Park logistics vehicle intelligent scheduling method and system based on multi-source data fusion

Abstract

The present application relates to the technical field of logistics management, in particular to a park logistics vehicle intelligent scheduling method and system based on multi-source data fusion. The present application predicts the number of queued vehicles and the work rate within a future preset time window of each work node through a long short-term memory network, and obtains a forward-looking comprehensive pressure coefficient by substituting the predicted value into a comprehensive pressure coefficient calculation formula. By directly incorporating the forward-looking comprehensive pressure coefficient into a coupling cost function with time and vehicle number compound dimensions, the vehicle is guided to the work node with the lowest queuing pressure at the "vehicle predicted arrival time" at the scheduling decision time, thereby achieving the purpose of keeping the scheduling instruction effective and real-time in the full time dimension, solving the problem that the decision information and the actual state of the physical world are out of sync due to the time lag between the scheduling decision time and the actual arrival time of the vehicle, and avoiding the situation that the originally optimal node is invalid due to the change of the queuing state when the vehicle arrives.

Description

Technical Field

[0001] This invention belongs to the field of logistics management technology, specifically relating to a method and system for intelligent scheduling of logistics vehicles in industrial parks based on multi-source data fusion. Background Technology

[0002] In industrial logistics parks, vehicle dispatch management involves multiple operational stages, including raw material procurement, finished product delivery, and intra-plant short-haul transfers, spanning multiple physical nodes such as access control, metering, warehousing, and loading / unloading. Currently, industrial parks typically deploy multiple independent information systems, such as transportation management systems, park management systems, access control systems, metering systems, and warehousing management systems, to manage transportation planning and execution, vehicle entry, exit and parking within the park, vehicle weighing, inventory and loading / unloading operations, and other business processes.

[0003] However, the aforementioned systems are usually built by different manufacturers at different times, and their data formats, communication protocols, and data update frequencies vary. They lack a unified data interaction standard and integration mechanism, which leads to scheduling decisions often relying on a single data source or manual on-site observation. This makes it impossible to comprehensively consider multi-dimensional dynamic factors such as the warehouse's real-time loading and unloading capacity, the current weighbridge load, the real-time saturation of the park's roads, and the priority of vehicle transportation tasks. This results in problems such as disorderly queuing of vehicles after entering the factory, uneven load of each operation node, and severe local congestion.

[0004] More critically, existing scheduling methods generally suffer from a neglected technical problem: a significant time lag, typically several minutes to tens of minutes, exists between the time the scheduling decision is made and the time the vehicle actually arrives at the work node. Within this time window, status parameters such as queue length and work rate at each work node within the park can change significantly. Because existing methods base scheduling decisions on node status data collected at the current moment, without considering the trend of node status changes upon vehicle arrival, the optimal node indicated in the scheduling instruction may no longer be truly optimal when the vehicle actually arrives. This means there is a disconnect between scheduling decision information and the actual physical state, severely reducing the real-time performance and effectiveness of the scheduling scheme. This cognitive bias caused by the time lag is particularly pronounced when vehicles travel long distances, routes are congested, or node work rates fluctuate significantly.

[0005] To address the aforementioned issues, this invention proposes a method and system for intelligent scheduling of logistics vehicles in industrial parks based on multi-source data fusion. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, the purpose of this invention is to provide a method and system for intelligent scheduling of logistics vehicles in industrial parks based on multi-source data fusion.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for intelligent scheduling of logistics vehicles in industrial parks based on multi-source data fusion includes: Real-time data from multiple sources is collected from various information systems within the park and then merged to form a fused dataset. The current comprehensive pressure coefficient of each work node is calculated based on the fused dataset, and the number of queuing vehicles and the operation rate of each work node in the future preset time window are predicted using a time series prediction model. The forward-looking comprehensive pressure coefficient of each work node at future time is calculated based on the predicted values. Based on the changing trends of vehicle arrival rate and factory exit rate in the fused dataset, as well as the changing trends of pressure coefficients at each work node, the current operating scenario of the park is identified, and weight coefficients matching the operating scenario are called according to the identification results. Based on the transportation task of the target vehicle, a set of candidate operation nodes is determined. For each candidate operation node, the estimated travel time and recommended travel route from the current location of the vehicle to the candidate operation node are calculated. Based on the forward-looking comprehensive pressure coefficient, estimated travel time, path tortuosity, road saturation coefficient and weight coefficient, the comprehensive scheduling cost of the candidate operation node is calculated. The candidate operation node with the lowest comprehensive scheduling cost is selected as the target operation node. The target operation node and the corresponding recommended travel route are encapsulated into a scheduling scheme and issued.

[0008] Preferably, a time series prediction model is used to predict the number of queuing vehicles and the operation rate of each operation node within a future preset time window. This includes: using the historical queuing vehicle number sequence, historical operation rate sequence, upstream vehicle arrival rate, estimated completion time of currently queuing vehicles, and remaining time of the traffic light phase at the park entrance as input features, a long short-term memory network is used to predict the number of queuing vehicles and the operation rate of the operation node at each moment within the future preset time window; based on the estimated travel time of the target vehicle to the operation node, the predicted value at the corresponding moment is selected from the prediction results and substituted into the comprehensive pressure coefficient calculation formula to obtain the forward-looking comprehensive pressure coefficient.

[0009] Preferably, identifying the current operating scenario of the park includes: determining whether the park is in a high-load or low-load scenario based on the changing direction of the park's vehicle arrival rate and departure rate, as well as the changing direction of the pressure coefficient of each work node; determining whether the park is in an abnormal congestion scenario based on the road saturation coefficient; and determining whether the park is in an equipment failure scenario based on the pressure coefficient of each work node.

[0010] Preferably, identifying the current operating scenario of the park also includes: when multiple scenario conditions are met simultaneously, determining a unique scenario according to the priority order of abnormal congestion scenario, equipment failure scenario, high load scenario, and low load scenario.

[0011] Preferably, calculating the estimated travel time and recommended travel path from the vehicle's current location to the candidate work node includes: taking the vehicle's current location as the starting point and the candidate work node as the ending point, and calculating the recommended travel path and the corresponding estimated travel time and path length based on the current road saturation coefficient and real-time traffic conditions using a path search algorithm; wherein, the heuristic function of the path search algorithm is dynamically adjusted according to the road saturation coefficient, and the higher the road saturation, the larger the heuristic function value.

[0012] Preferably, it also includes a cold start process for newly commissioned work nodes, including: Retrieve existing mature operation nodes, extract the physical feature vectors of each mature operation node, select mature nodes similar to the target new node based on physical feature similarity as the matching condition, and migrate the historical time series data of the mature node after scaling to become the initial training sample of the target new node for training the prediction model of the node; if no mature node matching the condition is found, the average historical data of all mature nodes in the park after global scaling is used as the initial training sample.

[0013] Preferably, after the scheduling plan is issued, the process also includes monitoring and handling of anomalies: obtaining the real-time location of the vehicle at preset time intervals and matching it with the recommended path; calculating the comprehensive matching degree based on the position deviation, direction deviation, and progress deviation; triggering the anomaly handling process when the comprehensive matching degree is lower than a preset threshold or the vehicle deviates from the recommended path for more than a preset time; identifying the anomaly type based on the deviation distribution; and performing corresponding autonomous recovery operations based on the anomaly type. The autonomous recovery operations include replanning the path or automatically updating the target work node.

[0014] Preferably, the autonomous recovery operation is performed according to the anomaly type, including: when the anomaly type is path deviation, the recommended driving route is recalculated and the scheduling plan is updated with the vehicle's current location as the starting point and the original target work node as the ending point; when the anomaly type is node change, it is assessed whether the node the vehicle is currently heading to is suitable for performing the vehicle's transportation task and whether the forward-looking comprehensive pressure coefficient is lower than a preset threshold. If the conditions are met, the node is updated as the target work node and the route is replanned. If the conditions are not met, external constraints are triggered; when autonomous recovery fails or the anomaly type does not meet the autonomous recovery conditions, the target work node input by the management personnel is received as a mandatory constraint, and the recommended driving route is recalculated and the scheduling plan is updated with the mandatory constraint as the ending point.

[0015] Preferably, it also includes closed-loop feedback: after the vehicle arrives at the target work node and completes the work, the actual work time is recorded and fed back to the data center to update the work rate parameters of the corresponding work node; it is determined whether there are any subsequent work nodes in the vehicle's transportation task. If there are, the scheduling plan is re-executed and generated. If all are completed, the vehicle is released to leave the factory.

[0016] This invention also discloses an intelligent dispatching system for park logistics vehicles based on multi-source data fusion, comprising: The multi-source data acquisition and fusion module is used to collect real-time data from various information systems in the park and perform format unification and time sequence alignment to form a fused dataset. The node status calculation and forward prediction module is used to calculate the current comprehensive pressure coefficient of each work node based on the fused dataset, and to use a time series prediction model to predict the number of queuing vehicles and the operation rate of each work node in the future preset time window. Based on the prediction values, the forward comprehensive pressure coefficient of each work node at future time is calculated. The scene recognition and weight adaptation module is used to identify the current operation scenario of the park based on the changing trends of vehicle arrival rate and factory exit rate in the fused dataset and the changing trends of pressure coefficients of each operation node, and to call up the weight coefficients that match the operation scenario based on the recognition results. The cold start transfer learning module is used to select mature nodes similar to the target new node based on physical feature similarity as the matching condition for newly put into use job nodes. The historical time series data of the mature node is then scaled and transferred to the target new node as the initial training sample for training the prediction model of the node. The integrated scheduling decision and route planning module is used to determine the set of candidate operation nodes based on the transportation task of the target vehicle, calculate the estimated travel time and recommended travel route for each candidate operation node, calculate the comprehensive scheduling cost based on the forward-looking comprehensive pressure coefficient, estimated travel time, route tortuosity, road saturation coefficient and weight coefficient, select the candidate operation node with the lowest comprehensive scheduling cost as the target operation node, encapsulate the target operation node and the corresponding recommended travel route into a scheduling scheme and issue it. The scheme distribution and execution monitoring module is used to distribute the scheduling scheme to each terminal, and after the scheduling scheme is distributed, it obtains the real-time location of the vehicle at a preset time interval and matches it with the recommended path. It calculates the comprehensive matching degree based on the position deviation, direction deviation and progress deviation. When the comprehensive matching degree is lower than the preset threshold or the vehicle deviates from the recommended path for more than the preset time, the execution status is determined to be abnormal. The anomaly handling and closed-loop feedback module is used to receive execution status anomaly information and identify the anomaly type according to the deviation distribution, and perform corresponding autonomous recovery operations according to the anomaly type; after the vehicle completes the operation, it feeds back the actual operation time to the node status calculation and forward prediction module, and determines whether there are any subsequent operation nodes in the vehicle's transportation task. If so, it triggers the integrated scheduling decision and route planning module to re-execute. Beneficial effects

[0017] 1. This invention uses a long short-term memory network to predict the number of vehicles queuing and the operation rate within a preset time window for each operation node, and substitutes the predicted values ​​into the comprehensive pressure coefficient calculation formula to obtain a forward-looking comprehensive pressure coefficient. By directly incorporating the forward-looking comprehensive pressure coefficient into a coupled cost function with the composite dimensions of time and vehicle number, vehicles are guided to the operation node with the lowest queuing pressure at the "expected arrival time of the vehicle" at the time of scheduling decision. This achieves the goal of maintaining the effectiveness and real-time performance of scheduling instructions throughout the entire time dimension, and solves the problem of the time lag between the scheduling decision time and the actual arrival time of the vehicle, which leads to the asynchrony between the decision information and the actual state of the physical world. It also avoids the situation where the originally optimal node has become invalid due to changes in the queuing state when the vehicle arrives.

[0018] 2. This invention calculates the rate of change of vehicle arrival rate and departure rate at the park entrance, and combines this with the trend of pressure coefficient changes at each work node to identify the current operating scenario in real time. At the same time, it automatically switches and calls the system after real-time scenario identification by using a combination of scenario-specific weight coefficients obtained offline through a genetic algorithm. This achieves the goal of automatically outputting adaptive scheduling strategies under different scenarios such as high load, low load, congestion, and equipment failure, thus avoiding the defect of scheduling strategies performing differently in different operating scenarios and being unable to adapt and adjust.

[0019] 3. This invention performs path calculation and cost evaluation simultaneously for each candidate node. By constructing a coupled cost function with physical dimensions, node queuing pressure, estimated travel time, path tortuosity, and road saturation are cross-coupled within the same decision-making framework. In addition, the heuristic function in the path search algorithm is dynamically adjusted according to the real-time road congestion status, so that the scheduling scheme is simultaneously optimal in both the dimensions of node suitability and path traffic efficiency. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system module diagram of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely for explaining the invention and are not intended to limit the scope of protection of the invention. Example 1

[0022] Please see Figure 1 As shown, this embodiment provides an intelligent scheduling method for logistics vehicles in a park based on multi-source data fusion. The method includes the following steps: S1. Acquisition and preprocessing of multi-source real-time data; Real-time data is collected from the park's transportation management system, park management system, access control system, metering system, and warehousing system. The transportation management system collects transportation task information for each vehicle, including vehicle identification, task type, task priority, and planned operation nodes. Task types include three categories: procurement entry, sales exit, and in-park transshipment. The park management system collects real-time location information for each vehicle and vehicle density for each road section. The access control system collects vehicle entry and exit records and the total number of vehicles in the park, thereby calculating the park entrance vehicle arrival rate, denoted as [data missing]. The unit is vehicles per minute. The system collects the operating status and number of weighbridges at each location from the metering system. Operating status includes three categories: idle, occupied, and faulty. The system also collects the inventory levels of each warehouse, the operational status of loading and unloading points, and the number of vehicles queuing from the warehousing system.

[0023] The collected system data are formatted and standardized, converted into a unified data dictionary format according to a preset field mapping table, and aligned with a unified time series label to form a fused dataset. Simultaneously, at fixed time intervals, such as every 30 seconds, the status data of each job node is written to a time series database for caching. The cached historical data covers at least the most recent 24 hours and is used for time series prediction and scene recognition in subsequent step S2.

[0024] S2. Calculate the forward-looking comprehensive pressure coefficient of each operation node and identify the current operation scenario based on the fused dataset; Based on the fused dataset obtained above, the current comprehensive pressure coefficient of each operation node within the park is calculated. Operation nodes include at least each loading / unloading point and each weighbridge.

[0025] Loading and unloading points Current comprehensive pressure coefficient The calculation formula is: ; in, This represents the current number of vehicles in the queue. The maximum queuing capacity is determined in advance based on the physical dimensions of the work lane and the safe distance between vehicles. The standard loading and unloading rate is based on the truncated average of historical data from the past thirty days during non-congested periods, and is updated daily. Non-congested periods refer to periods when the road saturation is below 0.6. The current actual loading and unloading rate is calculated from the average operation time of the most recently completed preset number of vehicles; This represents the total planned work time for the day. This represents the total accumulated work time. , , For the weighting coefficients, satisfying + + =1.

[0026] when When =0, Replace the calculation with a substitute to avoid division by zero errors.

[0027] Weighbridge Current comprehensive pressure coefficient The calculation formula is: ; in, This represents the current number of vehicles in the queue. To the maximum queuing capacity, This represents the average time taken for weighing recently. Standard weighing time, This is the working status coefficient. , , For the weighting coefficients, satisfying + + =1. The working status coefficient is 1 when the current working status is idle, 1.2 when it is occupied, and 10 when it is faulty.

[0028] Current road saturation coefficient in the park The calculation formula is: ; In the formula, For the first The vehicle density of a road segment is obtained by dividing the number of vehicles in that segment by the segment length. The road congestion density threshold is determined based on the physical speed limit of the park's roads, the average headway, and the safety reduction factor.

[0029] After calculating the current overall pressure coefficient of each work node, the current operational scenario of the park is further identified. Specifically: Calculate the vehicle arrival rate at the park entrance first derivative and the rate of vehicle departure from the park first derivative Simultaneously, the first derivative of the pressure coefficient at each work node is calculated; the judgment rules are as follows: when When the pressure coefficient at the procurement unloading point is greater than the pressure coefficient at the sales loading point, it is determined to be a procurement-driven high-load scenario. when When the pressure coefficient at the sales loading point increases at a rate greater than 0, it is determined to be a high-load scenario dominated by the factory. When the first derivative of the pressure coefficient of each node approaches zero in three consecutive sampling periods, each sampling period is 30 seconds, and the arrival rate of vehicles at the entrance is less than 60% of the moving average of the past 24 hours, it is judged as a low load scenario. When the road saturation coefficient When the threshold is exceeded, it is determined to be an abnormal congestion scenario; When any node's operating condition coefficient significantly increases and exceeds the pressure coefficient threshold, it is determined to be a device failure scenario. The preferred pressure coefficient threshold is 2.5, which can be adjusted within the range of 2.0 to 3.0 based on actual operating data from the park.

[0030] Furthermore, when multiple scenario conditions are met simultaneously, a unique scenario is determined according to the priority order of abnormal congestion, equipment failure, high load, and low load.

[0031] Based on this, a forward-looking comprehensive pressure coefficient is calculated using time series forecasting methods.

[0032] Specifically: Using cached historical state data of each job node as training data, a Long Short-Term Memory (LSTM) network is employed to predict the number of queuing vehicles and the job rate within a preset time window. The preset time window has three time granularities, such as 5 minutes, 10 minutes, and 15 minutes. In this embodiment, the LSTM network uses a two-layer stacked structure, with each layer containing 64 LSTM units. The Adam optimizer is used for training, with a learning rate of 0.001, a batch size of 32, and 200 training epochs. An early stopping strategy is adopted, terminating training when the validation set loss no longer decreases within 20 consecutive epochs. The training set and validation set are divided in an 8:2 ratio. The input features of the LSTM network include the following five dimensions: the historical sequence of the number of queuing vehicles at the node itself, the historical sequence of the job rate at the node, the upstream vehicle arrival rate at the current moment, the estimated completion time of the currently queuing vehicles, and the remaining time of the traffic light phase at the park entrance. This data is obtained in real-time from the traffic light control module of the park access control system. The network output consists of the predicted number of queuing vehicles and the predicted operation rate for each time window within the future. Substituting these values ​​into the pressure coefficient calculation formula above, a sequence of predicted comprehensive pressure coefficient values ​​for the node at each future time is obtained.

[0033] Based on the estimated travel time of the target vehicle, the predicted value at the corresponding moment is selected from the sequence as the forward-looking comprehensive pressure coefficient for that node. The estimated travel time is calculated separately for each candidate work node; the specific calculation method is described in step S3.

[0034] For newly deployed job nodes, i.e., nodes with insufficient cached historical data to meet the preset training sample threshold, the system introduces a transfer learning mechanism based on physical feature similarity to address the cold start problem. Specifically: The system automatically retrieves existing mature operation nodes, i.e. nodes whose cached data volume has reached a preset threshold, and extracts the physical feature vector of each mature node. The physical feature vector includes at least the platform physical dimensions, the number of operation channels, the type and rated power of loading and unloading equipment, and the historical average distribution of operation categories. The platform physical dimensions include length and width. The mature node most similar to the target new node is selected based on the similarity of the platform's physical size and the similarity of the historical average operation category. The similarity of the platform's physical size refers to the Euclidean distance between the two being less than a preset threshold, preferably greater than 90%. The similarity of the historical average operation category is based on Jaccard similarity, preferably greater than 80%. After a successful match, the historical time series data of the mature node is used as the source domain data. The ratio of the physical size of the target new node to the physical size of the source node is used as the scaling factor. The source domain data is linearly scaled and then transferred to the initial training samples of the target new node for training the LSTM prediction model of the node. If no suitable mature node is found, the average historical data of all mature nodes in the park, after global scaling, is used as the initial training sample. This ensures that the new node still has working forward-looking predictive capabilities during the cold start phase when it lacks its own historical data, and that the migration sample is gradually replaced by real data as the node's actual operating data accumulates.

[0035] S3. Generate an integrated scheduling scheme for target vehicles based on the forward-looking comprehensive pressure coefficient. The integrated scheduling scheme includes target operation node allocation and recommended driving route. Once a target vehicle enters the park or completes its current work node, its transportation task information is retrieved to determine the type of work to be performed. This then determines a set of available candidate work nodes, including procurement unloading points, sales loading points, or transshipment loading and unloading points. When the same task type corresponds to multiple candidate nodes, all of them are included in the candidate set.

[0036] For each candidate node, path calculation and cost evaluation are performed simultaneously. Specifically, starting from the vehicle's current position and ending at the candidate node, the cost is calculated based on the current road saturation coefficient. Based on real-time traffic conditions, the system uses the A* algorithm to calculate recommended driving routes and corresponding estimated travel times. and path length In the extension process of the A* algorithm, heuristic functions... Dynamically adjusted to: ; In the formula, This represents the Euclidean distance from the current node to the target node. The preset adjustment coefficient reflects the sensitivity of the dispatch system to road congestion. The value ranges from 0.3 to 0.8, with a default value of 0.5. It can be adjusted by fitting the actual operation data of the park.

[0037] Based on this, a coupling cost function with physical dimensions is constructed. : ; In the formula, The forward-looking comprehensive pressure coefficient of the candidate node obtained in step S2, in units of vehicles. The estimated travel time is expressed in minutes. The product of the two values ​​has the dimension of "vehicle-minutes", which physically represents the total number of minutes that the queue of vehicles is expected to wait when the vehicle arrives at that node. The path tortuosity factor is defined as the ratio of the recommended driving path length to the Euclidean distance from the start point to the end point. ≥1; This represents the current road saturation coefficient. To prevent division by zero, the value is set to 0.001; These are the weighting coefficients corresponding to the current scenario; This represents the physical speed limit within the park. This cost function incorporates a forward-looking pressure coefficient into the decision objective, ensuring that vehicles are directed to work nodes with lower queuing pressure at the "estimated arrival time of the vehicle" at the moment of scheduling decision-making. This directly solves the problem of scheduling information being out of sync with the physical world due to time lag.

[0038] The candidate node with the lowest overall scheduling cost is selected as the target job node, and the recommended travel path corresponding to the target job node in the aforementioned calculation is encapsulated together into an integrated scheduling scheme data package.

[0039] S4. Issue the integrated dispatch plan and monitor the vehicle execution status; The generated scheduling plan is distributed via wireless network to the driver's mobile terminal, the park's large-screen display terminal, and the access control system. The wireless network includes cellular networks and the park's WiFi network. The driver's terminal presents the target work node, queue number, and navigation route through text and voice announcements; the large-screen display terminal marks the vehicle's assigned location and route using the park's electronic map as the background; the access control system sets the corresponding access control to authorized access status according to the scheduling plan, automatically raising the barrier to allow authorized vehicles to pass, and refusing access to unauthorized vehicles and recording the anomaly.

[0040] After the scheduling plan is issued, vehicle coordinates are continuously acquired through real-time location data collection at preset time intervals, such as every 10 seconds, and matched with the recommended path. A multi-dimensional matching index is established for each vehicle, including three dimensions: position deviation, direction deviation, and progress deviation. Position deviation is the ratio of the straight-line distance between the vehicle's current position and the target node to the estimated remaining distance; direction deviation is the cosine of the angle between the current driving direction and the tangent direction of the recommended path at that position; progress deviation is the ratio of the travel time already traveled to the theoretical remaining time. The weighted sum of these three dimensions yields the comprehensive matching degree between the vehicle and the currently assigned node. An execution status abnormality is determined when the comprehensive matching degree falls below a preset threshold. The threshold is equal to three times the circular error of the vehicle's GPS module, and this threshold automatically adapts to different vehicle GPS hardware models. When the shortest distance between the vehicle's current position and the recommended driving path exceeds this threshold and continues for more than a preset time, such as 60 seconds, the vehicle is determined to have deviated from the recommended driving path.

[0041] S5. Based on the execution monitoring results, perform anomaly handling, autonomous recovery, and closed-loop feedback; The estimated travel time is multiplied by a tolerance factor, for example, 1.5, to determine the allowed execution time. Timing begins from the moment the scheduling plan is issued. A timeout detection is triggered when the timing exceeds the allowed execution time. If, within the allowed execution time, the overall matching degree is detected to be below a threshold or the vehicle deviates from the path, an anomaly handling process is triggered, including the following three stages: The first stage is automatic anomaly type identification. Based on the deviation distribution of the sub-items of the matching degree in step S4, the anomaly type is automatically identified. Specifically: when the position deviation is lower than the preset threshold while the direction deviation is higher than the preset threshold, it is determined to be a path deviation anomaly, indicating that the vehicle is not traveling along the recommended path but the direction is roughly correct; when the position deviation is higher than the preset threshold and the direction deviation is lower than the preset threshold, it is determined to be a node change anomaly, indicating that the vehicle actively drives to other unassigned work nodes; when the progress deviation is higher than the preset threshold and both the position deviation and direction deviation are normal, it is determined to be a stagnation anomaly, indicating that the vehicle has not moved for a long time; when multiple dimensions are abnormal at the same time, it is determined to be a composite anomaly.

[0042] The second phase is the autonomous recovery attempt. Based on the anomaly type identified in the first phase, corresponding autonomous recovery operations are executed. For path deviation anomalies, the recommended driving route is recalculated based on the current road saturation coefficient, starting from the vehicle's current location and ending at the originally assigned target work node. An updated scheduling plan is then automatically distributed to each terminal without manual intervention. For node change anomalies, the system assesses whether the node the vehicle is currently heading to is suitable for performing the vehicle's transportation task: if the node matches the vehicle's task type and the node's forward-looking comprehensive pressure coefficient is lower than a preset acceptable threshold, the node is automatically updated as the vehicle's target work node, and a new recommended driving route is replanned. An updated scheduling plan is then automatically distributed to each terminal. If the node does not match the vehicle's task type or the forward-looking comprehensive pressure coefficient exceeds a preset acceptable threshold, the system does not execute autonomous recovery but instead executes the third phase's external constraint intervention process.

[0043] For stationary anomalies, the anomaly information of the vehicle is pushed to the monitoring terminal of the park management personnel. The anomaly information includes the current location, stationary duration, and task type, which the management personnel verify to determine whether it is a normal parking situation such as vehicle malfunction or waiting for loading / unloading. For complex anomalies, the process directly proceeds to the third stage of external constraint intervention.

[0044] The third stage involves external constraint intervention and automatic system replanning. When a vehicle anomaly is classified as a node change anomaly but does not meet the conditions for autonomous recovery, a stagnation anomaly, or a composite anomaly, or when autonomous recovery attempts fail (e.g., the vehicle still cannot arrive on time after replanning the route), an alarm is pushed to the monitoring terminal of the park management personnel. Simultaneously, the system provides a mandatory command input interface. Park management personnel input the desired target work node ID as an external constraint through this interface. After this constraint is added to the optimization model, the system automatically replans the remaining routes and issues the scheduling scheme. Upon receiving the constraint, the system uses it as the mandatory endpoint and the vehicle's current location as the starting point, recalculates the recommended driving route based on the current road saturation coefficient, generates an updated scheduling scheme, and automatically issues it to each terminal, then returns to step S4 to continue execution. Whether the updated scheduling scheme is generated by the system's autonomous recovery or by automatic replanning after external constraint intervention, it is automatically synchronized to the driver's mobile terminal, the large-screen display terminal, and the access control management system to ensure data consistency across all terminals.

[0045] After a vehicle arrives at the target work node and completes loading, unloading, or weighing, the corresponding metering or warehousing system records the completion time. The system subtracts the vehicle arrival time from the completion time to obtain the actual work time, and feeds this actual work time back to the data center to update the forward-looking comprehensive pressure coefficient of the corresponding work node. Specifically, the actual work time is weighted and averaged with the average work time of the most recently completed preset number of vehicles to update the current actual loading / unloading rate or the average weighing time of the most recently completed vehicles. Then, the system returns to step S2 to recalculate the current comprehensive pressure coefficient and forward-looking comprehensive pressure coefficient of the work node, so that subsequent scheduling decisions for the node can be based on the updated status.

[0046] After a vehicle completes its current task, the system checks whether there are any unfinished subsequent task tasks in the vehicle's transportation task information. If there are, the system returns to step S3, and the integrated scheduling scheme is generated again, starting from the vehicle's current location. If all tasks for the vehicle have been completed, the system records the vehicle as ready to leave the factory and releases it through the access control system, ending the vehicle's scheduling process. Example 2

[0047] Please see Figure 2 As shown, this embodiment provides an intelligent dispatching system for park logistics vehicles based on multi-source data fusion, including the following modules: The multi-source data acquisition and fusion module interfaces with the park's transportation management system, park management system, access control system, metering system, and warehousing system, respectively, to retrieve or receive real-time data from each system according to a preset acquisition frequency. The collected data includes vehicle transportation task information, real-time vehicle location, vehicle density in road sections, vehicle entry and exit records, weighbridge operating status and weighing quantity, warehouse inventory, and the number of vehicles queuing at loading and unloading points. This module performs format standardization and time-series alignment on the collected data from each system to form a fused dataset, and writes the status data of each operation node into a time-series database for caching at fixed time intervals, providing real-time data support and historical time-series data support for subsequent modules.

[0048] The node status calculation and forward prediction module, based on the fused dataset output by the multi-source data acquisition and fusion module and cached historical data, calculates the current comprehensive pressure coefficient of each operation node within the park, including the comprehensive pressure coefficient of loading and unloading points and weighbridges, as well as the current road saturation coefficient of the park. Building upon this, the module uses a Long Short-Term Memory (LSTM) network to predict the number of queuing vehicles and the operation rate of each operation node within a preset future time window. Substituting these predicted values ​​into the comprehensive pressure coefficient calculation formula yields a sequence of predicted comprehensive pressure coefficients for each node at future times. The module then selects the corresponding predicted value as the forward-looking comprehensive pressure coefficient based on the expected travel time of the target vehicle. The output of this module is used for cost evaluation in the subsequent scheduling decision module.

[0049] The scene recognition and weight adaptation module, based on vehicle arrival rate and departure rate data output by the multi-source data acquisition and fusion module, calculates the first derivative of the vehicle arrival rate and departure rate at the park entrance. Simultaneously, it receives the pressure coefficients and their first derivatives for each work node from the node status calculation and forward prediction module. Following preset rules based on the dynamic change rate of physical traffic flow, it identifies the current operational scenario of the park in real time, including procurement-driven high-load scenarios, departure-driven high-load scenarios, low-load scenarios, abnormal congestion scenarios, and equipment failure scenarios. This module has a built-in weight coefficient combination corresponding to each scenario obtained through an offline optimization engine. Based on the real-time identified scenario, it automatically calls the matching weight coefficient and outputs it to the scheduling decision module.

[0050] The cold start transfer learning module addresses newly deployed work nodes (nodes with insufficient cached historical data to meet a preset training sample threshold). This module automatically retrieves existing mature work nodes, extracts their physical feature vectors, and selects the most similar mature node based on platform physical size similarity and historical average work category similarity. The historical time-series data of the most similar mature node is then scaled and transferred to serve as the initial training samples for training the node's LSTM prediction model. If no suitable mature node is found, the average historical data of all mature nodes in the park, scaled globally, is used as the initial training sample. This module ensures that new nodes retain working forward-looking predictive capabilities even during the cold start phase when they lack their own historical data.

[0051] The integrated scheduling decision-making and route planning module determines a set of candidate operation nodes based on the transportation task type of the target vehicle. For each candidate node, taking the vehicle's current location as the starting point and the candidate node as the ending point, it calculates the recommended driving route and the corresponding estimated travel time and path length using the A* algorithm based on the current road saturation coefficient and real-time traffic conditions. The heuristic function of the A* algorithm is dynamically adjusted according to the road saturation coefficient. This module also constructs a coupled cost function with physical dimensions. The forward-looking comprehensive pressure coefficient output by the node state calculation and forward prediction module and the scene weight output by the scene recognition and weight adaptation module are substituted into the cost function to calculate the comprehensive scheduling cost of each candidate node. The candidate node with the lowest comprehensive scheduling cost is selected as the target operation node, and the corresponding recommended driving route is encapsulated into an integrated scheduling scheme data package.

[0052] The scheduling plan distribution and execution monitoring module distributes the scheduling plan output by the integrated scheduling decision and route planning module to the driver's mobile terminal, the park's large-screen visualization terminal, and the access control system via a wireless network. The driver's terminal presents the target work node, queue number, and navigation route via text and voice broadcast; the large-screen terminal marks the vehicle's assigned location and route on the park's electronic map; and the access control system automatically raises the barrier to allow authorized vehicles to pass according to the scheduling plan. After the scheduling plan is distributed, this module continuously acquires vehicle coordinates at preset time intervals through real-time location acquisition and matches them with the recommended route. It establishes a multi-dimensional matching degree index, including position deviation, direction deviation, and progress deviation. The weighted sum is then used to obtain the comprehensive matching degree. When the comprehensive matching degree is lower than a preset threshold or the vehicle deviates from the recommended route for more than a preset time, the execution status is determined to be abnormal, and the monitoring results are output to the anomaly handling and feedback module. The preset threshold is equal to three times the circular error probability of the vehicle's GPS module and automatically adapts to different vehicle GPS hardware models.

[0053] The anomaly handling and closed-loop feedback module receives monitoring results from the scheme issuance and execution monitoring module. When the overall matching degree is detected to be below a threshold or the vehicle deviates from the path, the anomaly handling process is triggered. This module first automatically identifies the anomaly type based on the deviation distribution of each dimension of the matching degree, including path deviation, node change, stagnation, and composite anomalies. Then, it performs corresponding autonomous recovery attempts based on the anomaly type, including replanning the route and automatically updating the target node. When autonomous recovery fails or conditions are not met, the module provides a forced command input interface. Management personnel input the target node ID as an external constraint through this interface, and the system automatically completes route replanning and scheme issuance. After the vehicle completes its work, this module feeds back the actual work time to the node status calculation and forward prediction module, updating the standard rate parameters of the corresponding node. Simultaneously, it determines whether the vehicle has subsequent work nodes. If so, it triggers the integrated scheduling decision and route planning module to re-execute. If all tasks are completed, it records the vehicle's ready-to-leave status and releases it through the access control system.

[0054] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for intelligent scheduling of logistics vehicles in industrial parks based on multi-source data fusion, characterized in that, include: Real-time data from multiple sources is collected from various information systems within the park and then merged to form a fused dataset. The current comprehensive pressure coefficient of each work node is calculated based on the fused dataset, and the number of queuing vehicles and the operation rate of each work node in the future preset time window are predicted using a time series prediction model. The forward-looking comprehensive pressure coefficient of each work node at future time is calculated based on the predicted values. Based on the changing trends of vehicle arrival rate and factory exit rate in the fused dataset, as well as the changing trends of pressure coefficients at each work node, the current operating scenario of the park is identified, and weight coefficients matching the operating scenario are called according to the identification results. Based on the transportation task of the target vehicle, a set of candidate operation nodes is determined. For each candidate operation node, the estimated travel time and recommended travel route from the current location of the vehicle to the candidate operation node are calculated. Based on the forward-looking comprehensive pressure coefficient, estimated travel time, path tortuosity, road saturation coefficient and weight coefficient, the comprehensive scheduling cost of the candidate operation node is calculated. The candidate operation node with the lowest comprehensive scheduling cost is selected as the target operation node. The target operation node and the corresponding recommended travel route are encapsulated into a scheduling scheme and issued.

2. The intelligent scheduling method for logistics vehicles in industrial parks based on multi-source data fusion according to claim 1, characterized in that, A time series prediction model is used to predict the number of queuing vehicles and the operation rate of each work node within a future preset time window. This includes: using the historical queuing vehicle number sequence, historical operation rate sequence, upstream vehicle arrival rate, estimated completion time of the current queuing vehicles, and remaining time of the traffic light phase at the park entrance as input features, a long short-term memory network is used to predict the number of queuing vehicles and the operation rate of the work node at each moment within the future preset time window; based on the estimated travel time of the target vehicle to the work node, the predicted value at the corresponding moment is selected from the prediction results and substituted into the comprehensive pressure coefficient calculation formula to obtain the forward-looking comprehensive pressure coefficient.

3. The intelligent scheduling method for park logistics vehicles based on multi-source data fusion according to claim 1, characterized in that, Identify the current operational scenario of the park, including: determining whether the park is in a high-load or low-load scenario based on the changing direction of the park's vehicle arrival rate and departure rate, as well as the changing direction of the pressure coefficient of each work node; determining whether the park is in an abnormal congestion scenario based on the road saturation coefficient; and determining whether the park is in an equipment failure scenario based on the pressure coefficient of each work node.

4. The intelligent scheduling method for park logistics vehicles based on multi-source data fusion according to claim 3, characterized in that, Identifying the current operational scenario of the park also includes: when multiple scenario conditions are met simultaneously, determining a unique scenario according to the priority order of abnormal congestion scenario, equipment failure scenario, high load scenario, and low load scenario.

5. The intelligent scheduling method for park logistics vehicles based on multi-source data fusion according to claim 1, characterized in that, Calculate the estimated travel time and recommended travel path from the vehicle's current location to the candidate task node, including: taking the vehicle's current location as the starting point and the candidate task node as the ending point, and calculating the recommended travel path and the corresponding estimated travel time and path length based on the current road saturation coefficient and real-time traffic conditions using a path search algorithm; wherein, the heuristic function of the path search algorithm is dynamically adjusted according to the road saturation coefficient, and the higher the road saturation, the larger the heuristic function value.

6. The intelligent scheduling method for logistics vehicles in industrial parks based on multi-source data fusion according to claim 1, characterized in that, This also includes cold start processing for newly deployed work nodes, including: Retrieve existing mature operation nodes, extract the physical feature vectors of each mature operation node, select mature nodes similar to the target new node based on physical feature similarity as the matching condition, and migrate the historical time series data of the mature node after scaling to become the initial training sample of the target new node for training the prediction model of the node; if no mature node matching the condition is found, the average historical data of all mature nodes in the park after global scaling is used as the initial training sample.

7. The intelligent scheduling method for park logistics vehicles based on multi-source data fusion according to claim 1, characterized in that, After the scheduling plan is issued, it also includes execution monitoring and anomaly handling: real-time vehicle location is obtained at preset time intervals and matched with the recommended path, and the comprehensive matching degree is calculated based on the position deviation, direction deviation and progress deviation; when the comprehensive matching degree is lower than the preset threshold or the vehicle deviates from the recommended path for more than the preset time, the anomaly handling process is triggered, the anomaly type is identified according to the deviation distribution, and the corresponding autonomous recovery operation is performed according to the anomaly type. The autonomous recovery operation includes replanning the path or automatically updating the target work node.

8. The intelligent scheduling method for logistics vehicles in a park based on multi-source data fusion according to claim 7, characterized in that, Based on the anomaly type, corresponding autonomous recovery operations are performed, including: when the anomaly type is path deviation, the recommended driving route is recalculated and the scheduling plan is updated, starting from the vehicle's current location and ending at the original target work node; when the anomaly type is node change, it is assessed whether the node the vehicle is currently heading to is suitable for performing the vehicle's transportation task and whether the forward-looking comprehensive pressure coefficient is lower than a preset threshold. If the conditions are met, the node is updated as the target work node and the route is replanned; if not, external constraints are triggered; when autonomous recovery fails or the anomaly type does not meet the autonomous recovery conditions, the target work node input by the management personnel is received as a mandatory constraint, and the recommended driving route is recalculated and the scheduling plan is updated, with the mandatory constraint as the endpoint.

9. The intelligent scheduling method for logistics vehicles in industrial parks based on multi-source data fusion according to claim 1, characterized in that, It also includes closed-loop feedback: after the vehicle arrives at the target work node and completes the work, the actual work time is recorded and fed back to the data center to update the work rate parameters of the corresponding work node; it is determined whether there are any subsequent work nodes in the vehicle's transportation task. If so, the scheduling plan is re-executed and generated. If all are completed, the vehicle is released to leave the factory.

10. A smart dispatching system for logistics vehicles in industrial parks based on multi-source data fusion, characterized in that: include: The multi-source data acquisition and fusion module is used to collect real-time data from various information systems in the park and perform format unification and time sequence alignment to form a fused dataset. The node status calculation and forward prediction module is used to calculate the current comprehensive pressure coefficient of each work node based on the fused dataset, and to use a time series prediction model to predict the number of queuing vehicles and the operation rate of each work node in the future preset time window. Based on the prediction values, the forward comprehensive pressure coefficient of each work node at future time is calculated. The scene recognition and weight adaptation module is used to identify the current operation scenario of the park based on the changing trends of vehicle arrival rate and factory exit rate in the fused dataset and the changing trends of pressure coefficients of each operation node, and to call up the weight coefficients that match the operation scenario based on the recognition results. The cold start transfer learning module is used to select mature nodes similar to the target new node based on physical feature similarity as the matching condition for newly put into use job nodes. The historical time series data of the mature node is then scaled and transferred to the target new node as the initial training sample for training the prediction model of the node. The integrated scheduling decision and route planning module is used to determine the set of candidate operation nodes based on the transportation task of the target vehicle, calculate the estimated travel time and recommended travel route for each candidate operation node, calculate the comprehensive scheduling cost based on the forward-looking comprehensive pressure coefficient, estimated travel time, route tortuosity, road saturation coefficient and weight coefficient, select the candidate operation node with the lowest comprehensive scheduling cost as the target operation node, encapsulate the target operation node and the corresponding recommended travel route into a scheduling scheme and issue it. The scheme distribution and execution monitoring module is used to distribute the scheduling scheme to each terminal, and after the scheduling scheme is distributed, it obtains the real-time location of the vehicle at a preset time interval and matches it with the recommended path. It calculates the comprehensive matching degree based on the position deviation, direction deviation and progress deviation. When the comprehensive matching degree is lower than the preset threshold or the vehicle deviates from the recommended path for more than the preset time, the execution status is determined to be abnormal. The anomaly handling and closed-loop feedback module is used to receive execution status anomaly information and identify the anomaly type according to the deviation distribution, and perform corresponding autonomous recovery operations according to the anomaly type; after the vehicle completes the operation, it feeds back the actual operation time to the node status calculation and forward prediction module, and determines whether there are any subsequent operation nodes in the vehicle's transportation task. If so, it triggers the integrated scheduling decision and route planning module to re-execute.

Images