axle embedding and identify complex cheating behaviors; The final output includes the anomaly type, confidence level, and structured detection results involving axis position and time interval.
2. The multi-source sensing vehicle cheating detection method based on spatiotemporal graph neural network according to claim 1, characterized in that: In S2, noise suppression is achieved by combining wavelet transform and sliding window filtering on the multi-channel voltage signals of the dynamic truck scale. Then, the lateral position of the tires on the scale platform is obtained by weighting the voltage signals based on the spatial distribution of the weighing sensors. The calculation formula is as follows: Where, x t Let V be the lateral position of the tire on the weighing platform at time t, i.e., the distance from the left edge of the weighing platform. i Let x be the voltage value of the i-th load cell. i It corresponds to its horizontal position coordinates.
3. The multi-source sensing vehicle cheating detection method based on spatiotemporal graph neural network according to claim 1, characterized in that, In S2, the raw point cloud data acquired by the radar is first filtered out and clustered to separate the background point cloud, extract the vehicle target point set, and obtain the vehicle position sequence through target tracking. Then, the trajectory curvature feature is extracted by differential calculation of the vehicle centroid trajectory. The calculation formula is as follows: in, Let x(t) and y(t) represent the trajectory curvature characteristics of the vehicle's lateral sway on the weighing platform, where x(t) and y(t) are the vehicle's horizontal and vertical coordinates on the road surface coordinate system. , The first derivative of position with respect to time. , It is the second derivative.
4. The multi-source sensing vehicle cheating detection method based on spatiotemporal graph neural network according to claim 1, characterized in that: In S2, the video data captured by the camera is first subjected to perspective correction to eliminate lens distortion. Then, the tire region is extracted based on the tire detection model, and the minimum Euclidean distance between the tire and the edge of the weighing platform is calculated to quantify the edge pressing behavior. The calculation formula is as follows: Where, d edge denoted as the edge distance, p is the position of the tire in the coordinate system of the weighing platform, and B is the set of points on the edge of the weighing platform.
5. The multi-source sensing vehicle cheating detection method based on spatiotemporal graph neural network according to claim 1, characterized in that: The formula for calculating the initial embedding representation of nodes in S4 is: Among them, h i (0) Let i be the temporal characteristics of node i at the initial stage. For the k-th type of sensor features at node i, This is the corresponding feature mapping matrix. This is for normalized attention weights.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps of the multi-source sensing vehicle cheating detection method based on spatiotemporal graph neural network according to any one of claims 1 to 5.
7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the steps of the multi-source sensing vehicle cheating detection method based on spatiotemporal graph neural network according to any one of claims 1 to 5.