Hybrid expert model distributed computing system based on edge computing technology

Abstract

The invention relates to the technical field of artificial intelligence and edge computing, in particular to a hybrid expert model distributed computing system based on an edge computing technology. Comprising a central server, an edge device network, a data privacy protection module, an automatic model decomposition module, a performance feedback and adaptive learning module, a fault detection and self-healing module and a task priority scheduling module. The central server comprises a self-adaptive gating network, a task manager and a model manager; the self-adaptive gating network is used for receiving an input task, extracting task characteristics, intelligently identifying task data distribution and routing the task to the expert sub-model matched with the task according to an identification result; according to the method, the tasks are dynamically scheduled according to the computing power and the load condition of the edge device, the model updating and synchronization mechanism is imperfect, the consistency of the model in the distributed environment is ensured, the computing cost is remarkably reduced, and the dependence on special hardware facilities is reduced by utilizing idle edge device computing resources.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and edge computing, and specifically to a distributed computing system for a hybrid expert model based on edge computing technology. Background Art

[0002] With the rapid development of artificial intelligence technology, large language models (LLMs) have demonstrated powerful performance in fields such as natural language processing and computer vision. However, such models usually require a large amount of computing resources, leading to problems such as increased burden on data centers and increased energy consumption. Currently, the solutions mainly focus on two directions: One is to reduce the computational complexity of the model through techniques such as model compression and quantization; The other is to use distributed computing to improve computing efficiency; However, these solutions still have limitations. For example, model compression may lead to performance degradation, and traditional distributed computing requires the support of specialized hardware facilities; Although existing edge computing technologies can utilize dispersed computing resources, there are problems in the deployment of large AI models, such as the lack of an effective task allocation mechanism, the inability to dynamically schedule tasks according to the computing power and load conditions of edge devices, and the imperfect model update and synchronization mechanism, making it difficult to ensure the consistency of the model in a distributed environment. Therefore, a distributed computing system for a hybrid expert model based on edge computing technology is proposed to address the above problems. Summary of the Invention

[0003] The purpose of the present invention is to provide a distributed computing system for a hybrid expert model based on edge computing technology to solve the problems that although existing edge computing technologies can utilize dispersed computing resources, there are problems in the deployment of large AI models, such as the lack of an effective task allocation mechanism, the inability to dynamically schedule tasks according to the computing power and load conditions of edge devices, and the imperfect model update and synchronization mechanism, making it difficult to ensure the consistency of the model in a distributed environment.

[0004] To achieve the above purpose, the present invention provides the following technical solutions: A distributed computing system for a hybrid expert model based on edge computing technology, including a central server, an edge device network, a data privacy protection module, an automatic model decomposition module, a performance feedback and adaptive learning module, a fault detection and self-healing module, and a task priority scheduling module; The central server includes an adaptive gating network, a task manager, and a model manager; The adaptive gating network is used to receive input tasks, extract task features, and intelligently identify the task data distribution, and at the same time route the tasks to the matching expert sub-models according to the identification results; The task manager dynamically allocates tasks based on the recognition results of the adaptive gating network, combined with the computing power, current load status, and network connection quality of the edge device. The model manager is used to manage the version update, parameter synchronization, and consistency maintenance of each expert sub-model in the edge device network. The edge device network consists of multiple edge devices, and each edge device deploys an expert sub-model optimized for a specific task type.

[0005] As a further optimization of the present invention, wherein: the edge devices include personal computers, smartphones, tablets, and various Internet of Things devices, and each edge device is connected to the central server through a standard network protocol to form a widely distributed computing network.

[0006] As a further optimization of the present invention, wherein: during the task allocation process, the task manager not only relies on the task characteristics and data distribution identified by the adaptive gating network, but also comprehensively considers the real-time computing power, load status, and network connection quality of the edge device to achieve dynamic task scheduling and load balancing.

[0007] As a further optimization of the present invention, wherein: the update and synchronization of the expert sub-model by the model manager include the following steps: Regularly detect the version and status of the expert sub-model on each edge device; When it detects that the model parameters are updated or the version is upgraded, automatically distribute the updated model parameters to the relevant edge devices.

[0008] As a further optimization of the present invention, wherein: the data privacy protection module uses advanced data encryption technologies and secure communication protocols to protect sensitive information throughout the data collection, storage, and transmission processes.

[0009] As a further optimization of the present invention, wherein: the automatic model decomposition module is used to automatically decompose a large-scale artificial intelligence model into multiple expert sub-models according to task characteristics and data distribution. Each expert sub-model is optimized for a specific task type and automatically generates a model configuration suitable for distributed deployment.

[0010] As a further optimization of the present invention, wherein: the performance feedback and adaptive learning module is used to collect the performance data generated when each edge device processes tasks in real time, and online optimize the adaptive gating network and task allocation strategy based on the collected data.

[0011] As a further optimization of the present invention, wherein: the fault detection and self-healing module is used to monitor the operating status of each edge device, timely detect device failures or abnormal conditions, and automatically reallocate tasks and perform necessary model rollbacks when a failure occurs.

[0012] As a further optimization of the present invention, wherein: the task priority scheduling module ranks and dynamically schedules the tasks to be processed according to the task urgency, service quality agreement, and edge device resource status.

[0013] Compared with the prior art, the beneficial effects of the present invention are: 1. In the present invention, due to the imperfect dynamic scheduling of tasks, model updates, and synchronization mechanisms according to the computing power and load conditions of edge devices, the consistency of the model in a distributed environment is ensured, the computing cost is significantly reduced, the dependence on dedicated hardware facilities is reduced by utilizing the idle computing resources of edge devices, and at the same time, the service efficiency is improved. With the help of the mixture of experts model and dynamic task allocation mechanism, the optimal utilization of computing resources is achieved; 2. In the present invention, the quality of service is guaranteed, and the adaptive gating network ensures that tasks are routed to the most suitable expert sub-model for processing; 3. In the present invention, the system security is enhanced, and strict data privacy protection measures are adopted to ensure the data security on edge devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a system block diagram of a distributed computing system with a mixture of experts model based on edge computing technology according to the present invention; Figure 2 is a flowchart of the adaptive gating network of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0015] Embodiment 1: Please refer to Figure 1-2 , the present invention provides a technical solution: A distributed computing system with a mixture of experts model based on edge computing technology includes a central server, an edge device network, a data privacy protection module, an automatic model decomposition module, a performance feedback and adaptive learning module, a fault detection and self-healing module, and a task priority scheduling module; The central server includes an adaptive gating network, a task manager, and a model manager; The adaptive gating network is used to receive input tasks, extract task features, and intelligently identify the task data distribution, and at the same time route the tasks to the matching expert sub-models according to the identification results; The task manager dynamically allocates tasks based on the recognition results of the adaptive gating network, combined with the computing power, current load status, and network connection quality of the edge devices; The model manager is used to manage the version update, parameter synchronization, and consistency maintenance of each expert sub-model in the edge device network; The edge device network consists of multiple edge devices, and each edge device deploys an expert sub-model optimized for a specific task type. Through the collaborative work of the central server and the edge device network, efficient task allocation and processing are achieved, improving the overall system performance and resource utilization.

[0016] As a further technical solution of this scheme, the edge devices include personal computers, smartphones, tablets, and various Internet of Things devices, and each edge device is connected to the central server through standard network protocols, forming a widely distributed computing network. By using the widely distributed edge device network, the computing power of the system is extended to adapt to diverse task requirements; As a further technical solution of this scheme, during the task allocation process, the task manager not only relies on the task characteristics and data distribution identified by the adaptive gating network but also comprehensively considers the real-time computing power, load status, and network connection quality of the edge devices to achieve dynamic task scheduling and load balancing. Dynamic task scheduling and load balancing ensure that the system can still operate efficiently under high load, avoiding resource waste and task backlogs; As a further technical solution of this scheme, the update and synchronization of the expert sub-models by the model manager include the following steps: Regularly detect the versions and status of the expert sub-models on each edge device; When a model parameter update or version upgrade is detected, automatically distribute the updated model parameters to the relevant edge devices; It can ensure the continuous optimization of model performance in a distributed environment; As a further technical solution of this scheme, the data privacy protection module uses advanced data encryption technologies and secure communication protocols to protect sensitive information throughout the data collection, storage, and transmission processes. The data privacy protection module ensures the security of sensitive information during transmission and storage through encryption and secure communication protocols, in line with privacy protection regulations; As a further technical solution of this scheme, the automatic model decomposition module is used to automatically decompose large-scale artificial intelligence models into multiple expert sub-models according to task characteristics and data distribution. Each expert sub-model is optimized for a specific task type and automatically generates a model configuration suitable for distributed deployment. The automatic model decomposition module improves the flexibility of model deployment and task processing efficiency, adapting to different task types and data distributions; As a further implementation technical solution of this scheme, the performance feedback and adaptive learning module is used to collect in real time the performance data generated when each edge device processes tasks, and online optimize the adaptive gating network and task allocation strategy based on the collected data. Through real-time data collection and optimization, the performance feedback and adaptive learning module improves the overall system performance and task processing quality; As a further implementation technical solution of this scheme, the fault detection and self-healing module is used to monitor the running status of each edge device, timely detect device failures or abnormal conditions, and automatically re-allocate tasks and perform necessary model rollbacks when a failure occurs. The fault detection and self-healing module improves the reliability and self-healing ability of the system, and reduces the impact of faults on task processing; As a further implementation technical solution of this scheme, the task priority scheduling module sorts and dynamically schedules the tasks to be processed according to the task urgency, service quality agreement, and edge device resource status. The task priority scheduling module ensures that high-priority tasks are processed in a timely manner, meets the service quality requirements, and improves user satisfaction.

[0017] Workflow: S1: Task reception and feature extraction: The adaptive gating network of the central server receives the input task, extracts task features and identifies the data distribution, which is used for intelligent identification of task types and data distributions to ensure the accuracy of task routing; S2: Task allocation and scheduling: According to the recognition results of the adaptive gating network, the task manager dynamically allocates tasks in combination with the computing power, load status, and network connection quality of the edge devices, realizes dynamic task scheduling and load balancing, and improves the utilization rate of system resources; S3: Model management and synchronization: The model manager regularly detects the version and status of the expert sub-models on the edge devices, and automatically distributes the updated model parameters to ensure consistency, and guarantees the continuous optimization and consistency of model performance in a distributed environment; S4: Data privacy protection: The data privacy protection module adopts encryption technology and secure communication protocols during data collection, storage, and transmission to protect sensitive information, ensure data privacy and security, and comply with privacy protection regulations; S5: Model decomposition and deployment: The automatic model decomposition module decomposes the large-scale AI model into multiple expert sub-models according to task characteristics and data distributions, and generates a model configuration suitable for distributed deployment, improving the flexibility of model deployment and task processing efficiency; S6: Performance feedback and optimization: The performance feedback and adaptive learning module collects the performance data of edge devices in real time, optimizes the adaptive gating network and task allocation strategy, and improves the overall system performance and task processing quality through real-time feedback and optimization; S7: Fault Detection and Self-Healing: The fault detection and self-healing module monitors the operating status of edge devices, detects faults and automatically reallocates tasks and rolls back models, improving the reliability and self-healing ability of the system and reducing the impact of faults on task processing; S8: Task Priority Scheduling: The task priority scheduling module sorts and dynamically schedules the tasks to be processed according to the task urgency, service quality agreement and edge device resource status, ensuring that high-priority tasks are processed in a timely manner and meeting the service quality requirements.

[0018] Embodiment 2: Parts that are the same as those in Embodiment 1 will not be elaborated. The differences are as follows: First, a distributed edge computing architecture is constructed, and idle edge devices such as personal computers and smartphones are used to form a computing network, and efficient task processing is achieved through a dynamic task allocation mechanism, significantly reducing the computing cost. By utilizing the computing resources of idle edge devices, the dependence on dedicated hardware facilities is reduced; Second, a hybrid expert model architecture is designed, and a large language model is decomposed into multiple expert sub-models, and each sub-model is responsible for processing specific types of tasks. The computing efficiency is improved through a dynamic activation mechanism, and the service efficiency is improved. With the help of the hybrid expert model and the dynamic task allocation mechanism, the optimal utilization of computing resources is achieved; Third, an adaptive gating network (AGN) is introduced to achieve intelligent task identification and routing, enabling the expert sub-models to be specialized in specific tasks and ensuring the service quality. Through the adaptive gating network, tasks are ensured to be routed to the most suitable expert sub-model for processing; Fourth, an efficient model update and maintenance mechanism is developed to ensure the consistency and data security of the model in a distributed environment, enhance the system security, and adopt strict data privacy protection measures to ensure the data security on edge devices.

[0019] In this article, specific examples are used to elaborate on the principles and implementation manners of the present invention. The descriptions of the above examples are only used to help understand the method and its core idea of the present invention. The above is only the preferred implementation manner of the present invention. It should be noted that due to the limitation of literal expression and objectively existing infinite specific structures, for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements, embellishments or changes can be made, or the above technical features can be combined in an appropriate manner; these improvements, embellishments, changes or combinations, or directly applying the concept and technical solution of the invention to other occasions without improvement, shall all be regarded as the protection scope of the present invention.

Claims

1. A hybrid expert model distributed computing system based on edge computing technology, characterized in that: It includes a central server, an edge device network, a data privacy protection module, an automatic model decomposition module, a performance feedback and adaptive learning module, a fault detection and self-healing module, and a task priority scheduling module; The central server includes an adaptive gating network, a task manager, and a model manager; The adaptive gating network is used to receive input tasks, extract task features, and intelligently identify the distribution of task data, and route the tasks to the expert sub-model that matches them according to the identification results; The task manager dynamically allocates tasks based on the recognition results of the adaptive gating network and in combination with the computing power, current load status and network connection quality of the edge device; The model manager is used to manage version updates, parameter synchronization, and consistency maintenance of each expert sub-model in the edge device network; The edge device network is composed of multiple edge devices, and an expert sub-model optimized for a specific task type is deployed on each edge device.

2. According to claim 1, a hybrid expert model distributed computing system based on edge computing technology is characterized in that: The edge devices include personal computers, smart phones, tablets and various IoT devices, and each edge device is connected to the central server through a standard network protocol to form a widely distributed computing network.

3. The hybrid expert model distributed computing system based on edge computing technology according to claim 1, characterized in that: During the task allocation process, the task manager not only relies on the task characteristics and data distribution identified by the adaptive gating network, but also comprehensively considers the real-time computing power, load status and network connection quality of the edge device to achieve dynamic task scheduling and load balancing.

4. The hybrid expert model distributed computing system based on edge computing technology according to claim 1, characterized in that: The updating and synchronization of the expert sub-model by the model manager includes the following steps: Regularly check the version and status of the expert sub-model on each edge device; When a model parameter update or version upgrade is detected, the updated model parameters are automatically distributed to relevant edge devices.

5. The hybrid expert model distributed computing system based on edge computing technology according to claim 1, characterized in that: The data privacy protection module uses advanced data encryption technology and secure communication protocols to protect sensitive information throughout the process of data collection, storage and transmission.

6. The hybrid expert model distributed computing system based on edge computing technology according to claim 1, characterized in that: The automatic model decomposition module is used to automatically decompose a large-scale artificial intelligence model into multiple expert sub-models according to task characteristics and data distribution. Each expert sub-model is optimized for a specific task type and automatically generates a model configuration suitable for distributed deployment.

7. The hybrid expert model distributed computing system based on edge computing technology according to claim 1, characterized in that: The performance feedback and adaptive learning module is used to collect the performance data generated by each edge device when processing tasks in real time, and optimize the adaptive gating network and task allocation strategy online based on the collected data.

8. The hybrid expert model distributed computing system based on edge computing technology according to claim 1, characterized in that: The fault detection and self-healing module is used to monitor the operating status of each edge device, detect device failures or abnormal conditions in a timely manner, and automatically reallocate tasks and perform necessary model rollbacks when a failure occurs.

9. The hybrid expert model distributed computing system based on edge computing technology according to claim 1, characterized in that: The task priority scheduling module prioritizes and dynamically schedules the tasks to be processed according to the urgency of the tasks, the quality of service agreement, and the resource status of the edge device.

Images