Intelligent distribution equipment path control system

Abstract

The application relates to the technical field of intelligent distribution, in particular to an intelligent distribution equipment path control system, which comprises a space-time cognition graph module, the space-time cognition graph module is used for constructing a hierarchical space-time cognition graph, the hierarchical space-time cognition graph comprises a probability optimal solution route graph generated based on historical multiple walking path data and real-time environment data and a future obstacle intention trajectory predicted based on a generative AI world model, the hierarchical space-time cognition graph is constructed, the historical path entropy value algorithm and the generative AI world model are combined, forward-looking planning and local smooth correction of the path are realized, the problem of sudden stop and sudden turn caused by reactive planning of a traditional system is solved, a conflict pre-elimination mechanism is used for distributing key area time slices for multiple devices, a passive scheduling mode of negotiation after meeting is abandoned, intersection congestion and path deadlock are avoided, and the multiple-device collaborative traffic efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent delivery technology, specifically to an intelligent delivery equipment path control system. Background Technology

[0002] Existing intelligent delivery equipment path control technologies generally suffer from problems such as lagging path planning, conflicts in multi-device collaboration, weak dynamic task adaptability, insufficient robustness in complex environments, uneven path execution, and a lack of dynamic learning capabilities. Traditional systems employ step-by-step logic, which can only make reactive path adjustments based on the environmental information collected by current sensors. It cannot predict environmental changes in the future, leading to frequent sudden stops and turns of the equipment, affecting delivery efficiency and increasing energy consumption and mechanical wear.

[0003] In terms of multi-device collaboration, traditional systems employ a scheduling mode of negotiation after encounter, which easily leads to congestion at intersections and path deadlocks, failing to achieve globally optimal scheduling. Furthermore, the path planning algorithms of traditional systems have fixed weights, making them unable to dynamically adapt to sudden tasks. In severe weather conditions such as heavy rain and fog, LiDAR, visual sensors, and other sensors are prone to perception degradation, resulting in decreased positioning accuracy or even failure.

[0004] Therefore, an intelligent delivery equipment route control system is needed to improve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent delivery equipment path control system to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A smart delivery equipment route control system, comprising: The spatiotemporal cognitive graph module is used to construct a hierarchical spatiotemporal cognitive graph, which includes a probabilistic optimal solution route map generated based on historical multiple walking path data and real-time environmental data, and a future obstacle intention trajectory predicted based on a generative AI world model. The dynamic learning module, the spatiotemporal cognitive graph module includes a path database unit, a path analysis unit, and a model optimization unit. The path database unit is used to store multiple walking path data. The path analysis unit is used to systematically analyze the stored multiple walking path data to discover data patterns. The model optimization unit is used to dynamically optimize the spatiotemporal cognitive graph module based on the analysis results. The multi-agent collaborative scheduling module is used to allocate passage time slices for key areas to multiple delivery devices based on a conflict pre-resolution mechanism, so as to achieve conflict-free collaborative scheduling of multiple devices.

[0007] As a preferred embodiment of the present invention, the spatiotemporal cognitive map module includes: The global cognitive graph unit is used to call the historical multiple-walking path data stored in the path database unit, evaluate each potential path through the historical path entropy algorithm, and output a probabilistic optimal solution route map containing multiple potential paths and their evaluation parameters. The local prediction control unit is used to predict the intended trajectory of obstacles within the next 1-3 seconds by combining the generative AI world model based on the Transformer+Diffusion Policy architecture with the activity patterns of obstacles in similar scenarios stored in the path database unit, and to perform local smoothing correction on the probabilistic optimal solution route map.

[0008] As a preferred embodiment of the present invention, the multi-agent cooperative scheduling module includes: The time-space lock unit is used to allocate a unique passage time slice for each device to pass through intersections or narrow road sections at the moment of task allocation, in combination with the historical passage data of key areas stored in the path database unit, and the device passes according to the allocated time slice. The distributed negotiation unit is used to conduct distributed negotiation with surrounding devices to adjust the passage time slice or path when the device malfunctions or encounters a sudden obstacle. The digital twin simulation unit is used to build a digital twin model of the delivery scenario, simulate the path execution and collaborative scheduling process, and optimize the scheduling scheme.

[0009] As a preferred embodiment of the present invention, it further includes: The dynamic weight scheduling module is used to dynamically adjust the weight factors of the route planning based on the order priority. The dynamic weight scheduling module includes a priority identification unit and a weight dynamic adjustment unit. The weight dynamic adjustment unit adopts an attention mechanism and introduces dynamic weight factors including time cost weight, energy consumption cost weight, and path length weight. When an urgent order is detected, the time cost weight is increased, and local path replanning is forcibly triggered.

[0010] As a preferred embodiment of the present invention, the model optimization unit is further configured to dynamically adjust the initial value of the weight factor and the adjustment strategy of the dynamic weight scheduling module based on the analysis results of the path analysis unit.

[0011] As a preferred embodiment of the present invention, it further includes: The fault-tolerant control module is used to perform positioning compensation when the sensor's perception degrades or fails; The fault-tolerant control module includes a perception degradation detection unit and an inertial navigation compensation unit. The inertial navigation compensation unit is used to use historical motion state data collected by the IMU inertial measurement unit, combined with inertial navigation compensation parameters for similar scenarios stored in the path database unit, to predict the current position of the device through an inertial navigation algorithm, and to correct the positioning error by combining historical path data.

[0012] As a preferred embodiment of the present invention, it further includes: The semantic-level control module is used to convert path planning results into semantic-level motion commands to achieve smooth path execution. The semantic-level control module includes a motion primitive generation unit and an instruction optimization unit. The motion primitive generation unit is used to convert discrete path points into a semantic-level motion primitive sequence including uniform speed straight driving, smooth left turn, and slight deceleration to avoid obstacles. The instruction optimization unit is used to optimize the connection between each instruction and adjust the execution parameters of the motion primitives in combination with the device's own state data.

[0013] As a preferred embodiment of the present invention, the model optimization unit is further configured to optimize the motion primitive parameter settings of the semantic-level control module based on the analysis results of the path analysis unit.

[0014] As a preferred embodiment of the present invention, the path analysis unit uses the K-means clustering algorithm or the random forest algorithm to mine the data stored in the path database unit. The mining includes path efficiency analysis, scene feature analysis, equipment operation analysis, algorithm effect analysis, and dynamic trend analysis.

[0015] As a preferred embodiment of the present invention, it further includes: Cloud-edge collaborative units include cloud nodes and edge nodes; The cloud nodes are used for global path planning, multi-device collaborative scheduling, AI model training, and cloud backup of the path database unit; The edge nodes are integrated into the delivery equipment and are used for real-time processing of sensor data, local obstacle avoidance decisions, storage of the local path database at the edge, and fine-tuning of the local model. The cloud-edge collaboration unit dynamically allocates computing tasks based on network status, computing load, and order urgency, and achieves data and model synchronization between cloud and edge databases.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a hierarchical spatiotemporal cognitive map, combines a historical path entropy algorithm and a generative AI world model, and achieves forward-looking path planning and local smoothing correction. It solves the problem of sudden stops and turns caused by reactive planning in traditional systems. It adopts a conflict pre-resolution mechanism to allocate key area passage time slices for multiple devices, abandons the passive scheduling mode of negotiation after encounter, effectively avoids intersection congestion and path deadlock, and improves the efficiency of multi-device collaborative passage. This invention uses a fault-tolerant control module to predict the device position and correct positioning errors by using inertial navigation compensation when sensor perception degrades. This solves the problem of decreased positioning accuracy or even failure in complex environments, achieves smooth path execution, and reduces device energy consumption and mechanical wear. Attached Figure Description

[0017] Figure 1 This is an overall system block diagram of the present invention; Figure 2 This is a block diagram illustrating the principle of the dynamic learning module in this invention. Figure 3 This is a block diagram illustrating the principle of the multi-agent collaborative scheduling module in this invention. Figure 4 This is a schematic diagram of the cloud-edge collaborative unit in this invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] To facilitate understanding of the present invention, a more complete description will be given below with reference to relevant descriptions. Several embodiments of the invention are provided. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0020] Please see Figure 1-4 The present invention provides a technical solution: For an example, please refer to... Figure 1 , 2Items 3 and 4 represent a path control system for intelligent delivery equipment, suitable for delivery scenarios involving unmanned delivery vehicles in complex park environments in urban last-mile delivery. The system includes a perception module, a spatiotemporal cognitive graph module, a dynamic weighted scheduling module, a multi-agent collaborative scheduling module, a fault-tolerant control module, a semantic-level control module, an execution module, a dynamic learning module, and a cloud-edge collaborative unit. The modules are interconnected via a CAN bus to achieve data interaction and command transmission.

[0021] When an unmanned delivery vehicle receives a delivery order, the system initiates a complete path planning and control cycle. The perception module starts working first, and its multimodal sensor group includes a 32-line LiDAR, a binocular vision sensor, a 24GHz millimeter-wave radar, an IMU inertial measurement unit, an RTK-GNSS positioning module, and a UWB positioning module, which respectively collect environmental information, equipment status information, and order information.

[0022] LiDAR and binocular vision sensors scan obstacles, road markings, and intersections in front of and around the vehicle at a frequency of 10Hz. Millimeter-wave radar assists in detecting the speed of moving targets. An IMU (Inertial Measurement Unit) records the vehicle's speed, acceleration, and attitude angle in real time. RTK-GNSS and UWB positioning modules switch according to the scene, providing centimeter-level positioning in open outdoor areas and automatically switching to UWB positioning in GPS blind spots such as parks or residential areas. The data preprocessing unit uses an STM32H743 processor to timestamp and synchronize these multi-source sensor data. It then fuses the data using a Kalman filter algorithm to remove abnormal data caused by missing LiDAR point clouds or blurred vision sensors, outputting standardized environmental data, equipment status data, and order data.

[0023] These standardized data are simultaneously distributed to the spatiotemporal cognitive graph module, the dynamic weight scheduling module, the fault tolerance control module, and the dynamic learning module. The path database unit of the dynamic learning module immediately writes these raw data into the local database at the edge as raw material for subsequent analysis.

[0024] After receiving standardized data, the spatiotemporal cognitive graph module begins to construct a hierarchical spatiotemporal cognitive graph. This module uses an NVIDIA Jetson AGX Orin processor for data processing. Its global cognitive graph unit first sends a call request to the path database unit of the dynamic learning module to obtain multiple walking path data of all delivery tasks in the park over the past six months, including the congestion probability of each road segment at different times, the frequency of human intervention, and the traffic patterns of pedestrians and vehicles.

[0025] This unit constructs a spatiotemporal cognitive graph using a three-dimensional Cartesian coordinate system. The X-axis represents time information, the Y-axis represents spatial location information, and the Z-axis represents path evaluation parameters. It associates time entities, spatial entities, path evaluation entities, and order entities to form a dynamic knowledge graph with an update frequency of 1Hz.

[0026] Building upon this, the global cognitive graph unit introduces a historical path entropy algorithm to calculate the entropy of multiple potential paths retrieved from the database. Let the historical running data sample set of a certain path be X={x1, x2,, x...} n}, where x i Let H be the interference parameter for the i-th run. Then the entropy value H of this path is calculated as H = -p(x). i log2p(x) i ), where p(x) i The value is the ratio of the number of times the interference parameter occurs to the total number of runs.

[0027] A lower entropy value indicates a more stable path operation, less interference, and higher priority. Based on current real-time environmental data, this unit outputs a probabilistic optimal solution route map containing 3 to 5 potential paths and their evaluation parameters, clearly defining the priority of each path, and feeding the path planning results back to the dynamic learning module.

[0028] The local prediction control unit begins predicting future situations based on a generative AI world model. This unit employs a Transformer+Diffusion Policy architecture, drawing nearly 100 local scene samples from the path database unit to quickly adapt to a general model, keeping model inference latency below 0.3 seconds. Subsequently, this unit inputs real-time environmental data from the perception module, combines it with obstacle activity patterns from similar scenarios in the database, and uses the world model to predict the movement trajectories and intentions of each obstacle within the next 1-3 seconds.

[0029] Based on the prediction results, the local prediction control unit performs local smoothing corrections on the optimal path output by the global cognitive map unit. For example, when a pedestrian is predicted to cross the road, a slight deceleration avoidance command is added in advance, so that the vehicle naturally decelerates before the pedestrian crosses the road and resumes its original speed after the pedestrian has crossed, avoiding sudden stopping. The corrected path execution data is then fed back to the dynamic learning module.

[0030] During route planning, the dynamic weight scheduling module adjusts its parameters in real time based on order attributes. The priority identification unit of this module parses order information, identifies order priorities based on delivery time requirements and product type, and feeds back data such as order type and delivery results to the dynamic learning module. The dynamic weight adjustment unit employs an attention mechanism, introducing three dynamic weight factors: time cost weight, energy cost weight, and route length weight, the sum of which is 1. When an urgent order is detected, this unit automatically increases the time cost weight while decreasing the energy cost weight and route length weight, forcibly triggering local route replanning and prioritizing the route with the shortest delivery time. When there are no urgent orders, the weight factors are adjusted to achieve a globally optimal balance of time, energy consumption, and path length. The weight adjustment strategy and execution data are fed back to the dynamic learning module in real time for subsequent iterative optimization of the weight factors.

[0031] When multiple unmanned delivery vehicles are running simultaneously in the same park, the multi-agent collaborative scheduling module activates the collaborative scheduling mechanism. The spatiotemporal lock unit of this module first calls the historical traffic data of key areas from the path database unit of the dynamic learning module, including the distribution of congestion periods at intersections and the average traffic efficiency of each period.

[0032] At the moment of task allocation, the time-space lock unit allocates a unique passage time slice for each device to pass through the key area based on the path planning results and expected arrival time of each device, ensuring that multiple vehicles pass through in sequence without conflict and achieving conflict pre-resolution.

[0033] The digital twin simulation unit synchronously constructs a digital twin model of the delivery scenario, mapping the real-time status of each device, path planning results, and pedestrian and obstacle data in the environment to the virtual environment. Combined with the scenario feature patterns obtained from multiple walking path analyses in the database, the system rehearses the path execution and collaborative scheduling process, identifies potential conflicts in advance, and optimizes the scheduling scheme.

[0034] The distributed negotiation unit is responsible for handling emergencies. When a device malfunctions or encounters a sudden obstacle, the unit automatically conducts distributed negotiation with surrounding devices, adjusts the passage time slices or paths of each device according to the priority mechanism, and synchronizes the negotiation results to all relevant devices in real time to ensure the stability of global scheduling.

[0035] In terms of handling complex environments, the fault-tolerant control module continuously monitors the sensor status. The sensing degradation detection unit of this module monitors the working status of each sensor in real time. By comparing the consistency and stability of the data collected by multiple sensors, it determines whether the sensor has experienced sensing degradation.

[0036] When sensor degradation is detected, the unit feeds back information such as the scenario, time, and processing method of the degradation to the dynamic learning module. The inertial navigation compensation unit then starts, using historical motion state data collected by the IMU inertial measurement unit and inertial navigation compensation parameters for similar scenarios stored in the path database unit. The unit predicts the current position of the device through an inertial navigation algorithm until the sensor returns to normal. The unit combines historical path data and spatiotemporal cognitive map to correct the positioning error, ensuring that the positioning accuracy meets the delivery requirements, and feeds back the compensation effect data to the dynamic learning module.

[0037] After path planning is completed, the semantic-level control module transforms the planning results into executable motion commands. The motion primitive generation unit of this module transforms the discrete path points output by the global cognitive graph unit into a sequence of semantic-level motion primitives, including standardized command sequences such as uniform speed straight driving, gentle left turn, slight deceleration and obstacle avoidance. It also optimizes the parameter settings of the motion primitives by combining the analysis results of multiple walking paths output by the dynamic learning module.

[0038] The instruction optimization unit optimizes the motion primitive sequence to ensure smooth transitions between instructions and avoid frequent acceleration and deceleration. It also adjusts the execution parameters of the motion primitives by combining the device's own status data and the device operation loss data stored in the path database unit, thereby minimizing energy consumption and mechanical losses, and feeds the instruction execution effect back to the dynamic learning module.

[0039] The execution module receives motion commands output by the semantic-level control module. Its drive unit uses a DC brushless motor to drive the vehicle's power system, steering system, and braking system to perform actions such as constant speed, steering, and obstacle avoidance.

[0040] The status feedback unit uses an encoder and a GPS positioning module to collect real-time execution status data such as the vehicle's actual speed, position, attitude, energy consumption, and mechanical wear. This data is then fed back to the spatiotemporal cognitive graph module, the semantic-level control module, and the dynamic learning module to form a closed-loop control system, ensuring the accuracy of path execution.

[0041] The cloud-edge collaboration unit includes cloud nodes and edge nodes. The cloud nodes use Alibaba Cloud servers and are responsible for large-scale and complex computing tasks such as global path planning, collaborative scheduling of multiple unmanned delivery vehicles, training and updating of AI world models, and order aggregation optimization. They are also responsible for cloud backup of the path database, global data aggregation and analysis, and global model optimization.

[0042] Edge nodes are integrated into unmanned delivery vehicles and are responsible for real-time processing of sensor data, local obstacle avoidance, and millisecond-level decision-making. They can still operate autonomously even when the network is down. They are also responsible for storing the local path database at the edge, writing real-time data, and fine-tuning local models. The cloud-edge collaboration unit dynamically allocates computing tasks based on network status, computing load, and order urgency to achieve a balance between global optimization and local real-time response. At the same time, it realizes data synchronization and model synchronization between the cloud and edge databases.

[0043] The dynamic learning module runs through the entire system operation process. Its path database unit adopts a distributed database to store various types of data during system operation, including basic data of multiple walking paths, environmental and scene data, equipment operation data, planning and scheduling data, and order and execution data. This unit supports real-time writing, querying, updating and backup of data, adopts a data cleaning mechanism to remove invalid and abnormal data, and realizes data synchronization between the cloud and the edge through the cloud-edge collaboration unit.

[0044] The path analysis unit uses K-means clustering and random forest algorithms to systematically analyze the multiple walking path data and various related data stored in the database, including path efficiency analysis, scene feature analysis, equipment operation analysis, algorithm effect analysis, and dynamic trend analysis, and outputs a detailed analysis report.

[0045] Based on the analysis results of the path analysis unit, the model optimization unit dynamically iteratively optimizes the core algorithm model in the system. This includes adjusting the construction parameters and path evaluation weights of the spatiotemporal cognitive graph, fine-tuning the generative AI world model, optimizing the initial values ​​and adjustment strategies of the weight factors in the dynamic weight scheduling module, improving the time slice allocation strategy and distributed negotiation rules of the spatiotemporal lock unit, optimizing the perception degradation detection threshold and inertial navigation compensation parameters, and adjusting the parameter settings of semantic-level motion primitives. The optimized model parameters are synchronized to the corresponding modules and stored in the path database unit for subsequent iteration reference.

[0046] As delivery tasks are continuously executed and multiple walking path data are continuously accumulated, the system's path planning accuracy, collaborative scheduling efficiency, environmental adaptability, and equipment operation economy are continuously improved, realizing the system's self-learning, self-optimization, and self-adaptation.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A path control system for intelligent delivery equipment, characterized in that, include: The spatiotemporal cognitive graph module is used to construct a hierarchical spatiotemporal cognitive graph, which includes a probabilistic optimal solution route map generated based on historical multiple walking path data and real-time environmental data, and a future obstacle intention trajectory predicted based on a generative AI world model. The dynamic learning module, the spatiotemporal cognitive graph module includes a path database unit, a path analysis unit, and a model optimization unit. The path database unit is used to store multiple walking path data. The path analysis unit is used to systematically analyze the stored multiple walking path data to discover data patterns. The model optimization unit is used to dynamically optimize the spatiotemporal cognitive graph module based on the analysis results. The multi-agent collaborative scheduling module is used to allocate passage time slices for key areas to multiple delivery devices based on a conflict pre-resolution mechanism, so as to achieve conflict-free collaborative scheduling of multiple devices.

2. The intelligent delivery equipment path control system according to claim 1, characterized in that, The spatiotemporal cognitive map module includes: The global cognitive graph unit is used to call the historical multiple-walking path data stored in the path database unit, evaluate each potential path through the historical path entropy algorithm, and output a probabilistic optimal solution route map containing multiple potential paths and their evaluation parameters. The local prediction control unit is used to predict the intended trajectory of obstacles within the next 1-3 seconds by combining the generative AI world model based on the Transformer+Diffusion Policy architecture with the activity patterns of obstacles in similar scenarios stored in the path database unit, and to perform local smoothing correction on the probabilistic optimal solution route map.

3. The intelligent delivery equipment path control system according to claim 1, characterized in that, The multi-agent cooperative scheduling module includes: The time-space lock unit is used to allocate a unique passage time slice for each device to pass through intersections or narrow road sections at the moment of task allocation, in combination with the historical passage data of key areas stored in the path database unit, and the device passes according to the allocated time slice. The distributed negotiation unit is used to conduct distributed negotiation with surrounding devices to adjust the passage time slice or path when the device malfunctions or encounters a sudden obstacle. The digital twin simulation unit is used to build a digital twin model of the delivery scenario, simulate the path execution and collaborative scheduling process, and optimize the scheduling scheme.

4. The intelligent delivery equipment path control system according to claim 1, characterized in that, Also includes: The dynamic weight scheduling module is used to dynamically adjust the weight factors of the route planning based on the order priority. The dynamic weight scheduling module includes a priority identification unit and a weight dynamic adjustment unit. The weight dynamic adjustment unit adopts an attention mechanism and introduces dynamic weight factors including time cost weight, energy consumption cost weight, and path length weight. When an urgent order is detected, the time cost weight is increased, and local path replanning is forcibly triggered.

5. The intelligent delivery equipment path control system according to claim 4, characterized in that: The model optimization unit is also used to dynamically adjust the initial value of the weight factor and the adjustment strategy of the dynamic weight scheduling module based on the analysis results of the path analysis unit.

6. The intelligent delivery equipment path control system according to claim 1, characterized in that, Also includes: The fault-tolerant control module is used to perform positioning compensation when the sensor's perception degrades or fails; The fault-tolerant control module includes a perception degradation detection unit and an inertial navigation compensation unit. The inertial navigation compensation unit is used to use historical motion state data collected by the IMU inertial measurement unit, combined with inertial navigation compensation parameters for similar scenarios stored in the path database unit, to predict the current position of the device through an inertial navigation algorithm, and to correct the positioning error by combining historical path data.

7. The system according to claim 1, characterized in that, Also includes: The semantic-level control module is used to convert path planning results into semantic-level motion commands to achieve smooth path execution. The semantic-level control module includes a motion primitive generation unit and an instruction optimization unit. The motion primitive generation unit is used to convert discrete path points into a semantic-level motion primitive sequence including uniform speed straight driving, smooth left turn, and slight deceleration to avoid obstacles. The instruction optimization unit is used to optimize the connection between each instruction and adjust the execution parameters of the motion primitives in combination with the device's own state data.

8. The system according to claim 7, characterized in that, The model optimization unit is also used to optimize the motion primitive parameter settings of the semantic-level control module based on the analysis results of the path analysis unit.

9. The system according to claim 1, characterized in that, The path analysis unit uses the K-means clustering algorithm or the random forest algorithm to mine the data stored in the path database unit. The mining includes path efficiency analysis, scene feature analysis, equipment operation analysis, algorithm effect analysis, and dynamic trend analysis.

10. The system according to claim 1, characterized in that, Also includes: Cloud-edge collaborative units include cloud nodes and edge nodes; The cloud nodes are used for global path planning, multi-device collaborative scheduling, AI model training, and cloud backup of the path database unit; The edge nodes are integrated into the delivery equipment and are used for real-time processing of sensor data, local obstacle avoidance decisions, storage of the local path database at the edge, and fine-tuning of the local model. The cloud-edge collaboration unit dynamically allocates computing tasks based on network status, computing load, and order urgency, and achieves data and model synchronization between cloud and edge databases.

Images