Cloud-edge AI model coupled distributed cooperative training system and method

Abstract

The invention discloses a cloud-edge AI model coupled distributed cooperative training system and method. The system comprises a cloud computing power platform and a plurality of independent edge computing power platforms, the cloud computing power platform and the side end computing power platform are mutually coupled; the training data of any side end computing power platform is allocated and deployed to the other plurality of side end computing power platforms through the cloud computing power platform; and the side end computing power platform carries out information intercommunication through the cloud computing power platform, and simulates a distributed coordination state appearing in a task to carry out algorithm training when a multi-agent coordination task is trained in a complex environment. The problem that computing power resources of a single computing module are limited is solved, and processing and training of large-scale data such as large-scale image data and large models such as large language AI are accelerated. All the modules are independent of one another and can be coupled with one another through a switch, and training and application of a distributed lifetime learning algorithm can be achieved.

Description

Technical Field

[0001] The present invention belongs to the field of cloud-edge collaborative computing, and particularly relates to a distributed collaborative training system and method for coupling cloud-edge AI models. Background Art

[0002] With the growing demand for production automation in society, unmanned systems are being increasingly widely used in production and life. Intelligent agents such as unmanned aerial vehicles and unmanned vehicles play the role of executors. The simulation of intelligent agents before they are put into actual production use has been a widely discussed issue. Whether it is basic motion control or increasingly advanced and complex artificial intelligence algorithms, they all have to go through a long debugging process before being actually applied on intelligent agents. Moreover, when facing different scenarios, parameters in the algorithms, including the execution order, etc., will be different. Sequentially simulating and experimenting with different algorithms will consume a relatively long time. At the same time, the training rate of the algorithm model is limited by the computing power of a single computing platform. When the trained model is migrated to different platforms, it will also consume a certain amount of time. In order to improve the training and product implementation efficiency to a greater extent, instead of using a single computing unit for training, the cloud-edge collaborative method is chosen. It can not only independently complete the training of different algorithms at multiple edge ends. For the training of larger AI models, the cloud-edge collaborative method can also be used to split the large model tasks, make a reasonable task allocation according to the difference in cloud-edge computing power, and then couple the training results of each part (simple coupling or weighted coupling) to complete the training task with higher efficiency. Summary of the Invention

[0003] The technical problem to be solved by the present invention is: to provide a distributed collaborative training system and method for coupling cloud-edge AI models, which solves the problem of the difficulty of real test experiments in the prior art in complex environments.

[0004] The present invention adopts the following technical solutions to solve the above technical problems:

[0005] A distributed collaborative training system for coupling cloud-edge AI models includes a cloud computing power platform and several independent edge computing power platforms; the cloud computing power platform is coupled with the edge computing power platforms; the training data of any one edge computing power platform is distributed and deployed to multiple other edge computing power platforms through the cloud computing power platform; the edge computing power platforms communicate with each other through the cloud computing power platform, and when training the collaborative tasks of multiple intelligent agents in a complex environment, algorithm training is carried out by simulating the distributed collaborative state that appears in the tasks.

[0006] The cloud computing power platform and the edge computing power platform are connected through a switch. Each of the cloud computing power platform and the edge computing power platform is equipped with a corresponding display screen. The display screen of the cloud computing power platform is used to display the operation status of the entire system, and the display screen of the edge computing power platform is used to display the operation status and status of the intelligent agent corresponding to this module.

[0007] The switch is a switch that seamlessly switches between wired / wireless with a throughput greater than high. The edge computing power platform is equipped with multi-modal sensors including cameras, lidar, and rangefinders for collecting training data.

[0008] The distributed collaborative training method for cloud-edge AI model coupling includes five training methods, namely, self-training of edge intelligent agents, training of moving edge intelligent agents, collaborative training of cloud-edge intelligent agents, collaborative special training of cloud-edge intelligent agents, and multi-agent special collaborative training; among them,

[0009] Self-training of edge intelligent agents: The edge intelligent agents perform simulation training on the involved algorithms. When it comes to the collaborative part, the connection between different edge computing power platforms is established through a switch. After each edge computing power platform completes training, the data is uploaded to the cloud computing power platform for integration, and then the integrated data is deployed to the edge intelligent agents.

[0010] Training of moving edge intelligent agents: The moving edge computing power platform is connected to the cloud computing power platform as a remote edge, and then collaboratively trains with the cloud computing power platform. After the moving edge computing power platform collects data, it performs preliminary integration and sends it to the cloud computing power platform. The cloud computing power platform conducts model training and then returns the trained results to the moving edge computing power platform for status output.

[0011] Collaborative training of cloud-edge intelligent agents: The cloud computing power platform modularizes the large model in a specific way, and each module is sent to different edge computing power platforms for training. After the edge computing power platform completes training, the training results are sent to the cloud computing power platform, and the cloud computing power platform conducts model integration.

[0012] Collaborative special training of cloud-edge intelligent agents: Based on the large model training process of collaborative training of cloud-edge intelligent agents, for the planning and obstacle avoidance tasks of intelligent agents in specific complex scenarios or the autonomous decision-making tasks for complex environment exploration, combined with the semantic understanding ability of the language large model, the two are combined to conduct training for specific exploration and planning tasks.

[0013] Multi-agent special collaborative training: For multi-agent collaborative tasks, each edge computing power platform represents the corresponding intelligent agent, and uses communication means including local area network or switch to simulate the distribution of multiple intelligent agents in the task, conducts training on the multi-agent distributed communication method, and deploys the trained distributed communication method to the intelligent agent group to complete distributed collaboration.

[0014] For the planning and obstacle avoidance tasks of agents and the autonomous decision-making tasks of complex environment exploration in specific complex scenarios, in combination with the semantic understanding ability of the large language model, collaborative training is carried out using a cloud computing power platform. The planning and decision-making experience of complex link exploration is used as training data to train an AI for specific exploration planning tasks, which is carried on the cloud computing power platform. For the exploration needs of different agents in complex scenario tasks, the cloud large language model is requested to generate corresponding exploration planning instructions and directly apply them to the agents. Then, according to the actual task performance or simulation task performance of the agents, the human-machine interaction in the task-solving process is changed to machine-machine interaction.

[0015] For the distributed collaboration scheme of multiple agents in complex scenarios, one edge computing power platform represents one agent. A mirror is created in the cloud computing power platform for rendering the simulation environment and modeling and communication of multiple agents. Each edge computing power platform corresponds to one agent in the mirror rendering environment, and the same IP and communication group serial number as the computing module are adopted. According to the task requirements, with the assistance of the large language model for testing, the corresponding mirror of the training results is packaged and uploaded locally on the cloud computing power platform.

[0016] For the cloud-edge agent collaborative training, aiming at the problem that the training result deviation caused by the difference in training data of different training tasks assigned to the cloud and the edge makes it impossible to perform simple superposition fitting, a coupling method based on pruning and weighted superposition is proposed. The specific formula is as follows:

[0017]

[0018]

[0019]

[0020] Among them, is the result of the overall large model, is the result of the th sub-module, is the weight of each sub-module at the final coupling, The result after separate calculation by the sub-module minus the redundant or low-weight result , that is, the pruning operation. The weight is the proportion of the number of computing units of the corresponding module in the total number of computing units .

[0021] A computer device includes a memory and a processor. The memory stores a computer program. When the processor of the computer device executes the computer program, the distributed collaborative training method for cloud-edge AI model coupling as described above is implemented.

[0022] A computer-readable storage medium stores computer-readable instructions on which a computer program is stored. When the computer-readable instructions are executed by a processor, the distributed collaborative training method for cloud-edge AI model coupling as described above is invoked.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. The system and method of the present invention provide a parallel and efficient training solution for multi-task training, and can quickly realize the integration and deployment of algorithms.

[0025] 2. It solves the problem of limited computing power resources of a single computing module, and accelerates the processing and training of large-scale data such as a large number of image data and large models such as large language AI.

[0026] 3. Each module is independent of each other and can be coupled with each other through a switch. Using this feature, the training and application of distributed lifelong learning algorithms can be realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is the overall framework diagram of the present invention.

[0028] Figure 2 It is the algorithm training and deployment structure diagram of the cloud-edge computing power platform of the present invention.

[0029] Figure 3 It is the cloud-edge task training structure diagram of the present invention.

[0030] Figure 4 It is the cloud-edge large model training structure diagram of the present invention.

[0031] Figure 5 It is the large model decision-making task structure diagram of the present invention.

[0032] Figure 6 It is the cloud-edge distributed training and deployment structure diagram. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] The structure and working process of the present invention will be further described below with reference to the accompanying drawings.

[0034] A distributed collaborative training system with cloud-edge AI model coupling includes a cloud computing power platform and several independent edge computing power platforms; the cloud computing power platform is coupled with the edge computing power platforms; the training data of any edge computing power platform is distributed and deployed to multiple other edge computing power platforms through the cloud computing power platform; the edge computing power platforms communicate with each other through the cloud computing power platform, and when training the collaborative tasks of multiple intelligent agents in a complex environment, algorithm training is carried out by simulating the distributed collaborative state that appears in the task.

[0035] A distributed collaborative training method with cloud-edge AI model coupling includes five training methods, namely, self-training of edge intelligent agents, training of moving edge intelligent agents, collaborative training of cloud-edge intelligent agents, special collaborative training of cloud-edge intelligent agents, and special collaborative training of multiple intelligent agents; among them,

[0036] For the self-training of edge intelligent agents, the edge intelligent agents perform simulation training on the involved algorithms. When it comes to the collaborative part, the connection between different edge computing power platforms is established through a switch. After each edge computing power platform completes the training, the data is uploaded to the cloud computing power platform for integration, and then the integrated data is deployed to the edge intelligent agents.

[0037] For the training of moving edge intelligent agents, the moving edge computing power platform is connected to the cloud computing power platform as a remote edge, and then collaboratively trains with the cloud computing power platform. After the moving edge computing power platform collects data, it performs preliminary integration and sends it to the cloud computing power platform. The cloud computing power platform conducts model training and then returns the trained result to the moving edge computing power platform for status output.

[0038] For the collaborative training of cloud-edge intelligent agents, the cloud computing power platform modularizes the large model in a specific way, and each module is sent to different edge computing power platforms for training. After the edge computing power platforms complete the training, they send the training results to the cloud computing power platform, and the cloud computing power platform conducts model integration.

[0039] For the special collaborative training of cloud-edge intelligent agents, based on the large model training process of cloud-edge intelligent agent collaborative training, for the planning and obstacle avoidance tasks of intelligent agents in specific complex scenarios or for the autonomous decision-making tasks of exploring complex environments, combined with the semantic understanding ability of the language large model, the two are combined to carry out the training of specific exploration and planning tasks.

[0040] For the special collaborative training of multiple intelligent agents, for the collaborative tasks of multiple intelligent agents, each edge computing power platform represents the corresponding intelligent agent, and communication means including local area network or switch are used to simulate the distribution of multiple intelligent agents in the task, carry out the training of the multi-intelligent agent distributed communication method, and deploy the trained distributed communication method to the intelligent agent group to complete distributed collaboration.

[0041] Specific embodiments are as followsFigures 1 to 6 As shown

[0042] A distributed collaborative training system for cloud-edge AI model coupling, device hardware:

[0043] It consists of a cloud computing power platform and multiple edge computers (i.e., edge computing power platforms), each connected to a display screen to observe the virtual environment and the status of virtual agents, and they are connected using a switch. The server CPU uses Intel 8359P, and the GPU uses L40 with scene modeling acceleration and rendering functions to ensure that the server has powerful computing power and graphics rendering capabilities. The edge computing module uses orin agx to ensure its computing performance. The switch for connection is a high-throughput wired / wireless seamless switching switch to ensure the speed and quality of data transmission.

[0044] Cloud-edge interconnection:

[0045] The edge computing power platforms are connected through switches, local area networks or mesh networking. The switch is used for connection when the edge computing power platforms do not need to move, and both the local area network and the mesh module are used for connection when the edge computing power platforms have been mounted on mobile agents for training; the environment images of the edge computing power platforms are copied into the cloud computing power platform. By entering the device website of the cloud computing power platform, the corresponding modules can be viewed, and the modules can be operated by entering the images. Based on this principle, when a certain module has completed training, the module can package the complete result image and send it to the cloud computing power platform, and then it can be read through the cloud computing power platform and deployed to other blank modules.

[0046] Figure 2 The implementation solution for the self-training of the edge agents shown:

[0047] For visual recognition algorithms, laser mapping algorithms, etc. that may be used in complex underground scenarios, this embodiment provides a cloud-edge collaboration solution, specifically:

[0048] Each edge processor, which is orinagx in this embodiment, is equipped with different sensors according to different short hair requirements. Taking visual recognition as an example, a camera is equipped and camera driving and picture information reading are realized. The visual algorithm is modified and optimized on the edge module (i.e., the edge computing power platform). After optimization, the corresponding visual algorithm image is packaged. Here, docker is used for packaging. After creating the corresponding environment and files of docker, the cloud computing power platform is deployed to the edge module as a docker repository. Enter docker, package the image and submit it to the cloud computing power platform.

[0049] And so on, the algorithms obtained through independent or collaborative training of each edge are uploaded to the cloud in this way. Based on this upload process, there is a corresponding deployment process. Based on this process, high-efficiency dedicated intelligent agent computing modules can be produced.

[0050] Figure 3 A solution for implementing the training of edge intelligent agents representing motion:

[0051] For visual recognition algorithms, laser mapping algorithms, etc. that may be used in complex underground scenarios, when the algorithm complexity is too high and the required computing power exceeds the computing power of the edge computing module, the cloud-edge collaboration method is used for algorithm training and storage. Specifically:

[0052] When encountering some large-scale artificial intelligence algorithms or when the computing power conditions of the connected edge are not high, in addition to storing and scheduling functions, the cloud computing power platform can also exist as a computing module with the maximum computing power. Taking the access of a low-computing-power edge as an example, taking a visual processing algorithm such as the YOLO algorithm as an example, its real-time detection requires the assistance of a GPU. Assuming that the accessed edge has no GPU, it is impossible to perform detection and optimization smoothly. In response to this situation, the edge (edge computing power platform) and the cloud (cloud computing power platform) are connected through a switch. The image information received by the camera at the edge is transmitted to the cloud through the switch for processing, display, and optimization. Before this, the cloud will create a separate image for the above-mentioned receiving and computing operations. When the task is completed, similar to the edge uploading the image mentioned above, the cloud computing power platform packages and uploads the image on the local machine to complete the training and deployment of relatively complex algorithms under the condition of low edge computing power.

[0053] Figure 4 A solution for cloud-edge intelligent agent collaborative training:

[0054] To address the problem of insufficient computing power for large language model training that may occur, the method of cloud-edge collaborative distribution of sub-models is adopted to reduce the training burden of each module. The specific solution is:

[0055] In response to the computing power challenges brought about by the continuous increase in the scale of model operations, a distributed training method is used to reduce the load of individual hardware operations and improve the overall efficiency of model training. Multiple parallel training techniques such as data parallelism, pipeline parallelism, and tensor parallelism are adopted for distributed training of data. At the same time, schemes such as operator partitioning and pruning are adopted to divide the large model into small models and evenly distribute them to each edge computing module for data collection and iterative evolution. The parallel partitioning method can be divided into inter-layer parallelism and intra-layer parallelism. Using the calculation principle of block matrices, the parameters are sliced to different devices (slicing method) and mathematical consistency (mathematical equivalence) is ensured after slicing. At the same time, after slicing, dynamic programming and integer linear programming methods can be used within and between operators to select the optimal parallel strategy.

[0056] In response to the problem that the training result deviation caused by the difference in training data of different training tasks assigned to the cloud and the edge makes it impossible to perform simple superposition fitting, a coupling method based on pruning and weighted superposition is proposed to achieve a larger model with higher accuracy and robustness.

[0057] The specific formula is as follows:

[0058]

[0059]

[0060]

[0061] Among them, is the result of the overall large model, is the result of the th sub-module, is the weight corresponding to each sub-module during the final coupling, the result after separate calculation by the sub-module minus the redundant or low-weight result , that is, the pruning operation, and the weight is the proportion of the number of computing units of the corresponding module in the total number of computing units .

[0062] Figure 5 represents the solution for the collaborative special training of cloud-edge agents:

[0063] For problems such as path planning that may be used in complex underground scenarios, obstacle avoidance, and autonomous decision-making strategies such as full coverage strategies that may be considered during autonomous exploration, combined with the above large model training method, a role of using a large model AI to replace humans for programming control is proposed. Utilizing the language understanding of the language large model, the transformation from human-computer interaction to machine-machine interaction is realized, and further automation is achieved. The specific solution is as follows:

[0064] Use the above large model training solution to train language large models for specific functions such as planning and decision-making. Use the empirical data of agent planning in previous projects as training references to enable the large model to have the basic programming logic for solving specific problems. For the exploration needs of different agents in complex scenario tasks, request the cloud large model to generate corresponding exploration planning instructions and directly apply them to the agents, and then continuously optimize the decision-making ability of the large model based on the actual task performance or simulation task performance of the agents.

[0065] Figure 6 A solution representing the special collaborative training of multiple agents:

[0066] For the distributed collaborative tasks of multiple agents in complex environments, the present invention can be used to conduct communication tests for various distributed algorithms. The specific solution is as follows:

[0067] Connect each edge computing module using a local area network or mesh component. For communication tests, such as the distributed communication test of ROS2, under the condition of the same local area network or the same mesh network, test the distributed communication performance by means of full communication, group communication, attacking a certain node, etc. For the distributed collaborative scheme of multiple agents in complex scenarios, adopt the method of using one edge computing module to represent one agent. Create an image in the cloud server for rendering the simulation environment, modeling, and communication of multiple agents. Each edge computing module corresponds to one agent in the rendering environment of the image. Adopt the same IP and communication group serial number as the computing module. According to task requirements, such as collaborative exploration, conduct algorithm training and optimization. The language large model can be used to assist in the test, and the image corresponding to the training result is also packaged and uploaded locally in the cloud server.

[0068] Specifically, the distributed collaborative training method of cloud-edge AI model coupling includes five training methods, namely, self-training of edge agents, training of moving edge agents, collaborative training of cloud-edge agents, special collaborative training of cloud-edge agents, and special collaborative training of multiple agents;

[0069] The first category, assuming there are intelligent edge ends, each edge end independently uses its local data Train a certain algorithm. For example, for the exploration of underground spaces, algorithms such as visual recognition, lidar mapping, submap stitching, path planning, and distributed collaboration are required. Use each edge device to simulate and train the above algorithms. The general process is as follows:

[0070]

[0071] Among them is the model trained by the th edge device, The function is the training function of each edge device for a certain algorithm. The edge device uses the local dataset and the corresponding m to train the model. When it comes to the collaborative part, establish connections between different edge devices through a switch:

[0072]

[0073] represents the common output of several collaborative devices, The function is the integration function, which is responsible for integrating the information of the collaborative devices involved. After the training is completed, upload the data of each edge device to the cloud computing platform for integration, and then it can be directly deployed to the computing module of the intelligent agent:

[0074]

[0075] Among them, is the total model integrated after being uploaded to the server, and I is the integration function for integrating the computing platforms of each edge device.

[0076] Second, for the algorithm training that needs to cooperate with the movement of the intelligent agent, connect the intelligent agent to the cloud-edge system as a remote edge device through a local area network or a mesh network, etc. Facing problems such as insufficient computing power, slow training efficiency, and poor robustness of the intelligent agent's computing module, adopt the method of collaborative work between the edge computing platform and the cloud computing platform, and place the complex algorithm part such as image recognition and processing on the cloud computing platform. The intelligent agent computing module i equipped with the corresponding sensor is responsible for the acquisition and simple integration of the data involved in the movement process:

[0077]

[0078] Among them For the integrated data output of the corresponding agent module, I is the data integration function. After integrating and packaging the data, use wireless communication media such as local area network or mesh, or wired media such as switches, and use communication means such as socket or ROS topic communication to realize the transmission of the integrated data from the edge to the cloud:

[0079]

[0080] Here represents the edge device information input to the cloud, and the C function is the various communication methods mentioned above. The large computing power cloud server uses the required training model function Train() to perform more complex model training, and then the training results are returned to the agent for status output through another communication channel (to avoid conflicts with the communication data channel):

[0081]

[0082] .

[0083] The third category is for the training efficiency problem of ultra-large models Adopt the cloud-edge collaborative training method. The cloud server modularizes the large model in a specific way F(), such as simple model division, feature division, or model pruning division:

[0084]

[0085] Among them, the parameter represents the number of small models divided.

[0086] Each module is sent to different edge devices for training, that is:

[0087]

[0088] After the edge device training process ends, return the training results of the small models, and then the cloud server performs the final model integration, which can be expressed as:

[0089]

[0090] Among them is the integrated large model after training, is the integration function.

[0091] Taking this as a cycle, continuously update and iterate the large model, which is expressed as the following process:

[0092] . ​

[0093] The fourth category is the large model training process based on the third category of models. For the planning and obstacle avoidance tasks of agents in specific complex scenarios or the autonomous decision-making tasks for complex environment exploration, combined with the excellent semantic understanding ability of the language large model, the two are combined and the third category of large model training method is used. The planning and decision-making experience of complex link exploration is used as training data to train an AI model for specific exploration planning tasks. It is carried on the cloud server and is used for the exploration needs of different agents in complex scenario tasks. Request the cloud large model to generate corresponding exploration planning instructions and directly apply them to the agent, and then continuously optimize the decision-making ability of the large model according to the actual task performance or simulation task performance of the agent, simplify the task solving process, and change the human-machine interaction in the task solving process to machine-machine interaction. It can be roughly represented as the following process:

[0094]

[0095] Among them, O is the instruction output of the cloud model;

[0096]

[0097] Among them is the feedback after running at the edge, the function is the running function;

[0098]

[0099] Finally, the model upgrades itself through the feedback at the edge.

[0100] The fifth category is for multi-agent collaborative tasks. Each edge computer represents the corresponding agent and uses communication means such as local area network or switch to simulate the distribution of multiple agents in the task and conduct training on the multi-agent distributed communication method:

[0101]

[0102] Then, the above algorithm deployment process quickly deploys the algorithm to the agent to accelerate the training and application of the distributed algorithm used.

[0103] Those skilled in the art should understand that those skilled in the art can realize variations by combining the prior art and the above embodiments. Such variations do not affect the essence of the solution and will not be elaborated here.

[0104] It should be understood that the present solution is not limited to the above specific embodiments, and the devices and structures not described in detail should be understood to be implemented in a common manner in the art; any person skilled in the art can, without departing from the scope of the technical solution of the present solution, make many possible changes and modifications to the technical solution of the present solution by using the methods and technical contents disclosed above, or modify it into an equivalent embodiment with equivalent changes, which does not affect the essence of the present solution. Therefore, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present solution without departing from the content of the technical solution of the present solution still fall within the scope of protection of the technical solution of the present solution.

Claims

1. Distributed collaborative training system for cloud-edge AI model coupling, characterized by: It includes a cloud computing platform and several independent edge computing platforms; the cloud computing platform and the edge computing platform are coupled with each other; the training data of any edge computing platform is distributed and deployed to multiple other edge computing platforms through the cloud computing platform; the edge computing platforms exchange information through the cloud computing platform, and when training multi-agent collaborative tasks in complex environments, algorithm training is performed on the distributed collaborative states that appear in the simulated tasks.

2. The distributed collaborative training system for cloud-edge AI model coupling according to claim 1, characterized in that: The cloud computing platform and the edge computing platform are connected through a switch. The cloud computing platform and the edge computing platform are each equipped with a corresponding display screen. The display screen of the cloud computing platform is used to display the operation status of the entire system, and the display screen of the edge computing platform is used to display the operation status and status of the intelligent body corresponding to the module.

3. The distributed collaborative training system for cloud-edge AI model coupling according to claim 2, characterized in that: The switch is a switch that can seamlessly switch between wired / wireless with a throughput greater than high; the edge computing platform is equipped with multi-modal sensors including cameras, lidars, and rangefinders for collecting training data.

4. A distributed collaborative training method based on the cloud-edge AI model coupling of the system according to any one of claims 1 to 3, characterized in that: There are five training methods, namely, edge agent self-training, moving edge agent training, cloud-edge agent collaborative training, cloud-edge agent collaborative special training, and multi-agent special collaborative training. The edge agent trains itself. The edge agent simulates the training of the algorithms involved. When it comes to the collaborative part, the connection between different edge computing platforms is established through the switch. After the training of each edge computing platform is completed, the data is uploaded to the cloud computing platform for integration, and then the integrated data is deployed to the edge agent; For training of the edge intelligent body of the movement, the edge computing platform of the movement is connected to the cloud computing platform as the remote edge, and then trained in collaboration with the cloud computing platform. After the edge computing platform of the movement collects data, it performs preliminary integration and sends it to the cloud computing platform, which performs model training and then returns the trained results to the edge computing platform of the movement for status output; Cloud-edge intelligent collaborative training: The cloud computing platform modularizes the large model in a specific way, and each module is sent to a different edge computing platform for training. After the edge computing platform completes the training, the training results are sent to the cloud computing platform, which then integrates the model. Specialized training for cloud-edge intelligent agent collaboration: Based on the large-model training process of cloud-edge intelligent agent collaboration, it combines the semantic understanding ability of the language large model with the planning and obstacle avoidance tasks of intelligent agents in specific complex scenarios or the autonomous decision-making tasks for complex environment exploration to carry out training for specific exploration and planning tasks. Multi-agent special collaborative training, for multi-agent collaborative tasks, each edge computing platform represents the corresponding agent, and uses communication methods including local area networks or switches to simulate the distribution of multiple agents in the task, train multi-agent distributed communication methods, and deploy the trained distributed communication methods to the agent group to complete distributed collaboration.

5. The distributed collaborative training method of cloud-edge AI model coupling according to claim 4 is characterized in that: For the planning and obstacle avoidance tasks of intelligent agents in specific complex scenarios and the autonomous decision-making tasks of exploring complex environments, we use the semantic understanding ability of the language big model and collaborative training on the cloud computing platform. We use the planning and decision-making experience of exploring complex links as training data, and train AI to handle specific exploration and planning tasks. The AI ​​is installed on the cloud computing platform to meet the exploration needs of different intelligent agents in complex scenario tasks. We request the cloud big model to generate corresponding exploration planning instructions and directly apply them to the intelligent agent. Then, based on the actual task performance or simulated task performance of the intelligent agent, we transform the human-machine interaction in the task solving process into machine-machine interaction.

6. The distributed collaborative training method for cloud-edge AI model coupling according to claim 5, characterized in that: For the distributed collaboration solution of multiple agents in complex scenarios, an edge computing platform is used to represent an agent, and a mirror is created in the cloud computing platform for rendering the simulation environment and modeling and communication of multiple agents. Each edge computing platform corresponds to an agent in the mirror rendering environment, and uses the same IP and communication group number as the computing module. According to task requirements, the language large model is used to assist in testing, and the image corresponding to the training results is packaged and uploaded locally on the cloud computing platform.

7. The distributed collaborative training method of cloud-edge AI model coupling according to claim 4 is characterized in that: In the collaborative training of cloud-edge agents, a coupling method based on pruning and weighted superposition is proposed to solve the problem that the training results are biased due to the differences in training data of different training tasks assigned by the cloud-edge, which makes it impossible to perform simple superposition fitting. The specific formula is as follows: in, The result of the overall large model is For the The results of the submodules are is the weight of each submodule at the final coupling time, The results calculated by the submodules Subtract redundant or underweight results , that is, pruning operation, weight The number of computing units of the corresponding module In total computing unit The proportion of .

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor of the computer device executes the computer program, it implements a distributed collaborative training method for cloud-edge AI model coupling as described in any one of claims 4 to 7.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, on which a computer program is stored. When the computer-readable instructions are executed by the processor, the distributed collaborative training method for cloud-edge AI model coupling described in any one of claims 4 to 7 is called.

Images