A plant vehicle management and control method based on intelligent access control and unattended weighing

Abstract

This invention discloses a method for vehicle management in a factory area based on intelligent access control and unmanned weighing, belonging to the field of industrial intelligent management and control technology. The method includes: constructing a directed graph of the factory road network and collecting vehicle node traversal sequences to generate a vehicle spatiotemporal trajectory matrix; inputting the vehicle spatiotemporal trajectory matrix into a spatiotemporal graph convolutional network model, performing spatial graph convolution based on the road network topology and combining it with a temporal attention mechanism to output a probability vector of abnormal vehicle behavior; constructing a digital twin model of the factory area integrating GIS maps, weighing equipment status, and loading time, and combining it with a path optimization algorithm to generate a weighing equipment allocation sequence and a driving path sequence; controlling the opening and closing of access control gates based on the abnormal behavior probability vector and allocation sequence, and feeding back the execution results to update the digital twin model. This invention can achieve coordinated optimization of vehicle behavior prediction and scheduling control, improving traffic efficiency and safety management level.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent control technology, and in particular to a method for managing vehicles in a factory area based on intelligent access control and unmanned weighing. Background Technology

[0002] With the rapid development of industrial internet and smart logistics technologies, vehicle management in factory areas is gradually evolving from traditional manual dispatching and decentralized equipment management to an integrated approach of "intelligent access control + unmanned weighing + digital dispatching." Especially in mining, steel, chemical, and bulk material logistics scenarios, where vehicles frequently enter and exit, routes are complex, and weighing processes are frequent, traditional methods relying on manual registration, simple license plate recognition, and independent weighing systems are no longer sufficient to meet the demands of high concurrency, high security, and high-precision dispatching. While existing technologies have introduced license plate recognition, RFID, unmanned weighbridges, and simple route guidance systems, most remain at the level of "event-triggered control," lacking the ability to continuously model the entire vehicle behavior process and making it difficult to predict and proactively intervene in abnormal behaviors (such as detours, queue jumping, and illegal stops). Simultaneously, data silos between systems are prominent; access control, weighing dispatching, and route planning lack a unified spatiotemporal correlation modeling mechanism, leading to uneven resource allocation, low equipment utilization, and limited traffic efficiency. In addition, existing scheduling methods are mostly based on static rules or local optimization strategies, failing to combine the plant's road network topology with vehicle dynamic trajectories for global optimization, making it difficult to adapt to the real-time decision-making needs under complex operating conditions.

[0003] CN117972506A discloses a digital twin-driven unattended metering system. By constructing a physical entity layer, a virtual model layer, and a twin data layer, it achieves real-time mapping and data analysis of the operating status of weighing equipment, and performs prediction and early warning control based on the twin model. This method improves the automation and visualization level of the weighing system to some extent, but its core focus is on modeling the operating status of a single weighing device or metering system, lacking a unified modeling mechanism for the overall traffic flow and vehicle behavior in the factory area. In particular, it does not involve vehicle trajectory expression and abnormal behavior probability prediction based on road network structure, resulting in limited capabilities in multi-vehicle collaborative scheduling and risk management.

[0004] CN112731887B proposes a digital twin intelligent monitoring system for petrochemical loading and unloading lines. By constructing a digital twin of the loading and unloading line and combining it with neural network algorithms, it achieves equipment status monitoring, operation optimization, and fault prediction. This method enhances equipment-level collaboration and operation optimization capabilities; however, its focus remains on the loading and unloading equipment and process flow levels. It lacks in-depth modeling of the spatiotemporal evolution of vehicles within the plant's road network, fails to establish a linkage between vehicle behavior and access control and route scheduling, and does not introduce a probabilistic evaluation mechanism for abnormal behavior, making it difficult to achieve proactive identification and dynamic control of potential violations. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0006] In view of the common problems in existing vehicle management technologies in factory areas, such as the lack of ability to model the spatiotemporal behavior of vehicles based on road network topology, difficulty in timely prediction and intervention of abnormal behavior, disconnect between weighing resource scheduling and path planning, and insufficient intelligence of access control strategies, this invention is proposed.

[0007] Therefore, the problem to be solved by this invention is how to achieve structured modeling of vehicle running trajectories, probabilistic prediction of abnormal behavior, and collaborative optimization scheduling of weighing resources and traffic paths in complex factory environments.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a method for managing factory vehicles based on intelligent access control and unattended weighing, comprising, Using the road intersection nodes of the factory area GIS map as vertices and road segments as directed edges, a directed graph of the factory area road network is constructed, and the node passing sequence of each vehicle is collected through the collection unit to generate a vehicle spatiotemporal trajectory matrix. The vehicle spatiotemporal trajectory matrix is ​​input into a spatiotemporal graph convolutional network model, wherein the spatiotemporal graph convolutional network model performs spatial graph convolution operations based on the topology of the directed graph of the factory road network and performs weighted aggregation of the historical frame sequence of the vehicle spatiotemporal trajectory matrix through a temporal attention mechanism, and outputs the abnormal behavior probability vector of each vehicle within the prediction time window. Using the factory area GIS map, the real-time occupancy status of weighing equipment, and the remaining loading time at the loading dock as input parameters, a digital twin model of the factory area is constructed. The digital twin model of the factory area is then combined with a path optimization algorithm. Using the current position of each vehicle in the vehicle spatiotemporal trajectory matrix as a constraint, the weighing equipment allocation sequence and driving path sequence of each vehicle waiting to enter the site are output. According to the abnormal behavior probability vector and the weighing equipment allocation sequence, a passage control command is issued to the access control gate. When the abnormal probability of any vehicle in the abnormal behavior probability vector is greater than a preset probability threshold, a gate closure command is triggered for this vehicle, and the execution result of the weighing equipment allocation sequence is fed back to the factory area digital twin model to update the equipment occupancy status parameters in the factory area digital twin model.

[0009] Secondly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of the above-described method for managing factory vehicles based on intelligent access control and unattended weighing.

[0010] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the above-described method for managing factory vehicles based on intelligent access control and unattended weighing.

[0011] Compared with existing technologies, the advantages of this invention are as follows: By converting the factory area GIS map into a directed road network graph and generating a vehicle spatiotemporal trajectory matrix, a formal description and digital storage of the complex factory area road topology and vehicle historical trajectories are achieved; by constructing a spatiotemporal graph convolutional network model that integrates spatial graph convolution and temporal attention mechanisms, not only is the spatial neighborhood interaction characteristics of vehicles in the current road network accurately captured using the road network topology, but the key temporal information in the historical trajectory is also adaptively enhanced through the attention mechanism, thereby significantly improving the prediction accuracy and robustness of abnormal vehicle driving behavior; by constructing a digital twin model that combines real-time equipment status and vehicle position, and nesting it with a path optimization algorithm, using the current position of each vehicle as a dynamic constraint, the optimal weighing equipment allocation sequence and driving path sequence are output, realizing the dynamic allocation of global logistics resources. The algorithm effectively avoids equipment idling and road congestion by optimizing speed and path planning. By using the probability of abnormal behavior as a prerequisite for access control, it enables proactive interception of risky vehicles before they enter the site. Combined with the optimized allocation sequence, it issues departure instructions and feeds the execution results back to the digital twin model to update the equipment occupancy status, ensuring the accurate execution of control instructions and real-time self-calibration of model parameters. By adopting a path optimization algorithm based on simulated annealing and using the total vehicle waiting time as the optimization objective, it globally searches for the optimal equipment allocation scheme under multiple real-world constraints. This effectively solves the combined optimization problem under multiple vehicles, multiple devices, and multiple constraints, minimizing the overall queuing time of vehicles within the plant area. It provides rich and discriminative inputs for the abnormal behavior decoder, thereby comprehensively improving the intelligence, automation, and precision of vehicle management in the plant area. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of a factory vehicle management method based on intelligent access control and unmanned weighing. Detailed Implementation

[0013] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0014] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0015] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0016] As mentioned in the background section, existing unmanned weighing and factory vehicle management technologies generally suffer from the following problems: First, there is a lack of vehicle spatiotemporal trajectory modeling methods based on the factory road network topology, making it impossible to structurally express and deeply analyze vehicle behavior; second, there is a lack of abnormal behavior prediction mechanisms that integrate spatiotemporal features, making it difficult to achieve the transformation from "post-event processing" to "pre-event warning"; third, digital twin models are mostly limited to the equipment layer and fail to be deeply coupled with vehicle scheduling and route optimization, resulting in insufficient overall resource allocation efficiency; fourth, access control strategies are still mainly rule-driven, lacking a linkage mechanism with dynamic behavior analysis results. To address these problems, this invention provides a factory vehicle management method based on intelligent access control and unmanned weighing.

[0017] Reference Figure 1 , Figure 1 This is a flowchart illustrating a factory vehicle management method based on intelligent access control and unattended weighing, according to an embodiment of the present invention. Figure 1 As shown, a factory vehicle management method based on intelligent access control and unmanned weighing includes: S1: Using the road intersection nodes of the factory area GIS map as vertices and road segments as directed edges, construct a directed graph of the factory area road network, and collect the node passing sequence of each vehicle through the collection unit to generate a vehicle spatiotemporal trajectory matrix. S1.1: Extract the spatial topological relationship between road intersection nodes and road segments from the factory area GIS map, map the road intersection nodes as a set of vertices, and map the road segments as a set of directed edges according to the actual traffic direction of the factory area roads, and construct a directed graph of the factory area road network; Specifically, the spatial coordinates of road intersection nodes are read from the vector layer of the factory area GIS map. These intersection nodes are then labeled with node type tags according to their physical function, forming a vertex set. The start and end coordinates and traffic direction attributes of each road segment in the factory area GIS map are read. The start and end coordinates of each road segment are spatially matched with the corresponding road intersection nodes in the vertex set. Each road segment is converted into a directed edge pointing from the starting road intersection node to the ending road intersection node according to the actual traffic direction attributes of the factory roads. All directed edges form a directed edge set. Using the physical length of the road segment corresponding to each directed edge as the edge weight, the vertex set, the directed edge set, and the corresponding edge weights are combined to obtain a directed graph of the factory area road network. Preferably, the node type labels include access control entrance node type labels, weighing equipment node type labels, loading port node type labels, and road segment branch node type labels; based on the directed connection relationship between each vertex in the directed graph of the factory road network, a directed adjacency relationship table between vertices is constructed, which records the existence status of the directed edge between each pair of adjacent vertices and the corresponding edge weight; the direction of the directed edge in the directed graph of the factory road network is consistent with the one-way traffic direction of the factory roads; Furthermore, for road segments with slopes in the factory area GIS map, the slope angle of the road segment is introduced into the edge weight correction calculation. The corrected edge weight is obtained by dividing the horizontal projection length of the road segment by the cosine of the slope angle, so that the edge weight of each directed edge in the directed edge set reflects the actual driving cost of the heavy-duty tanker truck on the corresponding road segment. S1.2: Set up data acquisition units at the physical locations corresponding to each vertex of the directed graph of the factory road network to collect the sequence of each vehicle passing through the nodes of each vertex in the directed graph of the factory road network. Furthermore, a set of acquisition units is deployed at the physical location corresponding to each road intersection node in the vertex set. Each acquisition unit consists of three types of sub-acquisition units: a license plate recognition camera, a QR code scanner, and an inductive loop. The license plate recognition camera continuously acquires images of passing vehicles at a frame rate of no less than 25 frames per second. The QR code scanner reads the delivery code string held by the vehicle driver in a polling manner. The inductive loop outputs a trigger pulse signal when a vehicle enters the area. The trigger pulse signal of the inductive loop serves as the hard trigger signal source for the license plate recognition camera and the QR code scanner, synchronously activating image acquisition and QR code scanning when the vehicle arrives within the coverage area of ​​the acquisition unit, eliminating acquisition timing jitter caused by differences in vehicle speed. Specifically, clock synchronization calibration is performed on each acquisition unit, aligning the system clock of all acquisition units with the unified time server of the plant area, requiring the clock deviation to be no more than 10 milliseconds, and storing the vertex number, spatial coordinates and clock deviation parameters of each acquisition unit after calibration into the acquisition unit parameter library. Furthermore, when a vehicle arrives at the collection unit, the collection unit collects the vehicle's identity information and compares it with the qualified vehicle information database. If the comparison matches, a node passing record is generated. The local collection timestamp in the node passing record is then corrected according to the collection unit's parameter library. The corrected node passing record is then appended to the node passing sequence buffer. Specifically, this includes: When a vehicle reaches the physical location corresponding to a vertex in the vertex set, the inductive loop outputs a trigger pulse signal, simultaneously activating the license plate recognition camera and the QR code scanner. The license plate recognition camera captures images of the passing vehicle and uses a license plate recognition algorithm to extract the license plate string from the image, while the QR code scanner reads the delivery code string. The license plate string is compared with the vehicle information stored in the qualified vehicle information database. If they match, a node passage record is made. The node traversal record contains three fields: vehicle number, current vertex number, and local acquisition timestamp. Based on the clock offset parameter corresponding to the current vertex number in the acquisition unit parameter library, the local acquisition timestamp is subtracted from the corresponding clock offset parameter to obtain the alignment timestamp. The alignment timestamp replaces the local acquisition timestamp in the node traversal record, resulting in a corrected node traversal record. This corrected record is then appended to the node traversal sequence buffer, where the node traversal sequence buffer is grouped and stored by vehicle number. Each vehicle corresponds to an ordered corrected node traversal sequence arranged in ascending order of the alignment timestamp. The node traversal sequence records the vertex number and corresponding traversal timestamp of each vertex traversed by the vehicle in the directed graph of the factory road network. The system judges the difference in alignment timestamps of the two adjacent corrected nodes of the same vehicle at the same vertex. If the difference in alignment timestamps is less than the preset de-jitter threshold, only the corrected node with the smaller alignment timestamp is retained, and the repeated triggering noise caused by the low-speed passage of the inductive loop is filtered out. The preset de-jitter threshold is 3 seconds. S1.3: Based on the distribution of the node traversal sequence on the discrete time axis, store the historical position status of each vehicle row by row to generate the vehicle spatiotemporal trajectory matrix; S1.3.1: Validate the node transit sequence for each vehicle, identify node jump anomalies and speed anomalies, and store the anomaly records in the node transit sequence database after marking them. Specifically, this includes: Read the vertex number pairs and corresponding alignment timestamp differences of the two adjacent corrected node passage records in the passage sequence of each vehicle node; query the directed adjacency table between vertices to determine whether there is a directed edge between the previous vertex number and the next vertex number in the vertex number pair. If there is no directed edge, the adjacent record pair is determined to be a node jump anomaly, and the node jump flag is triggered for the subsequent corrected node passage record corresponding to the node jump anomaly; at the same time, calculate the average driving speed corresponding to this adjacent record pair based on the edge weight of the corresponding directed edge and the alignment timestamp difference in the directed adjacency table between vertices; if the average driving speed exceeds the factory area speed limit, the speed anomaly flag is triggered for the corresponding corrected node passage record; mark the corrected node passage records that trigger the node jump flag or speed anomaly flag, and store the marked vehicle node passage sequences into the node passage sequence database. It should be noted that the maximum speed limit in the factory area is set at 20 kilometers per hour; the node jump flag indicates that the vehicle passes through a road segment in the directed graph of the factory road network that is not directly connected by a directed edge. This situation indicates that the vehicle has left the designated road segment or that the data collection unit has missed data collection.

[0018] S1.3.2: Construct a discrete time axis with a preset time step interval, and interpolate the vertex numbers of each vehicle's marked node's traversal sequence between adjacent records according to the corresponding shortest path in the directed graph of the factory road network, to obtain the complete vertex number sequence of each vehicle on the discrete time axis, specifically including: Starting with the earliest entry alignment timestamp of all vehicles in the sequence database, and with a preset time step interval, each time step is arranged sequentially to construct a discrete time axis. For each vehicle, after marking the node, the preceding and following vertex numbers corresponding to the two adjacent corrected node passage records in the sequence are used. The shortest path search algorithm is performed on the directed adjacency table between vertices to obtain a list of vertices with the shortest path from the preceding vertex number to the following vertex number. For each time step on the discrete time axis that falls within the alignment timestamp interval of the two adjacent corrected node passage records, the time proportion of that time step within the alignment timestamp interval is calculated. Based on the time proportion, the vertex number at the corresponding position in the shortest path vertex list is taken as the interpolated vertex number for that time step. The interpolated vertex numbers of each time step are arranged in the order of the discrete time axis to obtain the complete vertex number sequence of this vehicle on the discrete time axis. For the time steps corresponding to the vehicle before entering the factory and after leaving the factory, the corresponding positions in the complete vertex number sequence are set to null placeholders. It should be noted that the preset time step is 30 seconds; the shortest path search algorithm is executed on the basis of the directed adjacency table between vertices at the cost of edge weights, thereby ensuring the continuity of the complete vertex number sequence in the time dimension and eliminating the interference of time step gaps caused by missed sampling by the acquisition unit on subsequent matrix operations. S1.3.3: Concatenate the complete vertex number sequence of all vehicles present row by row, and combine the node jump flag and speed anomaly flag to construct the vehicle spatiotemporal trajectory matrix; Furthermore, the complete vertex number sequence of all vehicles present in the factory area is concatenated row by row according to the vehicle number order to obtain a two-dimensional array with the vehicle number as the row index and the discrete time axis time step as the column index. This two-dimensional array is the vehicle spatiotemporal trajectory matrix. For the elements in the vehicle spatiotemporal trajectory matrix that are set to null placeholders, a mask matrix of the same size as the vehicle spatiotemporal trajectory matrix is ​​constructed. Preferably, the jump flag sequence and speed anomaly flag sequence of each vehicle node are aligned according to the same row and column indices as the vehicle spatiotemporal trajectory matrix to form a node jump flag matrix and a speed anomaly flag matrix of the same size as the vehicle spatiotemporal trajectory matrix; the vehicle spatiotemporal trajectory matrix, the node jump flag matrix, and the speed anomaly flag matrix are concatenated along the feature channel dimension to obtain the augmented spatiotemporal trajectory matrix; the augmented spatiotemporal trajectory matrix and the mask matrix are stored together in the spatiotemporal database, and the augmented spatiotemporal trajectory matrix and the mask matrix are used as input parameters for step S2, which are input to the spatiotemporal graph convolutional network model. The spatiotemporal graph convolutional network model, combined with the spatial topology defined by the directed adjacency relation table between vertices in step S1.1, performs spatial graph convolution operation and temporal attention mechanism weighted aggregation on the augmented spatiotemporal trajectory matrix, and outputs the probability vector of abnormal behavior of each vehicle within the prediction time window; It should be noted that each element in the vehicle spatiotemporal trajectory matrix stores the vertex number of the corresponding vehicle at the corresponding time step; the element at the null placeholder position in the mask matrix has a value of zero, and the element at the valid vertex number position has a value of one; the position with a value of zero in the mask matrix is ​​marked as an invalid time step; the augmented spatiotemporal trajectory matrix contains, in the feature channel dimension, a vertex number channel, a node jump flag channel, and a speed anomaly flag channel; the row index of the vehicle spatiotemporal trajectory matrix corresponds to the vehicle number, the column index corresponds to the discrete time step, and the matrix element is the vertex number of the vertex where the vehicle is located within the corresponding time step; time steps not within the coverage area of ​​the acquisition unit are occupied by null values; S2: Input the vehicle spatiotemporal trajectory matrix into the spatiotemporal graph convolutional network model, where the spatiotemporal graph convolutional network model performs spatial graph convolution operation based on the topology of the directed graph of the factory road network and performs weighted aggregation of the historical frame sequence of the vehicle spatiotemporal trajectory matrix through a temporal attention mechanism, and outputs the probability vector of abnormal behavior of each vehicle within the prediction time window. It should be noted that the spatiotemporal graph convolutional network model uses the mask matrix as the effective frame mask. During the calculation of the attention weight matrix in the temporal attention mechanism, invalid time steps corresponding to null placeholders in the augmented spatiotemporal trajectory matrix are masked to ensure that the attention weight matrix is ​​normalized only within the effective time steps. The dimension of the abnormal behavior probability vector is consistent with the number of discrete time steps within the prediction time window. The value of each dimension in the vector represents the probability estimate of the abnormal behavior of this vehicle within the corresponding time step. S2.1: Convert the discrete vertex number sequence in the augmented spatiotemporal trajectory matrix into node feature embedding vectors, perform spatial graph convolution operation using the normalized Laplacian matrix of the adjacency matrix as the graph structure operator, extract the spatial neighborhood context features of the current position of each vehicle, and obtain the spatial feature map; S2.1.1 converts the vertex numbers and additional feature channels of each vehicle at each time step in the augmented spatiotemporal trajectory matrix into node feature embedding vectors, which serve as the input feature representation for the spatiotemporal graph convolutional network model. Specifically, this includes: Using the size of the vertex set as the size of the embedding vocabulary, a learnable vertex embedding vector is initialized for each vertex number. The table is looked up from the augmented spatiotemporal trajectory matrix according to the vehicle number and time step index. The vertex number corresponding to each vehicle at each time step is replaced with the corresponding vertex embedding vector to obtain the vertex number embedding tensor. The vertex embedding vector is set to 64 dimensions; the shape of the vertex number embedding tensor is the number of vehicles multiplied by the total number of time steps multiplied by 64 dimensions; the vertex embedding vector is initialized using a random normal distribution, and the standard deviation is set to the square root of the inverse of the vertex set size to control the gradient scale consistency among the vertex embedding vectors during the initialization phase; for the time steps corresponding to the null placeholders in the augmented spatiotemporal trajectory matrix, the vertex embedding vectors at the corresponding positions are all set to zero, so that the masking operation of the mask matrix in subsequent calculations remains numerically consistent with the vertex number embedding tensor; Read the node type labels of each vertex stored in the vertex set, encode the node type labels with a four-dimensional one-hot vector to obtain the node type encoding vector, and project the node type encoding vector to sixty-four dimensions through a linear transformation to obtain the node type feature vector; fuse the node type feature vector with the vertex number embedding tensor element-wise in the feature dimension to obtain the initial node feature tensor. The node type labels include four categories: access control entrance node labels, weighing equipment node labels, loading port node labels, and road segment branch node labels. The purpose of introducing node type feature vectors is to enable the spatiotemporal graph convolutional network model to distinguish the semantic differences of different functional nodes during the spatial graph convolution operation, thereby giving higher anomaly sensitivity to the lingering behavior of vehicles near weighing equipment nodes and loading port nodes. Read the node jump flag and velocity anomaly flag stored in the additional feature channel of the augmented spatiotemporal trajectory matrix. Project the two types of flags into the 64-dimensional dimension through a linear transformation in the form of two-dimensional feature vectors to obtain the anomaly prior feature vector. Concatenate the anomaly prior feature vector with the initial node feature tensor in the feature dimension and compress it back to the 64-dimensional dimension through a linear transformation to obtain the fused node feature tensor. The fusion node feature tensor serves as the final input node feature for each vehicle at each time step in the spatiotemporal graph convolutional network model; the anomaly prior feature vector injects the detected local anomaly signals into the spatiotemporal graph convolutional network model in advance, so that the spatiotemporal graph convolutional network model has prior perception of local anomaly regions before performing graph convolution operations, thereby reducing the false negative rate of the anomaly behavior probability vector. S2.1.2: Based on the adjacency matrix, calculate the normalized Laplacian matrix of the directed graph of the factory road network, which serves as the graph structure operator for performing spatial graph convolution operations in the spatiotemporal graph convolutional network model. Specifically, this includes: Summing each row of the adjacency matrix yields the out-degree value of each vertex. Arranging all out-degree values ​​on the diagonal forms an out-degree diagonal matrix. Summing each column of the adjacency matrix yields the in-degree value of each vertex. Arranging all in-degree values ​​on the diagonal forms an in-degree diagonal matrix. For vertices in both the out-degree and in-degree diagonal matrices where the diagonal element is zero, the corresponding diagonal element is replaced with one to prevent division by zero errors in subsequent normalization operations. Multiply the adjacency matrix by the negative 1 / 2 power of the out-degree diagonal matrix on the left, and then multiply it by the negative 1 / 2 power of the in-degree diagonal matrix on the right to obtain the forward normalized adjacency matrix. Multiply the transpose of the adjacency matrix by the negative 1 / 2 power of the in-degree diagonal matrix on the left, and then multiply it by the negative 1 / 2 power of the out-degree diagonal matrix on the right to obtain the reverse normalized adjacency matrix. Summate the forward normalized adjacency matrix and the reverse normalized adjacency matrix and multiply by 1 / 2 to obtain the symmetric normalized adjacency matrix. Subtract the symmetric normalized adjacency matrix from the identity matrix to obtain the normalized Laplace matrix. The purpose of distinguishing between out-degree and in-degree diagonal matrices is to preserve the traffic flow direction information implied by the one-way traffic direction of roads in the directed graph of the factory road network. Compared with the undirected Laplace matrix, the directed normalized Laplace matrix can more accurately model the difference between forward and reverse driving of vehicles on one-way roads in the factory area, providing direction awareness for spatial graph convolution operations; the eigenvalues ​​of the normalized Laplace matrix range from zero to two. Using the normalized Laplace matrix as input, a sequence of Chebyshev polynomial expansion matrices of orders zero to K is calculated according to the Chebyshev recursive formula. The zero-order polynomial expansion matrix is ​​an identity matrix of the same size as the normalized Laplace matrix, the first-order polynomial expansion matrix is ​​a scaled Laplace matrix after linear scaling of the normalized Laplace matrix to the range of negative one to positive one, and the higher-order polynomial expansion matrices are calculated sequentially from the first two-order polynomial expansion matrices according to the recursive relationship. The K-order is set to three, resulting in four Chebyshev polynomial expansion matrices of orders zero, first, second, and third. The sequence of Chebyshev polynomial expansion matrices is stored in the graph basis cache. The K-order Chebyshev polynomial approximation limits the receptive field of graph convolution to the local neighborhood of K hops. When K is three, it corresponds to the spatial context of a vehicle within three adjacent road segments, which matches the typical number of road segment hops for a vehicle in the factory area from the access control node to the weighing equipment node. This not only covers the spatial neighborhood of key functional nodes, but also avoids introducing noise features from distant irrelevant nodes by using an excessively large K value. S2.1.3: Perform spatial graph convolution operation to extract spatial neighborhood context features; S2.1.3.1: Based on the Chebyshev polynomial expansion matrix sequence, spatial graph convolution operation is performed on the fused node feature tensor to extract the spatial neighborhood context features of each vehicle's current position, specifically including: The fused node feature tensor is expanded frame by frame along the time step dimension. For each time step, the node feature matrix is ​​multiplied sequentially with the zeroth, first, second, and third order polynomial expansion matrices in the Chebyshev polynomial expansion matrix sequence to obtain four sets of neighborhood aggregation feature matrices from zero to third order. The zeroth order neighborhood aggregation feature matrix corresponds to the retained terms of each vehicle's own features; the first to third order neighborhood aggregation feature matrices correspond to the feature propagation terms of one-hop, two-hop, and three-hop neighbor nodes, respectively. The above four sets of neighborhood aggregation feature matrices are concatenated along the feature dimension to obtain a multi-order neighborhood concatenated feature matrix. The feature dimension of the multi-order neighborhood splicing feature matrix is ​​256-dimensional; the multi-order neighborhood aggregation feature matrix uses the Chebyshev polynomial expansion matrix sequence as the graph structure operator, which does not require eigenvalue decomposition of the normalized Laplace matrix during forward propagation, thereby reducing the time complexity of spatial graph convolution operation from being proportional to the square of the vertex set size to being proportional to the number of non-zero elements in the adjacency matrix, which is suitable for real-time inference needs under the scale of factory road network. S2.1.3.2: Input the temporal context feature vector into the spatiotemporal feature fusion module, and extract high-order spatiotemporal interaction features through a multi-layer stacked feedforward network and residual connection structure to obtain a spatial feature map, specifically including: The multi-order neighborhood splicing feature matrix is ​​projected to 128 dimensions through a linear transformation with a bias term to obtain the spatial projection feature matrix. A linear correction unit activation function is applied element-wise to the spatial projection feature matrix. The activation result is then added element-wise to the residual term of the fused node feature tensor after linear transformation to 128 dimensions to obtain the spatial residual feature matrix. A layer normalization operation is applied to the spatial residual feature matrix to obtain the spatial normalized feature matrix. The spatial normalized feature matrices of each time step are then re-stacked in time-step order to obtain the spatial feature map. It should be noted that the shape of the spatial feature map is the number of vehicles multiplied by the total number of time steps multiplied by 128 dimensions; the introduction of the residual term makes the local features of the vehicles themselves in the fusion node feature tensor explicitly preserved after spatial graph convolution aggregation, preventing the features of each vehicle node from becoming overly smoothed to the neighboring nodes due to multi-order neighborhood aggregation, thereby maintaining the distinguishability of spatial features between different vehicles. S2.2: Input the historical frame sequence of spatial feature maps into the temporal attention module. The features at each time step in the historical frame sequence are weighted and aggregated using an attention weight matrix to obtain the temporal context feature vector, specifically including: Using the time step index of the spatial feature map as input, calculate the time position encoding vector corresponding to each time step according to the sine-cosine alternating encoding method; add the time position encoding vector to the spatial feature map element by element along the feature dimension to obtain the time encoding feature map; The temporal position encoding vector has the same dimension as the spatial feature map, both being 128-dimensional. The temporal encoding feature map preserves the absolute position information of each time step on the time axis, enabling the temporal attention module to perceive the sequential interval relationship between time steps when calculating the attention weight matrix. Odd-numbered dimensions in the temporal position encoding vector are encoded using a sine function, while even-numbered dimensions are encoded using a cosine function. The encoding frequency decreases exponentially with the increase of the dimension number, resulting in distinguishable periodic differences in different dimensions of the temporal position encoding vector for different time intervals. The feature vector corresponding to the last frame of the time-encoded feature map in the time step dimension is taken as the query source at the current time moment, and a query matrix is ​​obtained through linear transformation; the feature matrix of the time-encoded feature map over all historical time steps is taken through linear transformation to obtain the key matrix; the feature matrix of the time-encoded feature map over all historical time steps is taken through another independent linear transformation to obtain the value matrix; the weight matrices of the above three sets of linear transformations are independent of each other and are used for feature space projection of the query matrix, key matrix and value matrix, respectively, and the weight matrices of the three sets of linear transformations are jointly updated through gradient descent during model training; The feature dimension of the query matrix is ​​compressed to 64 dimensions; the feature dimension of the key matrix is ​​also compressed to 64 dimensions; the feature dimension of the value matrix remains at 128 dimensions; the query matrix is ​​generated only from the features at the current moment, reflecting the operation logic of the temporal attention module to backtrack and query the historical behavior sequence with the current vehicle status as the anchor point; the key matrix and the value matrix cover all historical time steps, and respectively undertake the two functions of similarity indexing and weighted content extraction of historical frames; Perform matrix multiplication on the transpose of the query matrix and the key matrix, and divide the result by the square root of the feature dimension value of the key matrix to obtain the scaled attention score matrix; read the mask matrix, fill the positions in the scaled attention score matrix corresponding to the mask matrix with negative infinity values ​​to obtain the masked attention score matrix; apply the Softmax normalization function to the masked attention score matrix along the historical time step dimension to obtain the attention weight matrix; The scaling attention score matrix is ​​shaped by the number of vehicles multiplied by the number of historical time steps; the sum of the elements in each row of the attention weight matrix is ​​one, and the weight values ​​corresponding to the original mask positions are close to zero; the scaling attention score matrix is ​​filled with negative infinity values ​​instead of being set to zero directly, so that the attention weights corresponding to invalid time steps after normalization by the Softmax function are close to zero rather than exactly equal to zero, thereby maintaining numerical stability during gradient backpropagation and avoiding the stagnation of attention weight matrix parameter updates due to gradient vanishing. The attention weight matrix and the value matrix are multiplied to obtain a single-head temporal weighted feature vector. Eight sets of query matrices, key matrices, and value matrices are generated independently using the same linear transformation method. The above weighting and aggregation operation is performed on each set to obtain eight sets of single-head temporal weighted feature vectors. The eight sets of single-head temporal weighted feature vectors are concatenated along the feature dimension to obtain a multi-head concatenated feature vector. The multi-head concatenated feature vector is projected back to 128 dimensions through a linear transformation, added to the residual term of the feature vector of the last frame of the temporal encoded feature map, and then normalized by a layer to obtain the temporal context feature vector. The single-head temporal weighted feature vector has a dimension of 128; the multi-head concatenated feature vector has a dimension of 1024; the eight-head attention mechanism enables the temporal attention module to independently capture multiple temporal dependency patterns in historical behavior sequences from eight different feature subspaces, including vehicle dwell time patterns at weighing equipment nodes, abnormal round-trip paths, and detour patterns. The parallel feature subspace of multi-head attention has a stronger ability to distinguish temporal abnormal patterns than single-head attention. S2.3: After extracting the temporal context feature vector through the multi-layer spatiotemporal feature fusion module, the input is given to the abnormal behavior decoder, and the output is the probability vector of abnormal behavior of each vehicle within the prediction time window. S2.3.1: High-order spatiotemporal interaction features are extracted through multi-layered stacked spatiotemporal fusion blocks to obtain a deep spatiotemporal fusion feature vector, specifically including: The temporal context feature vector is input into a position-wise feedforward network consisting of two linear transformation layers. The first linear transformation layer increases the feature dimension from 128 to 512 and applies a Gaussian error linear unit activation function. The second linear transformation layer compresses the feature dimension back to 128. The output of the position-wise feedforward network is added element-wise to the residual term of the temporal context feature vector and layer normalization is applied to obtain the feedforward normalized feature vector. The dimensionality increase factor of the position-wise feedforward network is set to four times, which is consistent with the standard Transformer architecture and achieves a balance between the number of model parameters and the feature representation capability. The Gaussian error linear unit activation function has a non-zero gradient in the negative region compared with the linear correction unit activation function, and is more sensitive to the negative region in the time series features that represent the low speed or stationary state of the vehicle. A spatiotemporal fusion block is formed by sequentially connecting the spatial graph convolution operation module, the temporal attention module, and the position-by-position feedforward network. Three spatiotemporal fusion blocks are stacked in the same structure. The input of each spatiotemporal fusion block is the feedforward normalized feature vector of the output of the previous layer. The input of the first spatiotemporal fusion block is the feature tensor of the fusion node, and the output of the third spatiotemporal fusion block is the deep spatiotemporal fusion feature vector, where the feature dimension of the deep spatiotemporal fusion feature vector is 128 dimensions.

[0019] The three-layer stacked spatiotemporal fusion block expands the receptive field layer by layer, so that the bottom spatiotemporal fusion block focuses on extracting local spatial anomalies of the vehicle in a single-hop neighborhood, the middle spatiotemporal fusion block focuses on the short-term behavior patterns of the vehicle in several time steps, and the top spatiotemporal fusion block captures global behavior anomalies of the vehicle throughout the entire presence period. The spatiotemporal features of the three granularities are hierarchically encoded in the deep spatiotemporal fusion feature vector. S2.3.2: The deep spatiotemporal fusion feature vector is input into the abnormal behavior decoder, mapped by a multilayer perceptron and a sigmoid activation function, and outputs the abnormal behavior probability vector of each vehicle within the prediction time window. This is then combined with a preset probability threshold to complete the anomaly determination, specifically including: The deep spatiotemporal fusion feature vector is input into a multilayer perceptron consisting of three linear transformations. The output dimensions of each layer are 64-dimensional, 32-dimensional, and the number of prediction time window steps, respectively. A linear correction unit activation function and a random dropout operation with a probability of 0.1 are applied between every two linear transformations. The final output of the multilayer perceptron is an anomaly score matrix whose shape is the number of vehicles multiplied by the number of prediction time window steps. Each row in the anomaly score matrix corresponds to the original anomaly score of a vehicle at each time step within the prediction time window. The three-layer decreasing dimension multilayer perceptron structure gradually compresses the high-dimensional semantic features in the 128-dimensional deep spatiotemporal fusion feature vector to a low-dimensional output corresponding to the number of prediction time window steps. At the same time, the Dropout random discard operation between the two linear transformations introduces regularization constraints during the training phase to suppress the overfitting of the model to specific license plate numbers or specific time period features in the training set. Apply the Sigmoid activation function to each element of the anomaly score matrix to map the range of values ​​of each element in the anomaly score matrix from negative infinity to positive infinity to the interval between zero and one, and obtain the anomaly behavior probability matrix; take the numerical sequence of the row corresponding to each vehicle in the anomaly behavior probability matrix as the anomaly behavior probability vector of this vehicle. The numerical values ​​of each element in the abnormal behavior probability vector represent the estimated probability of the vehicle exhibiting abnormal behavior within the corresponding prediction time step. The abnormal behavior probability vector, along with the corresponding vehicle number, is stored in the abnormal probability database. Abnormal behaviors include: vehicle circling the weighing station without being weighed before the weighing equipment node; vehicle returning to the factory entrance after completing the first weighing and undergoing repeated weighing; and abnormal lingering behavior of vehicle moving at low speed for an extended period beyond the preset time limit at the loading station node. The node jump flag and speed anomaly flag have been injected into the spatiotemporal graph convolutional network model through the abnormal prior feature vector. The aforementioned prior signals play an enhanced role in early warning during the generation of the abnormal behavior probability vector, enabling the spatiotemporal graph convolutional network model to detect the above three typical abnormal behaviors at a higher rate than the baseline model without prior signals. S3: Using the plant area GIS map, the real-time occupancy status of weighing equipment, and the remaining loading time at the loading dock as input parameters, construct a digital twin model of the plant area. Combine the digital twin model of the plant area with the path optimization algorithm. Using the current position of each vehicle in the vehicle spatiotemporal trajectory matrix as a constraint, output the weighing equipment allocation sequence and driving path sequence of each vehicle waiting to enter the plant. Preferably, the update frequency of the equipment layer status in the digital twin model of the plant area is consistent with the data reporting cycle of the weighing equipment controller, which is set to five seconds; the digital twin model of the plant area and the path optimization algorithm communicate unidirectionally through the equipment status prediction sequence; the solution result of the path optimization algorithm is fed back to the digital twin model of the plant area in a closed loop through the execution result of the weighing equipment allocation sequence. S3.1: The real-time occupancy status of the weighing equipment and the remaining loading time at the loading port are collected from the weighing equipment controller and the loading port material discharge control system. Combined with the topology of the directed graph of the plant road network and the attributes of each node, the three dynamic states of the plant equipment layer, road layer and vehicle layer are synchronously mapped in the plant digital twin model. Specifically, the raw data of the real-time occupancy status of each weighing device is collected from the weighing device controller, and outlier filtering and status encoding are performed on the raw data of the real-time occupancy status to obtain the weighing device status vector, which serves as the initial input of the equipment layer of the digital twin model of the plant area. Furthermore, through the RS-485 communication link established with the weighing equipment controller, each weighing equipment controller is polled every five seconds to read the current occupancy flag, the current vehicle number on the scale, and the cumulative weighing time of the current vehicle on the scale. When the occupancy flag is zero, it indicates that the weighing equipment is in an idle state, and when it is one, it indicates that the weighing equipment is in an occupied state. The above three data items, together with the collection timestamp, are combined into a real-time status record of a single weighing equipment and written into the weighing equipment status buffer queue according to the weighing equipment number. It should be noted that the weighing equipment status buffer queue retains historical status records within the most recent twelve collection cycles, that is, it retains historical data within the most recent sixty seconds; the historical window length of the weighing equipment status buffer queue is set to sixty seconds, which covers the longest expected time for a single vehicle to complete an unloaded weighing operation within the factory area, so that the factory digital twin model can obtain a complete sequence of equipment status evolution within a single weighing cycle when predicting the equipment status at the equipment level. Furthermore, real-time occupancy status anomaly filtering: Consistency verification is performed on the occupancy flag sequence of each weighing device in the weighing equipment status buffer queue for the most recent twelve acquisition cycles. If the occupancy flag of a weighing device changes from one to zero and then immediately jumps back to one within two consecutive acquisition cycles, and the cumulative weighing duration has not been reset within the aforementioned three cycles, then this occupancy flag jump is determined to be a noise spike caused by communication interference. The occupancy flag in the intermediate cycle is corrected to one, and the original record is overwritten in the weighing equipment status buffer queue. For the cumulative weighing duration data, the duration increment between two adjacent acquisition cycles is calculated. If the duration increment is negative and its absolute value is greater than 1.5 times the acquisition cycle duration, it is determined to be an abnormal timer reset. The cumulative weighing duration of the corresponding acquisition cycle is marked as invalid, and the cumulative weighing duration of the previous valid cycle plus the acquisition cycle duration is used as a substitute estimate to fill the gap. It should be noted that the above-mentioned noise filtering operation is completed before the data enters the digital twin model of the plant area to prevent false idle states caused by communication interference from triggering unnecessary weighing equipment allocation and scheduling, and to reduce the number of invalid scheduling instructions caused by data noise in the weighing equipment allocation sequence. Specifically, the status records of each filtered weighing device are numerically encoded: the occupancy flag is directly used as a binary feature; the cumulative weighing time is divided by the maximum rated weighing time for a single weighing session set by the plant (120 seconds) to obtain a normalized weighing progress value, ranging from zero to one; the normalized weighing progress value is concatenated with the occupancy flag to obtain the weighing device status vector of a single weighing device, which is two-dimensional; the weighing device status vectors of all weighing devices in the plant are arranged in order of weighing device number to obtain a weighing device status matrix, the shape of which is the number of weighing devices multiplied by two dimensions. Furthermore, the remaining loading time of each loading port is collected from the loading port unloading control system, and combined with the weighing equipment status matrix, the occupancy status changes of each weighing device and each loading port in the future time window are predicted by the equipment status prediction model, resulting in an equipment status prediction sequence, specifically including: Through the ModbusTCP communication link established with the loading port material feeding control system, the PLC controller of each loading port is polled every five seconds to read the current material feeding valve opening flag, the current loading vehicle number, the current cumulative loading weight, and the preset loading target weight of each loading port. The remaining loading weight is obtained by subtracting the current cumulative loading weight from the preset loading target weight. The remaining loading weight is obtained by dividing the remaining loading weight by the historical average material feeding rate of the loading port (which is statistically obtained from the historical records of the same type of loading port in the weighing equipment status buffer queue). The remaining loading time of each loading port is then written into the loading port status buffer queue according to the loading port number. Historical average feeding rates are statistically analyzed according to loading port number. The average ratio of net weight to feeding time of the most recent 20 loading operations at the corresponding loading port in the weighing equipment status buffer queue is taken to adapt to the differences in feeding rates caused by the different specifications of bulk loading machines at different loading ports. The remaining feeding time at the loading port is used as the initial constraint condition of the equipment status prediction sequence, and is input into the equipment status prediction model together with the weighing equipment status matrix. Using the weighing equipment state matrix and the remaining loading time of each loading port as input, an equipment state prediction model is constructed. The equipment state prediction model consists of a gated cyclic unit network and a linear output layer. The hidden state dimension of the gated cyclic unit network is 32 dimensions, and the input dimension is the number of weighing equipment multiplied by the sum of the number of loading ports in the two dimensions. The historical state records of the last 12 acquisition cycles in the weighing equipment state buffer queue are used as time series input to drive the gated cyclic unit network to gradually update the hidden state. The linear output layer maps the hidden state vector of the last time step of the gated cyclic unit network to the predicted value sequence of the occupancy flag bits of each weighing equipment in the next six acquisition cycles, resulting in the equipment state prediction sequence. The shape of the equipment state prediction sequence is the number of weighing equipment multiplied by six prediction steps.

[0020] The equipment status prediction model uses the complete occupancy time series of each weighing device in the historical operation data of the plant area as training samples. The binary cross-entropy loss function is used to optimize the deviation between the predicted value and the actual value of the occupancy flag. The prediction step is set to six steps, which means predicting the equipment status within the next thirty seconds. This time window matches the typical travel time of a vehicle from the plant entrance node to the weighing device node, ensuring that the weighing device allocation sequence has sufficient lead time before the vehicle actually arrives at the weighing device node. Physical constraint correction is applied to the equipment status prediction sequence: the remaining loading time of the loading port is read. If the remaining loading time of a certain loading port is greater than the total prediction time window (30 seconds), the occupancy flag prediction value of all six prediction steps in the equipment status prediction sequence corresponding to this loading port is forcibly set to one, covering the original prediction output of the gated cyclic unit network; if the occupancy flag prediction value in the equipment status prediction sequence changes from one to zero and then back to one; if the time step difference between the two changes is less than the number of steps corresponding to the glitch noise judgment threshold, the prediction value of the intermediate step is corrected to one, and the corrected equipment status prediction sequence is stored in the equipment prediction status library. Synchronization of three-layer states in the digital twin model of the factory area: Using the directed graph of the factory area road network as the spatial skeleton, and the weighing equipment state matrix, the equipment predicted state library and the vehicle spatiotemporal trajectory matrix as dynamic state inputs, the three types of dynamic states of equipment layer, road layer and vehicle layer are synchronously updated in the digital twin model of the factory area to obtain a real-time state snapshot of the entire factory area. The occupancy flag and normalized weighing progress value of each weighing device in the weighing equipment status matrix are written into the equipment status attribute field of the corresponding vertex in the directed graph of the plant road network according to the vertex number of the weighing device; the remaining loading time of each loading port is written into the loading status attribute field of the corresponding vertex in the directed graph of the plant road network according to the vertex number of the loading port; the above writing operation is triggered once every five-second acquisition cycle to keep the time synchronization error between the equipment layer status of the plant digital twin model and the actual status of the physical equipment no greater than the acquisition cycle duration; Using the latest column of the vehicle spatiotemporal trajectory matrix (corresponding to the vehicle location distribution at the current time step) as input, the number of vehicles on the road at the current time on each directed edge in the directed graph of the factory road network is counted. The number of vehicles on the road on each directed edge is divided by the rated traffic capacity of the road segment corresponding to that directed edge (calculated from the road width and average lane length occupied by vehicles in the factory GIS map) to obtain the road congestion coefficient of each directed edge. The road congestion coefficient is then superimposed on the physical length weight of each directed edge in the edge weight set in a linear weighted manner to obtain the comprehensive path cost of each directed edge. The comprehensive path cost is stored in the edge cost attribute field. It should be noted that both the equipment status attribute field and the loading status attribute field are mounted in the vertex attribute structure of the directed graph of the plant's road network, and stored in parallel with the node type label attribute field. This ensures that the graph structure data and equipment dynamic status data in the plant's digital twin model remain consistent within the same data object, avoiding increased indexing time during path optimization modeling due to cross-data structure queries. The linear weighting coefficient of the road congestion coefficient is set to the road congestion coefficient multiplied by 0.3 times the physical length weight of the corresponding directed edge. That is, the comprehensive path cost is equal to the physical length weight plus 0.3 times the physical length weight multiplied by the road congestion coefficient. This weighting coefficient is determined through regression analysis between vehicle waiting time and congestion coefficient in the plant's historical operating data. This is to reasonably reflect the impact of road congestion on travel time in path optimization, while avoiding the congestion coefficient from excessively dominating path selection, causing all vehicles to concentrate on detouring to suboptimal paths. Extract the vehicle vertex number sequence of the column corresponding to the current time step from the vehicle spatiotemporal trajectory matrix, and write the current vertex number of each on-site vehicle into the vehicle layer status table of the factory digital twin model; for vehicles with a mask value of zero corresponding to the current time step in the mask matrix (i.e., vehicles waiting to enter that are not currently in the factory area), mark them with a waiting-to-enter status flag in the vehicle layer status table, and record the pickup code and corresponding rated load of this vehicle in the qualified vehicle information database for use in establishing the allocation constraints of the weighing equipment; summarize the status information of all on-site vehicles and vehicles waiting to enter in the vehicle layer status table under the current time step to obtain a real-time status snapshot of the entire factory area. S3.2: Using the equipment status prediction sequence output by the digital twin model of the factory area as a constraint, construct a path optimization problem with the goal of minimizing the total vehicle waiting time. Call the path optimization algorithm to solve the path optimization problem and output the weighing equipment allocation sequence and driving path sequence for each vehicle waiting to enter the site. S3.2.1: Using a real-time snapshot of the entire factory area as input, a path optimization problem is established with the goal of minimizing the total vehicle waiting time and constraints on the capacity of each equipment node and the physical constraints of vehicle movement. The resulting path optimization problem model includes: The comprehensive path cost of each directed edge is read from the real-time status snapshot of the entire factory area. The shortest path algorithm for all nodes is executed on the directed graph of the factory road network (the Johnson algorithm is used to process directed graphs with non-negative weights). The shortest path cost between any two vertices in the vertex set and the corresponding shortest path node sequence are calculated to obtain the shortest path cost matrix and the dictionary of shortest path node sequences for all nodes. It should be noted that the row and column indices of the shortest path cost matrix for all nodes correspond to the vertex numbers in the vertex set, and the matrix elements are the shortest path cost values ​​between corresponding vertex pairs. The shortest path node sequence dictionary for all nodes uses vertex number pairs as keys and the corresponding shortest path node sequences as values, and is stored in the path lookup cache. The shortest path cost matrix and the shortest path node sequence dictionary for all nodes are recalculated synchronously every five-second collection cycle with the update of the comprehensive path cost value to reflect the impact of real-time road congestion on the shortest path selection. In directed graphs containing negative weights, the Johnson algorithm eliminates negative weights through potential function transformation. The comprehensive path cost value is obtained by linearly superimposing the physical length weight and the congestion coefficient, and is always a non-negative value. The Johnson algorithm here degenerates into executing Dijkstra's algorithm with each vertex as the source point on the directed graph of the factory road network. The time complexity is proportional to the square of the number of vertices multiplied by the logarithm, which is suitable for the scale of the factory road network. The objective function is to sum the total waiting times of all vehicles in the set of vehicles waiting to enter the facility. For a single vehicle, the total waiting time is defined as the total time from the current moment until it leaves the weighing equipment node after completing the empty weighing process. This includes the travel time to the assigned weighing equipment node and the queuing time at the weighing equipment node. The travel time is obtained by dividing the cost of the shortest path from the current vertex number of the vehicle to the vertex number of the target weighing equipment node in the shortest path cost matrix of all nodes by the average speed of the roads in the factory area (taken as 15 kilometers per hour). The queuing time is obtained by subtracting the current moment from the time step corresponding to the first predicted value of zero in the predicted sequence of the occupancy flag of the target weighing equipment in the equipment prediction state library. The objective of the path optimization problem is to search for the weighing equipment allocation scheme that minimizes the objective function value while satisfying the constraints. The following three types of constraints are defined for the path optimization problem: The first type is the capacity constraint, which states that at any given time, the number of vehicles on each weighing equipment node does not exceed one. This constraint is guaranteed by the logic that a new vehicle can only be allocated when the predicted value of the occupancy flag bit of the corresponding weighing equipment node in the equipment prediction state database is zero. The second type is the path reachability constraint, which states that there must be a directed reachable path between the target weighing equipment node to which each vehicle is to be allocated and the vertex where the vehicle is currently located in the directed graph of the factory road network. This constraint is determined by whether the corresponding element in the shortest path cost matrix of all nodes is a finite value. The third type is the load matching constraint, which states that the rated load of a vehicle allocated to a certain weighing equipment must be within the rated weighing range of that weighing equipment in the weighing equipment state matrix, to avoid oversized vehicles being scheduled to mismatched weighing equipment. The above three types of constraints are written into the path optimization problem model in the form of linear inequalities to obtain the complete path optimization problem model. S3.2.2: Taking the path optimization problem model as input, the path optimization algorithm is called to solve the problem, and the output is the weighing equipment allocation sequence and driving path sequence for each vehicle to enter the site, specifically including: Greedy Initial Solution Generation: Arrange the vehicles in the set of vehicles to be entered according to the order of their corresponding entry timestamps in the augmented spatiotemporal trajectory matrix to obtain the queue sequence of vehicles to be entered; according to the order of the queue sequence, select the weighing equipment with the shortest total waiting time from the currently idle or predicted earliest release weighing equipment for each vehicle, record this allocation relationship, and temporarily set the occupancy flag of the allocated weighing equipment after the corresponding prediction time step in the equipment prediction state database to one to reflect the impact of this allocation on the available equipment for subsequent vehicles; after traversing the queue sequence of vehicles to be entered, summarize the allocation relationships of all vehicles into an initial weighing equipment allocation scheme, which serves as the greedy initial solution for iterative optimization; It should be noted that the greedy initial solution generation method is based on the first-come, first-served principle. Its solution time complexity is linearly proportional to the number of vehicles waiting to enter multiplied by the number of weighing devices. It can generate a feasible initial solution in milliseconds, ensuring that the iterative optimization is completed within the time window constraint. Under the premise of satisfying the three types of constraints, the initial weighing device allocation scheme can usually directly reach a near-global optimal objective function value when the number of vehicles is small. Simulated annealing iterative optimization: Starting with the initial weighing equipment allocation scheme, the simulated annealing algorithm is used to iteratively optimize the path optimization problem model. The initial temperature is set to 0.1 times the value of the objective function calculated under the initial weighing equipment allocation scheme, the cooling coefficient is set to 0.95, and the number of iterations at each temperature is set to the square of the number of vehicles waiting to enter. In each iteration, the target weighing equipment numbers of two vehicles in the initial weighing equipment allocation scheme are randomly selected and swapped to generate candidate allocation schemes. The candidate allocation schemes are checked to see if they meet three types of constraints. For candidate allocation schemes that meet the constraints, the change in the corresponding objective function value is calculated. If the optimization objective... If the change in the objective function value is negative, the candidate allocation scheme is unconditionally accepted as the new current solution. If the change in the objective function value is positive, the acceptance probability is calculated using the current annealing temperature and the change in the objective function value according to the Boltzmann probability formula, and random numbers are uniformly sampled between zero and one. If the random number is less than the acceptance probability, the candidate allocation scheme is accepted. If the random number is greater than or equal to the acceptance probability, the candidate scheme is rejected, and the current scheme remains unchanged. When the temperature drops to the termination temperature threshold (set to one-thousandth of the initial temperature), the iteration stops, and the allocation scheme corresponding to the best historical solution continuously recorded during the iteration process is recorded as the optimized weighing equipment allocation scheme. It should be noted that when the path optimization problem is large in scale (more than 20 vehicles to enter), the simulated annealing algorithm has a lower time complexity than the exact solution algorithm, and can output a high-quality near-optimal solution within the factory scheduling response time requirement; in the optimized weighing equipment allocation scheme, the correspondence between each vehicle and the target weighing equipment is arranged in the order of vehicle number, and the weighing equipment allocation sequence of each vehicle to enter is obtained. Driving path sequence generation: For each vehicle to enter the weighing equipment allocation sequence, using the vertex number corresponding to the current time step of the vehicle in the vehicle's spatiotemporal trajectory matrix as the starting vertex and the vertex number of the target weighing equipment node corresponding to the vehicle in the weighing equipment allocation sequence as the ending vertex, query the dictionary of shortest path nodes for all nodes and retrieve the corresponding shortest path node sequence as the first segment of the driving path node sequence from the current position to the target weighing equipment node; using the vertex number of the target weighing equipment node as the starting vertex and the vertex number of the target loading port node corresponding to the vehicle's pickup code in the qualified vehicle information database as the ending vertex, query the dictionary of shortest path nodes for all nodes again and retrieve the corresponding shortest path node sequence. The shortest path node sequence is used as the second segment of the vehicle's travel path node sequence from the target weighing equipment node to the target loading port node; the first segment of the travel path node sequence and the second segment of the travel path node sequence are concatenated end to end (removing duplicate intermediate nodes) to obtain the complete travel path node sequence of the vehicle; the complete travel path node sequences of all vehicles waiting to enter are arranged in order of vehicle number to obtain the travel path sequence; the weighing equipment allocation sequence and the travel path sequence are stored in the scheduling result database for use in the access control gate release control logic, and the execution result of the weighing equipment allocation sequence is fed back to the plant area digital twin model to update the equipment occupancy status parameters in the plant area digital twin model; It should be noted that the vertex numbers in the driving path node sequence correspond exactly to the vertex numbers in the vertex set. The vertex coordinate attributes of the directed graph of the factory road network can be directly converted into the factory geographical coordinate sequence, and driving guidance information can be pushed to the corresponding vehicle drivers through voice broadcast or LED screen display. S4: Issue access control commands to the access control gate based on the abnormal behavior probability vector and the weighing equipment allocation sequence. When the abnormal probability of any vehicle in the abnormal behavior probability vector is greater than the preset probability threshold, trigger the gate closure command for this vehicle and feed back the execution result of the weighing equipment allocation sequence to the plant digital twin model to update the equipment occupancy status parameters in the plant digital twin model. Specifically, the access control release decision rule table uses the probability vector of abnormal behavior as the primary judgment basis and the allocation sequence of weighing equipment as the scheduling execution basis. A release control command can only be generated when both signals meet the release conditions. If either signal triggers the interception condition, a gate closure command is generated, forming a dual verification mechanism. The feedback writing operation of the execution result of the weighing equipment allocation sequence and the status update operation of the scheduling result database jointly maintain the real-time consistency of the equipment occupancy status parameters in the digital twin model of the plant area. S4.1: Using the abnormal behavior probability vector and the weighing equipment allocation sequence as dual input signals, construct an access control release decision rule table, and perform release qualification judgment on each vehicle in the set of vehicles to be entered; Specifically, before performing the release qualification determination, all the input information required for the current release determination cycle is read from each database in steps S1, S2 and S3, and summarized into a release determination information package, which serves as the unified input for the access control release decision rule table of the steps. Furthermore, the abnormal behavior probability vectors corresponding to all vehicles present and vehicles waiting to enter in the current judgment period are read from the abnormal probability database; the element value corresponding to the current prediction time step in the abnormal behavior probability vector is taken to obtain the current abnormal probability value of each vehicle at the current time, where the current abnormal probability value is a real number in the interval between zero and one; the current abnormal probability values ​​of all vehicles are arranged in order of vehicle number to obtain the current abnormal probability frame vector, where the dimension of the current abnormal probability frame vector is consistent with the total number of controlled vehicles in the current judgment period.

[0021] It should be noted that the extraction operation of the current frame vector of anomaly probability is triggered periodically with the time step of the discrete time axis, that is, it is updated every 30 seconds, which is not synchronized with the acquisition period (five seconds) of the weighing equipment status buffer queue. When the triggering times of the two do not coincide, the current frame vector of anomaly probability uses the most recent valid value of the previous judgment period until the spatiotemporal graph convolutional network model completes the next round of inference and updates the anomaly probability database. The positions of vehicles waiting to enter the factory with a mask matrix mask value of zero at the current time step in the current frame vector of anomaly probability retain their current anomaly probability value and do not replace it with a zero value, so as to ensure that the historical behavior characteristics of the vehicles waiting to enter the factory before entering the factory area have an impact on the current judgment result. Furthermore, the system reads the weighing equipment allocation sequence and driving path sequence corresponding to each vehicle waiting to enter the site within the current judgment period from the scheduling result database; for each allocation record in the weighing equipment allocation sequence, it extracts the target weighing equipment node vertex number and the corresponding vehicle number to generate an allocation mapping table from vehicle number to target weighing equipment node vertex number; for each path record in the driving path sequence, it extracts the corresponding vehicle number and the complete driving path node sequence to generate a path mapping table from vehicle number to complete driving path node sequence; the allocation mapping table and the path mapping table are merged to obtain the vehicle scheduling information table, in which each vehicle number in the vehicle scheduling information table corresponds to a unique target weighing equipment node vertex number and a unique complete driving path node sequence.

[0022] It should be noted that when the simulated annealing iterative optimization has not been solved by the trigger time of the current decision cycle, the vehicle scheduling information table uses the allocation mapping table and path mapping table corresponding to the initial weighing equipment allocation scheme, and uses the greedy initial solution as the scheduling basis for the current decision cycle to ensure that the issuance of access control commands is not interrupted due to the delay in optimization solution; when the simulated annealing iterative optimization is completed, the vehicle scheduling information table is updated with the optimized weighing equipment allocation scheme and takes effect when the next decision cycle is triggered. Specifically, the vertex number of each vehicle to be entered is read from the column corresponding to the current time step in the vehicle spatiotemporal trajectory matrix; for vehicles to be entered whose mask value is zero at the current time step in the mask matrix, the vehicle's entry status flag and corresponding pickup code are read from the vehicle layer status table; it is verified whether the pickup code is still valid in the qualified vehicle information database (i.e., not marked as used or revoked). The entry qualification valid flag of vehicles that pass the verification is set to one, and the entry qualification valid flag of vehicles that fail the verification is set to zero; the current vertex number, entry qualification valid flag, and pickup code of each vehicle to be entered are summarized, along with the current frame vector of the abnormal probability and the vehicle scheduling information table, to form a release judgment information packet, which is stored in the release judgment buffer. S4.2: Taking the release judgment information packet as input, perform release qualification judgment on each vehicle waiting to enter according to the access control release decision rule table, and output the release judgment result flag bit corresponding to each vehicle as the release control command; Furthermore, the statistical distribution of the current anomaly probabilities of all vehicles present and vehicles waiting to enter in the current judgment period is read from the anomaly probability database. The mean and standard deviation of the current anomaly probabilities for all controlled vehicles are calculated. Centered on a fixed benchmark threshold (0.7), if the mean of the current anomaly probabilities in the current judgment period is higher than 0.5, the preset probability threshold is lowered by 0.05 from the fixed benchmark threshold to obtain a tightened preset probability threshold. If the mean of the current anomaly probabilities is lower than 0.2, the preset probability threshold is raised by 0.05 from the fixed benchmark threshold to obtain a relaxed preset probability threshold. In other cases, the preset probability threshold remains unchanged from the fixed benchmark threshold. The preset probability threshold used in the current judgment period is recorded in the threshold adjustment log for use by management personnel for post-event auditing. It should be noted that the preset probability threshold adaptive adjustment mechanism takes the overall abnormal probability distribution of all controlled vehicles as a reference background. When the overall abnormality level is high, it automatically tightens the judgment criteria to reduce the risk of missed detection; when the overall abnormality level is low, it appropriately relaxes the judgment criteria to reduce the false interception rate of normal vehicles. The upward and downward adjustment amounts are both set to 0.05 to constrain the preset probability threshold to fluctuate within a narrow range of 0.65 to 0.75, and to prevent the judgment logic from failing due to excessive threshold deviation in extreme cases. Furthermore, for each vehicle waiting to enter in the release judgment information packet, an anomaly probability judgment is performed one by one: the current anomaly probability value corresponding to this vehicle in the current frame vector of anomaly probability is read and compared with the preset probability threshold used in the current judgment period; if the current anomaly probability value is less than the preset probability threshold, the anomaly probability judgment result for this vehicle is pass, and the anomaly probability judgment flag is set to zero; if the current anomaly probability value is greater than or equal to the preset probability threshold, the anomaly probability judgment result for this vehicle is fail, and the anomaly probability judgment flag is set to one; for vehicles with anomaly probability judgment flag of one, the node jump flag and speed anomaly flag corresponding to this vehicle in the node passing sequence database are read; if both the node jump flag and the speed anomaly flag are zero, the anomaly probability judgment flag is downgraded to a suspected anomaly flag to distinguish between interception triggered by pure model prediction and interception supported by actual measured anomaly records; Preferably, the purpose of distinguishing between the anomaly probability determination flag and the suspected anomaly flag is to provide a tiered handling basis for the abnormal vehicle early warning process: a vehicle with an anomaly probability determination flag of one and supported by a node jump flag or a speed anomaly flag triggers a Level 1 early warning, directly generating a gate closure command and simultaneously notifying security personnel to intervene; a vehicle with only a suspected anomaly flag triggers a Level 2 early warning, generating a gate temporary release command and initiating a manual review process. If the review result does not arrive within the preset review waiting time (set to sixty seconds), it is automatically downgraded to a Level 1 early warning. Specifically, for vehicles with an anomaly probability determination flag of zero, an entry qualification verification and scheduling readiness verification are performed: The entry qualification validity flag of the vehicle is read; if the entry qualification validity flag is zero, the release qualification rejection flag is set to one, terminating the subsequent verification process for this vehicle; if the entry qualification validity flag is one, the vehicle scheduling information table is read to verify whether the vehicle already has a valid target weighing equipment node vertex number and a complete travel path node sequence; if the corresponding record for this vehicle exists in the vehicle scheduling information table and the target weighing equipment node vertex number is valid... If the vehicle is valid, the scheduling ready flag is set to one. If the record corresponding to this vehicle in the vehicle scheduling information table is missing or the vertex number of the target weighing equipment node is invalid, the scheduling ready flag is set to zero, triggering an immediate supplementary calculation request for the path optimization problem model. This vehicle is added to the supplementary calculation queue, and the issuance of departure control instructions to this vehicle is temporarily suspended until the immediate supplementary calculation request is completed. For vehicles with an abnormal probability judgment flag of zero, an entry qualification valid flag of one, and a scheduling ready flag of one, the release judgment result flag is set to one, indicating that this vehicle has passed all release qualification judgments. S4.3: Based on the release determination result flag of each vehicle, generate corresponding release control instructions or gate closure instructions, send them to the corresponding gate controller via the access control bus, and trigger a graded early warning process for abnormal vehicles, specifically including: For each vehicle waiting to enter with a release decision flag of 1, the target weighing equipment node vertex number and the complete driving path node sequence corresponding to the vehicle are read from the vehicle dispatch information table. Using the vehicle number, target weighing equipment node vertex number, complete driving path node sequence, and current decision period timestamp as fields, a release control message is encapsulated according to the access control bus communication protocol. A driving guidance information field is added to the release control message. The driving guidance information field is composed of the node type label corresponding to each vertex in the complete driving path node sequence and the geographic coordinate sequence obtained by converting the spatial coordinate attributes of each vertex. This serves as the data source for pushing driving guidance information to the driver via LED screen or voice broadcasting equipment. The release control message is written into the release instruction sending queue and sent to the corresponding gate controller in order of vehicle number. The geographic coordinate sequence of the complete driving path node sequence in the release control message is directly read from the spatial coordinates stored in the vertex attributes of the directed graph of the factory road network, without the need for additional map interface calls; the node type label corresponding to each path node in the driving guidance information field enables the driver to distinguish whether the current guidance target is a weighing equipment node or a loading port node in the content displayed on the LED screen, reducing the probability of drivers mistakenly entering due to path confusion, thereby reducing the abnormally high road congestion coefficient caused by vehicles going the wrong way; The gate closure command generation and tiered early warning triggering process are as follows: For vehicles with a release determination result flag of zero, the system reads the corresponding anomaly probability determination flag, suspected anomaly flag, and release qualification rejection flag. For vehicles with an anomaly probability determination flag of one, a Level 1 gate closure command is generated. This Level 1 gate closure command is then sent to the corresponding gate controller via the access control bus, and a Level 1 early warning notification containing abnormal vehicle information is pushed to the security management terminal. For vehicles with only a suspected anomaly flag, a temporary release command is generated, and a 60-second manual review timer is started. If a review approval confirmation is received from the security management terminal before the manual review timer expires, the temporary release command is upgraded to a release control message and sent via the access control bus. If no review approval confirmation is received before the timer expires, the temporary release command is automatically upgraded to a Level 1 gate closure command. For vehicles with only a release qualification rejection flag, a qualification rejection gate closure command is generated. It should be noted that the Level 1 gate closure instruction includes an abnormal vehicle information field; this field contains the vehicle's ID, current vertex ID, current abnormal probability value, node jump flag, and speed abnormal flag; the qualification rejection gate closure instruction includes a rejection reason field; this field records the specific reason for the failure of the delivery code verification and pushes the content of the rejection reason field to the abnormal record table in the qualified vehicle information database; the design basis of the above-mentioned hierarchical early warning mechanism is as follows: the Level 1 gate closure instruction corresponds to situations with dual support from both measured abnormal behavior records and high-confidence model predictions, requiring immediate physical interception measures; the postponement instruction corresponds to situations with only model predictions but lacking measured abnormal records, introducing a manual review process to reduce the probability of normal vehicles being mistakenly intercepted, achieving a balance between security and traffic efficiency; the qualification rejection gate closure instruction corresponds to situations with failed document or delivery code verification, which is a pre-logic interception and does not involve the prediction results of the spatiotemporal graph convolutional network model, and is independent of the calculation logic of the abnormal behavior probability vector; After the access control message or the first-level barrier gate closure command is sent through the access control bus, a status query message is sent to the corresponding barrier gate controller at a polling interval of 200 milliseconds to read the execution status feedback signal returned by the barrier gate controller. The execution status feedback signal includes the current physical position status of the barrier gate (barrel raised, bar lowered, or in motion) and the command execution result code (execution successful, execution timeout, or execution failed). For barrier gates that return execution success and the barrier is raised, the timestamp of the raised position is recorded. For barrier gates that return execution timeout or execution failure, the original command is resent and recorded. If the resend attempt fails after three attempts, a fault warning for the barrier gate equipment is pushed to the operation and maintenance management terminal. At the same time, the comprehensive path cost value of the corresponding entrance node of the barrier gate in the edge cost attribute field is set to positive infinity, so that the entrance is excluded from the reachable path in the next recalculation of the shortest path cost matrix of all nodes, thus preventing subsequent vehicles from being dispatched to the faulty barrier gate entrance. The execution status feedback signals of each barrier gate are recorded in the execution status log according to the vehicle number and barrier gate number. The execution status log serves as the data source for generating the execution results of the weighing equipment allocation sequence. S4.4: Combine the execution status feedback signal after the barrier gate's execution action is completed with the weighing equipment allocation sequence to generate the weighing equipment allocation sequence execution result, and write the weighing equipment allocation sequence execution result into the equipment layer status update interface of the plant area digital twin model to trigger the rolling update of the equipment occupancy status parameters in the plant area digital twin model; Specifically, the system reads the timestamps of the gate lifting position and the command execution result codes corresponding to all executed release control messages within the current judgment period from the execution status log. For each vehicle with a successful command execution result code, the vehicle number, the corresponding target weighing equipment node vertex number in the weighing equipment allocation sequence, the timestamp of the gate lifting position, and the release success flag are used to construct the allocation execution record for this vehicle. For each vehicle with a timeout or failure command execution result code, the vehicle number, the target weighing equipment node vertex number, and the release failure flag are used to construct the allocation execution record for this vehicle. The allocation execution records of all vehicles waiting to enter are arranged in order of vehicle number to obtain the weighing equipment allocation sequence execution result. The ratio of the number of vehicles that successfully passed to the number of vehicles that failed to pass in the weighing equipment allocation sequence execution result is used as the gate release success rate indicator for the current judgment period and written into the operation statistics database for management personnel to query. It should be noted that the time stamp of the barrier lifting position in the allocation execution record serves as an auxiliary time reference for the vehicle layer status table. When the acquisition unit fails to acquire the record of the vehicle node passing through within the expected time, the estimated vertex number is filled into the corresponding position in the vehicle spatiotemporal trajectory matrix by adding the time stamp of the barrier lifting position to the expected travel time to the adjacent vertex. This prevents the first record of the node passing sequence from being missing due to the short road segment between the barrier gate entrance and the vertex covered by the first acquisition unit, which in turn causes the corresponding element of the mask matrix to remain at zero value after the vehicle actually enters the site. Furthermore, the vertex numbers of the target weighing equipment nodes corresponding to each vehicle that was successfully released in the weighing equipment allocation sequence execution result are written into the equipment layer state update interface of the plant digital twin model. The equipment layer state update interface forces the occupancy flag bit corresponding to the vertex number of the above target weighing equipment node to one in the subsequent prediction steps from the current moment in the equipment prediction state library, covering the idle prediction values ​​that may exist in the original prediction output of the gated cyclic unit network, and obtains the updated equipment prediction state library. The prediction sequence of the occupancy flag bit of each weighing equipment node in the updated equipment prediction state library is synchronously updated to the equipment state attribute field of the corresponding vertex in the directed graph of the plant road network, completing the current rolling update of the equipment occupancy state parameters in the plant digital twin model. The shortest path cost matrix of all nodes will be recalculated with the updated equipment prediction state library as input when the next five-second collection cycle is triggered. The simulated annealing iterative optimization will generate a new round of optimized weighing equipment allocation scheme based on the updated shortest path cost matrix of all nodes. It should be noted that the rolling update of the equipment occupancy status parameters only applies to the target weighing equipment nodes corresponding to vehicles that have been successfully released within the current judgment period. The predicted sequence of the occupancy flag bit for other weighing equipment nodes is not forcibly overwritten. The original prediction results of the gate control loop unit network based on the historical state sequence are retained to avoid excessive intervention in the natural evolution process of the prediction model. When the gate failure causes the release to fail, the predicted sequence of the occupancy flag bit for the corresponding target weighing equipment node is not forcibly set to one. The equipment status attribute field of the corresponding equipment node in the plant area digital twin model retains the original predicted value of the gate control loop unit network. After the actual occupancy status of the weighing equipment controller is read in the next five-second acquisition cycle, the actual acquired value is used to overwrite the equipment status attribute field, restoring the consistency between the equipment layer status of the plant area digital twin model and the actual status of the physical equipment.

[0023] In summary, this invention achieves a formal description and digital storage of complex factory road topology and historical vehicle trajectories by converting a factory area GIS map into a directed road network graph and generating a vehicle spatiotemporal trajectory matrix. By constructing a spatiotemporal graph convolutional network model that integrates spatial graph convolution and temporal attention mechanisms, it not only accurately captures the spatial neighborhood interaction features of vehicles in the current road network using the road network topology but also adaptively enhances key temporal information in historical trajectories through the attention mechanism, thereby significantly improving the accuracy and robustness of predicting abnormal vehicle driving behavior. Furthermore, by constructing a digital twin model combining real-time equipment status and vehicle location, and nesting it with a path optimization algorithm, using the current location of each vehicle as a dynamic constraint, it outputs the optimal weighing equipment allocation sequence and driving path sequence, realizing dynamic scheduling and path planning of global logistics resources. The system effectively avoids equipment idling and road congestion. By using the probability of abnormal behavior as a prerequisite for access control, it enables proactive interception of risky vehicles before they enter the site. Combined with the optimized allocation sequence, it issues departure instructions and feeds the execution results back to the digital twin model to update the equipment occupancy status, ensuring the accurate execution of control instructions and real-time self-calibration of model parameters. By adopting a path optimization algorithm based on simulated annealing and using the total vehicle waiting time as the optimization objective, it globally searches for the optimal equipment allocation scheme under multiple real-world constraints. This effectively solves the combined optimization problem under multiple vehicles, multiple devices, and multiple constraints, minimizing the overall queuing time of vehicles within the plant area. It provides rich and discriminative inputs for the abnormal behavior decoder, thereby comprehensively improving the intelligence, automation, and precision of vehicle management in the plant area.

[0024] This embodiment also provides a computer device applicable to the factory vehicle management method based on intelligent access control and unattended weighing, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the factory vehicle management method based on intelligent access control and unattended weighing as proposed in the above embodiment.

[0025] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0026] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the factory vehicle management method based on intelligent access control and unattended weighing proposed in the above embodiments.

[0027] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0028] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for managing factory vehicles based on intelligent access control and unattended weighing, characterized in that: include, Using the road intersection nodes of the factory area GIS map as vertices and road segments as directed edges, a directed graph of the factory area road network is constructed, and the node passing sequence of each vehicle is collected through the collection unit to generate a vehicle spatiotemporal trajectory matrix. The vehicle spatiotemporal trajectory matrix is ​​input into a spatiotemporal graph convolutional network model, wherein the spatiotemporal graph convolutional network model performs spatial graph convolution operations based on the topology of the directed graph of the factory road network and performs weighted aggregation of the historical frame sequence of the vehicle spatiotemporal trajectory matrix through a temporal attention mechanism, and outputs the abnormal behavior probability vector of each vehicle within the prediction time window. Using the factory area GIS map, the real-time occupancy status of weighing equipment, and the remaining loading time at the loading dock as input parameters, a digital twin model of the factory area is constructed. The digital twin model of the factory area is then combined with a path optimization algorithm. Using the current position of each vehicle in the vehicle spatiotemporal trajectory matrix as a constraint, the weighing equipment allocation sequence and driving path sequence of each vehicle waiting to enter the site are output. According to the abnormal behavior probability vector and the weighing equipment allocation sequence, a passage control command is issued to the access control gate. When the abnormal probability of any vehicle in the abnormal behavior probability vector is greater than a preset probability threshold, a gate closure command is triggered for this vehicle, and the execution result of the weighing equipment allocation sequence is fed back to the factory area digital twin model to update the equipment occupancy status parameters in the factory area digital twin model.

2. The factory vehicle management method based on intelligent access control and unattended weighing as described in claim 1, characterized in that: The method for generating the driving path sequence is as follows: For each vehicle waiting to enter the weighing equipment allocation sequence, the vertex number corresponding to the current time step of the vehicle in the vehicle's spatiotemporal trajectory matrix is ​​taken as the starting vertex, and the vertex number of the target weighing equipment node corresponding to the vehicle in the weighing equipment allocation sequence is taken as the ending vertex. The dictionary of shortest path node sequences of all nodes is queried, and the corresponding shortest path node sequence is taken out as the first segment of the driving path node sequence of the vehicle from the current position to the target weighing equipment node. Taking the vertex number of the target weighing equipment node as the starting vertex and the vertex number of the target loading port node corresponding to the vehicle's pickup code in the qualified vehicle information database as the ending vertex, the dictionary of shortest path nodes for all nodes is queried again, and the corresponding shortest path node sequence is taken out as the second segment of the vehicle's travel path node sequence from the target weighing equipment node to the target loading port node. By concatenating the first segment of the driving path node sequence with the second segment of the driving path node sequence, the complete driving path node sequence of this vehicle is obtained. Arrange the complete driving path node sequence of all vehicles waiting to enter the site in order of vehicle number to obtain the driving path sequence.

3. The factory vehicle management method based on intelligent access control and unattended weighing as described in claim 2, characterized in that: The weighing equipment allocation sequence is obtained by arranging the correspondence between each vehicle and the target weighing equipment in the optimized weighing equipment allocation scheme according to the vehicle number order.

4. The factory vehicle management method based on intelligent access control and unattended weighing as described in claim 3, characterized in that: The method for generating the optimized weighing equipment allocation scheme is as follows: Starting with the initial weighing equipment allocation scheme, the simulated annealing algorithm is used to iteratively optimize the path optimization problem model; Randomly select the target weighing equipment numbers of two vehicles in the initial weighing equipment allocation scheme and exchange them to generate candidate allocation schemes. Check whether the candidate allocation schemes meet the three types of constraints. Calculate the change in the corresponding optimization objective function value for the candidate allocation schemes that meet the constraints. If the change in the objective function value is negative, then the candidate allocation scheme is unconditionally accepted as the new current solution. If the change in the objective function value is positive, then according to the Boltzmann probability formula, the acceptance probability is calculated together with the current annealing temperature and the change in the objective function value, and random numbers are uniformly sampled between zero and one. If the random number is less than the acceptance probability, then accept this candidate allocation scheme; if the random number is greater than or equal to the acceptance probability, then reject the candidate scheme and keep the current scheme unchanged. When the temperature drops to the termination temperature threshold, the iteration stops, and the allocation scheme corresponding to the best historical solution continuously recorded during the iteration process is recorded as the optimized weighing equipment allocation scheme.

5. The factory vehicle management method based on intelligent access control and unattended weighing as described in claim 4, characterized in that: The optimization objective function is the sum of the total waiting times of all vehicles in the set of vehicles waiting to enter, and the total waiting times of all vehicles waiting to enter are summed.

6. The factory vehicle management method based on intelligent access control and unattended weighing as described in claim 1, characterized in that: The method for obtaining the abnormal behavior probability vector is as follows: The discrete vertex number sequence in the augmented spatiotemporal trajectory matrix is ​​converted into node feature embedding vectors. The normalized Laplacian matrix of the adjacency matrix is ​​used as the graph structure operator to perform spatial graph convolution operation, extract the spatial neighborhood context features of the current position of each vehicle, and obtain the spatial feature map. The historical frame sequence of the spatial feature map is input into the temporal attention module. The features of each time step in the historical frame sequence are weighted and aggregated by the attention weight matrix to obtain the temporal context feature vector. The temporal context feature vector is extracted by stacking multiple spatiotemporal feature fusion modules and then input into the abnormal behavior decoder to output the abnormal behavior probability vector of each vehicle within the prediction time window.

7. The factory vehicle management method based on intelligent access control and unattended weighing as described in claim 6, characterized in that: The method for obtaining the spatial feature map is as follows: Based on the Chebyshev polynomial expansion matrix sequence, spatial graph convolution operation is performed on the fusion node feature tensor to extract the spatial neighborhood context features of the current position of each vehicle. The temporal context feature vector is input into the spatiotemporal feature fusion module, and high-order spatiotemporal interaction features are extracted through a multi-layer stacked feedforward network and residual connection structure to obtain a spatial feature map.

8. The factory vehicle management method based on intelligent access control and unattended weighing as described in claim 6, characterized in that: The method for generating the vehicle spatiotemporal trajectory matrix is ​​as follows: The spatial topological relationship between road intersection nodes and road segments is extracted from the GIS map of the factory area. The road intersection nodes are mapped as a set of vertices, and the road segments are mapped as a set of directed edges according to the actual traffic direction of the factory roads, thus constructing a directed graph of the factory road network. Data acquisition units are deployed at the physical locations corresponding to each vertex of the directed graph of the factory road network to collect the sequence of each vehicle passing through the nodes of each vertex in the directed graph of the factory road network. Based on the distribution of the node traversal sequence on the discrete time axis, the historical position status of each vehicle is stored row by row to generate a vehicle spatiotemporal trajectory matrix.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the factory vehicle management method based on intelligent access control and unattended weighing as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the factory vehicle management method based on intelligent access control and unattended weighing as described in any one of claims 1 to 8.

Images