Artificial intelligence-based logistics path dynamic planning system

Abstract

The application discloses an artificial intelligence-based logistics path dynamic planning system, which comprises a continuous space mapping module, a tensor prediction regression module, a multi-phase field coupling solution module and a discretization adaptive control module.The continuous space mapping module maps a discrete logistics network into a continuous space manifold through a space interpolation algorithm, and establishes a scalar phase field distinguishing task priority.The tensor prediction regression module processes environmental characteristics by using a space-time graph convolution network, and outputs a basic anisotropic diffusion tensor subject to positive definite constraint.The multi-phase field coupling solution module constructs a reaction-diffusion partial differential equation set containing a cross-inhibition coupling term, and solves the equation set discretely, and outputs a global concentration field.The discretization adaptive control module extracts a local concentration negative gradient vector of a vehicle coordinate, matches a road section by using cosine similarity, and generates an instruction.The application converts discrete search into continuous partial differential equation solving, and effectively avoids mutual lockout phenomenon in multi-vehicle joint scheduling.

Description

Technical Field

[0001] This invention relates to the fields of intelligent logistics and artificial intelligence technology, specifically to a dynamic logistics path planning system based on artificial intelligence. Background Technology

[0002] With the continuous expansion of modern logistics systems, the complexity of logistics road networks has increased significantly. Traditional logistics route planning typically relies on discrete graph search algorithms. When faced with large-scale physical road networks and high-frequency joint scheduling of multiple vehicles at complex intersections, the number of computational nodes and edges in such discrete graph search algorithms grows exponentially, easily leading to a combinatorial explosion problem in computational power. This makes it difficult for the system to uniformly calculate the global macro-level traffic situation and the local micro-level vehicle control, resulting in low planning efficiency.

[0003] Meanwhile, modern logistics route planning requires the integration of multi-source dynamic environmental features such as road conditions, traffic flow, and weather information. Existing route planning systems, when introducing deep learning networks to extract these multimodal features, often directly apply the network output to the calculation of path resistance weights. Because the neural network model output itself lacks rigorous positive definite constraints on physical eigenvalues ​​and mechanisms for handling outlier extreme values, the network is prone to outputting abnormal values ​​when encountering sudden road closures or extreme weather conditions that cause abrupt changes in input data. These abnormal values ​​directly participate in the solution of underlying equations, easily leading to numerical divergence in the planning algorithm. This results in multi-source traffic flow and weather environmental features failing to accurately map to real physical traffic resistance, and the system's anti-interference capability is insufficient.

[0004] Furthermore, in actual logistics transportation, multiple transportation tasks with different priorities often coexist. Existing multi-priority logistics vehicle joint scheduling systems often use independent path planning calculations for each priority fleet. The evolution process of different priority tasks lacks nonlinear cross-inhibition and dynamic coupling mechanisms. This makes it impossible for low-priority vehicles to actively reduce their evolution speed and avoid high-priority vehicles based on the aggregation degree in physically narrow road sections or key traffic conflict areas. As a result, multiple fleets frequently lock up each other and experience local congestion at intersections, making it impossible to achieve system-level passive risk avoidance and multi-vehicle collaborative scheduling. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an AI-based dynamic planning system for logistics routes. It aims to solve the technical problems of a surge in computational load and combinatorial explosion in large-scale joint scheduling of discrete logistics networks, divergent solutions due to the lack of strict physical constraints in the extraction of multi-source environmental features, and local congestion and path locking caused by independent computation of multiple priority tasks.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An AI-based dynamic programming system for logistics routes includes: a continuous space mapping module, a tensor prediction and regression module, a multiphase field coupling solution module, and a discretized adaptive control module. Its key features are: The continuous space mapping module is used to acquire discrete logistics road network topology data, map the discrete logistics road network topology data into a continuous space manifold, establish multiple scalar phase fields on the continuous space manifold, and configure the logistics delivery destination as the boundary condition of the partial differential equation to complete the parameter initialization. The tensor prediction and regression module is used to collect multi-source environmental feature data of the road network, process the multi-source environmental feature data of the road network using a feature extraction network, and output the basic anisotropic diffusion tensor corresponding to the continuous spatial manifold. The multiphase field coupling solution module is used to read the scalar phase field parameters and the basic anisotropic diffusion tensor, construct a set of reaction-diffusion partial differential equations between different priority logistics tasks, and discretize the set of reaction-diffusion partial differential equations to output the global scalar concentration field distribution results. The discretized adaptive control module is used to obtain the physical positioning coordinates of the controlled logistics vehicle, calculate the local negative gradient vector based on the physical positioning coordinates in the global scalar concentration field distribution result, generate the next-hop driving command based on the directional similarity between the local negative gradient vector and the adjacent physical road segment, and send it to the vehicle terminal.

[0007] Preferably, the continuous spatial mapping module establishes a continuous topological distribution field using a spatial interpolation algorithm and a Gaussian kernel function; it allocates the controlled logistics vehicles to the first scalar phase field or the second scalar phase field according to the task priority attribute, and uses a spatial Dirac function to define the spatial concentration point source at the initial moment; at the same time, it applies homogeneous Dirichlet boundary conditions at continuous coordinate points to constrain the concentration value of the logistics delivery destination area to remain constant at zero.

[0008] Preferably, the tensor prediction regression module transforms multi-source environmental features including traffic flow, vehicle speed, road closure, and meteorological data into a multimodal environmental feature matrix, and combines them to generate a three-dimensional spatiotemporal feature tensor; it uses a spatiotemporal graph convolutional network to extract the spatial feature representation matrix and outputs the original regression matrix; thereby subjecting the basic anisotropic diffusion tensor to a symmetric positive definite matrix constraint, and resetting the corresponding feature values ​​to the lower limit value when the road is physically blocked.

[0009] Preferably, the multiphase field coupling solution module introduces a cross-suppression coupling term when constructing the reaction-diffusion partial differential equation system. It constructs a nonlinear cross-suppression factor by extracting the local concentration gradient magnitude of the first scalar phase field, calculates the dynamic equivalent diffusion tensor and performs eigenvalue lower bound truncation processing, and then constructs a coupled partial differential equation containing the evolution of the second scalar phase field of the dynamic equivalent diffusion tensor.

[0010] Preferably, the discretized adaptive control module obtains spatial boundary parameters to construct an affine transformation matrix of the coordinate system, calculates the local negative gradient vector using a second-order central difference algorithm, and calibrates and maps the local negative gradient vector back to the physical road network coordinate system; then, it performs magnitude normalization processing on the negative gradient vector of the physical direction to generate the desired guidance vector, and calculates the cosine similarity between the desired guidance vector and the unit direction vector of each adjacent physical road segment as the direction similarity.

[0011] This invention provides an artificial intelligence-based dynamic planning system for logistics routes. It offers the following advantages: 1. This invention maps a discrete logistics network topology to a continuous spatial manifold using a spatial interpolation algorithm, and establishes multiple parallel scalar phase fields on this continuous spatial manifold. This technical solution transforms the traditional discrete graph search problem into a numerical solution process of continuous partial differential equations, avoiding the computational power explosion problem that occurs when multiple vehicles are jointly scheduled in a large-scale road network, and improving the computational efficiency of the system in large-scale complex road networks.

[0012] 2. This invention utilizes a spatiotemporal graph convolutional network to process multi-source environmental feature matrices, outputting a fundamental anisotropic diffusion tensor, and applying eigenvalue positive definite constraints and extremum truncation to it. This technical solution eliminates the computational risk of partial differential equation divergence caused by abnormal values ​​in neural network outputs from the underlying mathematical logic, ensuring that multi-source traffic flow and meteorological features can be stably mapped to physical traffic resistance, and improving the resilience of dynamic path planning schemes to sudden environmental changes.

[0013] 3. In constructing the partial differential equation system, this invention introduces a nonlinear cross-suppression factor based on the magnitude of the concentration gradient of the high-priority phase field to dynamically adjust the equivalent diffusion tensor of the low-priority phase field. This technical solution ensures the rapid passage of high-priority logistics tasks while reducing the evolution speed of low-priority vehicles towards critical conflict areas by coupling the partial differential equations. This effectively avoids the phenomenon of multi-priority vehicle convoys locking up or congesting in narrow road sections, achieving system-level multi-vehicle collaborative scheduling. Attached Figure Description

[0014] Figure 1 This is a block diagram of the module structure of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0015] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see the appendix Figure 1-2 This invention provides an artificial intelligence-based dynamic planning system for logistics routes, including: a continuous space mapping module, a tensor prediction and regression module, a multiphase field coupling solution module, and a discretized adaptive control module.

[0017] At the physical architecture level, the above system relies on cloud servers, edge computing nodes, and vehicle terminals to form a distributed computing environment. The cloud servers and edge computing nodes work together to run the continuous space mapping module, tensor prediction and regression module, and multiphase field coupling solution module, which are responsible for numerical iteration of partial differential equations and forward propagation of graph networks. The vehicle terminals run the discretized adaptive control module, which is responsible for receiving field parameters from the cloud and executing local control logic. Data connections are established between the physical nodes through a communication network.

[0018] The process of performing topology mapping by the continuous space mapping module includes: Step 101: The continuous spatial mapping module obtains discrete logistics road network topology data of the target area from a preset geographic information system or logistics scheduling database through a communication network. The discrete logistics road network topology data is represented as a graph structure. ,in For the set of road network nodes, For the set of road network edges, the continuous spatial mapping module resolves the set of nodes. Each node in spatial coordinates .

[0019] Step 102: The continuous spatial mapping module determines the spatial boundary based on the extreme coordinates of the road network nodes, and constructs a two-dimensional or three-dimensional continuous spatial manifold. The continuous space mapping module is located in the continuous space manifold. The spatial interpolation kernel function is configured above, and the Gaussian kernel function is used to realize the transformation of discrete node features into a continuous spatial manifold. The numerical mapping of the Gaussian kernel function is calculated using the following formula: ; In the formula, For continuous space manifolds any continuous coordinate point Coordinates of discrete logistics network nodes The kernel mapping values ​​between them The smoothing bandwidth parameter is used to control the spatial influence range of discrete nodes' outward mapping. The value is determined based on the average spatial physical distance between adjacent nodes in the discrete logistics network topology data. For continuous coordinate points Coordinates of discrete logistics network nodes The Euclidean distance between them.

[0020] Step 103: The continuous space mapping module uses the aforementioned spatial interpolation kernel function to calculate the continuous space manifold. The fundamental mapping field function for each coordinate point within the continuous space is used. The continuous space mapping module establishes a continuous topological distribution field by superimposing the kernel mapping values ​​of each discrete logistics network node. This is applicable to continuous spatial manifolds. any coordinate point in Fundamental mapping field function The calculation formula is as follows: ; In the formula, For the set of road network nodes The total number of nodes in the system For discrete logistics network nodes The inherent topological weight coefficients, the continuous space mapping module is based on the discrete logistics network nodes In graph structure The degree of each node in the topology determines the inherent topological weight coefficient. The continuous spatial mapping module calculates the values ​​of the continuous spatial manifold according to the set spatial resolution. The basic mapping field function values ​​of the discretized grid points are output by the continuous spatial mapping module, which outputs the mapped continuous topological data matrix. This completes the underlying structural transformation from discrete logistics network topological data to continuous physical manifold, and provides the spatial basic data matrix required for solving partial differential equations.

[0021] The process of configuring the concentration diffusion source in the continuous spatial mapping module includes: Step 104: The continuous space mapping module reads the previously constructed two-dimensional or three-dimensional continuous space manifold. The continuous topology data matrix, the continuous space mapping module classifies logistics tasks according to their priority attributes, and is located in the continuous space manifold. Establish independent and parallel scalar phase fields covering the entire continuous spatial manifold.

[0022] Step 105: The continuous spatial mapping module establishes a first scalar phase field corresponding to the urgent logistics task and a second scalar phase field corresponding to the ordinary logistics task. The continuous spatial mapping module sets the spatial concentration distribution function of the first scalar phase field as follows: The continuous spatial mapping module sets the spatial concentration distribution function of the second scalar phase field as follows: In the above formula, For continuous space manifolds Continuous coordinate points within, As time variables, the first scalar phase field and the second scalar phase field share a continuous spatial manifold. The physical space boundary.

[0023] Step 106: The continuous spatial mapping module obtains the current physical location coordinates and task priority attributes of the logistics vehicles within the system, and maps the physical location coordinates to a continuous spatial manifold. The continuous coordinate points within the continuous spatial mapping module are used to assign each logistics vehicle to the first or second scalar phase field based on the task priority attribute, serving as the initial concentration diffusion source within each scalar phase field.

[0024] Step 107: The continuous space mapping module uses a grid-smoothed and regularized Dirac function to define the initial time of each logistics vehicle. For a spatial concentration point source, the continuous spatial mapping module calculates the initial time for the first scalar phase field. The concentration distribution formula is as follows: ; In the formula, This refers to the total number of express delivery vehicles within the system. For the first Initial load values ​​for each express delivery vehicle. For the first A special express delivery vehicle in continuous spatial manifold Mapped coordinates in For the spatial Dirac function, to accommodate the subsequent discretization and numerical computation of partial differential equations, the continuous space mapping module is used. The spatial Dirac function is regularized using a Gaussian distribution function within the neighborhood of the discretized grid point to eliminate numerical singularities.

[0025] Step 108: For the second scalar phase field, the continuous space mapping module calculates the initial time. The concentration distribution formula is as follows: ; In the formula, This represents the total number of ordinary logistics vehicles within the system. For the first The initial load value of a typical logistics vehicle. For the first A typical logistics vehicle in a continuous spatial manifold The mapping coordinates in the continuous spatial mapping module are calculated using the above formula to complete the initial configuration of the spatial parameters of the logistics vehicle load. The continuous spatial mapping module stores the calculated initial concentration field matrix into the system memory to provide the initial conditions for the subsequent time-domain evolution of partial differential equations.

[0026] The process of configuring boundary conditions in the continuous space mapping module includes: Step 109: The continuous spatial mapping module obtains the set of logistics delivery destinations for the target area. The continuous space mapping module resolves the various logistics delivery destinations in this set. The physical location coordinates are obtained, and the physical location coordinates are mapped to continuous spatial manifolds respectively. Internal corresponding continuous coordinate points .

[0027] Step 110, the continuous space mapping module maps the above continuous coordinate points Set as the reaction sink in the reaction-diffusion partial differential equation, and at this continuous coordinate point A homogeneous Dirichlet boundary condition is applied at the location. Through this boundary condition constraint, the system maintains the concentration value of the logistics distribution destination area at a constant zero, so that the coordinate location forms a local minimum point in the spatial concentration field.

[0028] Step 111: The continuous space mapping module constructs corresponding boundary condition mathematical expressions for the first scalar phase field and the second scalar phase field, respectively. For the first scalar phase field corresponding to the urgent logistics task... The homogeneous Dirichlet boundary condition formula is set as follows: ; For the second scalar phase field corresponding to ordinary logistics tasks The homogeneous Dirichlet boundary condition formula is set as follows: ; In the two formulas above, For the logistics delivery destination in a continuous spatial manifold Mapped coordinates in It is a time variable.

[0029] Step 112: The continuous spatial mapping module obtains each logistics delivery destination. Logistics throughput capacity calibration value The continuous spatial mapping module calibrates the throughput capacity of the logistics system based on a direct proportional mapping relationship. Converted to reaction-diffusion partial differential equations at this continuous coordinate point Spatial reaction consumption rate parameter The calculation formula is: ,in The system dissipation coefficient is set as the preset value. Subsequently, the continuous space mapping module inputs the constructed homogeneous Dirichlet boundary condition parameters and the spatial reaction consumption rate parameter into the system memory, establishes the spatial endpoint mathematical constraints of the partial differential equation evolution, and completes the initialization configuration of all parameters of the continuous space manifold.

[0030] The process of extracting the multi-source environmental feature matrix by the tensor prediction and regression module includes: Step 201: The tensor prediction and regression module collects multi-source environmental feature data of the road network through traffic sensors deployed in the physical road network and external meteorological data interface. The collected multi-source environmental feature data includes real-time traffic flow, average driving speed of each physical road segment, status of sudden road closure events, and meteorological precipitation data.

[0031] Step 202: The tensor prediction and regression module performs spatiotemporal alignment and numerical normalization on the aforementioned multi-source environmental feature data. It aligns feature data with different sampling frequencies to a unified time step set by the system using a numerical interpolation algorithm, and maps the values ​​of various feature data to a normalized value using an extreme value normalization algorithm. The dimensionless numerical range is used to eliminate the differences in dimensional calculations between multimodal environmental characteristic data. Step 203: The tensor prediction and regression module reads the spatial basic data matrix output by the aforementioned continuous spatial mapping module, obtains the spatial coordinates of each discretized grid point within the continuous spatial manifold, and assigns the normalized multi-source environmental feature data to the corresponding spatial location of the discretized grid point according to the spatial coordinate mapping relationship.

[0032] Step 204, for any discretized grid point within the continuous spatial manifold The tensor prediction and regression module concatenates the corresponding real-time traffic flow, average road speed, sudden road closure status, and meteorological precipitation data into vectors to generate the discretized grid point at the time step. One-dimensional feature vector The tensor prediction and regression module concatenates the one-dimensional feature vectors corresponding to all discrete grid points according to their spatial coordinates to construct the current time step. Multimodal environment feature matrix .

[0033] Step 205: The tensor prediction and regression module follows the set time window length. Extracting historical continuous The multimodal environment feature matrix at each time step, and the... The multimodal environmental feature matrices are stacked in the time dimension to generate a three-dimensional spatiotemporal feature tensor. The three-dimensional spatiotemporal feature tensor The mathematical space dimension is expressed as follows: ; In the formula, This represents the number of time steps contained within the time series window. This represents the total number of discretized grid points extracted within the continuous spatial manifold. For the attribute dimension of a single-point one-dimensional feature vector, the tensor prediction and regression module will construct a three-dimensional spatiotemporal feature tensor. It is stored in system memory as the standard data input for subsequent spatiotemporal graph convolutional network processing.

[0034] The process by which the tensor prediction regression module processes the feature matrix and outputs the initial state parameters includes: Step 206: The tensor prediction and regression module obtains the three-dimensional spatiotemporal feature tensor constructed above. The tensor prediction and regression module constructs a grid-based system based on the regular arrangement of discretized grid points within a continuous spatial manifold. Sparse topological adjacency matrix of neighborhood connectivity Subsequently, the topological adjacency matrix Add self-loops and normalize the degree matrix to generate a normalized adjacency matrix. It is used to characterize the spatial connectivity between discretized grid points within a manifold space.

[0035] Step 207, the tensor prediction and regression module converts the three-dimensional spatiotemporal feature tensor Expanding the sequence of time steps yields the input feature matrix corresponding to each time step. This module utilizes the graph convolutional layers within the spatiotemporal graph convolutional network to process the input feature matrix at each time step sequentially. The graph convolutional layer combines a normalized adjacency matrix. Perform node feature aggregation operations to extract spatial dimension correlation information of multimodal environmental features. The calculation formula is as follows: ; In the formula, For time steps The spatial feature representation matrix has dimensions of . ,in The total number of grid points. The number of output channels for graph convolution. The trainable weight matrix of the graph convolutional layer. For non-linear activation functions, this embodiment selects... function.

[0036] Step 208, the tensor prediction and regression module will extract the spatial feature representation matrix of the time step sequence. The system sequentially inputs gated recurrent unit layers within the spatiotemporal graph convolutional network, and then the matrix... The row vectors corresponding to each grid point are input into their respective gated recurrent units, and the internal weights are shared among all grid points. This gated recurrent unit layer performs temporal state updates in a forward order along the time dimension, extracting the dynamic evolution information of multimodal environmental features within the temporal window. The state update formula is as follows: ; In the formula, For time steps The hidden layer state feature matrix, This is the hidden layer state feature matrix from the previous time step. The tensor prediction and regression module extracts the final time step as the forward propagation function of the gated recurrent unit. Output top-level hidden state matrix It serves as a comprehensive encoding representation of the entire spatiotemporal feature sequence.

[0037] Step 209: The tensor prediction and regression module maps the top-level hidden state matrix through a fully connected mapping layer. A linear dimension transformation is performed, and the fully connected mapping layer maps the high-dimensional hidden layer features to low-dimensional tensor regression components, outputting the original regression matrix corresponding to each discretized grid point. Finally, the original regression matrix is ​​used as the initial state parameter of the basic anisotropic diffusion tensor without physical constraints, and input into the subsequent parameter constraint and truncation processing flow.

[0038] The process of performing positive definite constraints and truncation in the tensor prediction regression module includes: Step 210: The tensor prediction and regression module obtains the original regression matrix corresponding to each discretized grid point output by the fully connected mapping layer. The tensor prediction regression module performs symmetry processing on the original regression matrix, and the calculation formula is as follows: ; In the formula, This is the transpose of the original regression matrix; It is a transition symmetric matrix. Step 211, the tensor prediction regression module performs a test on the transition symmetric matrix. Eigenvalue decomposition is performed, and the decomposition formula is as follows: ; In the formula, It is an orthogonal eigenvector matrix, whose column vectors are eigenvectors, indicating the direction of the principal axis of the physical manifold in space; It is a diagonal matrix, and its main diagonal elements are the initial eigenvalues. This represents the basic capacity of logistics vehicles in the corresponding feature vector direction.

[0039] Step 212, the tensor prediction regression module processes the initial feature values. Applying positive definite constraints, under normal operating conditions where no road closures occur, the initial eigenvalues ​​are determined using nonlinear constraint operators. Mapped to a minimum positive real number greater than the system's default value. positive eigenvalues The calculation formula is as follows: ; This operation eliminates the risk of divergence in the solution of the diffusion equation due to the non-positive definiteness of the regression matrix.

[0040] Step 213: When the status data of a sudden road closure event indicates that the corresponding road segment is completely physically blocked, the tensor prediction and regression module triggers truncation processing, and the tensor prediction and regression module calculates the orthogonal eigenvector matrix. The cosine similarity between each column of feature vectors and the physical direction vector of the road segment is calculated. The column of feature vectors with the highest similarity is extracted, and its corresponding feature value is forcibly reset to the lower bound value of the limit. Subsequently, the tensor prediction regression module utilizes orthogonal eigenvector matrices With the corrected eigenvalue diagonal matrix The multiplication operation is reconstructed, and the calculation formula is as follows: ; The tensor prediction and regression module will calculate the underlying anisotropic diffusion tensor of each generated discretized grid point. Stored in system memory, it serves as the input of underlying environmental parameters for the multiphase field coupling solution module to construct the partial differential equation system.

[0041] The process of constructing high-priority phase-field partial differential equations by the multiphase field coupling solution module includes: Step 301: The multiphase field coupling solution module reads the first scalar phase field initialized by the continuous space mapping module and its corresponding spatial reaction consumption rate parameter, and obtains the basic anisotropic diffusion tensor of each discretized grid point in the continuous space manifold output by the tensor prediction regression module.

[0042] Step 302: Based on the business scheduling logic that express logistics tasks enjoy the highest right-of-way on the road network, the system sets the spatial diffusion process of the first scalar phase field to be unaffected by the concentrations of other logistics tasks. Therefore, the multiphase field coupling solution module constructs an independently evolving reaction-diffusion partial differential equation for the first scalar phase field. This equation consists of partial derivatives of concentration over time, anisotropic diffusion terms, and spatial reaction consumption terms. Step 303, for the first scalar phase field corresponding to the urgent logistics task, the specific calculation formula of the high-priority phase field reaction-diffusion partial differential equation constructed by the multiphase field coupling solution module is as follows: ; In the formula, The spatial concentration distribution function of the first scalar phase field time variables The partial derivatives; It is a spatial divergence operator used to calculate the flux divergence of a vector field; This is a spatial gradient operator used to calculate the spatial rate of change of the concentration field; For continuous coordinate points In time variable The corresponding fundamental anisotropic diffusion tensor, at this continuous coordinate point In the discretized numerical solution of the system, it is mapped to the aforementioned corresponding discretized grid points. ; For high-priority destinations at continuous coordinate points The spatial reaction consumption rate parameter is always zero in non-destination regions. The multiphase field coupling solution module configures the first scalar phase field concentration distribution and its gradient variable generated by the above equation as a globally shared parameter, serving as a mathematical reference benchmark for the subsequent construction of low-priority phase field cross-suppression coupling terms.

[0043] The process of constructing low-priority phase-field partial differential equations by the multiphase field coupling solution module includes: Step 304: The multiphase field coupling solution module reads the first scalar phase field concentration distribution function, which is configured as a globally shared parameter. And calculate the concentration distribution function at continuous coordinate points. Spatial gradient vector at Then, the magnitude of the spatial gradient vector is obtained, and the magnitude of the local concentration gradient is generated. The numerical value of the local concentration gradient modulus characterizes the degree of aggregation of urgent logistics tasks in the corresponding spatial location.

[0044] Step 305, the multiphase field coupling solution module is based on the above local concentration gradient magnitude Construct a nonlinear cross-suppression factor and incorporate it as a suppression term into the fundamental anisotropic diffusion tensor. In the middle, the dynamic equivalent diffusion tensor dedicated to low-priority phase fields is calculated. Dynamic equivalent diffusion tensor The specific calculation formula is as follows: ; In the formula, This is a preset system coupling strength coefficient used to control the absolute magnitude of cross-suppression. This is a sensitivity parameter used to adjust the mathematical response rate to changes in the spatial gradient of the first scalar phase field; Given a second-order identity matrix and a nonlinear function. The local concentration gradient magnitude is dimensionless and mapped to... In the numerical range, to prevent the aforementioned suppression terms from causing negative eigenvalues ​​in the equivalent diffusion tensor and thus leading to numerical solution divergence, the multiphase field coupled solution module performs calculations on the calculated values. Perform eigenvalue lower bound truncation, which is mandatory. All eigenvalues ​​are not less than the preset minimum positive real number. This ensures that the dynamic equivalent diffusion tensor strictly satisfies the mathematical conditions of a positive definite matrix.

[0045] Step 306: The multiphase field coupling solution module utilizes the processed dynamic equivalent diffusion tensor. A partial differential equation containing cross-suppression coupling terms is constructed for the second scalar phase field corresponding to ordinary logistics tasks. Through this coupling equation, the low-priority phase field is constrained by a dynamically decaying traffic tensor when it diffuses into the high-priority vehicle-dense area. The calculation formula of the coupled partial differential equation for the evolution of the second scalar phase field is as follows: ; In the formula, The spatial concentration distribution function of the second scalar phase field time variables The partial derivatives; It is a spatial divergence operator; For spatial gradient operators; For low-priority destinations at continuous coordinate points The spatial reaction consumption rate parameter at the location, and the continuous coordinate points in the above formula. When performing spatial discretization, the discretized grid points within the manifold After the multiphase field coupling solution module completes the construction of the coupled partial differential equation through one-to-one matching, it merges the equation with the aforementioned reaction-diffusion partial differential equations of independent evolution of the high-priority phase field to form a complete set of multi-priority coupled reaction-diffusion partial differential equations. Subsequently, the multiphase field coupling solution module inputs the set of partial differential equations into the numerical solver for subsequent discretization iterative calculations.

[0046] The discretization process of the multiphase field coupling solution module includes: Step 307: The multiphase field coupling solution module obtains the spatial resolution parameters of the discretized grid points within the continuous spatial manifold, and determines the lateral spatial step size of the two-dimensional spatial grid based on these spatial resolution parameters. With longitudinal spatial step size Meanwhile, the multiphase field coupling solution module calculates and determines the time step of the discretization iteration based on the numerical stability conditions of the partial differential equation solution. To prevent divergence in numerical iterative calculations, the system mandates this time step. Satisfying the von Neumann stability constraint for two-dimensional diffusion, the calculation formula is as follows: ,in It is the maximum value among the eigenvalues ​​of the fundamental anisotropic diffusion tensor corresponding to all grid points in the current manifold space.

[0047] Step 308: The multiphase field coupling solution module uses the finite difference method to replace the continuous partial derivative terms in the multi-priority coupled reaction-diffusion partial differential equation system with discrete difference quotients. Specifically, for the partial derivative terms of concentration over time, the system uses a first-order time-forward difference scheme for discretization; for the spatial divergence and gradient operation terms that include the anisotropic diffusion tensor, a second-order spatial center difference scheme is used for discretization.

[0048] Step 309: The multiphase field coupled solution module constructs the discretized numerical update equation. Taking the first scalar phase field corresponding to the urgent logistics task as an example, the multiphase field coupled solution module updates it at time step... and spatial grid points The discretization update formula at the point is set as follows: ; In the formula, For time steps The concentration values ​​at the first scalar phase field grid points; For the previous iteration time step The corresponding grid point concentration values; In time step And spatial grid points At this point, a second-order spatial central difference operator based on the fundamental anisotropic diffusion tensor is used. For the spatial reaction consumption rate parameter corresponding to the grid point, the second scalar phase field discretization update formula for ordinary logistics tasks has the same structure as the above formula. The multiphase field coupled solution module replaces its corresponding difference operator with one based on the dynamic equivalent diffusion tensor. The second-order spatial central difference operator.

[0049] Step 310: The multiphase field coupling solution module, combining the initial concentration distribution field matrices of each scalar phase field set by the aforementioned system and the homogeneous Dirichlet boundary conditions, initiates time-domain iterative calculations, and at each time step... After iteration, the scalar concentration values ​​of all discretized grid points in the stored manifold space are updated and overwritten. When the number of iterations reaches the preset control cycle iteration threshold, the multiphase field coupling solution module stops the equation iteration process within the control cycle and outputs the global scalar concentration field distribution results corresponding to each priority task. Subsequently, the multiphase field coupling solution module transmits the global scalar concentration field distribution results to the discretization adaptive control module to provide the data foundation required for subsequent local concentration negative gradient vector extraction.

[0050] The process of the discretized adaptive control module extracting local concentration features includes: Step 401: The discretized adaptive control module obtains the current physical positioning coordinates of the controlled logistics vehicle in the physical road network through the vehicle positioning receiver. The physical positioning coordinates are composed of longitude and latitude data generated by the global satellite navigation system and are used to characterize the absolute spatial position of the controlled logistics vehicle.

[0051] Step 402: The discretized adaptive control module obtains the spatial boundary parameters constructed by the aforementioned continuous spatial mapping module based on the extreme coordinates of the road network nodes. Based on these spatial boundary parameters, it calibrates the affine transformation matrix from the latitude and longitude coordinate system to the manifold coordinate system. Using this affine transformation matrix, the module maps the aforementioned physical positioning coordinates to the constructed two-dimensional or three-dimensional continuous spatial manifold, and calculates the corresponding continuous coordinate points. Subsequently, the system performs a Euclidean nearest neighbor search within the manifold space, traversing to find points with continuous coordinates. Discretized grid points with the smallest spatial distance And the extracted discretized grid points It is set as the mapping anchor point of the controlled logistics vehicle in the numerical computation space of partial differential equations.

[0052] Step 403: The discretized adaptive control module reads the global scalar concentration field distribution result output by the multiphase field coupling solution module and parses the task priority attribute inherent in the controlled logistics vehicle. Based on this task priority attribute, it matches the corresponding first or second scalar phase field in the global scalar concentration field distribution result. In the determined corresponding scalar phase field, the discretized adaptive control module uses the aforementioned mapping anchor point... Using the set grid search radius as the center point, extract the mapping anchor point. The set of discretized grid points in the neighborhood is obtained. Then, the scalar concentration values ​​of each point in the set of discretized grid points in the neighborhood are read from the corresponding scalar phase field, and combined to generate a local concentration field matrix. This local concentration field matrix is ​​stored in the system memory as the data basis for subsequent calculation of the local concentration negative gradient vector.

[0053] The process by which the discretized adaptive control module calculates the local concentration negative gradient vector includes: Step 404: The discretized adaptive control module reads the local concentration field matrix stored in the system memory and uses the second-order central difference algorithm to perform spatial differentiation on the discrete scalar concentration values ​​within the local concentration field matrix to calculate the mapping anchor point of the controlled logistics vehicle. The lateral and longitudinal partial derivatives of concentration at a given point are then used to synthesize the arrays of lateral and longitudinal partial derivatives to form the mapping anchor point. The local concentration gradient vector is obtained, and the negative one is multiplied by the local concentration gradient vector to generate the local negative gradient vector.

[0054] Step 405: Since the aforementioned local negative gradient vector is defined under the coordinate reference of a two-dimensional or three-dimensional continuous spatial manifold, in order to match it with the actual direction of the physical road network, the discretized adaptive control module retrieves the affine transformation matrix from the previously constructed latitude and longitude coordinate system to the manifold coordinate system, calculates the inverse matrix of the affine transformation matrix, generates the inverse physical calibration matrix, and then uses the inverse physical calibration matrix to perform a linear transformation multiplication operation on the local negative gradient vector in the manifold space, calibrating and mapping it back to the latitude and longitude coordinate system of the physical road network, and outputs the physical direction negative gradient vector.

[0055] Step 406: The discretized adaptive control module normalizes the magnitude of the transformed physical direction negative gradient vector to eliminate the influence of the absolute difference in gradient magnitude on subsequent direction similarity comparison calculations. To prevent the division by zero operation from crashing when the gradient is zero at a local minimum point, a minimum positive real number is introduced as a denominator protection term. The normalization calculation formula for the unit physical negative gradient vector is as follows: ; In the formula, This is the normalized unit physical negative gradient vector; This is the negative gradient vector of the physical direction after inverse coordinate system calibration; Let be the Euclidean space magnitude of the negative gradient vector in this physical direction; For a preset minimal positive real constant, the discretized adaptive control module will calculate the unit physical negative gradient vector. The instruction buffer of the current control cycle is entered as an angle reference for generating the next driving instruction.

[0056] The process of instruction generation and scheduling for the discretized adaptive control module includes: Step 407: The discretized adaptive control module determines the physical road network node where the controlled logistics vehicle is currently located in the discrete logistics road network topology data based on the current physical location coordinates of the controlled logistics vehicle, and extracts all adjacent physical road segments connected to the physical road network node that allow the vehicle to pass, and combines them to form a candidate road segment set.

[0057] Step 408: For each adjacent physical road segment in the candidate road segment set, the discretized adaptive control module obtains the spatial coordinates of the physical road network node where the vehicle is currently located and the far-end nodes of each adjacent physical road segment, and calculates the physical direction vector from the current node to the far-end node. In order to unify the calculation benchmark, the module performs modulus normalization processing on each physical direction vector to generate the unit direction vector corresponding to each adjacent physical road segment.

[0058] Step 409: The discretized adaptive control module extracts the previously generated unit physical negative gradient vector. As the desired guiding vector, the cosine similarity between this desired guiding vector and each unit direction vector in the candidate road segment set is calculated. The specific formula for calculating the cosine similarity is as follows: ; In the formula, For the desired guiding vector and the first Cosine similarity of adjacent physical road segments; For the first The unit direction vector of each adjacent physical road segment; It is a vector inner product operator; It is a positive integer, and , This represents the total number of adjacent physical road segments in the candidate road segment set, based on the fact that the magnitude of each vector involved in the calculation is in units. Given the constraints, the system directly calls the vector inner product operator to output the cosine similarity value.

[0059] Step 410: The discretized adaptive control module iterates through and compares the above calculation results, and selects the adjacent physical road segment with the largest cosine similarity value. The maximum cosine similarity value indicates that the spatial angle between the actual direction of the corresponding physical road segment and the concentration decrease direction output by the partial differential field is the smallest. Subsequently, the discretized adaptive control module converts the selected adjacent physical road segment into a next-hop driving command and sends it to the corresponding vehicle terminal through the wireless communication network. After the vehicle terminal receives and executes the command and updates the vehicle's physical positioning coordinates, the discretized adaptive control module triggers the system to return, using the updated vehicle position as a new concentration diffusion source, and enters the parameter initialization and equation solving loop of the next control cycle.

[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based dynamic logistics route planning system, including: The continuous space mapping module, tensor prediction and regression module, multiphase field coupling solution module, and discretized adaptive control module are characterized by: The continuous space mapping module is used to acquire discrete logistics road network topology data, map the discrete logistics road network topology data into a continuous space manifold, establish multiple scalar phase fields on the continuous space manifold, and configure the logistics delivery destination as the boundary condition of the partial differential equation to complete the parameter initialization. The tensor prediction and regression module is used to collect multi-source environmental feature data of the road network, process the multi-source environmental feature data of the road network using a feature extraction network, and output the basic anisotropic diffusion tensor corresponding to the continuous spatial manifold. The multiphase field coupling solution module is used to read the scalar phase field parameters and the basic anisotropic diffusion tensor, construct a set of reaction-diffusion partial differential equations between different priority logistics tasks, and discretize the set of reaction-diffusion partial differential equations to output the global scalar concentration field distribution results. The discretized adaptive control module is used to obtain the physical positioning coordinates of the controlled logistics vehicle, calculate the local negative gradient vector based on the physical positioning coordinates in the global scalar concentration field distribution result, generate the next-hop driving command based on the directional similarity between the local negative gradient vector and the adjacent physical road segment, and send it to the vehicle terminal.

2. The AI-based dynamic logistics route planning system according to claim 1, characterized in that: When the continuous spatial mapping module maps the discrete logistics road network topology data into a two-dimensional or three-dimensional continuous spatial manifold through a spatial interpolation algorithm, it analyzes the spatial coordinates of each node in the road network node set, determines the spatial boundary based on the extreme coordinates of the road network nodes to construct the continuous spatial manifold, and configures a Gaussian kernel function as a spatial interpolation kernel function on the continuous spatial manifold. The spatial interpolation kernel function is used to calculate the basic mapping field function of each discretized grid point. By superimposing the kernel mapping values ​​of each discrete logistics network node, a continuous topological distribution field is established, generating a mapped continuous topological data matrix.

3. The AI-based dynamic logistics route planning system according to claim 1, characterized in that: When the continuous space mapping module establishes multiple scalar phase fields and completes parameter initialization, the multiple scalar phase fields are used to distinguish logistics tasks of different priorities. Specifically, a first scalar phase field corresponding to an urgent logistics task and a second scalar phase field corresponding to an ordinary logistics task are established. The physical location coordinates and task priority attributes of the controlled logistics vehicles are obtained. The physical location coordinates are mapped to continuous coordinate points in the continuous spatial manifold. Based on the task priority attributes, each controlled logistics vehicle is assigned to the first scalar phase field or the second scalar phase field as an initial concentration diffusion source. The spatial concentration point source of each controlled logistics vehicle at the initial time is defined by the spatial Dirac function after grid smoothing and regularization, and the spatial parameter initialization configuration of the logistics vehicle load is completed.

4. The AI-based dynamic logistics route planning system according to claim 1, characterized in that: When the continuous space mapping module configures the logistics delivery destination as the boundary condition of the partial differential equation, it obtains the set of logistics delivery destinations in the target area, maps the physical location coordinates of each logistics delivery destination to the corresponding continuous coordinate points inside the continuous space manifold, and applies homogeneous Dirichlet boundary conditions at the continuous coordinate points to constrain the concentration value of the logistics delivery destination area to remain constant at zero. Simultaneously, the logistics throughput capacity calibration values ​​of each logistics delivery destination are obtained, and the logistics throughput capacity calibration values ​​are converted into spatial reaction consumption rate parameters at corresponding continuous coordinate points based on a positive proportional mapping relationship.

5. The AI-based dynamic logistics route planning system according to claim 1, characterized in that: The multi-source environmental feature data of the road network includes real-time traffic flow, average driving speed of road segments, status of sudden road closure events, and meteorological precipitation data. Before the tensor prediction regression module processes the multi-source environmental feature data of the road network using the feature extraction network, it transforms the multi-source environmental feature data of the road network into a multimodal environmental feature matrix. After aligning the multi-source environmental feature data of the road network in terms of spatiotemporal dimensions, it uses an extreme value normalization algorithm to map the values ​​to a dimensionless numerical range. For any discretized grid point within the continuous spatial manifold, the multi-source environmental feature data of the road network corresponding to the discretized grid point are concatenated to generate a one-dimensional feature vector. The feature vector is then constructed by splicing the discrete grid points according to their spatial coordinates. The feature vectors are stacked in the time dimension according to the set time window length to generate a three-dimensional spatiotemporal feature tensor.

6. The AI-based dynamic logistics route planning system according to claim 5, characterized in that: The feature extraction network used by the tensor prediction and regression module is a spatiotemporal graph convolutional network. When the tensor prediction and regression module outputs the basic anisotropic diffusion tensor corresponding to the continuous spatial manifold, it outputs the basic anisotropic diffusion tensor corresponding to each discretized grid point in the continuous spatial manifold, which is used to characterize the passage resistance in each direction in the continuous spatial manifold. A sparse topological adjacency matrix based on grid neighborhood connectivity is constructed according to the regular arrangement relationship of discretized grid points. Self-loops are added to the sparse topological adjacency matrix and degree matrix normalization is performed to generate a normalized adjacency matrix. After the three-dimensional spatiotemporal feature tensor is expanded into an input feature matrix according to the time step sequence, the input graph convolutional layer combines the normalized adjacency matrix to perform node feature aggregation operation to extract the spatial feature representation matrix. The spatial feature representation matrix is ​​input into the corresponding gated recurrent unit according to the node index, the top-level hidden state matrix output at the final time step is extracted, and linear dimension transformation is performed through a fully connected mapping layer to output the original regression matrix corresponding to each discretized grid point as the initial state parameters of the basic anisotropic diffusion tensor.

7. The AI-based dynamic logistics route planning system according to claim 6, characterized in that: When the tensor prediction regression module outputs the basic anisotropic diffusion tensor, it makes the basic anisotropic diffusion tensor subject to a symmetric positive definite matrix constraint, specifically including: adding the original regression matrix and the transpose of the original regression matrix and taking the arithmetic mean to generate a transitional symmetric matrix. The transition symmetric matrix is ​​decomposed into eigenvalues ​​to extract the diagonal matrix and orthogonal eigenvector matrix containing the initial eigenvalues. Under normal operating conditions, the initial eigenvalues ​​are mapped to positive eigenvalues ​​greater than the preset minimum positive real number using nonlinear constraint operators. When the status data of a sudden road closure event indicates that the corresponding road segment is completely physically blocked, calculate the cosine similarity between each column of feature vectors and the physical direction vector of the road segment, extract the column of feature vectors with the highest similarity, and reset the corresponding feature value to the lower limit value. The fundamental anisotropic diffusion tensor corresponding to each discretized grid point is generated by performing a reconstruction multiplication operation using the orthogonal eigenvector matrix and the corrected eigenvalue diagonal matrix.

8. The AI-based dynamic logistics route planning system according to claim 3, characterized in that: When constructing the reaction-diffusion partial differential equation system between logistics tasks of different priorities, the multiphase field coupling solution module introduces a cross-suppression coupling term. Specifically, for the first scalar phase field corresponding to the urgent logistics task, based on the business scheduling logic that the urgent logistics task is not affected by the concentration of other logistics tasks, an independently evolving reaction-diffusion partial differential equation is constructed, consisting of the partial derivative of concentration over time, anisotropic diffusion term, and spatial reaction consumption term. For the second scalar phase field corresponding to a general logistics task, the local concentration gradient magnitude of the first scalar phase field is extracted to construct a nonlinear cross-suppression factor. This nonlinear cross-suppression factor is then incorporated as a suppression term into the basic anisotropic diffusion tensor to calculate the dynamic equivalent diffusion tensor. The dynamic equivalent diffusion tensor is then subjected to eigenvalue lower bound truncation to ensure that it satisfies the positive definite matrix condition. This process constructs a coupled partial differential equation for the evolution of the second scalar phase field containing the dynamic equivalent diffusion tensor. The coupled partial differential equation is then merged with the reaction-diffusion partial differential equation of the first scalar phase field to form a set of reaction-diffusion partial differential equations containing cross-suppression coupling terms.

9. The AI-based dynamic logistics route planning system according to claim 1, characterized in that: The multiphase field coupling solution module uses the finite difference algorithm or the finite element algorithm to discretize and solve the reaction-diffusion partial differential equation system. When the finite difference algorithm is used, the horizontal and vertical spatial step sizes of the two-dimensional spatial grid are determined according to the spatial resolution parameters, and the time step size of the discretization iteration is determined based on the von Neumann stability constraint that satisfies two-dimensional diffusion. The continuous partial derivative terms in the reaction-diffusion partial differential equation system are replaced with discrete difference quotients using the finite difference method. The time partial derivative terms are discretized using a first-order forward time difference scheme, and the spatial divergence and gradient operation terms are discretized using a second-order central space difference scheme, thus constructing a spatiotemporal discretized numerical update equation. The global scalar concentration field distribution result is output by combining the initial concentration distribution field matrix and the homogeneous Dirichlet boundary conditions in the time domain iterative calculation.

10. The AI-based dynamic logistics route planning system according to claim 1, characterized in that: When the discretized adaptive control module calculates the local negative gradient vector and generates the next-hop driving command, it obtains the spatial boundary parameters of the continuous spatial manifold to construct the coordinate system affine transformation matrix, and maps the physical positioning coordinates of the controlled logistics vehicle to the continuous spatial manifold through the coordinate system affine transformation matrix and performs nearest neighbor search to lock the corresponding discretized grid points as mapping anchor points. Based on the task priority attribute, the corresponding scalar phase field is matched in the global scalar concentration field distribution result. The scalar concentration values ​​of the neighborhood discretized grid point set of the mapping anchor point are extracted to generate a local concentration field matrix. The local negative gradient vector at the mapping anchor point is calculated using the second-order central difference algorithm. The inverse matrix of the coordinate system affine transformation matrix is ​​retrieved to calibrate and map the local negative gradient vector back to the physical road network coordinate system to output the physical direction negative gradient vector. The negative gradient vector of the physical direction is normalized to generate the expected guidance vector. Adjacent physical road segments connected to physical road network nodes are extracted to form a candidate road segment set. The cosine similarity between the expected guidance vector and the unit direction vector of each adjacent physical road segment in the candidate road segment set is calculated as the direction similarity. The adjacent physical road segment with the largest cosine similarity value is selected as the next-hop driving command and issued. After the vehicle position is updated, the system returns to the parameter initialization stage.

Images