Logistics path control system based on machine learning
By using a machine learning-based multi-objective optimization system, the dynamic balance problem of multi-dimensional constraints in logistics route planning was solved, enabling optimal route adjustment under real-time traffic conditions, thereby improving delivery efficiency and reducing operating costs.
Patent Information
- Application Number
- CN202511131211.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-13
- Publication Date
- 2025-11-14
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing logistics route planning systems struggle to achieve a continuous and dynamic optimal balance while considering multiple constraints such as timeliness, service level, and vehicle utilization.
A machine learning-based multi-objective optimization system is adopted, which generates and adjusts the optimal logistics path in real time through data acquisition, feature processing, multi-objective optimization modules and scheduling execution modules, combined with deep reinforcement learning algorithms and dynamic weight adjustment strategies.
It enables multi-objective logistics path optimization under real-time traffic condition changes, improving delivery efficiency, reducing operating costs, and ensuring timely completion of delivery tasks.
Smart Images
Figure CN120952295A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of logistics technology and intelligent control technology, specifically a logistics path control system based on machine learning. Background Technology
[0002] Current logistics route planning generally adopts the classic vehicle routing problem framework, supplemented by heuristic or metaheuristic solution methods such as genetic algorithms, ant colony algorithms, and simulated annealing. Some advanced systems can integrate GPS positioning, real-time road congestion index, and warehouse and distribution scheduling information, and make rolling corrections to routes through rules or weighted coefficients; large platforms also access operational data such as historical orders, vehicle load, driver working hours, and delivery radius, and generate the optimal routes for the next day or the current batch through batch calculations at the dispatch center, which are then distributed to the vehicle terminals for execution.
[0003] However, the above solutions still have significant shortcomings: the objective function usually focuses on distance or fuel consumption, making it difficult to simultaneously consider multi-dimensional constraints such as timeliness, service level, and vehicle utilization. As a result, it is difficult to achieve a continuous and dynamic optimal balance between delivery efficiency, operating costs, and customer experience. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a machine learning-based logistics path control system. It demonstrates how a dynamic weight adjustment strategy enables multi-objective logistics path optimization based on real-time traffic conditions and road condition changes, thereby improving delivery efficiency and reducing costs.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based logistics route control system, comprising:
[0006] The data acquisition module is used to collect logistics-related data in real time and output raw, multi-source heterogeneous data.
[0007] The feature processing module is used to clean the source heterogeneous data, perform self-supervised spatiotemporal pre-training and feature construction, and generate a high-quality comprehensive feature vector for path optimization. The feature processing module includes a data cleaning unit, a spatiotemporal representation pre-training unit and a feature construction unit.
[0008] The multi-objective optimization module constructs a logistics path model based on a deep reinforcement learning algorithm and uses a dynamic weight adjustment strategy to jointly optimize the logistics path model, outputting the optimal path decision strategy.
[0009] The path generation module is used to generate the optimal driving path for the corresponding vehicle based on the optimal path decision strategy.
[0010] The scheduling and execution module is used to send the optimal driving route to the vehicle terminal.
[0011] Preferably, the logistics-related data includes order information, vehicle operation status data, traffic condition data, and weather data.
[0012] Preferably, the data cleaning unit performs data cleaning through the following steps:
[0013] 3.1 Identifying and removing outliers based on the Z-score method;
[0014] 3.2 Using Gaussian process regression to predict and complete missing values;
[0015] 3.3 All numerical features are mapped to the [0,1] interval using minimum-maximum normalization.
[0016] Preferably, the spatiotemporal representation pre-training unit adopts a masked node prediction method, which masks a set of randomly selected nodes in the road network graph and learns node embedding vectors by minimizing the self-supervised reconstruction error, thereby improving the ability to represent the spatiotemporal dependencies of traffic flow.
[0017] Preferably, the feature construction unit merges the static attribute vector and the spatiotemporal embedding vector through linear fusion to generate a comprehensive feature vector, so as to simultaneously reflect road attributes and real-time traffic conditions.
[0018] Preferably, the logistics route model includes driving distance, delivery time, and operating costs, and the formula for the logistics route model is:
[0019]
[0020] Where J is the total delivery cost, K is the total number of delivery vehicles, E is the edge set of the road network, and x ijk For binary decision variables, c takes the value 1 when the k-th vehicle passes through edge (i,j) and 0 otherwise. ij The comprehensive cost of edge (i,j) is defined as:
[0021] c ij =ω d d ij +ω t t ij ,
[0022] In the formula, d ij Let t be the geometric distance between nodes i and j. ij To estimate the travel time, ω d For distance penalty weights, ω t Time penalty weight.
[0023] Preferably, the dynamic weight adjustment strategy updates the weights of the k-th optimization objective based on the following model formula:
[0024]
[0025] in, Let k be the weight of the k-th optimization objective in the previous iteration. The updated weights for the k-th optimization objective, ΔJ k Let η be the performance increment of the k-th optimization objective since the previous iteration, η be the learning rate of weight updates used to control the weight adjustment rate, N be the total number of all optimization objectives in the multi-objective optimization module, and the denominator be the normalization term of all objective weight updates to ensure that the sum of the weights of each objective after the update is 1.
[0026] Preferably, the scheduling execution module monitors the Euclidean distance deviation between the vehicle's location and the planned location in real time, and automatically triggers online replanning when the deviation exceeds a preset threshold.
[0027] This invention provides a machine learning-based logistics route control system. It has the following beneficial effects:
[0028] This machine learning-based logistics route control system optimizes multi-objective logistics routes through deep reinforcement learning algorithms, effectively balancing delivery distance, delivery time, and operating costs. By dynamically adjusting objective weights, the system can automatically adjust optimization strategies based on real-time traffic conditions, weather changes, and other multi-dimensional factors, enabling delivery routes to respond to changes in real time and ensuring improved delivery efficiency while reducing total delivery costs.
[0029] By monitoring vehicle deviations and dynamic road conditions in real time, the system ensures that vehicles always follow the optimal route. Even in the event of sudden traffic jams, weather changes, or other unforeseen circumstances, the system can automatically trigger replanning and adjust routes in a timely manner to ensure that delivery tasks are completed on time, greatly enhancing the system's adaptability and flexibility. Attached Figure Description
[0030] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation
[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0032] Example 1
[0033] like Figure 1As shown, this embodiment of the invention provides a machine learning-based logistics route control system, including a data acquisition module for real-time acquisition of logistics-related data and output of raw multi-source heterogeneous data. The logistics-related data includes order information, vehicle operation status data, traffic condition data, and weather data.
[0034] The feature processing module is used to clean the heterogeneous source data, perform self-supervised spatiotemporal pre-training, and construct features to generate high-quality comprehensive feature vectors for path optimization. The feature processing module includes a data cleaning unit, a spatiotemporal representation pre-training unit, and a feature construction unit. The data cleaning unit cleans the data through the following steps:
[0035] 3.1 Identifying and removing outliers based on the Z-score method.
[0036] 3.2 Gaussian process regression is used to predict and complete missing values.
[0037] 3.3 All numerical features are mapped to the [0,1] interval using minimum-maximum normalization.
[0038] The spatiotemporal representation pre-training unit adopts a masked node prediction method, which masks a randomly selected set of nodes in the road network graph and learns node embedding vectors by minimizing the self-supervised reconstruction error, thereby improving the ability to represent the spatiotemporal dependencies of traffic flow.
[0039] The feature construction unit merges static attribute vectors and spatiotemporal embedding vectors through linear fusion to generate a comprehensive feature vector that simultaneously reflects road attributes and real-time traffic conditions. The static attribute vector contains basic road attributes such as road length, road type, and speed limit information. These static features do not change over time and are typically obtained from road databases or map services. The spatiotemporal embedding vector originates from the spatiotemporal representation pre-training unit and is learned through masked node prediction, reflecting the dynamic changes and spatiotemporal dependencies of nodes in traffic flow.
[0040] The multi-objective optimization module constructs a logistics route model based on a deep reinforcement learning algorithm and employs a dynamic weight adjustment strategy to jointly optimize the model, outputting the optimal route decision strategy. The logistics route model includes driving distance, delivery time, and operating costs. The formula for the logistics route model is:
[0041]
[0042] Where J is the total delivery cost, K is the total number of delivery vehicles, E is the edge set of the road network, and x ijk For binary decision variables, c takes the value 1 when the k-th vehicle passes through edge (i,j) and 0 otherwise. ij The comprehensive cost of edge (i,j) is defined as:
[0043] c ij =ω d d ij +ω t t ij ,
[0044] In the formula, d ij Let t be the geometric distance between nodes i and j. ij To estimate the travel time, ω d For distance penalty weights, ω t Time penalty weight.
[0045] The dynamic weight adjustment strategy updates the weights of the k-th optimization objective based on the following model formula:
[0046]
[0047] in, Let k be the weight of the k-th optimization objective in the previous iteration. The updated weights for the k-th optimization objective, ΔJ k Let η be the performance increment of the k-th optimization objective since the previous iteration, η be the learning rate of weight updates to control the weight adjustment rate, N be the total number of all optimization objectives in the multi-objective optimization module, and the denominator be the normalization term of all objective weight updates to ensure that the sum of the weights of each objective after the update is 1.
[0048] The route generation module is used to generate the optimal driving route for the corresponding vehicle based on the optimal route decision strategy.
[0049] The scheduling execution module is used to send the optimal driving route to the vehicle terminal. The module monitors the Euclidean distance deviation between the vehicle's current position and the planned position in real time, and automatically triggers online replanning when the deviation exceeds a preset threshold. The specific steps are as follows:
[0050] Route Distribution: The scheduling execution module receives the optimal driving route output by the route generation module and distributes the route information to each delivery vehicle via the vehicle terminal. The vehicle terminal will then navigate and execute the corresponding driving operations based on the received route information.
[0051] Location monitoring: The module monitors the location of each vehicle in real time using GPS or other positioning technologies. The system continuously compares the Euclidean distance deviation between the vehicle's current location and the planned path.
[0052] Suppose the vehicle's planned location is 100 meters from its current location, but the actual location deviates from the planned location by 120 meters. In this case, the system will identify the deviation and take appropriate action.
[0053] Online replanning: When the Euclidean distance deviation between the vehicle and the planned path exceeds a preset threshold (deviation exceeding 200 meters), the system will automatically trigger online replanning. Online replanning will recalculate the optimal path based on the current actual road conditions, traffic conditions, and vehicle position, and send the new path to the vehicle terminal.
[0054] Replanning trigger conditions: The replanning mechanism will be activated if the deviation exceeds 200 meters, or if there is temporary traffic congestion, road construction, or other similar situations.
[0055] Real-time scheduling adjustments: During execution, the scheduling module continuously adjusts the scheduling strategy based on real-time vehicle data to ensure vehicles deliver along the optimal route. If a vehicle deviates from the route during delivery, the system will update the route based on real-time data and guide the vehicle to continue along the adjusted route.
[0056] Example 2
[0057] This embodiment is based on a solution to the dynamic logistics route optimization problem, employing a deep reinforcement learning algorithm to optimize the logistics route. The specific steps are as follows:
[0058] 1. The goal of the logistics route model is to minimize the total delivery cost. Assume the total delivery cost is:
[0059]
[0060] Where: K is the total number of delivery vehicles, for example K = 10 vehicles.
[0061] E is the edge set of the road network, containing all possible delivery routes.
[0062] x ijk As a decision variable, when vehicle k travels along edge (i,j), x ijk =1, otherwise 0.
[0063] The overall cost of the path C ij It can be calculated using the following formula:
[0064] C ij =ω d d ij +ω t t ij ,
[0065] in:
[0066] d ij Let d be the geometric distance between node i and node j. Assume d ij = 50 kilometers.
[0067] t ij To estimate the travel time, assume tij = 1 hour.
[0068] ω d and ω t These are the penalty weights for distance and time, respectively, assuming ω... d =0.5 and ω t =0.3.
[0069] Therefore, the comprehensive cost of edge (i,j) is calculated as follows:
[0070] C ij =0.5×50 + 0.3×1 = 25 + 0.3 = 25.3
[0071] 2. Dynamic weight adjustment strategy
[0072] To achieve multi-objective optimization, a dynamic weight adjustment strategy is adopted. The weight of each optimization objective is adjusted according to the performance increment of the objective. The formula for weight update is:
[0073]
[0074] in: Let be the weight of the k-th objective in the previous iteration. Assume the initial weight of the first objective is .
[0075] ΔJ k Let ΔJ1 be the performance increment of target k since the last iteration. Assume ΔJ1 = 0.2.
[0076] η is the learning rate, which controls the rate at which the weights are adjusted. Assume η = 1.
[0077] According to this formula, the updated weights for:
[0078]
[0079] The new weights are calculated and all weights are normalized to ensure that the total target weights sum to 1.
[0080] 3. Path generation and scheduling execution
[0081] Based on the optimization results above, the optimal route for each vehicle is generated. For example:
[0082] For vehicle 1, the optimal path might be from the starting point A to B, then to C, and finally to the destination D.
[0083] For vehicle 2, the optimal route might be from the starting point A directly to the destination D.
[0084] The scheduling and execution module is responsible for monitoring the deviation between the vehicle's location and the planned route in real time. When the deviation exceeds a preset threshold, the system will automatically trigger online replanning to ensure that the task is completed on time.
[0085] 4. Performance Evaluation and Optimization
[0086] The model was simulated and tested, and the results were verified using the following data:
[0087] Assuming that the expected time for delivery vehicle 1 to travel from origin A to destination B is 2 hours, and the actual time is 1.8 hours, the route optimization is successful.
[0088] Assume that the expected time for delivery vehicle 2 to travel from origin A to destination D is 3 hours, and the actual time is 3.2 hours. Due to traffic issues, the route is replanned in real time.
[0089] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A machine learning-based logistics route control system, characterized in that, include: The data acquisition module is used to collect logistics-related data in real time and output raw, multi-source heterogeneous data. The feature processing module is used to clean the source heterogeneous data, perform self-supervised spatiotemporal pre-training and feature construction, and generate a high-quality comprehensive feature vector for path optimization. The feature processing module includes a data cleaning unit, a spatiotemporal representation pre-training unit and a feature construction unit. The multi-objective optimization module constructs a logistics path model based on a deep reinforcement learning algorithm and uses a dynamic weight adjustment strategy to jointly optimize the logistics path model, outputting the optimal path decision strategy. The path generation module is used to generate the optimal driving path for the corresponding vehicle based on the optimal path decision strategy. The scheduling and execution module is used to send the optimal driving route to the vehicle terminal.
2. The machine learning-based logistics route control system according to claim 1, characterized in that: The logistics-related data includes order information, vehicle operation status data, traffic condition data, and weather data.
3. The machine learning-based logistics route control system according to claim 1, characterized in that: The data cleaning unit performs data cleaning through the following steps: 3.1 Identifying and removing outliers based on the Z-score method; 3.2 Using Gaussian process regression to predict and complete missing values; 3.3 All numerical features are mapped to the [0,1] interval using minimum-maximum normalization.
4. A machine learning-based logistics route control system according to claim 1, characterized in that: The spatiotemporal representation pre-training unit uses a masked node prediction method to mask a randomly selected set of nodes in the road network graph and learns node embedding vectors by minimizing the self-supervised reconstruction error.
5. A machine learning-based logistics route control system according to claim 1, characterized in that: The feature construction unit merges the static attribute vector and the spatiotemporal embedding vector through linear fusion to generate a comprehensive feature vector.
6. A machine learning-based logistics route control system according to claim 1, characterized in that: The logistics route model includes driving distance, delivery time, and operating costs. The formula for the logistics route model is: Where J is the total delivery cost, K is the total number of delivery vehicles, E is the edge set of the road network, and x ijk For binary decision variables, c takes the value 1 when the k-th vehicle passes through edge (i,j) and 0 otherwise. ij The comprehensive cost of edge (i,j) is defined as: c ij =ω d d ij +oh t t ij , In the formula, d ij Let t be the geometric distance between nodes i and j. ij To estimate the travel time, ω d For distance penalty weights, ω t Time penalty weight.
7. A machine learning-based logistics route control system according to claim 1, characterized in that: The dynamic weight adjustment strategy updates the weights of the k-th optimization objective based on the following model formula: in, Let k be the weight of the k-th optimization objective in the previous iteration. The updated weights for the k-th optimization objective, ΔJ k Let η be the performance increment of the k-th optimization objective since the previous iteration, η be the learning rate of weight updates, N be the total number of all optimization objectives in the multi-objective optimization module, and the denominator be the normalization term of all objective weight updates.
8. A machine learning-based logistics route control system according to claim 1, characterized in that: The scheduling execution module monitors the Euclidean distance deviation between the vehicle's location and the planned location in real time, and automatically triggers online replanning when the deviation exceeds a preset threshold.