A decision analysis method for smart logistics

Abstract

The present application relates to the technical field of intelligent logistics, and particularly relates to a decision analysis method for intelligent logistics. The method comprises the following steps: acquiring logistics network topology data, order information, vehicle real-time state data and multi-dimensional dynamic state characteristics; generating a heuristic information matrix based on the logistics network topology data and the order information; combining the heuristic information matrix and the pheromone tensor to construct a candidate decision scheme, updating the pheromone tensor according to the comprehensive fulfillment cost of the candidate decision scheme, the vehicle real-time state data and the multi-dimensional dynamic state characteristics; projecting the candidate decision scheme to a decision hidden space through a variational autoencoder model according to a plurality of candidate decision schemes generated by the updating, performing optimization in the decision hidden space, and generating a final decision analysis scheme after decoding. The scheme of the present application can effectively avoid the defect that the traditional optimization is easy to fall into local optimization, generate a decision analysis scheme with better global performance, and significantly reduce the comprehensive fulfillment cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent logistics, and particularly relates to a decision analysis method for intelligent logistics. BACKGROUND

[0002] Under the background of modern economic globalization and rapid development of e-commerce, intelligent logistics has become a key infrastructure supporting the efficient operation of society. One of the core challenges is how to make dynamic, accurate and efficient decision analysis on complex vehicle routing problem (VRP). VRP aims to plan a vehicle distribution scheme that can minimize the total operation cost (such as driving distance, time, fuel consumption) for a series of customer orders with specific needs (such as delivery priority, service time window). In order to solve the above problem, researchers have proposed a variety of meta-heuristic algorithms, among which the ant colony optimization algorithm (ACO) has been widely used due to its strong parallel search ability and positive feedback mechanism. However, the traditional ACO and its variants have obvious shortcomings in dealing with modern logistics scenarios: first, the pheromone model of traditional ACO is usually a two-dimensional matrix based on path segments, which is difficult to effectively encode and utilize dynamic information such as time, vehicle state and other multi-dimensional information; second, the heuristic information relied on by the traditional ACO for path selection is mostly static and pre-set rules (such as the reciprocal of the distance between nodes), which cannot adaptively learn the deep relevance between nodes from complex logistics network topology data, nor can it effectively integrate real-time environmental factors such as traffic congestion and weather changes into the decision-making process, resulting in the algorithm being prone to local optimum and slow response to dynamic environment.

[0003] To overcome the limitations of traditional methods, in recent years, artificial intelligence technologies represented by deep learning have been gradually introduced into the field of logistics decision-making. For example, graph neural networks (GNN) and other models are used to represent the topological structure of the logistics network, so as to learn more informative node embedding representations to assist in path planning. This data-driven paradigm has shown great potential and can mine patterns from historical and real-time data that are difficult for traditional methods to capture. However, existing technologies still face challenges in deep integration and global optimization. On the one hand, the combination of deep learning models and meta-heuristic algorithms (such as ant colony algorithm) often stays at a shallow level, such as using neural networks to generate initial solutions or replacing static heuristic information, and fails to achieve deep coupling and co-evolution of the two in the iterative optimization process. On the other hand, when the algorithm generates a series of high-quality candidate decision schemes, how to efficiently explore and refine in the vast and discrete solution space to find the final optimal solution is still a technical bottleneck. Existing methods often lack effective modeling and utilization mechanisms for the internal structure of the high-quality solution group, making it difficult to further improve the decision scheme based on the existing good foundation. Therefore, there is an urgent need for a decision analysis method that can deeply integrate graph learning, dynamic environment perception, and advanced optimization strategies to meet the urgent needs of smart logistics for high efficiency, high robustness, and strong adaptability. SUMMARY

[0004] The present application aims to provide a decision analysis method for smart logistics to solve the problem of poor global performance and high comprehensive fulfillment cost of existing logistics decision analysis schemes. To this end, the present application provides a solution in the following aspect.

[0005] The application provides a decision analysis method for intelligent logistics, which comprises the following steps: acquiring logistics network topology data, order information containing delivery priority and time window, vehicle real-time state data, and multi-dimensional dynamic state characteristics including traffic conditions and weather; based on the logistics network topology data and the order information, a graph attention network is used to model the connection relationship between nodes in the logistics network to generate a heuristic information matrix; in the iterative path construction process, the heuristic information matrix and a pheromone tensor containing three dimensions of path segment, time segment and vehicle load rate are combined, and a probability selection model is used to determine the next travel path segment for the decision-making node in the path, so that a complete candidate decision scheme is constructed, and the comprehensive implementation cost thereof is calculated according to a preset cost model; according to the comprehensive implementation cost of the candidate decision scheme, the vehicle real-time state data and the multi-dimensional dynamic state characteristics are fused, and the pheromone tensor is updated; the path construction and pheromone updating process is repeated to generate multiple candidate decision schemes, the candidate decision schemes are vectorized and projected to a decision hidden space through a variational autoencoder model, optimization is performed in the decision hidden space according to a global cost function, and the optimized hidden vector is decoded to generate a final decision analysis scheme.

[0006] Compared with the prior art, the application deeply mines the deep correlation in the logistics network topology and the order information through the graph attention network, generates a heuristic information that can accurately reflect the complex connection relationship between nodes, provides higher quality guidance for path construction, and improves the quality of candidate solutions from the source. At the same time, a three-dimensional pheromone tensor containing path segment, time segment and vehicle load rate is introduced, so that the decision-making process can consider the time window and vehicle capacity and other key constraints in an integrated manner, and a more refined and economic scheduling scheme is obtained. Moreover, the real-time state characteristics of vehicles, traffic and weather are integrated in the pheromone updating link, which ensures that the optimization direction is highly aligned with the real world, and enhances the real feasibility of the decision scheme. The variational autoencoder is used to map multiple candidate decision schemes to a structured decision hidden space for global optimization, which effectively avoids the defect that the traditional optimization is easy to fall into local optimum, and finally generates a decision analysis scheme with better global performance, which significantly reduces the comprehensive implementation cost.

[0007] Preferably, the logistics network topology data comprises the geographical coordinates and connection relationship of warehouses and customer points; the vehicle real-time state data comprises vehicle position, speed and remaining power or oil quantity; and the multi-dimensional dynamic state characteristics comprise regional traffic congestion index and weather forecast in the next few hours.

[0008] Preferably, the connection relationship between nodes in the logistics network is modeled by using a graph attention network to generate a heuristic information matrix, including: all warehouses and customer points are taken as nodes of a graph, and feasible paths are taken as edges to construct a logistics network graph structure; the type, cargo demand, priority and service time window information of each logistics node are encoded into an initial node feature vector; the logistics network graph structure and the node feature vector are input into a pre-trained graph attention network model, the deep dependency between nodes is learned through a multi-head attention mechanism, an embedding representation is generated for each pair of nodes in the graph attention network, and a heuristic information value is calculated based on the similarity of the embedding representations of any two nodes to form the heuristic information matrix.

[0009] The heuristic information value is calculated based on the similarity of the node embedding representations to form the heuristic information matrix, which can provide efficient guidance information for path construction. This matrix not only reflects the static relationship between nodes, but also incorporates dynamic state features (such as priority, time window), thereby significantly improving the relevance of path optimization.

[0010] Preferably, the connection relationship between nodes in the logistics network is modeled by using a graph attention network to generate a heuristic information matrix, including: the geographic location, cargo demand and service time window information of each logistics node are encoded into an initial node feature vector; the initial node feature vector is input into a multi-layer graph attention network model, the connection weight between any two nodes is calculated and output through an attention mechanism to form the heuristic information matrix.

[0011] The connection weight between any two nodes is calculated by using a multi-layer graph attention network model through an attention mechanism, which can dynamically learn the deep dependency and interaction intensity between nodes to generate a high-precision heuristic information matrix, thereby improving the relevance and accuracy of path planning.

[0012] Preferably, in the probability selection model, the transition probability of a vehicle selecting the next unvisited node at the current node is proportional to the weighted product of the pheromone concentration and the heuristic information corresponding to the path segment; the weights of the pheromone concentration and the heuristic information are adjusted by preset parameters a and β.

[0013] The pheromone concentration can reflect the advantages and disadvantages of historical paths, and the heuristic information can provide guidance based on network topology and business demand. The weighted product of the two enables the vehicle to consider historical experience and prior knowledge when selecting the next node, thereby improving the efficiency and quality of path construction.

[0014] Preferably, the comprehensive fulfillment cost is composed of at least two of the driving distance cost, the time-related cost and the penalty cost, wherein the time-related cost includes order waiting time and vehicle service time, and the penalty cost includes a fee caused by violating a customer time window or a vehicle capacity limit.

[0015] Preferably, the updating of the pheromone tensor comprises: performing a global evaporation operation on all values in the three-dimensional pheromone tensor; selecting one or several candidate decision schemes with the lowest comprehensive fulfillment cost in the current iteration, and increasing the pheromone concentration at the positions in the three-dimensional pheromone tensor corresponding to the path segments, time segments and vehicle load rate intervals in the candidate decision schemes; the increased pheromone concentration value is inversely proportional to the comprehensive fulfillment cost of the scheme, and the increased pheromone concentration value is dynamically adjusted by weighting the traffic congestion index and the vehicle energy consumption data corresponding to the path segments.

[0016] The global evaporation operation periodically reduces the pheromone concentration, and the pheromone increment updating of the lowest comprehensive fulfillment cost scheme dynamically strengthens the high-quality path, effectively avoids falling into a local optimal solution, and improves the quality of global path optimization. Moreover, the three-dimensional pheromone tensor (path segment, time segment, vehicle load rate) combined with the weighted adjustment of the traffic congestion index and the energy consumption data can comprehensively integrate multi-dimensional logistics information, and ensure that the updated pheromone tensor can accurately reflect the complex dynamic characteristics of the logistics network.

[0017] Preferably, the updating of the pheromone tensor comprises: performing an evaporation operation on the pheromone, and the updating rule is related to a preset evaporation rate; and enhancing the pheromone of the path segments in the candidate decision scheme according to the comprehensive fulfillment cost of the candidate decision scheme, and the enhancement amount is proportional to a preset pheromone intensity coefficient and inversely proportional to the comprehensive fulfillment cost.

[0018] Preferably, the vectorization of the candidate decision scheme and the projection of the vectorized candidate decision scheme to the decision latent space by the variational autoencoder model comprise: processing each vehicle path in the candidate decision scheme into a unified node sequence; converting the node sequence into a fixed-dimensional numerical vector as the input of the variational autoencoder model, and outputting the parameters defining the probability distribution in the decision latent space from the encoder of the variational autoencoder model.

[0019] Preferably, the optimization in the decision latent space according to the global cost function comprises: generating an initial latent vector by sampling from the probability distribution in the decision latent space; and performing iterative updating on the initial latent vector according to a preset number of iterations and a learning rate by using a gradient optimization algorithm to minimize the global cost function.

[0020] Compared with the prior art, the application deeply mines the deep correlation between the logistics network topology and the order information through the graph attention network, generates heuristic information that can accurately reflect the complex connection relationship between nodes, provides higher quality guidance for path construction, and improves the quality of candidate solutions from the source. At the same time, a three-dimensional pheromone tensor containing path segments, time segments and vehicle load rates is introduced, so that the decision process can integrate the time window and vehicle capacity and other key constraints for consideration, and obtain a more refined and economic scheduling scheme. Moreover, real-time state features such as vehicles, traffic and weather are integrated into the pheromone update link to ensure that the optimization direction is highly aligned with the real world, and the real feasibility of the decision scheme is enhanced. The use of a variational autoencoder maps multiple good candidate solutions into a structured decision hidden space for global optimization, effectively avoiding the defect that traditional optimization is prone to local optimization. Through overall exploration and refinement of the solution space, a decision analysis scheme with better global performance is finally generated, significantly reducing the overall fulfillment cost. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A step flowchart of the decision analysis method for intelligent logistics in the embodiment is schematically shown;

[0022] Figure 2 A three-dimensional pheromone tensor structure diagram is schematically shown. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application.

[0024] As Figure 1 shown, the intelligent security image recognition method based on deep learning in the embodiment includes the following steps:

[0025] Step S1, acquiring logistics network topology data, order information containing delivery priority and time window, vehicle real-time state data, and multi-dimensional dynamic state features including traffic conditions and weather.

[0026] Specifically, the logistics network topology data, such as the geographic coordinates and connection relationship of warehouses and customer points, is derived from a geographic information system; the order information is accessed in real time from an enterprise resource planning or order management system; the vehicle real-time state data such as position, speed and remaining power or oil is reported through a vehicle-mounted Internet of Things terminal; the multi-dimensional dynamic state features such as regional traffic congestion index and weather forecast in the next few hours are obtained by calling the data API interface of third-party map service providers and weather services, and all data is stored in a database after cleaning and format unification for calling.

[0027] In step S2, based on the logistics network topology data and order information, a graph attention network is used to model the connection relationship between nodes in the logistics network to generate a heuristic information matrix.

[0028] Specifically, all warehouses and customer points are taken as nodes of a graph, and feasible paths are taken as edges to construct a logistics network graph structure; each node is encoded into an initial node feature vector according to its type, cargo demand, priority, service time window and other information; then, the logistics network graph structure and the initial node feature vector are input into a pre-trained graph attention network model, the model learns the deep dependency between nodes through a multi-head attention mechanism, finally generates an embedding representation for each pair of nodes in the graph attention network, and calculates a heuristic information value based on the similarity of the embedding representations of any two nodes, all these values together constitute the heuristic information matrix.

[0029] In an optional embodiment, modeling the connection relationship between nodes in the logistics network by using a graph attention network includes: encoding the geographic location, cargo demand, service time window and other information of each logistics node into a node initial feature vector; inputting the initial feature vector into a multi-layer graph attention network model to calculate and output the connection weight between any two nodes through an attention mechanism, forming a heuristic information matrix.

[0030] Specifically, this process first features each node. For example, a distribution center node A has a geographic coordinate of longitude 116.4, latitude 39.9, a cargo demand of 200 pieces on the current day, and a service time window of 8 a.m. to 10 a.m. These information, i.e. 116.4, 39.9, 200, 8, 10, will be integrated and input into an embedding layer to be converted into an initial node feature vector of 128 dimensions. Similarly, customer nodes B, C, D, etc. also generate their respective node feature vectors in this way to form an initial feature matrix, providing basic data input for subsequent network analysis.

[0031] The initial node feature vector containing rich node information is sent to a graph attention network model with a three-layer structure for deep processing. In the first layer of the network, the model calculates the attention score between node A and all other nodes, which reflects the importance of other nodes in determining the connection strategy of node A. Nodes that are geographically adjacent and have compatible time windows will get higher attention weights. After nonlinear transformation and information propagation through three layers of network, the network finally outputs an N by N matrix, where N is the total number of nodes. Each element value in the matrix, for example, the value in the i-th row and j-th column, represents the heuristic value of the path segment from node i directly to node j, and the higher the value, the greater the potential value of the path segment in constructing a high-quality route solution. This matrix is the heuristic information matrix.

[0032] Step S3, in the iterative path construction process, the heuristic information matrix is combined with a pheromone tensor containing three dimensions of path segments, time segments and vehicle load rates, and a probability selection model is used to determine the next travel path segment for the pending decision node in the path, so as to construct a complete candidate decision scheme, and a preset cost model is used to calculate the comprehensive fulfillment cost thereof.

[0033] Specifically, as shown in Figure 2 At each step of path construction, the decision maker located at the current node, i.e. the virtual ant, selects the next customer node to be visited according to a transition probability formula; the transition probability is proportional to the product of the value of the target path segment in the heuristic information matrix and the concentration value of the corresponding current time segment and vehicle load interval in the pheromone tensor; this process is repeated until all orders are assigned to the path, forming a set of vehicle paths covering all orders, i.e. a candidate decision scheme; subsequently, a preset cost model calculates the total travel time, the penalty cost of overtime window and the vehicle energy consumption cost of the candidate decision scheme, and obtains the comprehensive fulfillment cost thereof.

[0034] In an optional embodiment, the next travel path segment is determined by a probability selection model, wherein the transition probability of the vehicle in the current node to select the next unvisited node is proportional to the weighted product of the pheromone concentration and the heuristic information corresponding to the path segment; the weights of the pheromone concentration and the heuristic information are adjusted by preset parameters a and β.

[0035] Specifically, when a delivery vehicle is located at the current node i and needs to select the next destination from a set of candidate nodes that have not been visited, a transition probability is calculated for each possible path segment. The calculation of the transition probability integrates two kinds of information: historical experience, i.e. pheromone concentration τ(i,j), and prior knowledge given by the graph attention network, i.e. heuristic information η(i,j). The transition probability P(i,j) is proportional to the product of the a-th power of τ(i,j) and the β-th power of η(i,j). Parameters a and β are preset hyperparameters used to balance exploration and exploitation.

[0036] For example, set a equal to 1 and b equal to 2, which means the guiding role of heuristic information is placed in a more important position. Assuming that the vehicle is at node i, it can choose to go to node j or node k. The pheromone concentration of path i to j is 0.8, and the heuristic information value is 0.9; the pheromone concentration of path i to k is 0.7, and the heuristic information value is 0.6. Then, the tendency score of choosing node j is 0.8 raised to the power of 0.9 raised to the power of 2, i.e. 0.648. The tendency score of choosing node k is 0.7 raised to the power of 0.6 raised to the power of 2, i.e. 0.252. According to the relative size of the two scores, the next node is randomly determined by using roulette selection method. Obviously, the probability of choosing node j is much greater than that of choosing node k.

[0037] In an optional embodiment, the comprehensive fulfillment cost is composed of at least two of the driving distance cost, the time-related cost and the penalty cost, wherein the time-related cost includes order waiting time and vehicle service time, and the penalty cost includes fees generated due to violation of customer time window or vehicle capacity limit.

[0038] Specifically, the pros and cons of a candidate decision scheme are quantified by a multi-dimensional cost function. This function is a weighted sum of various costs, and the formula can be expressed as total cost equal to w1 multiplied by driving distance cost, plus w2 multiplied by time-related cost, plus w3 multiplied by penalty cost. For example, the driving distance cost can be set at 2.5 yuan per kilometer, and if the total driving distance of the scheme is 300 kilometers, this cost is 750 yuan. The time-related cost calculates the waiting time of the vehicle due to early arrival and the service loading and unloading time at the customer's point. Assuming that the cost per hour of people and vehicles is 60 yuan, and if the total waiting and service time is 5 hours, this cost is 300 yuan.

[0039] The penalty cost is used to quantify the deficiency of service quality. For example, a customer requires delivery between 2 pm and 4 pm, but the vehicle arrives at 4:15 pm, resulting in a penalty related to the delay time, such as a fine of 10 yuan per minute of delay, which is 150 yuan. Similarly, if a truck with a rated load of 2 tons actually carries 2.2 tons of goods, which is 0.2 tons of overload, it may be calculated at a standard of 500 yuan per 0.1 tons of overload, resulting in a fine of 1000 yuan. Finally, by setting the weights w1, w2, w3, such as 1.0, 0.8, 1.5, the various costs are aggregated to obtain a comprehensive fulfillment cost as the core indicator for evaluating candidate decision schemes.

[0040] Step S4, updating the pheromone tensor according to the comprehensive fulfillment cost of the candidate decision scheme and integrating the real-time state data of the vehicle and the multi-dimensional dynamic state characteristics.

[0041] Specifically, a global evaporation operation is performed on all values in the three-dimensional pheromone tensor, that is, multiplying by a volatility coefficient less than 1; then, one or several candidate decision schemes with the lowest comprehensive implementation cost in the current iteration are selected, and the pheromone concentration at the position corresponding to the time period and vehicle load rate interval when the path segment in the candidate decision scheme is accessed in the three-dimensional pheromone tensor of the path segment is increased; the increased concentration value is inversely proportional to the comprehensive implementation cost of the scheme, and the increased concentration value is dynamically weighted and adjusted by using the real-time traffic congestion index and actual energy consumption data of the path segment in the real world, so that the accumulation of pheromones can better reflect the optimal choice in the real world.

[0042] In an optional embodiment, the updating of the pheromone tensor includes: performing an evaporation operation on the pheromones, the updating rule of which is related to a preset evaporation rate; and enhancing the pheromones of the path segments contained in the candidate decision scheme according to the comprehensive implementation cost of the candidate decision scheme, the enhancement amount being proportional to a preset pheromone intensity coefficient and inversely proportional to the comprehensive implementation cost.

[0043] Specifically, after each round of iteration, the entire pheromone tensor is globally updated, and the updating includes two steps: evaporation and enhancement. First, the evaporation, the pheromones on all path segments are attenuated according to a fixed evaporation rate ρ, for example, ρ is set to 0.1. Then the new pheromone concentration of each path segment (i, j) will become 0.9 times the original concentration. This operation simulates the evaporation of pheromones in nature, helps to avoid the algorithm from falling into local optimum too early, can forget the poor path selection, and thus leaves room for exploring new possibilities.

[0044] Second, enhancement, the paths passed by the optimal or multiple optimal schemes found in the current iteration are rewarded according to the comprehensive implementation cost of the schemes. The enhancement rule is that the pheromone enhancement amount is inversely proportional to the comprehensive implementation cost of the scheme. For example, a pheromone intensity coefficient Q is preset to 100. If the total cost of the optimal scheme A found in the current iteration is 2000 yuan, the pheromone enhancement amount obtained by all path segments in the scheme A is 100 divided by 2000, that is, 0.05. If the cost of another optimal scheme B is 2500 yuan, the pheromone enhancement amount obtained by the path segments in the scheme B is 100 divided by 2500, that is, 0.04. In this way, the high-quality path with lower cost will receive stronger positive feedback, so that the probability of being selected in the subsequent iteration is higher.

[0045] Step S5, repeat the path construction and pheromone updating process to generate a plurality of candidate decision schemes, vectorize the candidate decision schemes, project the vectorized candidate decision schemes to a decision latent space through a variational autoencoder model, perform optimization in the decision latent space according to a global cost function, and decode the optimized latent vector to generate a final decision analysis scheme.

[0046] Specifically, each of the multiple high-quality candidate decision schemes generated by the ant colony optimization algorithm is encoded into a standardized sequence vector; using these vectors as a training set, a variational autoencoder model composed of an encoder and a decoder is trained, the variational autoencoder model learns to map discrete sequence vectors into a low-dimensional, continuous and smooth decision latent space; then, in the decision latent space, a global cost of the decoded scheme is used as the objective function, and a continuous optimization algorithm such as gradient descent or Bayesian optimization is used to search for the optimal latent vector; finally, the optimal latent vector is input into the decoder of the variational autoencoder, and the decoder reconstructs it into a new and more optimal scheme vector, which is then parsed into specific vehicle scheduling instructions to form the final decision analysis scheme.

[0047] In an optional embodiment, the candidate decision scheme is projected into the decision latent space by the variational autoencoder model after being vectorized, including: processing each vehicle path in the candidate decision scheme into a unified node sequence; converting the node sequence into a fixed-dimensional numerical vector as the input of the variational autoencoder model, and outputting the parameters defining the probability distribution in the decision latent space by the encoder thereof.

[0048] Specifically, in order to enable the deep learning model to process discrete solutions in combinatorial optimization problems, a complete distribution scheme is vectorized. A decision scheme may contain multiple vehicle paths, for example, the path of vehicle 1 is 0-2-5-0, and the path of vehicle 2 is 0-1-4-3-0, where 0 represents the distribution center. The two paths are spliced into a single node sequence, such as 0-2-5-0-1-4-3-0. In order to adapt to the needs of problems of different scales, the node sequence is padded or truncated to a preset fixed length, such as 100, and the insufficient part is padded with a special identifier such as -1.

[0049] The fixed-length node sequence is then sent to the encoder part of the variational autoencoder. Inside the encoder, each node ID is first converted to a high-dimensional floating-point number vector by an embedding layer, and then the entire node sequence is processed by structures such as recurrent neural networks to capture the internal structure and order information of the path. The final output of the encoder is not a single vector, but two parameters that describe a Gaussian distribution in the decision latent space: mean and variance. For example, for a 32-dimensional latent space, the encoder will output a 32-dimensional mean vector and a 32-dimensional variance vector, which together define a probability cloud representing a flexible representation of the input scheme in a low-dimensional continuous space.

[0050] In an optional embodiment, the optimization is performed in the decision latent space, including: generating an initial latent vector by sampling from a probability distribution in the decision latent space; and iteratively updating the latent vector by a gradient optimization algorithm according to a preset number of iterations and a learning rate, to minimize the global cost function.

[0051] Specifically, the optimization process starts from the decision latent space. An initial latent vector z is sampled from a defined Gaussian distribution by a reparameterization trick, using the mean and variance of the previous encoder output. This vector z is a specific instance of the original complex route scheme in a low-dimensional space, and can be regarded as the starting point of optimization. Then, the latent vector z is fed into the decoder of the variational autoencoder, which maps it back to a specific, executable route scheme. The comprehensive fulfillment cost of the scheme can be calculated, and the gradient of the cost with respect to the latent vector z can also be calculated due to the continuous differentiability of the entire model.

[0052] The latent vector z is iteratively updated by a gradient optimization algorithm such as Adam or SGD. For example, the learning rate is set to 0.001 and the number of iterations is set to 200. In each iteration, the cost gradient corresponding to the current latent vector z is calculated, and z is updated in the opposite direction of the gradient, i.e.

[0053] ;

[0054] In the formula, z_new is the updated latent vector, z_old is the latent vector before updating, learning_rate is the learning rate, and gradient is the gradient.

[0055] This process is equivalent to "going down the mountain" in a smooth latent space to find the point that minimizes the cost. Each time z is updated, a new route scheme is decoded and its cost is evaluated. After 200 iterations, the route scheme corresponding to the final latent vector is the improved solution after continuous space optimization.

[0056] In the description of the present specification, the meaning of "a plurality of" is at least two, such as two, three or more, unless otherwise explicitly specified.

[0057] Although the present specification has shown and described several embodiments of the present application, it will be apparent to those skilled in the art that many modifications, changes and substitutions can be made without departing from the spirit and scope of the present application.

Claims

1. A decision analysis method for smart logistics, characterized in that: The following steps are involved: Obtain logistics network topology data, order information including delivery priorities and time windows, real-time vehicle status data, and multi-dimensional dynamic status characteristics including traffic conditions and weather; Based on the logistics network topology data and order information, a graph attention network is used to model the connection relationship between nodes in the logistics network to generate a heuristic information matrix; During the iterative path construction process, the heuristic information matrix is ​​combined with a pheromone tensor containing three dimensions: path segment, time period, and vehicle load factor. The next path segment is determined for the node to be decided in the path through a probabilistic selection model, thereby constructing a complete candidate decision plan and calculating its comprehensive implementation cost according to a preset cost model. updating the pheromone tensor based on the comprehensive implementation cost of the candidate decision solutions and integrating the real-time vehicle status data with multi-dimensional dynamic status features; The path construction and pheromone update processes are repeated to generate multiple candidate decision solutions. The candidate decision solutions are vectorized and projected into the decision latent space through a variational autoencoder model. Optimization is performed in the decision latent space based on a global cost function, and the optimized latent vector is decoded to generate the final decision analysis solution.

2. The decision analysis method for smart logistics according to claim 1, characterized in that: The logistics network topology data includes the geographic coordinates and connection relationships between warehouses and customer points; the vehicle real-time status data includes vehicle location, speed, and remaining power or fuel; and the multi-dimensional dynamic status characteristics include regional traffic congestion index and weather forecast for the next few hours.

3. The decision analysis method for smart logistics according to claim 1, characterized in that: The graph attention network is used to model the connection relationship between nodes in the logistics network to generate a heuristic information matrix, including: All warehouses and customer points are regarded as nodes of the graph, and feasible paths are regarded as edges to construct a logistics network graph structure; the type, cargo demand, priority and service time window information of each logistics node are encoded as an initial node feature vector; the logistics network graph structure and node feature vector are input into a pre-trained graph attention network model, and the deep dependencies between nodes are learned through the multi-head attention mechanism. An embedding representation is generated for each pair of nodes in the graph attention network, and the heuristic information value is calculated based on the similarity of the embedding representations of any two nodes to form the heuristic information matrix.

4. The decision analysis method for smart logistics according to claim 1, characterized in that: The graph attention network is used to model the connection relationship between nodes in the logistics network to generate a heuristic information matrix, including: The geographical location, cargo demand, and service time window of each logistics node are encoded into an initial node feature vector. The initial node feature vector is input into a multi-layer graph attention network model, and the connection weight between any two nodes is calculated and output through the attention mechanism to form the heuristic information matrix.

5. The decision analysis method for smart logistics according to claim 1, characterized in that: In the probabilistic selection model, the probability of a vehicle selecting the next unvisited node at the current node is proportional to the weighted product of the pheromone concentration and heuristic information corresponding to the path segment; the weights of the pheromone concentration and heuristic information are adjusted by preset parameters α and β.

6. The decision analysis method for smart logistics according to claim 1, characterized in that: The comprehensive fulfillment cost is composed of a weighted combination of at least two of the travel distance cost, time-related cost, and penalty cost, wherein the time-related cost includes order waiting time and vehicle service time, and the penalty cost includes the fees incurred due to violation of customer time windows or vehicle capacity restrictions.

7. The decision analysis method for smart logistics according to claim 2, characterized in that: Performing an update on the pheromone tensor, including: A global volatilization operation is performed on all values ​​in the three-dimensional pheromone tensor; one or several candidate decision plans with the lowest comprehensive execution cost in the current iteration are selected, and the pheromone concentration is increased at the three-dimensional pheromone tensor positions corresponding to the path segments, time periods and vehicle load rate intervals contained in these candidate decision plans; the increased pheromone concentration value is inversely proportional to the comprehensive execution cost of the plan, and the increased pheromone concentration value is dynamically weighted and adjusted using the traffic congestion index and vehicle energy consumption data corresponding to the path segments.

8. The decision analysis method for smart logistics according to claim 1, characterized in that: Performing an update on the pheromone tensor, including: A volatilization operation is performed on the pheromone, and its update rule is related to the preset volatilization rate; and according to the comprehensive implementation cost of the candidate decision plan, the pheromone of the path segment contained in the candidate decision plan is enhanced, and the enhancement amount is proportional to the preset pheromone intensity coefficient and inversely proportional to the comprehensive implementation cost.

9. The decision analysis method for smart logistics according to claim 1, characterized in that: The vectorizing of the candidate decision solutions and then projecting them into the decision latent space through a variational autoencoder model includes: Each vehicle path in the candidate decision solution is processed into a unified node sequence; the node sequence is converted into a numerical vector of fixed dimension as the input of the variational autoencoder model, and the encoder output defines the parameters of the probability distribution in the decision latent space.

10. The decision analysis method for smart logistics according to claim 9, characterized in that: The performing optimization in the decision latent space according to the global cost function includes: An initial latent vector is generated by sampling from a probability distribution in the decision latent space; and a gradient optimization algorithm is used to iteratively update the initial latent vector according to a preset number of iterations and a learning rate to minimize the global cost function.

Images