Edge AI collaboration method and system based on cloud big model enhancement

Abstract

The invention relates to an edge AI collaboration method and system based on cloud big model enhancement. The method comprises the steps of analyzing a natural language request through a natural language processing technology and converting the natural language request into sub-tasks; the processing capacity of each edge device is evaluated, and the subtasks are unloaded to the corresponding edge devices in a coordinated mode; edge AI model training is carried out on each edge device in a distributed mode; and performing synchronization mechanism and consistency constraint on each preliminary edge AI model by using a cross-device consistency maintenance algorithm, and aligning model parameters of the preliminary edge AI models among edge devices to complete edge AI collaboration. A natural language processing technology is utilized to deeply analyze a natural language request of a user, the natural language request is accurately converted into a sub-task which can be executed by an edge device, real-time dynamic allocation is carried out, the resource utilization rate can be optimized, and task execution delay is reduced; by performing synchronization mechanism and consistency constraint on the model, the overall performance and generalization ability of the model can be improved.

Description

Technical Field

[0001] The present invention relates to the field of edge artificial intelligence technology, and in particular to an edge AI collaboration method and system based on cloud-based large model enhancement. Background Art

[0002] In the field of edge AI, with the standardization and global deployment of 5G network technology, the world is rapidly moving towards a future network era defined by connected intelligence. In this vision, seamless connectivity between humans, objects, and intelligence is the core goal. Edge AI, as a key technology to achieve this goal, undertakes the task of realizing data collection, processing, transmission, and consumption at the edge of the network. Although edge AI systems have shown great potential in providing high-quality, low-latency, and privacy-protected AI services, existing technologies still face many challenges and shortcomings in achieving the comprehensive goal of connected intelligence.

[0003] Existing edge AI systems are deficient in many aspects. First, in terms of user interface design, traditional methods require users to interact with AI models through applications and be familiar with the functions of each application, which limits the smoothness and intuitiveness of the user experience. Secondly, there is the problem of integrating sensors and AI models. In order to meet user needs, suppliers must integrate sensors and AI models from different providers, which requires tremendous efforts to achieve seamless coordination between multiple sensors and AI models. In addition, the demand for customization of AI models is growing, but existing systems are obviously insufficient in the level of automated organization, adaptation, and optimization of themselves to meet the diverse needs of users, especially in terms of privacy protection. How to ensure user privacy during data training and model customization is an urgent problem that existing edge AI systems need to solve.

[0004] Therefore, traditional edge AI is often relatively single and limited, has high computing resource requirements, and has problems with low collaborative processing flexibility and scalability. Summary of the invention

[0005] Based on this, in order to solve the above technical problems, an edge AI collaboration method and system based on cloud-based large model enhancement is provided, which can reduce the demand for computing resources, improve flexibility and scalability, and is suitable for multiple intelligent system application scenarios.

[0006] An edge AI collaboration method based on cloud-based large model enhancement, the method comprising: Collect natural language requests based on a large cloud model, parse the natural language requests through natural language processing technology, obtain user needs and convert them into various subtasks; Evaluate the processing capability of each edge device, and coordinate and offload each of the subtasks to the corresponding edge devices according to the processing capability; The cloud-based large model is used to automatically generate the code required for joint learning, and the edge AI model training is performed on each edge device in a distributed manner based on the code required for joint learning to obtain each preliminary edge AI model; Using a cross-device consistency maintenance algorithm to perform synchronization mechanism and consistency constraints on each of the preliminary edge AI models, aligning model parameters of the preliminary edge AI model between each of the edge devices, and obtaining a target edge AI model; Complete edge AI collaboration of each of the edge devices based on each of the target edge AI models.

[0007] In one embodiment, the natural language request is parsed by natural language processing technology to obtain user requirements and convert them into various subtasks, including: Input the natural language request into a pre-trained cloud-based large model, calculate the score of the next word according to the natural language request through natural language processing technology, and perform word prediction based on the score to obtain a word prediction result; Determine user behavior according to the word prediction result, extract entity information in the natural language request, and use the user behavior and entity information as user demand; The user requirements are converted into various subtasks through a conversion function.

[0008] In one embodiment, evaluating the processing capability of each edge device includes: Determine evaluation factors for each edge device; the evaluation factors include processor speed, memory capacity, storage space, and network connection of the device; Determining a computing capability indicator of the edge device according to the evaluation factor; The computing capability indicator is converted into a vector representation.

[0009] In one embodiment, the coordination of offloading each of the subtasks to corresponding edge devices according to the processing capability includes: Convert the natural language request into a structured request representation by using natural language processing technology to obtain a user request; Using a matching function to match the user request with the processing capability of each edge device to obtain a matching result; Setting a task offloading strategy based on the matching result; The subtasks are dynamically coordinated and offloaded to corresponding edge devices according to the task offloading strategy.

[0010] In one embodiment, the cloud-based large model is used to automatically generate the code required for joint learning, including: Obtaining data distribution on each of the edge devices and obtaining hyperparameters of the joint learning algorithm; According to the data distribution and hyperparameters, use a code generation function to generate the code required for the federated learning; Among them, differential privacy technology is used to achieve privacy protection in the joint learning process.

[0011] In one embodiment, edge AI model training is performed on each edge device in a distributed manner based on the code required for the joint learning to obtain each preliminary edge AI model, including: Determine the number of edge devices, and calculate the allocation weight according to the data volume of the natural language request or the processing capacity; Design a distributed architecture according to the number of edge devices and the allocation weights, and each edge device works collaboratively based on the distributed architecture; Using the code required for the joint learning, edge AI model training is performed on each edge device on the distributed architecture, and the parameters in each edge AI model are adjusted to obtain each preliminary edge AI model.

[0012] In one embodiment, a cross-device consistency maintenance algorithm is used to synchronize the preliminary edge AI models and implement consistency constraints, including: Obtaining each parameter of each of the preliminary edge AI models on each of the edge devices, determining a global average model parameter of each of the preliminary edge AI models based on each of the parameters, and obtaining differences between each of the parameters; Compare each of the parameters with the global average model parameters respectively, and update the parameters to complete parameter synchronization; Based on the differences of each of the parameters and the global average model parameters, consistency constraints are determined, and consistency constraints are performed on each of the preliminary edge AI models according to the consistency constraints.

[0013] In one embodiment, the method further comprises: Extracting each of the parameters, performing weighted averaging on each of the parameters, and obtaining a target parameter; The global average model parameters are updated based on the target parameters.

[0014] An edge AI collaborative system based on cloud-based large model enhancement, the system comprising: The user intention understanding module is used to collect natural language requests based on the cloud-based large model, parse the natural language requests through natural language processing technology, obtain user needs and convert them into various subtasks; An intelligent coordination and task offloading module is used to evaluate the processing capability of each edge device and coordinate and offload each of the subtasks to the corresponding edge devices according to the processing capability; An automated joint learning module, used to automatically generate the code required for joint learning using the cloud-based large model, and to perform edge AI model training on each of the edge devices in a distributed manner based on the code required for joint learning to obtain each preliminary edge AI model; A cross-device consistency maintenance module, used to use a cross-device consistency maintenance algorithm to perform synchronization mechanisms and consistency constraints on each of the preliminary edge AI models, align model parameters of the preliminary edge AI model between each of the edge devices, and obtain a target edge AI model; The edge AI collaboration module is used to complete the edge AI collaboration of each of the edge devices based on each of the target edge AI models.

[0015] The above-mentioned edge AI collaborative method and system based on cloud-based large model enhancement, by using natural language processing technology to deeply analyze the user's natural language requests, accurately convert them into subtasks executable by edge devices, and dynamically allocate them in real time, can optimize resource utilization and reduce task execution delays; by training models on edge devices in a distributed manner, the joint training and knowledge sharing of models can be achieved; by synchronizing the model with a mechanism and consistency constraints, the effective alignment of model parameters between different devices can be ensured, thereby improving the overall performance and generalization ability of the model, and improving flexibility and scalability, which is suitable for multiple intelligent system application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is an application environment diagram of an edge AI collaboration method based on cloud big model enhancement in one embodiment; Figure 2 A flowchart of an edge AI collaboration method based on cloud-based large model enhancement in one embodiment; Figure 3 It is a structural block diagram of an edge AI collaborative system based on cloud big model enhancement in one embodiment; Figure 4 It is a structural block diagram of an edge AI collaborative system based on cloud big model enhancement in another embodiment; Figure 5 A schematic diagram of the timing of sending and receiving signaling between modules in one embodiment; Figure 6 A schematic diagram of the structure of a user intention understanding module in one embodiment; Figure 7 A schematic diagram of the structure of an intelligent coordination and task offloading module in one embodiment; Figure 8 A schematic diagram of the structure of an automated joint learning module in one embodiment; Fig. 9 A schematic diagram of the structure of a cross-device consistency maintenance module in one embodiment; Fig.10 A schematic diagram of a process applied to the field of health monitoring in an embodiment; Fig.11 A schematic diagram of a process applied to the field of intelligent education in an embodiment; Fig.12 A schematic diagram of a process applied to the field of industrial automation in an embodiment; Fig.13 A schematic diagram of a process applied to the field of smart home in an embodiment; Fig.14 1 is a diagram of the internal structure of a cloud device in one embodiment. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0018] The edge AI collaboration method based on cloud-based large model enhancement provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Figure 1As shown, the application environment includes a cloud device 110 and each edge device 120, and the cloud device 110 is connected and communicated with each edge device 120 respectively. The cloud device 110 can collect natural language requests based on the cloud big model, parse the natural language requests through natural language processing technology, obtain user needs and convert them into various subtasks; the cloud device 110 can evaluate the processing capabilities of each edge device 120, and coordinate and unload each subtask to the corresponding edge device 120 according to the processing capabilities; the cloud device 110 can use the cloud big model to automatically generate the code required for joint learning, and perform edge AI model training on each edge device 120 in a distributed manner based on the code required for joint learning to obtain each preliminary edge AI model; the cloud device 110 can use the cross-device consistency maintenance algorithm to synchronize the mechanism and consistency constraints of each preliminary edge AI model, align the model parameters of the preliminary edge AI model between each edge device, and obtain the target edge AI model; the cloud device 110 can complete the edge AI collaboration of each edge device 120 based on each target edge AI model. The cloud device 110 may be, but is not limited to, various personal computers, laptop computers, smart phones, robots, unmanned aerial vehicles and other devices.

[0019] In one embodiment, Figure 2 As shown, an edge AI collaboration method based on cloud large model enhancement is provided, including the following steps: Step 202: Collect natural language requests based on the cloud-based big model, parse the natural language requests through natural language processing technology, obtain user needs and convert them into various subtasks.

[0020] The cloud device can rely on a powerful large model to parse the user's natural language request. Specifically, the cloud device can accurately understand the user's needs through advanced natural language processing technology, and convert these needs into subtasks that can be executed by the edge device. Among them, the large model can be a GPT model, which deeply analyzes the user's natural language request to achieve accurate grasp and rapid response to user needs. In this embodiment, this process not only improves the processing speed of user requests, but also enhances the responsiveness and flexibility of the system.

[0021] In one embodiment, a provided edge AI collaboration method based on cloud-based large model enhancement may also include a process of understanding user intent and dividing subtasks. The specific process includes: inputting a natural language request into a pre-trained cloud-based large model, calculating the score of the next word based on the natural language request through natural language processing technology, and performing word prediction based on the score to obtain a word prediction result; determining user behavior based on the word prediction result, extracting entity information from the natural language request, and using user behavior and entity information as user needs; and converting user needs into subtasks through conversion functions.

[0022] In this embodiment, the user's intention can be accurately understood by using advanced natural language processing technology. Specifically, complex patterns in language can be captured by pre-trained large models. Large models are usually trained on large-scale text data sets to learn the grammar, semantics, and contextual information of the language. For example, given a user's natural language request x, the model can predict the next most likely word or phrase y. This process can be expressed by the formula: ;in, represents the scoring function of the model, which calculates the score of the next word l based on the input context x; V is the size of the vocabulary.

[0023] The core of understanding user intent is to parse and convert user requests. In this embodiment, the user's natural language request can be first decomposed into a series of subtasks. This process involves two main steps: intent recognition and slot filling. The purpose of intent recognition is to determine the main behavior of the user request, while slot filling is to extract key information from the request. For example, assuming that the user request can be expressed as , where w i is the i-th word in the request, then the formula for obtaining the intent I and the slot set S can be expressed as: ; where i represents the possible intent category, s j Represents the entity slot, v j Represents slots j The value of k represents the slot s j The index of .

[0024] In this embodiment, converting user requests into subtasks is the key to achieving fast response. Each subtask is a specific operation that can be performed by the edge device. The generation process of subtasks can be expressed as: ;in, It is a transformation function that generates the corresponding subtask T according to the recognized intent I and slot information S.

[0025] The generated subtasks need to be further optimized and scheduled to ensure that they can be efficiently executed on edge devices. The optimization process takes into account the dependencies, execution order, and resource requirements of the tasks. The scheduling strategy can be expressed as: ;in, is the sequence of subtasks after scheduling, , and They represent the resource consumption cost (power consumption), task execution time and dependency (time required for the predecessor task) of subtask T respectively. , and is the weight parameter.

[0026] Step 204 , evaluating the processing capability of each edge device, and coordinating and offloading each subtask to each corresponding edge device according to the processing capability.

[0027] Cloud devices can intelligently coordinate edge AI models based on user requests and edge device capabilities, and offload tasks to the most appropriate device. By evaluating the computing power, storage capacity, and network status of edge devices, tasks can be dynamically allocated to achieve efficient use of edge computing resources.

[0028] Specifically, in one embodiment, a cloud-based large model-enhanced edge AI collaboration method may also include a process for evaluating edge device processing capabilities, the specific process including: determining evaluation factors for each edge device; the evaluation factors include the device's processor speed, memory capacity, storage space, and network connection; determining the computing power index of the edge device based on the evaluation factors; and converting the computing power index into a vector representation.

[0029] Before intelligent coordination and task offloading, cloud devices first need to evaluate the computing power of edge devices, which can be achieved by considering factors such as processor speed, memory capacity, storage space, and network connection of edge devices. The computing power of edge devices can be expressed as a vector , where C ei Indicates the computing capability indicator of the i-th device.

[0030] In this embodiment, user request analysis aims to understand user needs and convert them into tasks that can be executed by edge devices. This step can be achieved through natural language processing technology to convert the user's natural language request into a structured request representation: , where r j Represents an aspect of a user request.

[0031] In one embodiment, a provided edge AI collaboration method based on cloud-based large model enhancement may also include a process of coordinated task offloading, the specific process including: converting a natural language request into a structured request representation through natural language processing technology to obtain a user request; using a matching function to match the user request with the processing capabilities of each edge device to obtain a matching result; setting a task offloading strategy based on the matching result; and dynamically coordinating and offloading each subtask to the corresponding edge devices according to the task offloading strategy.

[0032] The task-device matching process aims to find the most suitable device for each subtask. This process can be achieved by defining a matching function M(R, C e ) is implemented by comparing user requests with device capabilities and outputting the optimal device allocation solution.

[0033] The task offloading strategy needs to consider the execution efficiency, energy consumption and network communication cost of the task. The optimization problem of task offloading can be expressed as: ; Where D is a matrix representing the decision of task offloading, d ij The decision variable for the i-th device to perform the j-th subtask, c ij is the execution cost, t ij is the execution time, is the communication cost weight, S commcost (D) is the communication cost associated with the offloading decision.

[0034] Dynamic resource allocation is to adjust resource allocation according to the current task offloading decision and device status. It can be achieved by defining a resource allocation function A(D, C e ) to achieve this by solving the optimal resource allocation problem Implement dynamic resource allocation, where Represents the Frobenius norm, which is used to measure the degree of match between resource allocation and device capabilities. This problem can be solved using optimization algorithms such as Lagrange multiplier method and linear programming to dynamically adjust the allocation of CPU, memory and storage resources.

[0035] Step 206, using the cloud-based large model to automatically generate the code required for joint learning, and based on the code required for joint learning, edge AI model training is performed on each edge device in a distributed manner to obtain each preliminary edge AI model.

[0036] Without sharing user data, cloud devices can use large cloud models to automatically generate the code required for federated learning, and train key parts of AI models on edge devices in a distributed manner, which can achieve joint model training and knowledge sharing while protecting user privacy.

[0037] In one embodiment, a provided edge AI collaboration method based on cloud large model enhancement may also include a code generation process, the specific process including: obtaining the data distribution on each edge device and obtaining the hyperparameters of the joint learning algorithm; according to the data distribution and hyperparameters, using the code generation function to generate the code required for joint learning; wherein, differential privacy technology is used to achieve privacy protection in the joint learning process.

[0038] Among them, the cloud-based large model is responsible for automatically generating the code required for joint learning. This process can be expressed by the formula: ; where C is the generated code, is the code generation function, G is the global model, is the data distribution on edge devices, is a hyperparameter of the federated learning algorithm.

[0039] In one embodiment, a provided edge AI collaboration method based on cloud-based large model enhancement may also include an edge model training process, the specific process including: determining the number of edge devices, and calculating the allocation weight based on the amount of data or processing power requested by natural language; designing a distributed architecture based on the number of edge devices and the allocation weight, and each edge device works collaboratively based on the distributed architecture; using the code required for joint learning, training the edge AI model of each edge device on the distributed architecture, and adjusting the parameters in each edge AI model to obtain each preliminary edge AI model.

[0040] In this embodiment, a distributed architecture is adopted to allow multiple edge devices to work together while keeping data local. The distributed architecture can be expressed as a variant of the federated averaging algorithm, where the global model G is periodically updated to all edge device models {L i The weighted average of} is: ; where K is the number of participating edge devices, w i It is a weight assigned according to the data volume or computing power of the device.

[0041] In order to protect user privacy, differential privacy technology is used in this embodiment to ensure that sensitive information of any single user will not be leaked during the joint learning process. Differential privacy can be achieved by adding noise, which can be expressed as: ;in, It's the privacy budget. The function limits the size of the gradient, is the model loss on edge device i.

[0042] In the process of automated federated learning, the model on the edge device needs to be fine-tuned based on local data. The fine-tuning process can be achieved through stochastic gradient descent, which can be expressed as: ;in, is the learning rate, is the model loss on edge device i, is the dataset on device i.

[0043] Step 208, use a cross-device consistency maintenance algorithm to perform synchronization mechanisms and consistency constraints on each preliminary edge AI model, align the model parameters of the preliminary edge AI model between each edge device, and obtain the target edge AI model.

[0044] In order to ensure the consistency of training across different edge devices, a cross-device consistency maintenance algorithm is introduced in this embodiment. Through synchronization mechanism and consistency constraints, the effective alignment of model parameters between different devices is ensured, thereby improving the overall performance and generalization ability of the model.

[0045] In one embodiment, an edge AI collaboration method based on cloud large model enhancement is provided, which may also include a process for performing a synchronization mechanism and consistency constraints. The specific process includes: obtaining each parameter of each preliminary edge AI model on each edge device, determining the global average model parameters of each preliminary edge AI model based on each parameter, and obtaining the difference of each parameter; comparing each parameter with the global average model parameter respectively, and updating the parameters to complete parameter synchronization; determining the consistency constraints based on the differences of each parameter and the global average model parameters, and performing consistency constraints on each preliminary edge AI model according to the consistency constraints.

[0046] In order to maintain consistency across devices, in this embodiment, it is necessary to first define consistency constraints, which can be achieved by limiting the differences in model parameters on different devices. The consistency constraints can be mathematically expressed as: ;in, is the consistency loss, W i is the model parameter on the i-th device, is the global average model parameter, is the Frobenius norm, is the weight coefficient.

[0047] In this embodiment, parameter synchronization is a key step to achieve consistency. After each round of training, the model parameters on each device need to be compared with the global model parameters and updated appropriately. The synchronization process can be expressed as: ;in, and are the model parameters of the i-th device at time t+1 and time t, respectively. is the learning rate, is the local loss on the ith device.

[0048] In one embodiment, an edge AI collaboration method based on cloud-based large model enhancement is provided, which may also include a process for updating global model parameters. The specific process includes: extracting each parameter, performing weighted averaging on each parameter to obtain a target parameter; and updating the global average model parameter based on the target parameter.

[0049] Global model parameters It is necessary to update the model parameters on all devices, which can be achieved by weighted averaging, as shown in the following figure: ;in, It is the weight assigned according to the data volume or computing power of device i.

[0050] Step 210: Complete edge AI collaboration of each edge device based on each target edge AI model.

[0051] In one embodiment, the provided edge AI collaboration method based on cloud-based large model enhancement may also include a process of optimizing communication overhead. In order to reduce communication overhead, compressed communication technology may be used to reduce the size of model parameters that need to be synchronized. In addition, synchronization is performed only when the change in model parameters exceeds a certain threshold, which can be specifically expressed as: ;in, is a Boolean function indicating whether to synchronize. is the synchronization threshold.

[0052] This application provides an edge AI collaboration method based on cloud-based large model enhancement, which has achieved significant technological breakthroughs and performance improvements in the field of edge computing and artificial intelligence. Specifically, the cloud-based large model assists in understanding user intent and uses advanced natural language processing technology to significantly improve the ability to parse user natural language requests. In specific experiments, on the public SQuAD dataset, the method of this application achieved an accuracy of 90% in the intent recognition task, which is about 5% higher than the existing technology; this application uses intelligent coordination and task offloading of edge AI models to dynamically allocate tasks according to the real-time status of edge devices, effectively improving the utilization rate of edge computing resources; without sharing user data, the joint training and knowledge sharing of the model are realized through the automated joint learning module; the synchronization mechanism and consistency constraints ensure the effective alignment of model parameters between different devices. Compared with the method without consistency maintenance, while maintaining the model performance, the performance loss caused by inconsistent parameters between devices is reduced. The experimental results on the GSM8K dataset show that the performance loss is reduced by 10%; the synchronization mechanism and consistency constraints ensure the effective alignment of model parameters between different devices, thereby improving the overall performance and generalization ability of the model; when performing AI tasks, edge devices can more efficiently utilize local and cloud resources, reducing dependence on single-point high-performance computing resources; it supports rapid adaptation to new application scenarios and demand changes. The introduction of large cloud models provides the system with powerful language understanding and task planning capabilities, enabling the system to flexibly respond to diverse user requests. At the same time, edge AI The model's intelligent coordination and task offloading ensure the scalability of the system, and demonstrate excellent adaptability and stability in scenarios such as smart homes and smart cities.

[0053] Through the above innovations, this application not only solves some key problems in the prior art, but also provides an efficient, secure and user-friendly edge AI system solution, which has important practical application value and broad market prospects.

[0054] It should be understood that, although the various steps in the above-mentioned flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned flow chart may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0055] In one embodiment, Figure 3 As shown, an edge AI collaboration system based on cloud large model enhancement is provided, including: a user intention understanding module 310, an intelligent coordination and task offloading module 320, an automated joint learning module 330, a cross-device consistency maintenance module 340 and an edge AI collaboration module 350, wherein: The user intention understanding module 310 is used to collect natural language requests based on the cloud-based large model, parse the natural language requests through natural language processing technology, obtain user needs and convert them into various subtasks; The intelligent coordination and task offloading module 320 is used to evaluate the processing capability of each edge device and coordinate and offload each subtask to each corresponding edge device according to the processing capability; The automated joint learning module 330 is used to automatically generate the code required for joint learning by using the cloud large model, and to perform edge AI model training on each edge device in a distributed manner based on the code required for joint learning to obtain each preliminary edge AI model; A cross-device consistency maintenance module 340 is used to use a cross-device consistency maintenance algorithm to perform synchronization mechanisms and consistency constraints on each preliminary edge AI model, align model parameters of the preliminary edge AI model between each edge device, and obtain a target edge AI model; The edge AI collaboration module 350 is used to complete the edge AI collaboration of each edge device based on each target edge AI model.

[0056] In one embodiment, an edge AI collaborative system based on cloud-based large model enhancement can also be provided as follows: Figure 4 As shown, it includes: user interface, cloud large model processing unit, edge AI model coordination unit, privacy protection joint learning unit, cross-device consistency maintenance unit, and task execution unit, among which: The user interface allows users to submit requests in natural language and is the front-end interface for interaction between the system and users. The cloud-based large model processing unit receives requests submitted by the user interface and uses the cloud-based large model for preliminary processing and understanding of user intent. The edge AI model coordination unit receives the results of the cloud-based processing unit and intelligently coordinates task offloading based on the capabilities of the edge device and user requests. The privacy-preserving joint learning unit is responsible for performing automated joint learning on edge devices while ensuring that the privacy of user data is protected. The cross-device consistency maintenance unit ensures that model parameters trained on different edge devices remain consistent, improving the overall performance of the model. The task execution unit performs specific AI tasks on edge devices, such as image recognition, language processing, etc.

[0057] exist Figure 4In the figure, arrows are used to represent data flow and information transmission in each part. The output of the user interface points to the cloud-based large model processing unit through an arrow, indicating the transmission of user requests; the output of the cloud-based large model processing unit is passed to the edge AI model coordination unit through an arrow, indicating the transmission of user intent understanding results; the output of the edge AI model coordination unit is passed to the privacy-preserving joint learning unit and the cross-device consistency maintenance unit through an arrow, indicating the transmission of task offloading and coordination decisions; the output of the privacy-preserving joint learning unit and the cross-device consistency maintenance unit is passed to the task execution unit through an arrow, indicating the transmission of fine-tuned models and evaluation results.

[0058] In one embodiment, in the end-cloud collaborative learning environment, the timing of sending and receiving signaling between modules is as follows: Figure 5 As shown, it specifically includes: cloud module, edge module, and automated evaluation module, among which: The cloud module is responsible for generating and managing signaling, including task allocation and performance feedback; the edge module is responsible for receiving signaling from the cloud and performing corresponding tasks based on the signaling; the automated evaluation module is responsible for collecting the execution results of the edge module and feeding them back to the cloud module for performance evaluation.

[0059] exist Figure 5 In the figure, the cloud module sends a signal to the edge module through the "send task" arrow. After receiving the signal, the edge module feeds back the signal to the cloud module through the "task confirmation" arrow. After the task is completed, the edge module sends the result to the automation evaluation module through the "execution result" arrow. The automation evaluation module then feeds back the evaluation result to the cloud module through the "performance feedback" arrow.

[0060] In one embodiment, the user intent understanding module 310 is also used to: input the natural language request into a pre-trained cloud-based large model, calculate the score of the next word according to the natural language request through natural language processing technology, and perform word prediction based on the score to obtain a word prediction result; determine user behavior based on the word prediction result, extract entity information in the natural language request, and use the user behavior and entity information as user needs; convert user needs into various subtasks through conversion functions.

[0061] The structural diagram of the user intention understanding module is as follows Figure 6 As shown in Figure 1, the user intent understanding module relies on a large model (such as GPT) deployed in the cloud, which is responsible for parsing the user's natural language requests and converting these requests into subtasks that can be executed by edge devices. Figure 6 As shown in the figure, the user intention understanding module includes: natural language processing technology, intention recognition and slot filling, subtask generation, subtask optimization and scheduling, module collaboration and performance evaluation, among which: Natural language processing technology is responsible for using advanced natural language processing technology to understand users' natural language requests; intent recognition and slot filling are used to identify the main intent and key information from users' requests; subtask generation generates subtasks executable by edge devices based on the identified intent and slot information; subtask optimization and scheduling optimizes and schedules the generated subtasks to ensure that the tasks can be efficiently executed on edge devices; module collaboration and performance evaluation is responsible for evaluating the performance of the entire module and providing feedback to optimize the collaborative work of the modules.

[0062] exist Figure 6 In the example, the output of natural language processing technology is directed to intent recognition and slot filling through "understanding user intent", indicating that the understanding result of user requests is passed to the intent recognition module. The output of intent recognition and slot filling is directed to subtask generation through "generating subtasks", indicating that the recognized intent and slot information are used to generate specific subtasks. The output of subtask generation is directed to subtask optimization and scheduling through "task optimization and scheduling", indicating that the subtask is passed to the optimization and scheduling module for further processing. The output of subtask optimization and scheduling is directed to the edge device through "executing subtasks", indicating that the optimized subtask is sent to the edge device for execution. Module collaboration and performance evaluation play a monitoring and evaluation role in the whole process, and its results are fed back to natural language processing technology to optimize the performance of the understanding module.

[0063] In one embodiment, the intelligent coordination and task offloading module 320 is also used to determine the evaluation factors of each edge device; the evaluation factors include the device's processor speed, memory capacity, storage space, and network connection; determine the computing power index of the edge device based on the evaluation factors; and convert the computing power index into a vector representation.

[0064] In one embodiment, the intelligent coordination and task offloading module 320 is also used to convert natural language requests into structured request representations through natural language processing technology to obtain user requests; use matching functions to match user requests with the processing capabilities of each edge device to obtain matching results; set a task offloading strategy based on the matching results; and dynamically coordinate and offload each subtask to the corresponding edge devices according to the task offloading strategy.

[0065] The structural diagram of the intelligent coordination and task offloading module is as follows: Figure 7 As shown in Figure 1, the intelligent coordination and task offloading module is responsible for intelligently coordinating the edge AI model and offloading tasks to the most appropriate device based on user requests and the capabilities of edge devices. Figure 7 As shown in Figure 1, the intelligent coordination and task offloading module includes: device capability evaluation, user request analysis, task and device matching, task offloading strategy, dynamic resource allocation, task execution monitoring, performance feedback and optimization, among which: Device capability assessment is responsible for evaluating the computing power, storage capacity and network status of edge devices to provide a basis for task allocation; user request analysis is used to understand user needs and convert them into tasks executable by edge devices; task and device matching intelligently matches tasks and devices based on user requests and device capabilities; task offloading strategies make task offloading decisions, taking into account execution efficiency, energy consumption and network communication costs; dynamic resource allocation dynamically adjusts resource allocation based on task offloading decisions and device status; task execution monitoring monitors the task execution status in real time and triggers rescheduling or resource reallocation when necessary; performance feedback and optimization provide performance feedback based on monitoring results to optimize device capability assessment and task offloading strategies.

[0066] exist Figure 7 In the example, the output of device capability evaluation points to task-device matching through "evaluation result", indicating that the result of device capability evaluation is used for task-device matching. The output of user request analysis points to task-device matching through "user request analysis", indicating that the analyzed user request is used for task matching. The output of task-device matching points to task offloading strategy through "matching result", indicating that the result after task-device matching is passed to the task offloading strategy module. The output of task offloading strategy points to dynamic resource allocation through "offloading decision", indicating that the result of task offloading strategy is passed to the dynamic resource allocation module. The output of dynamic resource allocation points to task execution monitoring through "resource allocation", indicating that the result after resource allocation is used for task execution monitoring. The output of task execution monitoring points to performance feedback and optimization through "execution status", indicating that the result of task execution monitoring is provided to the performance feedback and optimization module. The output of performance feedback and optimization is fed back to device capability evaluation and task offloading strategy through "optimization instruction", forming a closed-loop optimization process.

[0067] In one embodiment, the automated federated learning module 330 is also used to obtain the data distribution on each edge device and obtain the hyperparameters of the federated learning algorithm; based on the data distribution and hyperparameters, a code generation function is used to generate the code required for federated learning; wherein, differential privacy technology is used to achieve privacy protection during the federated learning process.

[0068] In one embodiment, the automated federated learning module 330 is also used to determine the number of edge devices, and calculate the allocation weights based on the amount of data or processing power of the natural language request; design a distributed architecture based on the number of edge devices and the allocation weights, and each edge device works together based on the distributed architecture; use the code required for federated learning to train the edge AI model for each edge device on the distributed architecture, and adjust the parameters in each edge AI model to obtain each preliminary edge AI model.

[0069] The structural diagram of the automated joint learning module is as follows: Figure 8As shown in the figure, the automated federated learning module uses the cloud-based large model to automatically generate the code required for federated learning, and trains the AI ​​model on edge devices in a distributed manner while protecting user privacy. Figure 8 As shown in the figure, the automated federated learning module includes: federated learning architecture design, data privacy protection, automatic code generation, model fine-tuning, cross-device consistency maintenance, and model update, among which: The federated learning architecture design adopts a distributed architecture, allowing multiple edge devices to work together while keeping data localized; data privacy protection introduces differential privacy technology to protect user data by adding noise to prevent the leakage of sensitive information during the federated learning process; automatic code generation The cloud-based large model automatically generates the code required for federated learning based on the global model, data distribution, and algorithm hyperparameters; model fine-tuning The model on the edge device is fine-tuned based on local data to adapt to specific data sets and tasks; cross-device consistency maintenance ensures the effective alignment of model parameters on different devices through synchronization mechanisms and consistency constraints; model update generates a new global model based on the global model and the updated model on the edge device to achieve continuous improvement and adaptability of the model.

[0070] exist Figure 8 In the example, the output of the joint learning architecture design is directed to automatic code generation through "distributed architecture design", indicating that the architecture design is used to guide automatic code generation. The output of data privacy protection is directed to model fine-tuning through "privacy protection", indicating that privacy protection needs to be considered during model fine-tuning. The output of automatic code generation is directed to model fine-tuning through "generated code", indicating that the automatically generated code is used for model fine-tuning. The output of model fine-tuning is directed to cross-device consistency maintenance through "model update", indicating that the fine-tuned model is used for cross-device consistency maintenance. The output of cross-device consistency maintenance is directed to model update through "consistency model", indicating that the model after consistency maintenance is used for model update. The output of model update is fed back to the joint learning architecture design and automatic code generation, forming a closed-loop optimization process.

[0071] In one embodiment, the cross-device consistency maintenance module 340 is also used to obtain various parameters of each preliminary edge AI model on each edge device, determine the global average model parameters of each preliminary edge AI model based on the various parameters, and obtain the differences of each parameter; compare each parameter with the global average model parameter, and update the parameters to complete parameter synchronization; determine the consistency constraints based on the differences of each parameter and the global average model parameters, and perform consistency constraints on each preliminary edge AI model according to the consistency constraints.

[0072] In one embodiment, the cross-device consistency maintenance module 340 is further used to extract various parameters, perform weighted average on the various parameters, and obtain target parameters; and update the global average model parameters based on the target parameters.

[0073] The structural diagram of the cross-device consistency maintenance module is as follows: Fig. 9 As shown in Figure 1, the cross-device consistency maintenance module ensures that the model parameters trained on different edge devices remain consistent, improving the overall performance and generalization ability of the model. Fig. 9 As shown, the cross-device consistency maintenance module includes: consistency constraint definition, parameter synchronization mechanism, global model parameter update, communication overhead optimization, dynamic consistency adjustment, performance feedback and optimization, among which: The consistency constraint definition is responsible for defining consistency constraints across devices and limiting the differences in model parameters on different devices; the parameter synchronization mechanism implements parameter synchronization to ensure that the model parameters on each device are consistent with the global model parameters; the global model parameter update updates the global model parameters based on the model parameters on all devices; the communication overhead optimization uses compressed communication technology to reduce the size of model parameters that need to be synchronized and reduce communication overhead; dynamic consistency adjustment dynamically adjusts the strength of consistency constraints based on real-time performance feedback; performance feedback and optimization collects performance data and provides feedback to optimize consistency maintenance strategies.

[0074] exist Fig. 9 In the example, the output of the consistency constraint definition is directed to the parameter synchronization mechanism through "Define consistency constraints", indicating that the consistency constraints are used to guide the parameter synchronization mechanism. The output of the parameter synchronization mechanism is directed to the global model parameter update through "Synchronize parameters", indicating that the synchronized parameters are used to update the global model parameters. The output of the global model parameter update is directed to the communication overhead optimization through "Update global model", indicating that the communication overhead needs to be optimized after the global model is updated. The output of the communication overhead optimization is directed to the dynamic consistency adjustment through "Reduce communication overhead", indicating that the consistency can be adjusted dynamically after communication optimization. The output of the dynamic consistency adjustment is directed to the performance feedback and optimization through "Adjust consistency", indicating that the performance feedback after dynamic adjustment is used to further optimize the consistency maintenance strategy. The output of the performance feedback and optimization is fed back to the consistency constraint definition, forming a closed-loop optimization process.

[0075] In one embodiment, the provided edge AI collaboration method and system based on cloud-based large model enhancement can be applied to the field of health monitoring to improve the efficiency of disease prevention and management through intelligent analysis and early warning systems. Fig.10 The specific implementation process is as follows: Health data collection: collect users’ physiological data, such as heart rate, blood pressure, activity level, etc., through wearable devices and medical sensors; Data preprocessing and privacy protection: Preprocess the collected health data and use differential privacy technology to protect user privacy; User intention understanding: The user intention understanding module assisted by a large cloud model analyzes the user's health monitoring needs; Task allocation and model coordination: The intelligent coordination and task offloading module allocates health analysis tasks based on user needs and device capabilities; Automated federated learning: Distributed model training on local devices while ensuring data privacy; Cross-device consistency maintenance: ensure that model parameters trained on different devices remain consistent; Model fine-tuning and performance evaluation: The model evaluation module optimizes model performance based on monitoring results; Health warning and user feedback: The system provides health warnings based on analysis results and receives user feedback for model iteration.

[0076] In one embodiment, the provided edge AI collaboration method and system based on cloud-based large model enhancement can be applied to the field of intelligent education to improve learning efficiency and education quality through a personalized learning path recommendation system. Fig.11 The specific implementation process is as follows: Learning data collection: Collect students’ learning behavior data through the intelligent education platform, such as video viewing progress, homework completion, test scores, etc. Data preprocessing and analysis: Perform preprocessing operations such as cleaning and normalization on the collected learning data, and analyze students' learning habits and abilities; User intent understanding: The user intent understanding module assisted by the cloud-based big model analyzes students’ learning goals and needs; Learning task generation: intelligently generate personalized learning tasks and learning plans based on students’ learning goals; Learning resource recommendation: The intelligent coordination and task offloading module recommends suitable learning resources and tools based on learning tasks; Personalized learning strategy development: Develop personalized learning strategies on local devices while protecting student data privacy; Learning effect monitoring and evaluation: The cross-device consistency maintenance module monitors learning effects and evaluates the effectiveness of learning strategies; Model fine-tuning and optimization: The model evaluation module continuously fine-tunes the learning recommendation system based on learning effect data; Learning feedback collection and system iteration: Collect students’ feedback and learning outcomes, and continuously iterate and optimize the education model.

[0077] In one embodiment, the provided edge AI collaboration method and system based on cloud-based large model enhancement can be applied to the field of industrial automation to improve production efficiency and equipment operation stability through intelligent monitoring and maintenance systems. Fig.12 The specific implementation process is as follows: Industrial data collection: Collect data on the production line, such as temperature, pressure, vibration, etc., through industrial sensors and monitoring equipment; Data preprocessing: preprocess the collected industrial data, including denoising and standardization, to improve the accuracy of data analysis; User intent understanding: The user intent understanding module assisted by the cloud-based large model analyzes the operator's instructions and monitoring requirements; Task allocation and model coordination: The intelligent coordination and task offloading module allocates analysis tasks according to monitoring requirements and the capabilities of edge devices; Automated federated learning: Distributed model training on edge devices while ensuring the privacy of production data; Cross-device consistency maintenance: Ensure that model parameters trained on different devices remain consistent to support collaborative work; Model fine-tuning and performance evaluation: The model evaluation module optimizes model performance based on monitoring results and evaluates device status; Equipment maintenance warning: The system provides equipment maintenance warnings based on analysis results and iterates the model based on operator feedback.

[0078] In one embodiment, the provided edge AI collaboration method and system based on cloud-based large model enhancement can be applied to the field of smart homes to improve living comfort and energy efficiency through intelligent environmental control systems. Fig.13 The specific implementation process is as follows: Home environment data collection: Collect environmental data such as temperature, humidity, light intensity, etc. through smart home sensors; Data preprocessing: preprocess the collected environmental data to ensure data quality; User intention understanding: The user intention understanding module assisted by the cloud-based large model analyzes the user's living preferences and control instructions; Task allocation and model coordination: The intelligent coordination and task offloading module allocates control tasks based on user preferences and home device status; Automated federated learning: Distributed model training on local devices while protecting user privacy; Cross-device consistency maintenance: Ensure that the environmental control policies on different devices are consistent; Model fine-tuning and performance evaluation: The model evaluation module adjusts the control strategy according to changes in the living environment; Environmental control execution: The system automatically adjusts the home environment according to the optimized control strategy, such as adjusting the temperature, opening and closing curtains, etc.

[0079] In one embodiment, a cloud device is provided, whose internal structure diagram can be as follows: Fig.14 As shown. The cloud device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the cloud device is used to provide computing and control capabilities. The memory of the cloud device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the cloud device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, an edge AI collaboration method based on cloud large model enhancement is implemented. The display screen of the cloud device can be a liquid crystal display or an electronic ink display screen, and the input device of the cloud device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the shell of the cloud device, or an external keyboard, touchpad or mouse, etc.

[0080] Those skilled in the art will understand that Fig.14 The structure shown in the figure is only a block diagram of a part of the structure related to the scheme of the present application, and does not constitute a limitation on the cloud device to which the scheme of the present application is applied. The specific cloud device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0081] In one embodiment, a cloud device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, steps of an edge AI collaborative method based on cloud large model enhancement are implemented.

[0082] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the edge AI collaboration method based on cloud large model enhancement are implemented.

[0083] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0084] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0085] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. An edge AI collaboration method based on cloud-based large model enhancement, characterized in that: The method comprises: Collect natural language requests based on a large cloud model, parse the natural language requests through natural language processing technology, obtain user needs and convert them into various subtasks; Evaluate the processing capability of each edge device, and coordinate and offload each of the subtasks to the corresponding edge devices according to the processing capability; The cloud-based large model is used to automatically generate the code required for joint learning, and the edge AI model training is performed on each edge device in a distributed manner based on the code required for joint learning to obtain each preliminary edge AI model; Using a cross-device consistency maintenance algorithm to perform synchronization mechanism and consistency constraints on each of the preliminary edge AI models, aligning model parameters of the preliminary edge AI model between each of the edge devices, and obtaining a target edge AI model; Complete edge AI collaboration of each of the edge devices based on each of the target edge AI models.

2. The edge AI collaboration method based on cloud-based large model enhancement according to claim 1 is characterized in that: The natural language request is parsed through natural language processing technology to obtain user needs and convert them into various subtasks, including: Input the natural language request into a pre-trained cloud-based large model, calculate the score of the next word according to the natural language request through natural language processing technology, and perform word prediction based on the score to obtain a word prediction result; Determine user behavior according to the word prediction result, extract entity information in the natural language request, and use the user behavior and entity information as user demand; The user requirements are converted into various subtasks through a conversion function.

3. The edge AI collaboration method based on cloud-based large model enhancement according to claim 1 is characterized in that: Evaluate the processing capabilities of each edge device, including: Determine evaluation factors for each edge device; the evaluation factors include processor speed, memory capacity, storage space, and network connection of the device; Determining a computing capability indicator of the edge device according to the evaluation factor; The computing capability indicator is converted into a vector representation.

4. The edge AI collaboration method based on cloud-based large model enhancement according to claim 1 is characterized in that: Coordinating and offloading each of the subtasks to corresponding edge devices according to the processing capability includes: Convert the natural language request into a structured request representation by using natural language processing technology to obtain a user request; Using a matching function to match the user request with the processing capability of each edge device to obtain a matching result; Setting a task offloading strategy based on the matching result; The subtasks are dynamically coordinated and offloaded to corresponding edge devices according to the task offloading strategy.

5. The edge AI collaboration method based on cloud-based large model enhancement according to claim 1 is characterized in that: The cloud-based large model is used to automatically generate the code required for joint learning, including: Obtaining data distribution on each of the edge devices and obtaining hyperparameters of the joint learning algorithm; According to the data distribution and hyperparameters, use a code generation function to generate the code required for the federated learning; Among them, differential privacy technology is used to achieve privacy protection in the joint learning process.

6. The edge AI collaboration method based on cloud-based large model enhancement according to claim 1 is characterized in that: Based on the code required for the joint learning, edge AI model training is performed on each edge device in a distributed manner to obtain each preliminary edge AI model, including: Determine the number of edge devices, and calculate the allocation weight according to the data volume of the natural language request or the processing capacity; Design a distributed architecture according to the number of edge devices and the allocation weights, and each edge device works collaboratively based on the distributed architecture; Using the code required for the joint learning, edge AI model training is performed on each edge device on the distributed architecture, and the parameters in each edge AI model are adjusted to obtain each preliminary edge AI model.

7. The edge AI collaboration method based on cloud-based large model enhancement according to claim 1 is characterized in that: A cross-device consistency maintenance algorithm is used to synchronize the mechanisms and consistency constraints of each of the preliminary edge AI models, including: Obtaining each parameter of each of the preliminary edge AI models on each of the edge devices, determining a global average model parameter of each of the preliminary edge AI models based on each of the parameters, and obtaining differences between each of the parameters; Compare each of the parameters with the global average model parameters respectively, and update the parameters to complete parameter synchronization; Based on the differences of each of the parameters and the global average model parameters, consistency constraints are determined, and consistency constraints are performed on each of the preliminary edge AI models according to the consistency constraints.

8. The edge AI collaboration method based on cloud-based large model enhancement according to claim 7 is characterized in that: The method further comprises: Extracting each of the parameters, performing weighted averaging on each of the parameters, and obtaining a target parameter; The global average model parameters are updated based on the target parameters.

9. An edge AI collaborative system based on cloud-based large model enhancement, characterized in that: The system comprises: The user intention understanding module is used to collect natural language requests based on the cloud-based large model, parse the natural language requests through natural language processing technology, obtain user needs and convert them into various subtasks; An intelligent coordination and task offloading module is used to evaluate the processing capability of each edge device and coordinate and offload each of the subtasks to the corresponding edge devices according to the processing capability; An automated joint learning module, used to automatically generate the code required for joint learning using the cloud-based large model, and to perform edge AI model training on each of the edge devices in a distributed manner based on the code required for joint learning to obtain each preliminary edge AI model; A cross-device consistency maintenance module, used to use a cross-device consistency maintenance algorithm to perform synchronization mechanisms and consistency constraints on each of the preliminary edge AI models, align model parameters of the preliminary edge AI model between each of the edge devices, and obtain a target edge AI model; The edge AI collaboration module is used to complete the edge AI collaboration of each of the edge devices based on each of the target edge AI models.

10. The edge AI collaborative system based on cloud-based large model enhancement according to claim 9 is characterized in that: The user intention understanding module is also used to: input the natural language request into a pre-trained cloud-based large model, calculate the score of the next word according to the natural language request through natural language processing technology, and predict words based on the score to obtain word prediction results; determine user behavior based on the word prediction results, extract entity information in the natural language request, and use the user behavior and entity information as user needs; and convert the user needs into various subtasks through conversion functions.

Images