Artificial intelligence-based oversize load transportation path dynamic planning method

Abstract

The application discloses an over-limit large piece transportation path dynamic planning method based on artificial intelligence and relates to the technical field of artificial intelligence path planning. The method comprises the following steps: acquiring a dynamic space constraint set of an over-limit large piece transportation task and a multi-modal state tensor of a real-time road network; performing an accessible domain filtering operation on the multi-modal state tensor based on the dynamic space constraint set to generate a candidate path topology graph; constructing a load-right-of-way collaborative mapping relationship through an edge load sequence in the candidate path topology graph and a node passing cost matrix; performing dynamic backtracking incentive adjustment on the candidate path topology graph according to the mapping relationship and a preset time window relaxation coefficient to obtain an initial planning path set; and performing iterative advantage evaluation on the initial planning path set by using a multi-objective reinforcement learning network to output an optimal dynamic planning path.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence route planning technology, specifically to a dynamic planning method for transporting oversized and heavy cargo based on artificial intelligence. Background Technology

[0002] Route planning for oversized and heavy cargo transportation involves a complex matching problem between dynamic spatial constraints such as cargo dimensions and axle load distribution and multi-dimensional traffic capacity parameters in the real-time road network, including height limits, width limits, load-bearing capacity, and radius of curvature. Existing route planning technologies typically rely on static road network attribute data, simplifying oversized constraints into a single-dimensional traffic condition filter, and then performing route search based on the shortest path or minimum time cost. These approaches have limitations when dealing with dynamically changing traffic flow conditions, accessibility determination under multi-dimensional spatial constraints, and the collaborative effects between load and right-of-way.

[0003] In existing technical solutions, the representation of road network status is mostly based on single-modality or low-dimensional attribute tables, which cannot simultaneously integrate status information from multiple dimensions such as timestamps, spatial curvature, and traffic flow. This leads to a severe mismatch between the assessment of the actual capacity of road arc segments and the spatial envelope requirements of multiple characteristic load distribution surfaces of oversized cargo. Furthermore, the load status and passage costs of edges and nodes in the road network are often treated in isolation, lacking a quantitative construction of the collaborative mapping relationship between historical load fluctuations and transportation time windows and signal control cycles. This makes it difficult to accurately reflect the impact of load accumulation on right-of-way changes during path expansion. In addition, when selecting the final path from candidate paths, existing methods mostly use single-objective evaluation functions, lacking a mechanism for multi-objective joint evaluation and dynamic trade-off of comprehensive transportation costs and passage risks. The generated paths fail to achieve a balance between transportation economy and traffic safety.

[0004] To address the aforementioned issues, it is necessary to solve the challenges of how to utilize dynamic spatial constraint sets and dimension-wise matching of real-time road network multimodal state tensors to achieve high-precision reachability domain screening and establish load-right-of-way collaborative mapping relationships, as well as how to conduct parallel evaluation and advantage ranking of transportation costs and risks through multi-objective reinforcement learning under multi-path schemes. Summary of the Invention

[0005] This invention provides a dynamic planning method for transporting oversized and overweight cargo based on artificial intelligence. The purpose is to solve the problems in the existing technology that the single dimension of road network state representation cannot accurately match the multi-feature spatial constraints of oversized cargo, and the lack of multi-objective joint evaluation of transportation costs and traffic risks in the route optimization process.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a dynamic planning method for the transportation route of oversized and heavy-duty items based on artificial intelligence, applied to the route planning node in a logistics scheduling system. The method includes the following steps: The process involves obtaining a dynamic spatial constraint set for the oversized and heavy-duty transport task and a multimodal state tensor of the real-time road network; performing reachability domain elimination on the multimodal state tensor of the real-time road network based on the dynamic spatial constraint set to generate a candidate path topology; constructing a load-right-of-way collaborative mapping relationship for the oversized and heavy-duty transport task using the edge load sequence and node passage cost matrix in the candidate path topology; dynamically adjusting the candidate path topology based on the load-right-of-way collaborative mapping relationship and a preset time window relaxation coefficient to obtain an initial planned path set for the oversized and heavy-duty transport task; and using a multi-objective reinforcement learning network to iteratively evaluate the advantages of the initial planned path set to output the optimal dynamic planning path for the oversized and heavy-duty transport task.

[0007] As a preferred embodiment of the present invention, the process of obtaining the dynamic spatial constraint set includes: analyzing the cargo outline dimension sequence and axle load spectrum distribution parameters of the oversized cargo transportation task; decomposing the cargo outline dimension sequence into multiple characteristic load distribution surfaces according to the coupling point coordinates of the transport vehicle; extracting the spatial envelope vector of each characteristic load distribution surface from the multiple characteristic load distribution surfaces, and taking the set of all spatial envelope vectors as the dynamic spatial constraint set. This process can transform the complex geometric shape and load-bearing characteristics of oversized cargo into multi-dimensional spatial vectors, providing accurate constraint basis for subsequent road network selection.

[0008] The process of obtaining the real-time road network multimodal state tensor further includes: using the largest spatial envelope vector in the dynamic spatial constraint set as a scanning window, sliding and collecting the capacity parameter set of each road arc segment on the basic road network layer; stacking the capacity parameter sets of each road arc segment into tensors according to the timestamp dimension, spatial curvature dimension, and traffic flow dimension to generate the real-time road network multimodal state tensor; wherein, the capacity parameter set includes height limit threshold, width limit threshold, load-bearing capacity level, and radius of curvature. Through this preferred scheme, the road network data can integrate spatiotemporal multidimensional information to accurately reflect the road's adaptability to overloaded transportation at different times.

[0009] When performing reachability domain filtering on a multimodal state tensor based on a dynamic spatial constraint set, one technical solution of this invention is as follows: traverse the capacity parameter set of each road arc segment in the multimodal state tensor of the real-time road network; compare the capacity parameter set of each road arc segment with each feature load distribution surface in the dynamic spatial constraint set dimension by dimension; when any one-dimensional spatial parameter of any feature load distribution surface exceeds the threshold of the corresponding dimension in the capacity parameter set, mark the road arc segment as an unreachable arc segment; remove all unreachable arc segments from the multimodal state tensor of the real-time road network, and form the candidate path topology graph with the remaining road arc segments and their connection relationships. This filtering operation can eliminate all road segments that cannot meet the over-limit passage requirements in advance, so that subsequent path search is only performed in the high-feasibility subset, significantly improving planning efficiency.

[0010] Preferably, during the dimension-by-dimensional comparison, if the difference in any dimension is less than a preset tolerance threshold, then that dimension is determined not to exceed the threshold. Introducing a tolerance mechanism can avoid erroneously eliminating feasible edge segments due to data measurement errors or minor exceedances, thereby improving the integrity and robustness of the path topology.

[0011] The construction of the load-right-of-way collaborative mapping relationship further includes: extracting the historical load fluctuation sequence of each edge and the signal control cycle sequence of each node in the candidate path topology map; performing a convolution operation on the historical load fluctuation sequence of each edge with the axle load spectrum distribution parameters of the oversized and heavy-duty transport task to obtain the dynamic load response value of each edge; cross-correlatedly linking the signal control cycle sequence of each node with a preset transport time window to obtain the spatiotemporal passage cost of each node; and combining the dynamic load response values ​​of all edges and the spatiotemporal passage costs of all nodes to form the load-right-of-way collaborative mapping relationship. This mapping relationship realizes the coordinated quantification of road segment carrying capacity and node passage rights at the time-series level, providing a dynamic cost evaluation benchmark for path optimization.

[0012] When performing convolution operations, it is preferable to use a one-dimensional dilated convolution kernel to match the load response characteristics at different time scales. Dilated convolution can increase the receptive field without increasing the computational load, enabling the load response value to comprehensively reflect the coupled impact of short-term and long-term historical load fluctuations on the current transportation task.

[0013] When generating the initial planned path set, the technical solution of this invention includes: taking the starting point of the oversized and heavy-duty transport task as the root node, performing breadth-first expansion in the candidate path topology graph according to the dynamic load response values ​​of the edges in the load-right-of-way collaborative mapping relationship from smallest to largest, generating multiple expansion paths; when any expansion path reaches the intersection node in the candidate path topology graph, calculating the cumulative spatiotemporal cost of the expansion path from the starting point to the intersection node; determining whether the cumulative spatiotemporal cost exceeds the product of the time window relaxation coefficient and the theoretical shortest cost; if so, applying a backtracking incentive penalty to the expansion path and deleting the terminal branch of the expansion path; and using the set of all expansion paths that have not been deleted as the initial planned path set. This dynamic backtracking incentive mechanism can automatically prune inferior branches whose costs significantly deviate from the optimal value during the expansion process, retain candidate paths with competitive transportation timeliness, and effectively constrain the search space.

[0014] The backtracking incentive penalty is calculated as follows: the difference between the accumulated spatiotemporal cost and the product of the time window relaxation coefficient and the theoretical shortest cost is multiplied by a preset penalty coefficient, and this penalty value is added to the accumulated spatiotemporal cost of the expansion path. This quantitative penalty mechanism guides the search backtracking to low-cost regions, avoiding premature entrapment in local depth expansion.

[0015] When using a multi-objective reinforcement learning network for iterative evaluation, a preferred technical solution includes: encoding each initial planned path in the initial planned path set as a state feature vector sequence; inputting the state feature vector sequence into the value evaluation branch of the multi-objective reinforcement learning network to calculate the estimated comprehensive transportation cost of each initial planned path; inputting the state feature vector sequence into the risk evaluation branch of the multi-objective reinforcement learning network to calculate the traffic risk confidence level of each initial planned path; sorting the initial planned path set in descending order based on the weighted difference between the estimated comprehensive transportation cost and the traffic risk confidence level, and selecting the initial planned path at the top of the sorted list as the optimal dynamic planning path. This network can simultaneously perform forward-looking evaluations from both transportation cost and traffic safety dimensions, balancing transportation economy and reliability through a weighted decision-making mechanism.

[0016] The process of encoding the initial planned path further includes: extracting the capacity parameter set of the road arc to which each node belongs and the traffic flow density at each node in sequence according to the node order in the initial planned path; concatenating the capacity parameter set and traffic flow density of each node to form the node feature vector of that node; and arranging the node feature vectors of all nodes into the state feature vector sequence according to the node order in the initial planned path. This encoding method completely preserves the topological order of the path and the local state information of the nodes.

[0017] The processing steps of the value assessment branch include: convolutional downsampling of the state feature vector sequence to extract the local spatial features of each node; bidirectional long short-term memory recursion along the node dimension to generate the forward and backward hidden states of each node; feature fusion of the forward and backward hidden states of each node and inputting them into a fully connected layer to obtain the node cost component of each node; accumulating the node cost components of all nodes in the initial planning path and using the accumulated result as the estimated comprehensive transportation cost of the initial planning path. Through the combination of convolution and bidirectional recursion, the value assessment branch can capture the cumulative cost features of bidirectional dependencies in the path sequence.

[0018] The risk assessment branch process includes: performing a self-attention transformation on the state feature vector sequence to generate a global dependency weight for each node relative to the entire path; reshaping the node feature vector of the corresponding node according to the global dependency weight of each node to obtain the risk feature expression of each node; inputting the risk feature expression of each node into a graph convolutional network to aggregate the risk features of neighboring nodes and output the local risk coefficient of each node; and using the maximum value of the local risk coefficients of all nodes in the initial planned path as the travel risk confidence of the initial planned path. The self-attention mechanism can capture the potential risk association between any two nodes on the path, while the graph convolutional network aggregates the neighborhood risk features, enabling the travel risk confidence to reflect the most prominent bottleneck risk in the entire path.

[0019] The technical effects and advantages provided by the present invention in the above technical solution are as follows: Each feature load distribution surface in the dynamic spatial constraint set is compared dimension-by-dimensionally with the capacity parameter set of each road arc in the real-time road network multimodal state tensor. When any one-dimensional spatial parameter of any feature load distribution surface exceeds the corresponding dimensional threshold, the road arc is marked as an unreachable arc and removed, thereby generating a candidate path topology map that satisfies multidimensional constraints such as cargo outline dimensions and axle load distribution. This process avoids accessibility misjudgments caused by single-dimensional filtering, ensuring that the path search space strictly matches the actual spatial envelope and carrying capacity requirements of oversized cargo, making the removed arcs more accurate and reducing the computational scale of subsequent path expansion.

[0020] The historical load fluctuation sequence of each edge in the candidate path topology is extracted and convolved with the axle load spectrum distribution parameters of the oversized and heavy-duty transport task. The signal control cycle sequence of each node is also extracted and cross-correlated with the transport time window. This combines the edge-level dynamic load response value with the node-level spatiotemporal passage cost into a load-right-of-way collaborative mapping relationship. Based on this mapping relationship, dynamic backtracking incentive adjustments are made during path expansion. When the cumulative spatiotemporal cost of the expanded path exceeds the product of the preset time window relaxation coefficient and the theoretical shortest cost, a backtracking incentive penalty is applied, and the terminal branch is deleted, resulting in the initial planned path set. This approach ensures that the path generation process is guided by both dynamic changes in edge load and node passage time windows. The backtracking penalty mechanism can truncate path branches with uncontrollable costs in the early stages, leaving paths that are both load feasible and time window adaptable.

[0021] A multi-objective reinforcement learning network is used to perform convolutional downsampling and bidirectional long short-term memory recursion on the state feature vector sequence of the initial planned path, accumulating the node cost components to obtain the comprehensive transportation cost estimate. Simultaneously, a risk assessment branch performs self-attention transformation and graph convolutional aggregation on the state feature vector sequence, taking the maximum value of the local risk coefficients of all nodes as the traffic risk confidence score. The weighted difference between the transportation cost estimate and the traffic risk confidence score is used to rank the paths based on their advantages, outputting the optimal dynamic programming path. By evaluating transportation cost and traffic risk in parallel and using risk confidence score to correct the cost ranking, the drawback of selecting a path based solely on a single cost indicator, which may introduce high-risk road segments, is avoided. The output path maintains a low transportation cost while possessing a controllable risk level. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0023] Figure 1 This is a flowchart of a dynamic planning method for transporting oversized and heavy cargo based on artificial intelligence. Figure 2 It is a flowchart of breadth-first expansion of oversized and large-item transportation routes based on cumulative spatiotemporal costs and backtracking incentive penalties; Figure 3 This is a matching analysis diagram of the road arc width limit threshold and the maximum lateral occupancy width; Figure 4 This is a schematic diagram of the historical load fluctuation sequence and dynamic load response value of the candidate path topology graph edges; Figure 5It is a scatter plot of the multi-objective reinforcement learning evaluation of the initial planning path and the optimal dynamic programming path. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] See Figure 1 This invention provides an artificial intelligence-based dynamic planning method for oversized and overweight cargo transportation routes, applied to route planning nodes in a logistics scheduling system. The method includes: obtaining a dynamic spatial constraint set for the oversized and overweight cargo transportation task and a multimodal state tensor of the real-time road network; performing reachability domain elimination on the multimodal state tensor of the real-time road network based on the dynamic spatial constraint set to generate a candidate path topology; constructing a load-right-of-way collaborative mapping relationship for the oversized and overweight cargo transportation task using the edge load sequence and node passage cost matrix in the candidate path topology; dynamically backtracking and adjusting the candidate path topology based on the load-right-of-way collaborative mapping relationship and a preset time window relaxation coefficient to obtain an initial planned path set for the oversized and overweight cargo transportation task; and using a multi-objective reinforcement learning network to iteratively evaluate the advantages of the initial planned path set and output the optimal dynamic planning path for the oversized and overweight cargo transportation task.

[0026] Example 1:

[0027] In practical implementation, the sequence of cargo dimensions and the axle load spectrum distribution parameters for oversized and heavy-duty transport tasks are analyzed. The cargo dimension sequence includes the total length, maximum width, maximum height of the transported cargo, and the segment dimensions at each cross-section. The axle load spectrum distribution parameters include the axle load value, axle group type, and arrangement of adjacent axle spacing for each axle.

[0028] The cargo's external dimensions are decomposed into multiple characteristic load distribution surfaces based on the vehicle's attachment point coordinates. The vehicle's attachment point coordinates include the three-dimensional coordinates of the tractor's saddle center point, the three-dimensional coordinates of the hinge points between each trailer and the front cargo box, and the three-dimensional coordinates of the trailer's outrigger positions. Using these attachment point coordinates, the overall cargo space is segmented along the vehicle's longitudinal direction. Each segmented region corresponds to a section of cargo above the vehicle body, and the boundary of this region is described by a set of three-dimensional contour vertices, forming a characteristic load distribution surface.

[0029] The spatial envelope vector of each characteristic load distribution surface is extracted from multiple characteristic load distribution surfaces. For a single characteristic load distribution surface, the minimum and maximum coordinates of all vertices within the surface in the longitudinal, lateral, and vertical directions in the vehicle coordinate system are calculated to obtain a six-dimensional spatial envelope vector. Each component of the six-dimensional spatial envelope vector represents the minimum and maximum occupied coordinates of the characteristic load distribution surface in the longitudinal, transverse, and vertical directions, respectively. The set of spatial envelope vectors of all characteristic load distribution surfaces is defined as the dynamic spatial constraint set.

[0030] When acquiring the multimodal state tensor of the real-time road network, the maximum spatial envelope vector is determined from the dynamic spatial constraint set. The method for determining the maximum spatial envelope vector is as follows: traverse each spatial envelope vector in the dynamic spatial constraint set, and find the minimum vertical component among all spatial envelope vectors. minimum value Longitudinal maximum component maximum value Horizontal minimum component minimum value Horizontal maximum component maximum value Vertical minimum component minimum value and the vertical maximum component maximum value The above six extreme values ​​constitute the maximum spatial envelope vector. The maximum spatial envelope vector defines the minimum three-dimensional clearance envelope size required for the overall transportation of oversized and heavy equipment during road traffic.

[0031] Using the maximum spatial envelope vector as the scanning window, sliding data acquisition is performed along road arcs on the basic road network layer. The basic road network layer is a spatial database storing road centerline geometry, cross-sectional clearance dimensions, bridge and tunnel clearances, and pavement load-bearing capacity. During the sliding data acquisition process, for each road arc in the basic road network layer, the maximum spatial envelope vector is geometrically bounded to the allowable clearance envelope of the road arc cross-section. When the maximum spatial envelope vector is completely contained within the clearance envelope of the road arc, the capacity parameter set for that road arc is extracted from the basic road network layer. The capacity parameter set consists of four parameters: height limit threshold, width limit threshold, load-bearing capacity, and radius of curvature. The height limit threshold is the maximum vertical height that can be traversed on the road arc; the width limit threshold is the maximum horizontal width that can be traversed on the road arc; the load-bearing capacity is the allowable axle load or total weight of the pavement and bridge / culvert structure on the road arc; and the radius of curvature is the horizontal curve radius of the road arc centerline.

[0032] After collecting the capacity parameter sets for each road arc segment, the capacity parameter sets for each road arc segment are stacked as tensors according to the timestamp dimension, spatial curvature dimension, and traffic flow dimension to generate a multimodal state tensor of the real-time road network. The generated multimodal state tensor... Represented using a four-dimensional tensor, specifically in the form of: ,in This indicates the number of sampling time periods divided along the timestamp dimension. This represents the number of curvature intervals divided along the spatial curvature dimension. This indicates the number of density levels divided along the traffic flow dimension. The fourth dimension has a size of 4, corresponding to the height limit threshold, width limit threshold, load capacity level, and radius of curvature, respectively. For The settings will schedule the start of the oversized and overweight cargo transportation mission according to a fixed time step. Divided into multiple sampling periods, if If set to 15 minutes, then .for The settings involve collecting the curvature radius values ​​of all road arc segments in the basic road network layer, and dividing the curvature radius from its minimum to maximum value range using a logarithmic, equally spaced method. A spatial curvature interval, The value is 5, and the five spatial curvature intervals correspond to extremely high curvature curves, high curvature curves, medium curvature curves, low curvature curves, and straight sections, respectively. For The system is configured to obtain traffic flow density values ​​for each road arc segment during each sampling period from a real-time traffic monitoring system, and then divide the range of traffic flow density values ​​for the entire day into equal intervals. Density levels, The value is 10, and each density level corresponds to a density value range.

[0033] During the stacking process, for each road arc segment acquired, the timestamp at which the road arc segment was acquired during the sliding acquisition is determined, and this timestamp is mapped to the corresponding sampling time period number. , Obtain the radius of curvature value of the road arc segment, and map the radius of curvature value to... One of the spatial curvature intervals is used to obtain the spatial curvature index. , Read the real-time traffic flow density value of the road arc segment during the sampling period, and map the density value to... One of the density levels is used to obtain the density level number. , The four parameters of the road arc segment—height limit threshold, width limit threshold, load-bearing capacity, and radius of curvature—are combined into a four-dimensional vector, which is then filled into the multimodal state tensor of the real-time road network. The middle position is elements When multiple road arcs are mapped to the same When determining the location, the maximum value is taken for each element in the height limit, width limit, and load-bearing level dimensions, while the minimum value is taken for the radius of curvature dimension, to ensure that the most stringent passage constraints are stored at that location. If there exists a certain If no road arcs were captured at this location, then this location... Fill in the four-dimensional vector with all zeros. After filling in all road arcs, the complete multimodal state tensor of the real-time road network is obtained. This multimodal state tensor of the real-time road network is used for subsequent reachability domain filtering operations.

[0034] Example 2:

[0035] In practice, each road arc in the multimodal state tensor of the real-time road network is traversed. For each traversed road arc, the set of traffic capacity parameters stored in the multimodal state tensor of the real-time road network for that road arc is obtained. Each parameter in the set of traffic capacity parameters corresponds to a traffic dimension, namely, the vertical clearance traffic dimension corresponding to the height limit threshold, the lateral clearance traffic dimension corresponding to the width limit threshold, the axle load traffic dimension corresponding to the load-bearing level, and the curve passability dimension corresponding to the radius of curvature.

[0036] Each characteristic load distribution surface is extracted from the dynamic spatial constraint set. For a given characteristic load distribution surface, its spatial parameters in each travel dimension are calculated using its spatial envelope vector. The spatial envelope vector is a six-dimensional vector. This is converted into spatial parameters with four passage dimensions: vertical occupancy height. Depend on The calculation shows that the horizontal width is [missing information]. Depend on The calculated length of the longitudinal section is... Depend on The maximum axle load was calculated. The maximum rated axle load of the axle group corresponding to the characteristic load distribution surface is obtained by extracting the axle load distribution parameters of the oversized and heavy-duty transport mission.

[0037] The capacity parameter set for each road arc segment is compared dimension-by-dimensionally with the spatial parameters of the four traffic dimensions of each characteristic load distribution surface. During the comparison, the vertical occupancy height is considered. Compared with the height limit threshold, the horizontal width occupied will be determined. Compared with the width limit threshold, the vertical occupied length is determined. By combining the radius of curvature, the cornering performance was compared, and the maximum axle load was considered. Comparison with load-bearing capacity. Specifically, the curve passability comparison method is as follows: based on the radius of curvature of the road arc and the road surface widening value, the maximum allowable cargo length under that radius of curvature is found from a pre-calibrated curve geometric passability comparison table, and the longitudinal occupancy length is calculated. Compare with the maximum allowed cargo length.

[0038] If any characteristic load distribution surface satisfies the following condition in at least one of the four travel dimensions, the road arc segment is marked as an unreachable arc segment, and subsequent comparisons of that road arc segment are immediately stopped:

[0039] in: This represents the total number of characteristic load distribution surfaces in the dynamic space constraint set. The index of the characteristic load distribution surface; Indicates the pass dimension number. Corresponding to the vertical clearance dimension, Corresponding to the horizontal clearance dimension, Corresponding to the cornering capability dimension, Corresponding axle load passage dimension; Indicates the first Vertical occupancy height of each characteristic load distribution surface , Indicates the first Lateral occupancy width of each characteristic load distribution surface , Indicates the first Longitudinal occupancy length of each characteristic load distribution surface , Indicates the first Maximum shaft load on each characteristic load distribution surface ; This indicates the height limit threshold for road arc segments. This indicates the width limit threshold for road arc segments. This indicates the maximum length of cargo that can pass through a road arc segment at this radius of curvature. Indicates the load-bearing capacity of the road arc segment; The vertical clearance tolerance threshold is set to 0.05 meters, based on the maximum dynamic bounce height margin of the trailer suspension system during driving. The lateral clearance tolerance threshold is set at 0.05 meters, based on the maximum lateral offset caused by lateral wind deflection and road lateral slope. The threshold for curve passability is set at 0.1 meters, based on the maximum deviation between the trailer trajectory and the rigidity assumption during the turning process. This is the axle load tolerance threshold, set to 0 tons, indicating that the load-bearing capacity comparison is not allowed to exceed any limits.

[0040] After traversing and comparing all road arcs in the multimodal state tensor of the real-time road network, all road arcs marked as unreachable are removed. The remaining road arcs, along with their original spatial connections in the base road network layer, together form a candidate path topology graph. Nodes in the candidate path topology graph are the endpoints of the remaining road arcs, and edges are the remaining road arcs themselves. This candidate path topology graph is used for subsequent construction of load-right-of-way collaborative mapping relationships.

[0041] See Figure 3 In the figure, the horizontal axis represents the width restriction threshold of the road arc segment in meters, and the vertical axis represents the maximum lateral width occupied by the characteristic load distribution surface corresponding to the oversized and heavy-duty transport task, also in meters. Solid dots in the figure are marked as "accessible arc segments," indicating that the width restriction threshold of the corresponding road arc segment meets the lateral occupancy width requirement of the characteristic load distribution surface, conforming to the comparison conditions of the lateral clearance dimension between the road arc segment and the characteristic load distribution surface in Example 2. Crosses are marked as "inaccessible arc segments," indicating that the width restriction threshold of the corresponding road arc segment fails to meet the lateral occupancy width requirement of the characteristic load distribution surface, resulting in the road arc segment being marked as inaccessible and removed. The dashed line in the figure represents the relationship curve between the width restriction threshold and the tolerance threshold, drawn according to the lateral clearance tolerance threshold of 0.05 meters set in Example 2. The area above the dashed line is the inaccessible area, and the area below the dashed line is the accessible area.

[0042] As shown in the figure, most solid dots (reachable arcs) are located below the dashed line, indicating that the road arc is considered reachable when the lateral occupancy width meets the tolerance condition of the road width limit threshold. Conversely, a large number of crosses (unreachable arcs) are located above the dashed line, indicating that the maximum lateral occupancy width of the characteristic load distribution surface exceeds the tolerance range of the road arc's width limit threshold, making the road arc unreachable. The horizontal axis ranges approximately from 1.8 meters to 8 meters, and the vertical axis ranges approximately from 2 meters to 6 meters, indicating the matching between the lateral occupancy width of goods and the real-time road network's lateral clearance threshold in the dynamic spatial constraint set of this embodiment. This figure intuitively reflects the reachability screening effect of the spatial constraint judgment rules set in Embodiment 2 on road arcs in the real-time road network, providing data support for the subsequent construction of candidate path topology maps.

[0043] Example 3:

[0044] In practice, the historical load fluctuation sequence corresponding to each edge is extracted from the candidate path topology graph. For each edge in the candidate path topology graph, the edge corresponds to a preserved road arc segment in the basic road network. The historical transportation database of the logistics scheduling system is accessed, and load records of the road arc segment over multiple consecutive time windows are extracted with a fixed time sampling step to form the historical load fluctuation sequence. ,in This represents the index of an edge in the candidate path topology graph. This indicates the length of the historical load fluctuation sequence, in days. A value of 30 indicates that historical payload data from the past 30 days will be collected. Indicates the first The edge at the 1st The sum of the equivalent total axle load mass of all vehicles passing through the road arc segment to which the edge belongs within a given time window.

[0045] Simultaneously, the signal control cycle sequence for each node is extracted from the candidate path topology map. For a node in the candidate path topology map, this node corresponds to an intersection where two or more retained road arcs in the basic road network intersect. The signal timing scheme for this intersection within a preset transport time window coverage area is obtained from the traffic signal control system, and the signal control cycle sequence is extracted. ,in This represents the index of a node in the candidate path topology graph. This represents the total number of signal periods in the signal control period sequence. Indicates the first The node at the th The duration of a signal cycle, in seconds. Indicates the first The node at the th The duration of the green light in each signal cycle, in seconds, for the direction in which transport vehicles are permitted to travel.

[0046] The dynamic load response value of each edge is obtained by convolving the historical load fluctuation sequence of each edge with the axle load spectrum distribution parameters of the oversized and heavy-duty transport task. The axle load spectrum distribution parameters are obtained from the oversized and heavy-duty transport task and include the axle load values ​​and axle spacing arrangement of each axle of the transport vehicle. The axle load spectrum distribution parameters are encoded into a convolution kernel. Specifically, the axle loads of each axle of the transport vehicle are arranged sequentially from front to back along the longitudinal direction to form a one-dimensional axle load sequence. This one-dimensional axle load sequence is then filled into the convolution kernel window according to the time sequence of the vehicle's arrival at the road arc segment. The weight value of each position within the convolution kernel is equal to the corresponding axle load value. Simultaneously, a one-dimensional dilated convolution kernel is used to match the load response characteristics at different time scales. The dilation rate of the one-dimensional dilated convolution kernel is... The settings are based on the total length and average operating speed of the transport vehicles. The calculation method is as follows ,in This indicates the total length of the transport vehicle, in meters. This indicates the average estimated operating speed of transport vehicles in the urban road network, expressed in meters per second. This represents the time sampling step size of the historical load fluctuation sequence, in seconds. This represents the floor operation. The encoded one-dimensional dilated convolution kernel is then processed at the 1st... Historical load fluctuation sequence of the strip Slide along the upper time dimension to perform a one-dimensional dilated convolution operation, calculate the convolution response value step by step, and take the maximum value among all convolution response values ​​as the first value. Dynamic load response value of the strip .

[0047] The signal control cycle sequence of each node is cross-correlated with a preset transportation time window to obtain the spatiotemporal passage cost of each node. The preset transportation time window defines the start and end times of the oversized and heavy-duty transport task that are allowed to pass through each node. This transportation time window is issued by the logistics scheduling system according to the task plan. The transportation time window is mapped onto the time axis where the signal control cycle sequence is located to determine the number of signal cycles covered by the transportation time window. For each signal cycle covered by the transportation time window, the expected delay cost within that signal cycle is calculated. Specifically, the calculation method is: based on the cycle length of the signal cycle... Subtract the green light duration for the direction in which transport vehicles are allowed to travel. The red light waiting time for each cycle is obtained, and then multiplied by the vehicle start-up loss time coefficient preset at the node. The node that obtains the signal period is delayed. The value is determined according to the first The geometric dimensions of the intersection corresponding to each node are determined, when the intersection is a large, irregular intersection. The value is set to 1.5 when the intersection is a standard cross or T-junction. The value is set to 1.0 when the intersection is a small channelized intersection. The value is set to 0.75. The average value of the nodes in all signal cycles within the transportation time window coverage area is calculated using the delay, yielding the [value]. The spatiotemporal travel cost of each node The unit is seconds.

[0048] The dynamic load response values ​​of all edges and the spatiotemporal travel costs of all nodes are combined to form a load-right-of-way collaborative mapping relationship. This load-right-of-way collaborative mapping relationship employs a two-layer mapping table structure: the first layer is an edge-load mapping table, storing the edge identifier of each edge and its dynamic load response value. The mapping between nodes; the second layer is a node-cost mapping table, which stores the node identifier of each node and the spatiotemporal travel cost of that node. The mapping between the two layers of tables is linked together through the connection relationships between edges and nodes in the candidate path topology graph. This allows for the simultaneous indexing of the dynamic load response value of each edge and the spatiotemporal passage cost of each node for any path in the candidate path topology graph composed of several edges and nodes, thus forming a complete load-right-of-way collaborative mapping relationship. This load-right-of-way collaborative mapping relationship is provided for subsequent dynamic backtracking incentive adjustments.

[0049] See Figure 4 In the figure, the horizontal axis represents time in 15-minute intervals, covering approximately 3000 sampling periods, and the vertical axis represents load in tons. In the legend, the solid curve represents the historical load fluctuation sequence, the dashed curve represents the corresponding convolutional response value sequence, and the hollow dots mark the location of the maximum response value.

[0050] As shown in the figure, the historical load fluctuation sequence exhibits periodic fluctuation characteristics, with load values ​​varying between approximately 600 and 1200 tons, and a period of about 700 to 800 sampling periods, reflecting the dynamic changes in vehicle axle loads on the road arc segment over the past 30 days. The convolutional response value sequence closely follows the fluctuation trend of the historical load fluctuation sequence, showing the dynamic load response obtained after performing a one-dimensional dilated convolution operation between the convolution kernel and the historical load sequence. The response value is generally slightly higher than the peak of the historical load value at the corresponding moment, reflecting the superposition effect of axle load spectrum distribution parameters and historical load fluctuations.

[0051] The location of the maximum response value in the graph is clearly marked by a hollow circle, situated at approximately the 2880th sampling time interval. This corresponds to a peak load exceeding 1200 tons, indicating that the road arc segment was most significantly affected by the axle load of vehicles involved in the oversized and heavy-duty transport mission at this moment, reaching its peak dynamic load response. This maximum response value is used as the dynamic load response value for the corresponding edge in the candidate path topology graph. Calculation is an important parameter for the load-rights cooperative mapping relationship.

[0052] Overall, the data in the figure accurately reflects the dynamic load response calculation process based on historical load fluctuation sequences and axle load spectrum distribution parameters in Example 3. Through continuous convolution operations in the time dimension, the load fluctuation characteristics of the road arc segment within different time windows are effectively captured, providing accurate edge load information support for subsequent dynamic retrospective excitation adjustments.

[0053] Example 4:

[0054] In specific implementation, please refer to Figure 2Using the starting point of the oversized and heavy-duty transport task as the root node, a breadth-first expansion is performed on the candidate path topology graph. The root node is added to the expansion queue, and nodes are sequentially retrieved from the head of the queue for expansion. For each expansion operation, the dynamic load response value of the corresponding edge between the current node and the adjacent node to be expanded is read from the load-right-of-way cooperative mapping relationship. All adjacent edges are sorted in ascending order of dynamic load response value, and breadth-first expansion is performed along the edge with the smaller dynamic load response value to the adjacent node, generating multiple expansion paths. For each expansion path generated, the edge sequence and node sequence of the expansion path are recorded in the path list.

[0055] When any extended path reaches an intersection node in the candidate path topology graph, calculate the cumulative spatiotemporal cost of that extended path from its origin to that intersection node. (Cumulative spatiotemporal cost) The calculation method is as follows:

[0056] in: This represents the total number of edges traversed by the extended path from the starting point to the current intersection node. This represents the traversal index of the edges in the extended path. Indicates the first in the extended path The dynamic load response value of the strip edge, which is read from the load-right-of-way coordination mapping relationship, is in tons; Indicates the first in the extended path The conversion factor from the dynamic load response value of the strip to the time cost. The value is determined by: the total mass of the oversized and overweight transport task and the first... The ratio of the load-bearing capacity of the road arc segment to which the strip belongs is used to determine the load-bearing capacity. When the ratio is less than 0.5... The value is set to 0.8, and when the ratio is between 0.5 and 0.8... The value is set to 1.0, and when the ratio is between 0.8 and 1.0... The value is set to 1.5; when the ratio is greater than 1.0... The value is set to 3.0. The above threshold is calibrated based on the empirical relationship between the road bearing capacity safety factor and the traffic speed. This represents the total number of nodes traversed by the extended path from the starting point to the current intersection node. This represents the traversal index of the node in the extended path. Indicates the first in the extended path The spatiotemporal travel cost of each node is read from the load-rights cooperative mapping relationship and is expressed in seconds.

[0057] Determine the cumulative spacetime cost Does it exceed the time window relaxation coefficient? Compared with the theoretical shortest cost The product of the time window relaxation coefficients. These are preset parameters. The value range is from 1.0 to 1.5. The specific value is set by the logistics scheduling system based on the urgency of the oversized and overweight cargo transportation task; the higher the urgency of the task, the higher the value. The closer the value is to 1.0, the lower the task urgency. The closer the value is to 1.5, the lower the theoretical shortest cost. The method for obtaining the shortest path is as follows: In the candidate path topology graph, Dijkstra's shortest path algorithm is used to calculate the theoretical shortest path from the starting point to the current intersection node, and the dynamic load response value of all edges on the theoretical shortest path is multiplied by their respective conversion coefficients. The sum of these costs, plus the sum of the spacetime travel costs of all nodes on the theoretical shortest path, yields the theoretical shortest cost. .

[0058] If the cumulative spacetime cost relaxation coefficient exceeding the time window Compared with the theoretical shortest cost The product of these factors applies a backtracking incentive penalty to the extended path. The backtracking incentive penalty is calculated as follows:

[0059] in: This represents the backtracking incentive penalty value. This represents the cumulative spacetime cost. Indicates the relaxation coefficient of the time window. This represents the theoretical shortest cost. This indicates the preset penalty coefficient. Penalty Coefficient The value is set to 2.0 because, when penalizing the expansion path, the accumulated time and space cost after the penalty should generate a sufficient ranking disadvantage, thus being naturally eliminated in subsequent expansions. This backtracking incentive penalty value... Cumulative spatiotemporal cost added to this extended path The cumulative spatiotemporal cost of the extended path is updated to... Then, all terminal branches of the extended path that continue to extend forward after reaching the current intersection node are removed from the extended path list, and no further breadth-first expansion is performed.

[0060] After traversing and evaluating all extended paths, the set of all extended paths that were not deleted is used as the initial planning path set. Each initial planning path in the initial planning path set represents a feasible candidate path for transporting oversized and heavy cargo under the constraints of dynamic load response, node spatiotemporal passage cost, and time window relaxation. This initial planning path set is used for subsequent iterative advantage evaluation of the multi-objective reinforcement learning network.

[0061] Example 5:

[0062] In practice, a multi-objective reinforcement learning network is used to iteratively evaluate the advantages of the initial planned path set. The core architecture of the multi-objective reinforcement learning network consists of four parts: an input encoding layer, a value assessment branch, a risk assessment branch, and a difference ranking module. The input encoding layer transforms the initial planned paths into a sequence of state feature vectors; the value assessment branch outputs a comprehensive transportation cost estimate based on the state feature vector sequence; the risk assessment branch outputs a passage risk confidence level based on the state feature vector sequence; and the difference ranking module calculates a weighted difference between the comprehensive transportation cost estimate and the passage risk confidence level, and sorts the initial planned path set in descending order based on the weighted difference.

[0063] For each initially planned path in the initial planned path set, the initial planned path is encoded as a sequence of state feature vectors. The encoding process is as follows: according to the node order in the initial planned path, the capacity parameter set of the road segment to which each node belongs and the traffic flow density at each node are extracted sequentially. The capacity parameter set includes four parameters: height limit threshold, width limit threshold, load capacity level, and radius of curvature. The traffic flow density is obtained from the real-time traffic monitoring system, specifically the quantified value of the traffic flow density level corresponding to the timestamp when the oversized and overweight transport task is expected to pass through the node. The four parameters from the capacity parameter set of each node are concatenated with the traffic flow density at that node to form the node feature vector of that node. The node feature vector is a five-dimensional vector, with the five dimensions being the height limit threshold, width limit threshold, load capacity level, radius of curvature, and traffic flow density, respectively. According to the order of the nodes in the initial planned path, the node feature vectors of all nodes are arranged into a sequence of state feature vectors. The length of the state feature vector sequence is equal to the total number of nodes contained in the initial planned path.

[0064] The state feature vector sequence is input into the value evaluation branch of the multi-objective reinforcement learning network to calculate the estimated comprehensive transportation cost for each initially planned path. The value evaluation branch is designed with the following hierarchical structure: a convolutional downsampling layer, a bidirectional long short-term memory recursive layer, a feature fusion layer, and a fully connected layer. In the value evaluation branch, the state feature vector sequence is convolutionally downsampled. The convolutional downsampling layer consists of three stacked one-dimensional convolutions. The first one-dimensional convolution has a kernel length of 3, a stride of 1, and 16 output channels; the second one-dimensional convolution has a kernel length of 3, a stride of 2, and 32 output channels; and the third one-dimensional convolution has a kernel length of 3, a stride of 2, and 64 output channels. All three convolutions use linear rectified units as activation functions. The convolutional downsampling layer extracts the local spatial features of each node from its local neighborhood, outputting a 64-dimensional local spatial feature vector for each node.

[0065] The local spatial feature vector of each node is input into a bidirectional long short-term memory (LSTM) recursive layer along the node dimension. The bidirectional LSM recursive layer contains one forward LSM unit and one backward LSM unit. The forward LSM unit processes the local spatial feature vector sequence from front to back according to the node order, while the backward LSM unit processes the local spatial feature vector sequence from back to front according to the node order. The hidden state dimension of both the forward and backward LSM units is set to 128 dimensions. For each node, the forward LSM unit outputs the 128-dimensional forward hidden state of that node, and the backward LSM unit outputs the 128-dimensional backward hidden state of that node.

[0066] The 128-dimensional forward and backward hidden states of each node are input into a feature fusion layer. This layer concatenates the forward and backward hidden states along their feature dimensions to generate a 256-dimensional fused feature vector for that node. This 256-dimensional fused feature vector is then input into a fully connected layer with 128 neurons to obtain the 128-dimensional node cost feature. The node cost feature is then input into an output fully connected layer with one neuron to obtain the node cost component. The node cost components of all nodes in the initial planned path are summed, and the sum is used as the estimated comprehensive transportation cost for the initial planned path.

[0067] The state feature vector sequence is input into the risk assessment branch of the multi-objective reinforcement learning network to calculate the passability risk confidence of each initially planned path. The risk assessment branch is designed with the following hierarchical structure: a self-attention transformation layer, a weighted reshaping layer, a graph convolutional network layer, and a max pooling layer. In the risk assessment branch, a self-attention transformation is performed on the state feature vector sequence. The self-attention transformation layer is implemented using a multi-head attention mechanism with eight self-attention heads. Each self-attention head linearly maps the state feature vector sequence into a query matrix, a key matrix, and a value matrix, respectively, with each feature dimension set to 64. The output of each self-attention head is generated by multiplying the inner product of the transpose of the query matrix and the key matrix, followed by scaling and flexible max normalization, and then multiplying it with the value matrix. The outputs of the eight self-attention heads are concatenated along the feature dimensions and then mapped back to the original dimensions of the state feature vector sequence through a linear transformation, resulting in the global dependency weight matrix of each node relative to the entire path.

[0068] The node feature vector of each node is reshaped by weighting it according to the global dependency weight matrix. The weighting method involves element-wise multiplying the weight vector corresponding to that node in the global dependency weight matrix with the node feature vector to obtain the risk feature representation of that node. The dimension of the risk feature representation is the same as the dimension of the node feature vector, both being five dimensions.

[0069] The risk feature representation of each node is input into a graph convolutional network (GCNN). The graph structure of the GCNN is constructed from the node connection order of the initial planned path. The adjacency matrix is ​​set so that edges exist only between adjacent nodes, and the weight of each edge is set to 1. The GCNN consists of two stacked graph convolutional layers. The first graph convolutional layer maps the five-dimensional risk feature representation to 32-dimensional hidden layer risk features, and the second graph convolutional layer maps the 32-dimensional hidden layer risk features to 1-dimensional local risk coefficients. The activation function of both graph convolutional layers is a linear rectified unit (LRC). The GCNN aggregates the risk features of adjacent nodes at each node and outputs the local risk coefficient of each node.

[0070] The maximum value of the local risk coefficient of all nodes in the initial planned path is taken, and this maximum value is used as the confidence level of the travel risk of the initial planned path.

[0071] The initial planned routes are sorted in descending order based on the weighted difference between the estimated comprehensive transportation costs and the confidence level of traffic risk. The weighted difference is calculated as follows:

[0072] in: This represents an initial planned path from the set of initial planned paths. Indicates the initial planned path The overall advantage assessment value, Indicates the initial planned path The estimated total transportation cost, Indicates the initial planned path The confidence level of the passage risk The weighting factor represents the estimated overall transportation cost. The weighting coefficient represents the confidence level of the passage risk. and The values ​​of all values ​​are between 0 and 1, and satisfy the following conditions: . and The specific value is set by the logistics scheduling system based on the risk preference configuration file for oversized and overweight cargo transportation tasks. When the task is set to safety priority mode... The value is set to 0.4 and The value is set to 0.6 when the task is set to efficiency-first mode. The value is set to 0.7 and The value is set to 0.3 when the task is set to balanced mode. and All values ​​are set to 0.5.

[0073] All initial planning paths in the initial planning path set are evaluated according to their comprehensive advantages. Sort the items in descending order of size, and use the initial planned path that is first in the sorted order as the optimal dynamic planning path for the transportation of oversized and heavy items.

[0074] The training process of the multi-objective reinforcement learning network was completed in the offline training environment of the logistics scheduling system. Training data was collected from historical records of oversized and heavy-duty transport tasks. Each training sample included a sequence of state feature vectors for a historical initial planned path, along with an actual comprehensive transportation cost label and an actual traffic risk label for that path. The actual comprehensive transportation cost label was the total transit time (in seconds) recorded during the actual transport process for that historical initial planned path. The actual traffic risk label was a binary marker indicating whether a traffic obstruction event occurred during the actual transport process for that historical initial planned path; a marker of 1 indicated a traffic obstruction, and a marker of 0 indicated no obstruction. The training loss function consisted of a weighted sum of two losses: a value assessment loss and a risk assessment loss. The value assessment loss used mean squared error loss, and the risk assessment loss used binary cross-entropy loss. The weighting coefficients for both losses were set to 1.0. The adaptive moment estimation optimizer was used, with an initial learning rate of 0.001. The learning rate decreased to 0.5 times its original value every 50 training epochs. The batch size was set to 32, and the total number of training epochs was set to 200. After training, the multi-objective reinforcement learning network is deployed to the path planning node of the logistics scheduling system for dynamic path planning of oversized and heavy-duty transport during the online phase.

[0075] See Figure 5 The horizontal axis of the graph represents the estimated comprehensive transportation cost of the initial planned route, in seconds, ranging from approximately 300 to 1300 seconds, reflecting the estimated transportation time of the route. The vertical axis represents the confidence level of the traffic risk, ranging from 0 to 1, indicating the confidence level of the route's traffic safety risk. Multiple initial planned route distribution points are marked with solid black dots in the graph. Overall, the graph shows a trend where the confidence level of the traffic risk increases with the increase of the estimated comprehensive transportation cost, indicating that routes with higher transportation costs are usually accompanied by greater traffic risks. The curve distribution is relatively continuous and shows an upward trend. Within the range where the estimated comprehensive transportation cost increases from approximately 300 seconds to 1000 seconds, the risk confidence level gradually rises from near zero to near 1.0, demonstrating that the multi-objective reinforcement learning network can effectively distinguish between good and bad routes in its value assessment and risk assessment.

[0076] The optimal dynamic programming path is marked with a black pentagram in the figure. The estimated comprehensive transportation cost of this path is approximately 1200 seconds, and the risk confidence level is 1.0, placing it at the position with the highest comprehensive advantage evaluation value among all initial planned paths. This optimal path reflects the final path selected after iterative advantage evaluation by the multi-objective reinforcement learning network in Example 5. It comprehensively considers the weighted difference between transportation cost and risk confidence level, demonstrating the optimal planning of oversized and heavy-duty cargo transportation paths while balancing efficiency and safety.

[0077] This figure visually demonstrates the advantage ranking effect after encoding, evaluating value, and assessing risk of the initial planned path set through a multi-objective reinforcement learning network. It verifies that the method described in Example 5 can effectively balance and rank the two objectives of the comprehensive transportation cost estimate and the confidence level of the passage risk of the initial planned path, providing a basis for the optimal dynamic planning path selection that balances safety and efficiency for oversized and heavy cargo transportation tasks.

[0078] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A dynamic planning method for transporting oversized and heavy cargo based on artificial intelligence, characterized in that, The method, applied to a path planning node in a logistics scheduling system, includes: Obtain the dynamic spatial constraint set of the oversized and heavy-duty transport task and the multimodal state tensor of the real-time road network; Based on the dynamic spatial constraint set, reachability domain elimination operation is performed on the multimodal state tensor of the real-time road network to generate a candidate path topology graph; The load-right-of-way collaborative mapping relationship of the oversized and heavy-duty transport task is constructed by using the edge load sequence and node passage cost matrix in the candidate path topology graph. Based on the load-right-of-way collaborative mapping relationship and the preset time window relaxation coefficient, the candidate path topology is dynamically backtracked and adjusted to obtain the initial planned path set for the oversized and heavy-duty transport task. A multi-objective reinforcement learning network is used to iteratively evaluate the advantages of the initial planned path set and output the optimal dynamic planning path for the oversized and heavy-duty transport task.

2. The method for dynamic planning of transportation routes for oversized and heavy-duty items based on artificial intelligence according to claim 1, characterized in that, The dynamic spatial constraint set for obtaining the oversized and heavy cargo transportation task further includes: Analyze the cargo outline dimension sequence and axle load spectrum distribution parameters of the oversized and heavy-duty transport mission; The sequence of cargo outline dimensions is decomposed into multiple characteristic load distribution surfaces according to the coordinates of the vehicle's attachment point; Extract the spatial envelope vector of each of the multiple characteristic load distribution surfaces, and use the set of all spatial envelope vectors as the dynamic spatial constraint set.

3. The method for dynamic planning of transportation routes for oversized and heavy-duty items based on artificial intelligence according to claim 2, characterized in that, The acquisition of the multimodal state tensor of the real-time road network further includes: Using the largest spatial envelope vector in the dynamic spatial constraint set as the scanning window, the traffic capacity parameter set of each road arc segment is collected by sliding on the basic road network layer; The traffic capacity parameter sets of each road arc segment are stacked as tensors according to the timestamp dimension, spatial curvature dimension and traffic flow dimension to generate the multimodal state tensor of the real-time road network. The traffic capacity parameter set includes height limit threshold, width limit threshold, load capacity level, and radius of curvature.

4. The method for dynamic planning of transportation routes for oversized and heavy-duty items based on artificial intelligence according to claim 1, characterized in that, Based on the dynamic spatial constraint set, performing reachability domain elimination on the multimodal state tensor of the real-time road network to generate a candidate path topology graph further includes: Traverse the capacity parameter set of each road arc segment in the multimodal state tensor of the real-time road network; The traffic capacity parameter set of each road arc segment is compared dimension by dimension with each characteristic load distribution surface in the dynamic spatial constraint set; When any one-dimensional spatial parameter of any characteristic load distribution surface exceeds the threshold of the corresponding dimension in the traffic capacity parameter set, the road arc segment is marked as an unreachable arc segment. All unreachable arc segments are removed from the multimodal state tensor of the real-time road network, and the remaining road arc segments and their connection relationships are used to form the candidate path topology graph.

5. The method for dynamic planning of transportation routes for oversized and heavy-duty items based on artificial intelligence according to claim 4, characterized in that, When comparing the traffic capacity parameter set of each road arc segment with each characteristic load distribution surface in the dynamic spatial constraint set dimension by dimension, if the difference of any dimension is less than the preset tolerance threshold, then it is determined that the dimension has not exceeded the threshold.

6. The method for dynamic planning of transportation routes for oversized and heavy-duty items based on artificial intelligence according to claim 1, characterized in that, The step of constructing the load-right-of-way collaborative mapping relationship for the oversized and heavy-duty transport task by using the edge load sequence and node passage cost matrix in the candidate path topology graph further includes: Extract the historical load fluctuation sequence of each edge and the signal control cycle sequence of each node in the candidate path topology graph; The historical load fluctuation sequence of each edge is convolved with the axle load spectrum distribution parameters of the oversized and heavy-duty transport task to obtain the dynamic load response value of each edge. The signal control cycle sequence of each node is cross-correlated with a preset transportation time window to obtain the spatiotemporal passage cost of each node. The dynamic load response values ​​of all edges and the spatiotemporal passage costs of all nodes are combined to form the load-right-of-way collaborative mapping relationship.

7. The method for dynamic planning of transportation routes for oversized and heavy-duty items based on artificial intelligence according to claim 6, characterized in that, When performing convolution operations on the historical load fluctuation sequence of each edge with the axle load spectrum distribution parameters of the oversized and heavy-duty transport task, a one-dimensional dilated convolution kernel is used to match the load response characteristics at different time scales.

8. The method for dynamic planning of transportation routes for oversized and heavy-duty items based on artificial intelligence according to claim 6, characterized in that, The step of dynamically backtracking and adjusting the candidate path topology based on the load-right-of-way collaborative mapping relationship and a preset time window relaxation coefficient to obtain the initial planned path set for the oversized and heavy-duty transport task further includes: Taking the starting point of the oversized and heavy-duty transport task as the root node, according to the dynamic load response values ​​of the edges in the load-right-of-way collaborative mapping relationship in ascending order, a breadth-first expansion is performed on the candidate path topology graph to generate multiple expansion paths; When any extended path reaches the intersection node in the candidate path topology graph, calculate the cumulative time and space cost of the extended path from the starting point to the intersection node. Determine whether the accumulated spatiotemporal cost exceeds the product of the time window relaxation coefficient and the theoretical shortest cost. If so, apply a backtracking incentive penalty to the extended path and delete the terminal branch of the extended path. The set of all extended paths that have not been deleted is taken as the initial planned path set.

9. The method for dynamic planning of transportation routes for oversized and heavy-duty items based on artificial intelligence according to claim 8, characterized in that, The calculation method of the backtracking incentive penalty is as follows: multiply the difference between the product of the cumulative spatiotemporal cost and the product of the time window relaxation coefficient and the theoretical shortest cost by a preset penalty coefficient, and add the penalty value to the cumulative spatiotemporal cost of the extended path.

10. The method for dynamic planning of transportation routes for oversized and heavy-duty items based on artificial intelligence according to claim 8, characterized in that, The step of using a multi-objective reinforcement learning network to iteratively evaluate the advantages of the initial planned path set and output the optimal dynamic planning path for the oversized and heavy-duty transport task further includes: Each initial planning path in the initial planning path set is encoded as a sequence of state feature vectors; The state feature vector sequence is input into the value evaluation branch of the multi-objective reinforcement learning network to calculate the estimated comprehensive transportation cost for each initial planned path; The state feature vector sequence is input into the risk assessment branch of the multi-objective reinforcement learning network to calculate the passability risk confidence of each initial planned path; Based on the weighted difference between the estimated comprehensive transportation cost and the confidence level of the traffic risk, the initial planned path set is sorted in descending order, and the initial planned path that ranks first after sorting is taken as the optimal dynamic planning path.

Images