An intelligent logistics conveying system scheduling method based on an automation algorithm

Abstract

The application discloses an intelligent logistics conveying system scheduling method based on an automatic algorithm, which comprises the following steps: step 1, collecting operation data in a logistics conveying system; step 2, preprocessing the operation data to form a standardized logistics data set; step 3, inputting the standardized logistics data set into an improved GraphGPS network, performing edge-enhanced local propagation and sparse global modeling operations, and generating a node embedding feature tensor; step 4, based on the node embedding feature tensor, constructing a multi-objective scheduling optimization function set; step 5, solving the multi-objective scheduling optimization function set by using an improved seal optimization algorithm, and outputting an optimal scheduling scheme; step 6, issuing the optimal scheduling scheme to an execution unit of the logistics conveying system; and step 7, collecting real-time operation feedback data of the execution unit, and updating the improved GraphGPS network and the optimal scheduling scheme. The application improves the response efficiency and execution stability of the logistics conveying system.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial scheduling optimization technology, and in particular to a scheduling method for an intelligent logistics transportation system based on automated algorithms. Background Technology

[0002] With the continuous improvement of intelligent manufacturing and logistics automation, the scheduling problem of logistics transportation systems in complex production workshops, warehousing centers, and distribution networks has gradually become a crucial aspect of industrial intelligent decision-making. Existing logistics scheduling methods largely rely on rule engines or heuristic algorithms for path planning and resource allocation, making it difficult to effectively adapt to the real-time decision-making needs driven by dynamic, multi-source data. In scenarios with intensive multi-task operations, complex resource constraints, and high equipment heterogeneity, the following problems commonly exist:

[0003] In logistics transportation systems, there are complex nonlinear interactions between warehouses, automated guided vehicles (AGVs), conveying equipment, and order tasks. Traditional scheduling methods lack the ability to model global semantic relationships, making it difficult to accurately characterize the state evolution process. The inconsistent temporal order and varied spatial structure of the operational data result in insufficient generalization ability of scheduling models to time changes and spatial topology. Current scheduling optimization methods typically optimize a single metric, such as the shortest path or the shortest time, failing to simultaneously consider the multi-objective collaborative requirements of warehouse load balancing, path transportation efficiency, equipment utilization, and task response fairness. Existing genetic algorithms and particle swarm optimization algorithms are prone to getting trapped in local optima in high-dimensional complex search spaces and lack dynamic strategy adjustment capabilities.

[0004] Therefore, how to provide a scheduling method for an intelligent logistics transportation system based on automated algorithms is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a scheduling method for an intelligent logistics transportation system based on an automated algorithm. This invention fully integrates graph neural network modeling and multi-objective intelligent optimization technology, describes in detail the embedding modeling process of the improved GraphGPS network in complex logistics scenarios, and solves the multi-objective scheduling strategy based on the improved walrus optimization algorithm. It has the advantages of strong global perception, high scheduling efficiency, and optimized resource utilization.

[0006] A scheduling method for an intelligent logistics transportation system based on an automated algorithm, according to an embodiment of the present invention, includes the following steps:

[0007] Step 1: Collect operational data from the logistics and transportation system;

[0008] Step 2: Preprocess the operational data to form a standardized logistics dataset;

[0009] Step 3: Input the standardized logistics dataset into the improved GraphGPS network, which generates node embedding feature tensors by performing edge-enhanced local propagation operations and path-bias-based sparse global modeling operations.

[0010] Step 4: Construct a multi-objective scheduling optimization function set based on the node embedding feature tensor;

[0011] Step 5: Use the improved walrus optimization algorithm to solve the multi-objective scheduling optimization function set, calculate the comprehensive fitness value based on the weighted Tchebycheff aggregation function, and output the optimal scheduling scheme;

[0012] Step Six: Send the optimal scheduling plan to the execution unit of the logistics transportation system;

[0013] Step 7: Collect real-time operational feedback data from the execution unit, combine the operational feedback data with the standardized logistics dataset, and update the improved GraphGPS network and the optimal scheduling scheme.

[0014] Optionally, the operational data includes warehouse node data, conveying path data, automated guided vehicle (AGV) operating status data, conveying equipment capacity data, and order task data.

[0015] Optionally, step two specifically includes:

[0016] The running data is cleaned by: filling missing values ​​with linear interpolation, removing outliers with Z-Score anomaly detection, unifying categorical data types with one-hot encoding, and converting the timestamp field into a time feature vector.

[0017] The cleaned operational data is aligned to a time series according to a preset time granularity.

[0018] Based on the time-aligned operational data, system units with unique spatial identifiers or functional attributes are used as graph nodes. These graph nodes include warehouse nodes, automated guided vehicle nodes, conveying station nodes, loading and unloading operation nodes, and order task nodes.

[0019] The attribute information of each graph node is extracted according to the time step and concatenated into a node feature vector. The attribute information includes node number, region code, inventory level, task queue length, running status, power status, task priority and time feature vector.

[0020] Based on the transportation path data, identify the transportation connection relationships between nodes, construct path edges, and extract path attribute data to construct edge feature vectors;

[0021] The path attribute data includes path length, historical transport time, path capacity, path availability, path congestion coefficient, path start node identifier, and path destination node identifier.

[0022] An adjacency matrix is ​​constructed based on the connection relationship between the starting node and the target node of the path. The adjacency matrix is ​​a two-dimensional sparse matrix that represents the connectivity relationship between the nodes in the graph. If there is a connection between node i and node j, the corresponding element of the adjacency matrix is ​​assigned a value of 1; otherwise, it is assigned a value of 0.

[0023] The node feature vectors and edge feature vectors are subjected to min-max normalization according to their feature dimensions, and the adjacency matrix is ​​subjected to symmetric normalization.

[0024] The graph structure corresponding to each time step is organized into a standardized logistics dataset according to the time step. The graph structure includes graph nodes and normalized node feature vectors, path edges and normalized edge feature vectors and normalized adjacency matrices.

[0025] Optionally, step three specifically includes:

[0026] Input the graph structure of the current time step in the standardized logistics dataset into the improved GraphGPS network;

[0027] For the graph structure at the current time step, all neighboring nodes j that are connected to graph node i by an edge constitute the neighboring node set;

[0028] The node feature vector of neighbor node j and the edge feature vector between neighbor node j and graph node i are concatenated to form a joint feature vector. The joint feature vector is then input into the multilayer perceptron to generate the message vector of graph node i receiving from neighbor node j.

[0029] The message vectors received by graph node i from the set of neighbor nodes are weighted and aggregated to form the edge-enhanced local feature vector of graph node i.

[0030] For each path edge in the graph structure, extract the path length, path direction identifier, node topological distance, and relative position identifier, and concatenate them to form a path bias vector.

[0031] The path bias vector is mapped to a position guidance vector of a set dimension through a multilayer perceptron.

[0032] Organize the position guidance vectors of all path edges into a position guidance tensor according to the graph structure;

[0033] Based on the adjacency matrix and the position guidance tensor, a sparse attention matrix is ​​constructed: if the corresponding element of the adjacency matrix is ​​1 or the norm of the position guidance vector between graph node i and graph node j exceeds the set sparse connection threshold, then the corresponding element of the sparse attention matrix is ​​1; otherwise, the corresponding element of the sparse attention matrix is ​​0.

[0034] For each graph node i, form a sparse adjacency set of all graph nodes j whose elements in the sparse attention matrix have a value of 1;

[0035] Based on the sparse adjacency set, the node feature vector of each graph node j is used to generate a value vector through a trainable linear mapping matrix;

[0036] Based on the position guidance vector, calculate the attention weight between graph node i and each graph node j in the sparse adjacency set;

[0037] The value vectors of all graph nodes j in the sparse adjacency set are weighted and aggregated using attention weights to generate the global feature vector of graph node i.

[0038] The edge-enhanced local feature vector of graph node i is summed element-wise with the global feature vector, and the node embedding feature vector of the target dimension is output through a fully connected layer.

[0039] The node embedding features of all graph nodes are stacked and concatenated according to the node dimension to form a node embedding feature tensor.

[0040] Optionally, step four includes:

[0041] Based on the node embedding feature tensor, the node embedding feature vectors of warehouse nodes, automated guided vehicle nodes and order task nodes are extracted according to the node type, and stacked according to the numbering order of each type of node to form the warehouse state matrix, vehicle state matrix and task state matrix respectively.

[0042] Based on the warehouse state matrix, the inventory load value of each warehouse node is extracted, and the absolute difference between the inventory load value of each warehouse node and the average inventory load value of all warehouse nodes is summed to construct the warehouse load balancing objective function.

[0043] Based on the vehicle state matrix and the task state matrix, the start time of each order task, the estimated transportation time of the corresponding transportation route, and the scheduling delay time estimated based on route congestion and vehicle state are extracted.

[0044] The starting time, transportation time, and scheduling delay time of each order task are summed and averaged to construct the objective function for path transportation efficiency;

[0045] Based on the capacity data of the conveying equipment, the real-time utilization rate of each conveying equipment is extracted, and the corresponding semantic state feature vector is extracted from the node embedded feature tensor.

[0046] The real-time utilization rate is increased in dimensionality through a linear layer and concatenated with the semantic state feature vector. It is then mapped to a device score value through a multilayer perceptron. The objective function of device utilization rate is constructed by averaging all device scores.

[0047] Based on the task state matrix, task nodes with priority labels higher than a set threshold are selected, and the mean embedding feature vector of the task node is obtained. The Euclidean distance between the embedding feature vector of the task node and the mean embedding feature vector is weighted and summed to construct the task response fairness objective function.

[0048] The objective functions of warehouse load balancing, route transportation efficiency, equipment utilization, and task response fairness are combined into a multi-objective scheduling optimization function set.

[0049] Optionally, step five specifically includes:

[0050] Initialize several scheduling individuals to form a scheduling population, and set a maximum number of iterations. The position vector of the scheduling individual is formed by linearly concatenating the warehouse state matrix, vehicle state matrix, task state matrix and semantic state feature vector in the feature dimension.

[0051] For each scheduling entity, calculate the objective function values ​​for warehouse load balancing, path transportation efficiency, equipment utilization, and task response fairness.

[0052] Define a target weight vector and a desired reference point, where the desired reference point represents the theoretical reference value of the corresponding objective function;

[0053] Based on the target weight vector, the absolute difference between each objective function value and the expected reference point is used to calculate the comprehensive fitness value through the weighted Tchebycheff aggregation function;

[0054] In each iteration, the scheduling individual with the smallest overall fitness value in the current scheduling population is selected as the optimal scheduling individual;

[0055] Based on the maximum number of iterations, the position update of the scheduling population is divided into a global exploration phase, a local development phase, and a fine-tuning phase, as follows:

[0056] During the global exploration phase, a global step size coefficient is introduced. The difference between the position vector of the current scheduled individual and the position vector of the optimal scheduled individual is multiplied by the global step size coefficient and a uniformly random number in the interval [0,1], and then added to the current position vector to obtain the updated position vector.

[0057] During the local development phase, a local perturbation coefficient is introduced. The difference between the position vector of the current scheduled individual and the position vector of the optimal scheduled individual is multiplied by the local perturbation coefficient and the standard normal distribution random variable, and then added to the current position vector to obtain the updated position vector.

[0058] During the fine-tuning phase, the average position vector of all scheduled individuals in the current scheduling population is taken as the population mean. The difference between the position vector of the current scheduled individual and the position vector of the optimal scheduled individual, and the difference between the position vector of the current scheduled individual and the population mean are obtained respectively. These differences are multiplied by the fine-tuning control coefficient and added to the current position vector to obtain the updated position vector.

[0059] After completing the maximum number of iterations, the scheduling individual with the smallest overall fitness value is selected from the last round of scheduling population as the optimal scheduling scheme.

[0060] Optionally, step six specifically includes:

[0061] The optimal scheduling scheme is structured and analyzed to extract the mapping relationship between task nodes and corresponding warehouse nodes, transportation routes and conveying equipment, generate a structured control instruction set and transmit it synchronously to the execution unit of the logistics conveying system.

[0062] The execution unit includes a conveyor belt controller, an automated guided vehicle scheduling module, and a warehouse loading and unloading module;

[0063] The structured control instruction set specifically includes:

[0064] Send path allocation instructions to the conveyor belt controller to control the conveyor belt system to start and switch paths according to the set path priority;

[0065] The automated guided vehicle (AGV) scheduling module issues a transportation task instruction to execute route scheduling and resource allocation. The transportation task instruction includes a starting node identifier, a target node identifier, and task payload parameters.

[0066] The operation queue instructions issued to the warehousing and loading / unloading module are used to execute sorting and inbound / outbound operations. The operation queue instructions include the order task number, the target warehouse number, material attribute parameters, and the scheduling time window.

[0067] Optionally, the operational feedback data specifically includes:

[0068] The path start / stop status, operating speed, switching delay time, and conveyor belt load rate of the conveyor belt system;

[0069] The actual driving path of the automated guided vehicle (AGV) from the starting node to the target node, the task completion time, the energy consumption level, the obstacle avoidance frequency, and the error trajectory.

[0070] The task response time, operation time, material handling accuracy, and real-time inventory changes of the warehousing and loading / unloading module.

[0071] The beneficial effects of this invention are:

[0072] First, this invention comprehensively characterizes the complex relationships between various entities in the logistics transportation system, such as warehouse nodes, automated guided vehicle nodes, and order task nodes, by constructing a standardized logistics dataset.

[0073] Secondly, this invention employs an improved GraphGPS network, integrating edge-enhanced local propagation and sparse global modeling strategies. This significantly improves the semantic integrity and global expressive power of the node embedding feature tensor, effectively solving the problem of insufficient information propagation caused by the sparse adjacency structure in traditional graph neural networks in logistics scheduling scenarios. Based on the generated node embedding feature tensor, warehouse state matrices, vehicle state matrices, and task state matrices are further constructed. Combined with real-time equipment utilization and semantic state feature vectors, a multi-objective scheduling optimization function set is built, encompassing warehouse load balancing, route transportation efficiency, equipment utilization, and task response fairness, comprehensively characterizing the multi-dimensional optimization needs in scheduling tasks.

[0074] Furthermore, this invention introduces an improved walrus optimization algorithm, which linearly concatenates the position vectors of scheduling individuals along the feature dimension using the warehouse state matrix, vehicle state matrix, task state matrix, and semantic state feature vector. The optimization process is divided into a global exploration phase, a local development phase, and a fine-tuning phase. For each phase, global step-size perturbation, local normal perturbation, and mean-guided fine-tuning strategies are employed, improving the convergence speed and global optimum capability of the optimization process. The optimal scheduling scheme is then mapped into a control instruction set through a structured analytical method and distributed to the conveyor belt controller, automated guided vehicle scheduling module, and warehouse loading / unloading module, achieving precise mapping and efficient collaboration between task nodes and system resources.

[0075] In summary, this invention enables automated modeling of complex scheduling tasks in logistics transportation systems, improves the accuracy of operational data representation, the coordination of multi-objective scheduling functions, and the real-time execution of control commands, thereby enhancing the real-time perception and adaptive capabilities during intelligent logistics operations. Attached Figure Description

[0076] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0077] Figure 1 This is a schematic diagram of a scheduling method for an intelligent logistics transportation system based on an automated algorithm proposed in this invention;

[0078] Figure 2 This is a flowchart of the improved GraphGPS network modeling in the intelligent logistics transportation system scheduling method based on automated algorithms proposed in this invention;

[0079] Figure 3 This is a flowchart illustrating the construction of a multi-objective optimization function and the solution of an improved walrus optimization algorithm in a scheduling method for an intelligent logistics transportation system based on automated algorithms proposed in this invention. Detailed Implementation

[0080] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0081] refer to Figure 1-3 A scheduling method for an intelligent logistics delivery system based on an automated algorithm includes the following steps:

[0082] Step 1: Collect operational data from the logistics and transportation system;

[0083] Step 2: Preprocess the operational data to form a standardized logistics dataset;

[0084] Step 3: Input the standardized logistics dataset into the improved GraphGPS network, which generates node embedding feature tensors by performing edge-enhanced local propagation operations and path-bias-based sparse global modeling operations.

[0085] Step 4: Construct a multi-objective scheduling optimization function set based on the node embedding feature tensor;

[0086] Step 5: Use the improved walrus optimization algorithm to solve the multi-objective scheduling optimization function set, calculate the comprehensive fitness value based on the weighted Tchebycheff aggregation function, and output the optimal scheduling scheme;

[0087] Step Six: Send the optimal scheduling plan to the execution unit of the logistics transportation system;

[0088] Step 7: Collect real-time operational feedback data from the execution unit, combine the operational feedback data with the standardized logistics dataset, and update the improved GraphGPS network and the optimal scheduling scheme.

[0089] In this embodiment, the operational data includes warehouse node data, conveying path data, automated guided vehicle (AGV) operating status data, conveying equipment capacity data, and order task data;

[0090] The warehouse node data includes warehouse geographical location, inventory capacity, real-time inventory level, processing capacity, and operation time window; the conveying path data includes path length, transportation time, path capacity, path availability, path congestion coefficient, path start node identifier, and path destination node identifier; the automated guided vehicle (AGV) operating status data includes current location, real-time speed, remaining battery power, load status, and operation priority; the conveying equipment capacity data includes maximum conveyor belt capacity, sorting equipment processing capacity, lifting equipment rated load, and real-time utilization rate; and the order task data includes order arrival time, destination node, cargo volume, cargo weight, and task priority.

[0091] In this embodiment, step two specifically includes:

[0092] The running data is cleaned by: filling missing values ​​with linear interpolation, removing outliers with Z-Score anomaly detection, unifying categorical data types with one-hot encoding, and converting the timestamp field into a time feature vector.

[0093] The cleaned operational data is aligned to a time series according to a preset time granularity.

[0094] Based on the time-aligned operational data, system units with unique spatial identifiers or functional attributes are used as graph nodes. These graph nodes include warehouse nodes, automated guided vehicle nodes, conveying station nodes, loading and unloading operation nodes, and order task nodes.

[0095] The attribute information of each graph node is extracted according to the time step and concatenated into a node feature vector. The attribute information includes node number, region code, inventory level, task queue length, running status, power status, task priority and time feature vector.

[0096] Based on the transportation path data, identify the transportation connection relationships between nodes, construct path edges, and extract path attribute data to construct edge feature vectors;

[0097] The path attribute data includes path length, historical transport time, path capacity, path availability, path congestion coefficient, path start node identifier, and path destination node identifier.

[0098] An adjacency matrix is ​​constructed based on the connection relationship between the starting node and the target node of the path. The adjacency matrix is ​​a two-dimensional sparse matrix that represents the connectivity relationship between the nodes in the graph. If there is a connection between node i and node j, the corresponding element of the adjacency matrix is ​​assigned a value of 1; otherwise, it is assigned a value of 0.

[0099] The node feature vectors and edge feature vectors are subjected to min-max normalization according to their feature dimensions, and the adjacency matrix is ​​subjected to symmetric normalization.

[0100] The graph structure corresponding to each time step is organized into a standardized logistics dataset according to the time step. The graph structure includes graph nodes and normalized node feature vectors, path edges and normalized edge feature vectors and normalized adjacency matrices.

[0101] In this embodiment, step three specifically includes:

[0102] Input the graph structure of the current time step in the standardized logistics dataset into the improved GraphGPS network;

[0103] For the graph structure at the current time step, all neighboring nodes j that are connected to graph node i by an edge constitute the neighboring node set;

[0104] The node feature vector of neighbor node j and the edge feature vector between neighbor node j and graph node i are concatenated to form a joint feature vector. The joint feature vector is then input into the multilayer perceptron to generate the message vector of graph node i receiving from neighbor node j.

[0105] The message vectors received by graph node i from the set of neighbor nodes are weighted and aggregated to form the edge-enhanced local feature vector of graph node i.

[0106] For each path edge in the graph structure, extract the path length, path direction identifier, node topological distance, and relative position identifier, and concatenate them to form a path bias vector.

[0107] The path bias vector is mapped to a position guidance vector of a set dimension through a multilayer perceptron.

[0108] Organize the position guidance vectors of all path edges into a position guidance tensor according to the graph structure;

[0109] Based on the adjacency matrix and the position guidance tensor, a sparse attention matrix is ​​constructed: if the corresponding element of the adjacency matrix is ​​1 or the norm of the position guidance vector between graph node i and graph node j exceeds the set sparse connection threshold, then the corresponding element of the sparse attention matrix is ​​1; otherwise, the corresponding element of the sparse attention matrix is ​​0.

[0110] For each graph node i, form a sparse adjacency set of all graph nodes j whose elements in the sparse attention matrix have a value of 1;

[0111] Based on the sparse adjacency set, the node feature vector of each graph node j is used to generate a value vector through a trainable linear mapping matrix;

[0112] Calculate the attention weights between graph node i and each graph node j in the sparse adjacency set:

[0113] ;

[0114] in, Let i represent the node feature vector of node i in the graph. This represents the node feature vector of node j in the graph. and Let be a trainable linear mapping matrix. This represents the positional guidance vector between graph node i and graph node j. Let represent the sparse adjacency set of graph node i, and Softmax be the normalization function;

[0115] The value vectors of all graph nodes j in the sparse adjacency set are weighted and aggregated using attention weights to generate the global feature vector of graph node i.

[0116] The edge-enhanced local feature vector of graph node i is summed element-wise with the global feature vector, and the node embedding feature vector of the target dimension is output through a fully connected layer.

[0117] The node embedding features of all graph nodes are stacked and concatenated according to the node dimension to form a node embedding feature tensor.

[0118] In this invention, the improved GraphGPS network effectively enhances the granularity and contextual representation of node neighborhood information interaction by introducing edge feature concatenation and multilayer perceptron processing within the local neighborhood. At the same time, it constructs a path bias vector using path length, direction, topological distance, and relative position identifiers and maps it to a position guidance tensor. Combined with a sparse attention matrix, it captures potential dependencies between long-distance nodes and optimizes sparse connections, significantly improving the global perception and representation capabilities of the node embedded feature tensor, and improving the scheduling optimization accuracy, convergence speed, and overall scheduling scheme stability.

[0119] In this embodiment, step four includes:

[0120] Based on the node embedding feature tensor, the node embedding feature vectors of warehouse nodes, automated guided vehicle nodes and order task nodes are extracted according to the node type, and stacked according to the numbering order of each type of node to form the warehouse state matrix, vehicle state matrix and task state matrix respectively.

[0121] Based on the warehouse state matrix, the inventory load value of each warehouse node is extracted, and the absolute difference between the inventory load value of each warehouse node and the average inventory load value of all warehouse nodes is summed to construct the warehouse load balancing objective function.

[0122] Based on the vehicle state matrix and the task state matrix, the start time of each order task, the estimated transportation time of the corresponding transportation route, and the scheduling delay time estimated based on route congestion and vehicle state are extracted.

[0123] The starting time, transportation time, and scheduling delay time of each order task are summed and averaged to construct the objective function for path transportation efficiency;

[0124] Based on the capacity data of the conveying equipment, the real-time utilization rate of each conveying equipment is extracted, and the corresponding semantic state feature vector is extracted from the node embedded feature tensor.

[0125] The real-time utilization rate is increased in dimensionality through a linear layer and concatenated with the semantic state feature vector. It is then mapped to a device score value through a multilayer perceptron. The objective function of device utilization rate is constructed by averaging all device scores.

[0126] In this invention, the equipment utilization objective function integrates the real-time utilization of the conveying equipment with the semantic state feature vector in the node embedded feature tensor. At the numerical level, it reflects the actual load state of equipment resources, and at the semantic level, it describes the historical behavior patterns and state evolution characteristics of equipment operation. The real-time utilization is then combined with the semantic state feature vector through linear dimensionality increase. The nonlinear relationship between equipment utilization and state semantics is effectively captured by a multilayer perceptron, thereby constructing a more reasonable and discriminative equipment scoring system. This significantly improves the measurement accuracy of equipment resource utilization efficiency and the resource matching rationality of scheduling results in scheduling optimization.

[0127] Based on the task state matrix, task nodes with priority labels higher than a set threshold are selected, and the mean embedding feature vector of the task node is obtained. The Euclidean distance between the embedding feature vector of the task node and the mean embedding feature vector is weighted and summed to construct the task response fairness objective function.

[0128] The objective functions of warehouse load balancing, route transportation efficiency, equipment utilization, and task response fairness are combined into a multi-objective scheduling optimization function set. The optimization objective of the multi-objective scheduling optimization function set is to simultaneously minimize the difference in warehouse inventory pressure and average transportation time, maximize the utilization of transportation equipment, and improve the response fairness of high-priority tasks.

[0129] In this embodiment, step five specifically includes:

[0130] Initialize several scheduling individuals to form a scheduling population, and set a maximum number of iterations. The position vector of the scheduling individual is formed by linearly concatenating the warehouse state matrix, vehicle state matrix, task state matrix and semantic state feature vector in the feature dimension.

[0131] For each scheduling entity, calculate the objective function values ​​for warehouse load balancing, path transportation efficiency, equipment utilization, and task response fairness.

[0132] Define a target weight vector and a desired reference point, where the desired reference point represents the theoretical reference value of the corresponding objective function;

[0133] Based on the target weight vector, the absolute difference between each objective function value and the expected reference point is used to calculate the overall fitness value through the weighted Tchebycheff aggregation function:

[0134] ;

[0135] in, Represents the first element in the target weight vector. Each weight value Indicates the first The expected reference value of each objective function Represents the first function in the multi-objective scheduling optimization function set. The objective function value, This indicates taking the maximum value among all results;

[0136] In each iteration, the scheduling individual with the smallest overall fitness value in the current scheduling population is selected as the optimal scheduling individual;

[0137] Based on the maximum number of iterations, the position update of the scheduling population is divided into a global exploration phase, a local development phase, and a fine-tuning phase, as follows:

[0138] During the global exploration phase, a global step size coefficient is introduced. The difference between the position vector of the current scheduled individual and the position vector of the optimal scheduled individual is multiplied by the global step size coefficient and a uniformly random number in the interval [0,1], and then added to the current position vector to obtain the updated position vector.

[0139] ;

[0140] in, This represents the position vector of the currently scheduled individual. This represents the position vector of the optimally scheduled individual. Represents the global step size coefficient. Represents a uniformly random number in the interval [0,1]. This represents the position vector of the scheduled individual after the update;

[0141] During the local development phase, a local perturbation coefficient is introduced. The difference between the position vector of the current scheduled individual and the position vector of the optimal scheduled individual is multiplied by the local perturbation coefficient and a standard normally distributed random variable, and then added to the current position vector to obtain the updated position vector.

[0142] ;

[0143] in, This represents the local disturbance coefficient. Represents a standard normally distributed random variable;

[0144] During the fine-tuning phase, the average position vector of all scheduled individuals in the current scheduling population is used as the population mean.

[0145] The differences between the current scheduled individual's position vector and the optimal scheduled individual's position vector, and between the current scheduled individual's position vector and the population mean, are obtained separately. These differences are multiplied by the fine-tuning control coefficient and added to the current position vector to obtain the updated position vector.

[0146] ;

[0147] in, and This indicates the fine-tuning control coefficient. Indicates the population mean;

[0148] After completing the maximum number of iterations, the scheduling individual with the smallest overall fitness value is selected from the last round of scheduling population as the optimal scheduling scheme.

[0149] For example, in a scheduling task at a digital warehouse, the initial scheduling population size is set to 60, and the maximum number of iterations is set to 120. The position vector of each scheduling individual is constructed by linearly concatenating the warehouse state matrix, vehicle state matrix, task state matrix, and semantic state feature vector along the feature dimension. Under the initial average scheme, the performance of each indicator is as follows: warehouse load difference is 0.198, average route transportation time is 13.41 minutes, equipment score is 0.744, and task response standard deviation is 0.079. The objective function weight vector is set to {0.25, 0.30, 0.25, 0.20}, and the expected reference point is {0.10, 10.00, 1.00, 0.00}. The comprehensive fitness value of the scheduling individual is calculated based on the weighted Tchebycheff aggregation function.

[0150] In the global exploration phase, a global step size coefficient of 0.6 is set. This phase rapidly narrows the distance between the scheduling individuals and the globally optimal solution, improving the overall fitness convergence speed. In the local development phase, a local perturbation coefficient of 0.3 is introduced, and a small perturbation is applied to the scheduling individuals using standard normal random variables. This process guides the scheduling individuals to perform a fine search along the current optimal direction, avoiding getting trapped in local optima. In the fine-tuning phase, the average position vector of all individuals is calculated, and the vector differences between the scheduling individuals and the globally optimal individual, as well as the population mean, are multiplied by fine-tuning control coefficients of 0.15 and 0.10, respectively, for fine updates. This phase strengthens the diversity and collaborative adjustment among individuals within the understanding space, improving the stability and generalization ability of the final result. After 120 iterations, the scheduling individuals gradually converge, and the various indicators of the final optimal scheduling scheme are significantly improved: the warehouse load difference is reduced to 0.124, the path transportation time is shortened to 11.73 minutes, the equipment score is improved to 0.884, and the task response fairness is improved to 0.041. Compared with the initial average scheme, the overall fitness value is improved by 36.2%.

[0151] In this embodiment, step six specifically includes:

[0152] The optimal scheduling scheme is structured and analyzed to extract the mapping relationship between task nodes and corresponding warehouse nodes, transportation routes and conveying equipment, generate a structured control instruction set and transmit it synchronously to the execution unit of the logistics conveying system.

[0153] The execution unit includes a conveyor belt controller, an automated guided vehicle scheduling module, and a warehouse loading and unloading module;

[0154] The structured control instruction set specifically includes:

[0155] Send path allocation instructions to the conveyor belt controller to control the conveyor belt system to start and switch paths according to the set path priority;

[0156] The automated guided vehicle (AGV) scheduling module issues a transportation task instruction to execute route scheduling and resource allocation. The transportation task instruction includes a starting node identifier, a target node identifier, and task payload parameters.

[0157] The operation queue instructions issued to the warehousing and loading / unloading module are used to execute sorting and inbound / outbound operations. The operation queue instructions include the order task number, the target warehouse number, material attribute parameters, and the scheduling time window.

[0158] In this embodiment, the operational feedback data specifically includes:

[0159] The path start / stop status, operating speed, switching delay time, and conveyor belt load rate of the conveyor belt system;

[0160] The actual driving path of the automated guided vehicle (AGV) from the starting node to the target node, the task completion time, the energy consumption level, the obstacle avoidance frequency, and the error trajectory.

[0161] The task response time, operation time, material handling accuracy, and real-time inventory changes of the warehousing and loading / unloading module.

[0162] Example 1

[0163] To verify the feasibility of this invention in practice, it was applied to the intelligent logistics conveying system of a smart manufacturing enterprise. This intelligent logistics conveying system is deployed in the enterprise's digital warehousing center and consists of 24 conveyor belt systems, 16 automated guided vehicles (AGVs), and 6 intelligent warehousing areas. It processes over 4,800 orders daily. The internal logistics paths of the intelligent logistics conveying system are complex and intersecting, with high task response requirements and close relationships between equipment operation and scheduling. Traditional scheduling methods struggle to achieve efficient coordination and real-time response, easily leading to problems such as path conflicts, warehouse load imbalances, and scheduling delays, severely restricting overall operational efficiency.

[0164] During implementation, industrial controllers and edge acquisition terminals deployed in various units of the intelligent logistics conveyor system are used to collect real-time data on the conveyor system's path start / stop status, operating speed, switching delay time, and load rate; the actual driving paths, energy consumption levels, obstacle avoidance frequency, and error trajectories of the automated guided vehicles (AGVs) at their starting and target nodes; the task response time of the warehousing and loading / unloading modules and real-time warehouse inventory changes; and the priority, status, and scheduling time of order tasks. Subsequently, preprocessing operations are performed on the multi-source operational data to construct a structured and standardized logistics dataset. Node and edge feature information, including path length, historical transportation time, path capacity, path congestion coefficient, current AGV power consumption, order task priority, and warehouse inventory level, is extracted, and an adjacency matrix is ​​constructed.

[0165] A standardized logistics dataset is input into an improved GraphGPS network. During edge-enhanced local propagation, the features of each node's neighborhood are aggregated and interacted, generating edge-enhanced local feature vectors. A path-bias-driven sparse global modeling module captures potential dependencies between long-distance nodes, outputting global feature vectors, which are then fused to generate a node embedding feature tensor. Based on this node embedding feature tensor, a multi-objective scheduling optimization function set is constructed, covering warehouse load balancing, route transportation efficiency, equipment utilization, and task response fairness. An improved walrus optimization algorithm is introduced to dynamically adjust the individual scheduling position vectors, outputting the optimal scheduling scheme.

[0166] The optimal scheduling scheme is structured and parsed into a set of control instructions, including start / stop instructions for the transport path, AGV transport task parameters, and warehouse operation scheduling instructions, which are then distributed to each execution unit for scheduling response. During scheduling execution, feedback information is collected in real time to evaluate the system's performance and dynamically update model parameters.

[0167] To comprehensively evaluate the superiority of this invention in practical applications, it is compared with rule-based heuristic scheduling methods currently used by enterprises. The experiment ran for 12 working days with a sampling interval of 10 seconds, and the experimental results are shown in Table 1.

[0168] Table 1. Performance Comparison of the Invention and Comparative Solutions in Intelligent Logistics Conveying Systems

[0169] ;

[0170] As shown in Table 1, the method of this invention significantly outperforms the comparative scheme in several key performance indicators. Regarding warehouse load balancing, the standard deviation of warehouse load decreased from 79.5 units to 34.7 units, with a significant reduction in fluctuation, effectively alleviating the problem of some warehouses being overloaded while others are idle. The average order response time and transportation route time decreased to 32.1 seconds and 21.5 seconds respectively, indicating stronger real-time performance and agility in scheduling efficiency and route allocation. The average utilization rate of conveying equipment reached 87.2%, an improvement of over 38% compared to traditional scheduling systems, fully maximizing resource efficiency.

[0171] Regarding task response fairness, the fairness deviation value decreased by more than 60%, indicating that the method of this invention effectively reduced scheduling deviation and improved the rationality of task allocation by calculating the Euclidean distance deviation between the embedding vector of high-priority tasks and the global mean vector. The average energy consumption of the method of this invention decreased by more than 15%, indicating that the method of this invention has significant energy-saving effect. In terms of average scheduling success rate, the method of this invention reached 98.7%, which is significantly better than the comparative scheme.

[0172] This embodiment fully verifies the deployability and scalability of the present invention in actual complex logistics transportation systems, and can balance energy efficiency, scheduling fairness and system stability while maintaining high scheduling accuracy and high response.

[0173] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A scheduling method for an intelligent logistics transportation system based on an automated algorithm, characterized in that, Includes the following steps: Step 1: Collect operational data from the logistics and transportation system; Step 2: Preprocess the operational data to form a standardized logistics dataset; Step 3: Input the standardized logistics dataset into the improved GraphGPS network, which generates node embedding feature tensors by performing edge-enhanced local propagation operations and path-bias-based sparse global modeling operations. Step 4: Construct a multi-objective scheduling optimization function set based on the node embedding feature tensor; Step 5: Solve the multi-objective scheduling optimization function set using the improved walrus optimization algorithm, calculate the comprehensive fitness value based on the weighted Tchebycheff aggregation function, and output the optimal scheduling scheme, specifically including: Initialize several scheduling individuals to form a scheduling population, and set a maximum number of iterations. The position vector of the scheduling individual is formed by linearly concatenating the warehouse state matrix, vehicle state matrix, task state matrix and semantic state feature vector in the feature dimension. For each scheduling entity, calculate the objective function values ​​for warehouse load balancing, path transportation efficiency, equipment utilization, and task response fairness. Define a target weight vector and a desired reference point, where the desired reference point represents the theoretical reference value of the corresponding objective function; Based on the target weight vector, the absolute difference between each objective function value and the expected reference point is used to calculate the comprehensive fitness value through the weighted Tchebycheff aggregation function; In each iteration, the scheduling individual with the smallest overall fitness value in the current scheduling population is selected as the optimal scheduling individual; Based on the maximum number of iterations, the position update of the scheduling population is divided into a global exploration phase, a local development phase, and a fine-tuning phase, as follows: During the global exploration phase, a global step size coefficient is introduced. The difference between the position vector of the current scheduled individual and the position vector of the optimal scheduled individual is multiplied by the global step size coefficient and a uniformly random number in the interval [0,1], and then added to the current position vector to obtain the updated position vector. During the local development phase, a local perturbation coefficient is introduced. The difference between the position vector of the current scheduled individual and the position vector of the optimal scheduled individual is multiplied by the local perturbation coefficient and the standard normal distribution random variable, and then added to the current position vector to obtain the updated position vector. During the fine-tuning phase, the average position vector of all scheduled individuals in the current scheduling population is taken as the population mean. The difference between the position vector of the current scheduled individual and the position vector of the optimal scheduled individual, and the difference between the position vector of the current scheduled individual and the population mean are obtained respectively. These differences are multiplied by the fine-tuning control coefficient and added to the current position vector to obtain the updated position vector. After completing the maximum number of iterations, select the scheduling individual with the smallest comprehensive fitness value from the last round of scheduling population as the optimal scheduling scheme; Step Six: Send the optimal scheduling plan to the execution unit of the logistics transportation system; Step 7: Collect real-time operational feedback data from the execution unit, combine the operational feedback data with the standardized logistics dataset, and update the improved GraphGPS network and the optimal scheduling scheme.

2. The scheduling method for an intelligent logistics transportation system based on an automated algorithm according to claim 1, characterized in that, The operational data includes warehouse node data, conveying path data, automated guided vehicle (AGV) operating status data, conveying equipment capacity data, and order task data.

3. The scheduling method for an intelligent logistics transportation system based on an automated algorithm according to claim 1, characterized in that, Step two specifically includes: The running data is cleaned by: filling missing values ​​with linear interpolation, removing outliers with Z-Score anomaly detection, unifying categorical data types with one-hot encoding, and converting the timestamp field into a time feature vector. The cleaned operational data is aligned to a time series according to a preset time granularity. Based on the time-aligned operational data, system units with unique spatial identifiers or functional attributes are used as graph nodes. These graph nodes include warehouse nodes, automated guided vehicle nodes, conveying station nodes, loading and unloading operation nodes, and order task nodes. The attribute information of each graph node is extracted according to the time step and concatenated into a node feature vector. The attribute information includes node number, region code, inventory level, task queue length, running status, power status, task priority and time feature vector. Based on the transportation path data, identify the transportation connection relationships between nodes, construct path edges, and extract path attribute data to construct edge feature vectors; The path attribute data includes path length, historical transport time, path capacity, path availability, path congestion coefficient, path start node identifier, and path destination node identifier. An adjacency matrix is ​​constructed based on the connection relationship between the starting node and the target node of the path. The adjacency matrix is ​​a two-dimensional sparse matrix that represents the connectivity relationship between the nodes in the graph. If there is a connection between node i and node j, the corresponding element of the adjacency matrix is ​​assigned a value of 1; otherwise, it is assigned a value of 0. The node feature vectors and edge feature vectors are subjected to min-max normalization according to their feature dimensions, and the adjacency matrix is ​​subjected to symmetric normalization. The graph structure corresponding to each time step is organized into a standardized logistics dataset according to the time step. The graph structure includes graph nodes and normalized node feature vectors, path edges and normalized edge feature vectors and normalized adjacency matrices.

4. The scheduling method for an intelligent logistics transportation system based on an automated algorithm according to claim 1, characterized in that, Step three specifically includes: Input the graph structure of the current time step in the standardized logistics dataset into the improved GraphGPS network; For the graph structure at the current time step, all neighboring nodes j that are connected to graph node i by an edge constitute the neighboring node set; The node feature vector of neighbor node j and the edge feature vector between neighbor node j and graph node i are concatenated to form a joint feature vector. The joint feature vector is then input into the multilayer perceptron to generate the message vector of graph node i receiving from neighbor node j. The message vectors received by graph node i from the set of neighbor nodes are weighted and aggregated to form the edge-enhanced local feature vector of graph node i. For each path edge in the graph structure, extract the path length, path direction identifier, node topological distance, and relative position identifier, and concatenate them to form a path bias vector. The path bias vector is mapped to a position guidance vector of a set dimension through a multilayer perceptron. Organize the position guidance vectors of all path edges into a position guidance tensor according to the graph structure; Based on the adjacency matrix and the position guidance tensor, a sparse attention matrix is ​​constructed: if the corresponding element of the adjacency matrix is ​​1 or the norm of the position guidance vector between graph node i and graph node j exceeds the set sparse connection threshold, then the corresponding element of the sparse attention matrix is ​​1; otherwise, the corresponding element of the sparse attention matrix is ​​0. For each graph node i, form a sparse adjacency set of all graph nodes j whose elements in the sparse attention matrix have a value of 1; Based on the sparse adjacency set, the node feature vector of each graph node j is used to generate a value vector through a trainable linear mapping matrix; Based on the position guidance vector, calculate the attention weight between graph node i and each graph node j in the sparse adjacency set; The value vectors of all graph nodes j in the sparse adjacency set are weighted and aggregated using attention weights to generate the global feature vector of graph node i. The edge-enhanced local feature vector of graph node i is summed element-wise with the global feature vector, and the node embedding feature vector of the target dimension is output through a fully connected layer. The node embedding features of all graph nodes are stacked and concatenated according to the node dimension to form a node embedding feature tensor.

5. The scheduling method for an intelligent logistics transportation system based on an automated algorithm according to claim 1, characterized in that, Step four includes: Based on the node embedding feature tensor, the node embedding feature vectors of warehouse nodes, automated guided vehicle nodes and order task nodes are extracted according to the node type, and stacked according to the numbering order of each type of node to form the warehouse state matrix, vehicle state matrix and task state matrix respectively. Based on the warehouse state matrix, the inventory load value of each warehouse node is extracted, and the absolute difference between the inventory load value of each warehouse node and the average inventory load value of all warehouse nodes is summed to construct the warehouse load balancing objective function. Based on the vehicle state matrix and the task state matrix, the start time of each order task, the estimated transportation time of the corresponding transportation route, and the scheduling delay time estimated based on route congestion and vehicle state are extracted. The starting time, transportation time, and scheduling delay time of each order task are summed and averaged to construct the objective function for path transportation efficiency; Based on the capacity data of the conveying equipment, the real-time utilization rate of each conveying equipment is extracted, and the corresponding semantic state feature vector is extracted from the node embedded feature tensor. The real-time utilization rate is increased in dimensionality through a linear layer and concatenated with the semantic state feature vector. It is then mapped to a device score value through a multilayer perceptron. The objective function of device utilization rate is constructed by averaging all device scores. Based on the task state matrix, task nodes with priority labels higher than a set threshold are selected, and the mean embedding feature vector of the task node is obtained. The Euclidean distance between the embedding feature vector of the task node and the mean embedding feature vector is weighted and summed to construct the task response fairness objective function. The objective functions of warehouse load balancing, route transportation efficiency, equipment utilization, and task response fairness are combined into a multi-objective scheduling optimization function set.

6. The scheduling method for an intelligent logistics transportation system based on an automated algorithm according to claim 1, characterized in that, Step six specifically includes: The optimal scheduling scheme is structured and analyzed to extract the mapping relationship between task nodes and corresponding warehouse nodes, transportation routes and conveying equipment, generate a structured control instruction set and transmit it synchronously to the execution unit of the logistics conveying system. The execution unit includes a conveyor belt controller, an automated guided vehicle scheduling module, and a warehouse loading and unloading module; The structured control instruction set specifically includes: Send path allocation instructions to the conveyor belt controller to control the conveyor belt system to start and switch paths according to the set path priority; The automated guided vehicle (AGV) scheduling module issues a transportation task instruction to execute route scheduling and resource allocation. The transportation task instruction includes a starting node identifier, a target node identifier, and task payload parameters. The operation queue instructions issued to the warehousing and loading / unloading module are used to execute sorting and inbound / outbound operations. The operation queue instructions include the order task number, the target warehouse number, material attribute parameters, and the scheduling time window.

7. The scheduling method for an intelligent logistics transportation system based on an automated algorithm according to claim 1, characterized in that, The operational feedback data specifically includes: The path start / stop status, operating speed, switching delay time, and conveyor belt load rate of the conveyor belt system; The actual driving path of the automated guided vehicle (AGV) from the starting node to the target node, the task completion time, the energy consumption level, the obstacle avoidance frequency, and the error trajectory. The task response time, operation time, material handling accuracy, and real-time inventory changes of the warehousing and loading / unloading module.

Images