An implementation method and system of edge computing and cloud large model collaborative inference

Abstract

The application discloses an implementation method and system for edge computing and cloud large model collaborative inference, and belongs to the technical field of artificial intelligence, edge computing and Internet of Things. The method builds a double-layer inference architecture in eight processes, unifies the model capabilities of both ends by relying on knowledge distillation and model pruning, and completes inference and decision by cooperating with dynamic computing power distribution. The system is composed of six modules such as an edge computing unit and a private cloud large model unit, adopts private deployment in an intranet, stores core data locally, can independently run in a network interruption state, and also supports joint iterative optimization of the models at both ends. The application overcomes the defects of pure edge and pure cloud inference intelligence, such as high time delay, offline failure, weak data security and the like, and balances real-time performance and intelligence, with excellent resource utilization. The scheme is suitable for industries such as industry, coal mine and electric power, can meet the requirements of intelligent monitoring and control with high reliability and high safety, and can also provide bottom-layer architecture support for a private AIoT platform.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, edge computing and Internet of Things, specifically to a method and system for implementing collaborative reasoning between edge computing and large cloud models. Background Technology

[0002] AIoT inference architecture is the core underlying support for IoT intelligent monitoring and control scenarios. In existing technologies, such architectures mostly adopt a single deployment mode of pure cloud or pure edge, which has significant technical defects: the pure cloud inference mode is highly dependent on network transmission, has high data inference latency, and the system is completely unavailable when the network is interrupted. In addition, data uploaded to the public cloud poses a risk of leakage and cannot meet the privacy protection requirements of high-security industries; the pure edge inference mode is subject to the hard limitation of edge hardware computing power, and can only support simple rule-based judgments. It cannot realize deep feature analysis and global correlation inference of complex time series data, and its intelligence is insufficient.

[0003] Existing edge-cloud collaborative inference solutions lack dedicated model distillation mechanisms tailored to the computing power and latency characteristics of IoT scenarios. This results in inconsistent inference capabilities between edge and cloud models, leading to a lack of consistency in inference results. Furthermore, rigid task allocation strategies fail to dynamically adjust inference nodes based on data type and computing load, resulting in low system resource utilization. Moreover, most collaborative solutions are deployed using public cloud architectures, neglecting the private deployment needs of industries such as manufacturing, coal mining, and power. Their offline fault tolerance is insufficient, making them unsuitable for the aforementioned high-reliability and high-security application scenarios.

[0004] In summary, there is an urgent need for an edge-cloud collaborative inference solution that balances low latency, high intelligence, data security, and offline autonomy to address industry technical challenges such as insufficient edge computing power, high cloud latency, lack of model collaboration, offline unavailability, and weak security. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for implementing collaborative inference between edge computing and cloud-based large models. Its core objective is to solve the technical problems of insufficient intelligence, high inference latency, offline unavailability, weak data security, and inconsistent model capabilities in traditional pure edge and pure cloud inference architectures and existing collaborative solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for collaborative reasoning between edge computing and cloud-based large models includes eight core steps: architecture deployment, model building, task scheduling, edge reasoning, cloud reasoning, result fusion, iterative optimization, and offline assurance. It achieves the migration of cloud-based large model capabilities to the edge through knowledge distillation and model pruning, allocates different types of data to the edge or cloud for processing using dynamic task scheduling strategies, ensures data security by adopting an intranet private deployment mode, supports autonomous edge operation in offline conditions, and completes joint iterative optimization of edge and cloud models based on edge feedback data.

[0007] The corresponding edge computing and cloud-based large-scale model collaborative inference system includes six functional units: edge computing unit 1, task scheduling unit 2, private cloud-based large-scale model unit 3, fusion decision unit 4, model iteration unit 5, and offline support unit 6. Each unit is interconnected and complementary in function, constructing a two-layer inference architecture to achieve the organic combination of low-latency edge inference and high-intelligence cloud analysis. The system is deployed privately on the internal network, supports access to multiple protocol devices and cross-scenario reuse, and adapts to the intelligent monitoring and control needs of multiple industries.

[0008] Preferably, the edge computing unit 1 has data caching, network interruption resume, and local storage functions, and can receive control signals from the offline protection unit 6 when the network is interrupted, and automatically switch to offline autonomous mode.

[0009] Preferably, the cloud-based large model in the private cloud large model unit 3 adopts the Transformer deep neural network architecture and completes training and fine-tuning based on industry time series data, fault case library, industry historical data and knowledge graph.

[0010] Preferably, the model iteration unit 5 can collect the running feedback data and labeled samples of the edge computing unit 1 daily, and jointly fine-tune the edge lightweight model and the cloud-based large model to continuously improve the inference accuracy.

[0011] Preferably, the dynamic allocation strategy of the task scheduling unit 2 is as follows: regular data is processed in real time at the edge, and abnormal and complex data are uploaded to the cloud for deep inference, and the allocation is dynamically adjusted according to data latency requirements, computing load and anomaly level.

[0012] Beneficial effects 1. This invention constructs a two-layer inference architecture consisting of a lightweight edge model and a private large cloud model, thereby combining low-latency edge inference with high-intelligence cloud analysis. This ensures real-time response in data processing in IoT scenarios while enabling deep feature extraction and global correlation analysis of complex time-series data, balancing response speed and analysis accuracy.

[0013] 2. This invention uses knowledge distillation and model pruning techniques to migrate the core capabilities of large cloud models to the edge, generating lightweight models adapted to edge hardware computing power. This significantly reduces the computing power requirements of edge models while minimizing model performance loss, thus achieving unified capabilities between edge and cloud models.

[0014] 3. This invention adopts a private deployment mode on the enterprise intranet. The cloud-based large model and core business systems both run on the enterprise intranet, and the original data and training data are kept in the local environment throughout the process. This eliminates the risk of data leakage from the deployment architecture, ensuring strong data security and compliance, and is suitable for high-security industry requirements.

[0015] 4. This invention sets up a dynamic task scheduling strategy, which allocates different types of data to corresponding inference nodes according to data latency requirements, computing load and anomaly level. Regular data is processed locally at the edge, and complex data is deeply analyzed in the cloud. The computing power allocation is reasonable and greatly improves the overall resource utilization of the system.

[0016] 5. The present invention is equipped with an offline protection unit 6 and an edge computing unit 1 that integrates data caching, network interruption resume transmission and local storage functions. When the network is interrupted, the edge node automatically switches to offline autonomous mode to ensure continuous and stable operation of the business. It can be adapted to harsh industrial environments such as industrial plants and coal mines.

[0017] 6. This invention sets up a model iteration unit 5, which performs daily joint fine-tuning and optimization of the lightweight edge model and the large cloud model based on the operational feedback data and labeled samples at the edge, so that the model inference accuracy continues to improve with the usage time, and realizes the self-optimization and self-upgrading of the model.

[0018] 7. The method and system of the present invention adopt a standardized architecture design, support the access of multi-protocol devices, and can be reused across multiple industry scenarios such as industrial monitoring, coal mine safety, and power operation and maintenance. It provides underlying patent and architecture support for private AIoT platforms and has broad industry application value. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall architecture of edge and cloud collaborative reasoning in this invention; Figure 2 This is a flowchart illustrating the collaborative reasoning method of the present invention; Figure 3 This is a schematic diagram of the offline protection and computing power scheduling structure of the present invention; In the diagram: 1. Edge computing unit; 2. Task scheduling unit; 3. Private cloud large model unit; 4. Fusion decision unit; 5. Model iteration unit; 6. Offline support unit. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: Please see Figures 1-3 This invention provides a method and system for implementing collaborative reasoning between edge computing and cloud-based large-scale models. The system includes an edge computing unit 1, a task scheduling unit 2, a private cloud-based large-scale model unit 3, and a fusion decision unit 4, which are sequentially connected in communication. It also includes a model iteration unit 5, which is connected in communication with the edge computing unit 1 and the private cloud-based large-scale model unit 3, and an offline support unit 6, which is connected in communication with the edge computing unit 1. Each unit works together to execute all the steps of the implementation method and complete collaborative reasoning and decision control in multi-industry intelligent monitoring and control scenarios.

[0022] In this embodiment, the edge computing unit 1 adopts an IoT edge gateway hardware architecture and is deployed in the IoT field. It can collect the operating data of terminal devices and environmental time-series data, and integrate data preprocessing, data caching, local storage and network interruption resume functions. The private cloud large model unit 3 is deployed in the enterprise intranet private cloud. The cloud large model is built based on the Transformer deep neural network architecture and is trained and fine-tuned by industry time-series data, fault case library, industry historical data and knowledge graph. The entire system is deployed in the enterprise intranet private environment, does not rely on external public cloud services, supports the access of devices with multiple protocols such as Modbus, MQTT, TCP / IP, etc., and can be directly adapted to application scenarios such as industrial monitoring, coal mine safety, power operation and maintenance.

[0023] In this embodiment, the specific execution steps of the collaborative reasoning implementation method are as follows: 1. Architecture Deployment: Deploy edge computing nodes corresponding to edge computing unit 1 at the IoT site, and deploy cloud-based large model inference units corresponding to private cloud large model unit 3 in the enterprise intranet private cloud. Establish an intranet encrypted communication link between the edge and the cloud through an encrypted communication protocol to ensure data transmission security. 2. Model Building: Based on industry time-series data and fault case library, complete the basic training of the large cloud Transformer architecture model, combine industry knowledge graph to fine-tune the large cloud model, and then transfer the core feature extraction capability of the large cloud model to the edge through knowledge distillation and model pruning technology to generate a lightweight model adapted to edge hardware computing power. 3. Task scheduling: Task scheduling unit 2 monitors the system's computing load, data latency requirements, and data anomaly level in real time, and executes a dynamic allocation strategy: Normal data without anomalies is processed in real time by edge computing unit 1, while abnormal data and complex time-series data are uploaded to the private cloud large model unit 3 for deep inference via an encrypted link. 4. Edge inference: Edge computing unit 1 performs preprocessing operations such as cleaning and normalization on the collected terminal data, completes the initial judgment of data anomalies through a lightweight model, and directly generates local control commands for routine anomalies or real-time control needs, realizing low-latency local response and closed-loop control. 5. Cloud-based inference: Edge computing unit 1 uploads abnormal samples and complex time-series data fragments to private cloud-based large model unit 3. The cloud-based large model performs time-series feature extraction and multi-dimensional correlation analysis on the data to generate in-depth decision-making results, clarifying the root cause of the anomaly, its development trend, and global control strategies. 6. Result Fusion: The fusion decision unit 4 receives the edge inference results and the cloud inference results, performs cross-validation and logical fusion on the two sets of results, eliminates contradictory results, optimizes the control strategy, forms the final optimal control command, and sends it to the terminal execution device; 7. Iterative optimization: Model iteration unit 5 collects daily operation feedback data, actual control effect data and manually labeled samples from edge computing unit 1, and uses them as supplementary training data to jointly fine-tune and optimize the lightweight edge model and the large cloud model, continuously improving the model's anomaly recognition accuracy and decision rationality. 8. Offline Assurance: The offline assurance unit 6 monitors the communication link status between the edge and the cloud in real time. When a communication interruption is detected, it immediately sends a control signal to the edge computing unit 1. The edge computing unit 1 automatically switches to offline autonomous mode and completes independent inference and control based on local cached data and a lightweight model to ensure continuous operation of services. After communication is restored, the edge computing unit 1 automatically performs network interruption resume transmission and uploads the operation data during the offline period to the cloud to complete data synchronization.

[0024] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for implementing collaborative reasoning between edge computing and large cloud-based models, characterized in that, Includes the following steps: S1 architecture deployment: deploy edge computing nodes at the IoT site, deploy cloud-based large model inference units in the enterprise's private cloud, and establish encrypted communication links within the internal network; S2 Model Construction: Train a large cloud-based Transformer architecture model based on industry time-series data and a fault case library, and generate a lightweight edge model through knowledge distillation and model pruning; S3 task scheduling: Based on data latency requirements, computing load and anomaly level, regular data is allocated to the edge for real-time processing, while abnormal and complex data are uploaded to the cloud for deep inference. S4 Edge Inference: Edge nodes preprocess the collected data and perform lightweight model initial judgment to achieve low-latency local response and control; S5 Cloud Inference: The edge device uploads abnormal samples and time series data to the cloud-based large model for time series feature extraction and correlation analysis to generate in-depth decision results; S6 Result Fusion: Verify and fuse edge inference results with cloud inference results to form the final control command for execution; S7 Iterative Optimization: Joint fine-tuning of the lightweight edge model and the large cloud model based on edge feedback data; S8 offline protection: When communication with the central node is interrupted, the edge node automatically switches to offline autonomous mode to ensure continuous operation of services. The edge computing node also has data caching, network interruption resume and local storage functions.

2. The method for implementing collaborative reasoning between edge computing and large cloud models according to claim 1, characterized in that, The cloud-based large model and core business systems both run on the enterprise intranet, with raw data and training data retained in the local environment throughout the process, adopting an intranet private deployment mode.

3. The method for implementing collaborative reasoning between edge computing and large cloud models according to claim 1, characterized in that, The method is applicable to intelligent monitoring and control scenarios such as industrial monitoring, coal mine safety, power operation and maintenance, smart aquaculture, fire early warning, and smart wearables.

4. A collaborative inference system for edge computing and cloud-based large-scale models, characterized in that, It includes an edge computing unit 1, a task scheduling unit 2, a private cloud large model unit 3, and a fusion decision unit 4 that are connected in sequence, as well as a model iteration unit 5 that is connected in communication with the edge computing unit 1 and the private cloud large model unit 3 respectively, and an offline support unit 6 that is connected in communication with the edge computing unit 1. The edge computing unit 1 is used for real-time data processing, lightweight model inference, and local control; the private cloud large model unit 3 is used for in-depth analysis of time-series data and global decision-making. The task scheduling unit 2 is used to allocate inference tasks according to data real-time requirements and system computing load. The fusion decision unit 4 is used to verify and fuse the inference results from the edge and the cloud. The model iteration unit 5 is used for the joint optimization of the edge lightweight model and the cloud large model; The offline support unit 6 is used to control the edge computing unit 1 to achieve autonomous operation in the event of a network interruption.

5. The edge computing and cloud-based large-model collaborative inference system according to claim 4, characterized in that, The edge lightweight model is obtained from the cloud-based large model through knowledge distillation and pruning, and the edge lightweight model retains the core feature extraction capability of the cloud-based large model.

6. The edge computing and cloud-based large-model collaborative inference system according to claim 4, characterized in that, The cloud-based large model in the private cloud-based large model unit 3 adopts the Transformer deep neural network architecture, and the cloud-based large model completes fine-tuning training based on industry historical data and knowledge graph.

7. The edge computing and cloud-based large-model collaborative inference system according to claim 4, characterized in that, The system is deployed entirely within a private enterprise intranet environment, without relying on external public cloud services, and supports multi-protocol device access and cross-industry scenario reuse.

8. The edge computing and cloud-based large-model collaborative inference system according to claim 4, characterized in that, The edge computing unit 1 integrates data acquisition, data preprocessing, data caching, local storage, and offline resume transmission functions, and can receive control signals from the offline protection unit 6 to switch to offline autonomous mode.

9. The edge computing and cloud-based large-model collaborative reasoning system according to claim 4, characterized in that, The model iteration unit 5 can collect the running feedback data and labeled samples of the edge computing unit 1, and perform daily joint fine-tuning and optimization of the edge lightweight model and the cloud large model.

10. The edge computing and cloud-based large-model collaborative inference system according to any one of claims 4-9, characterized in that, The system is adapted to the implementation method of any one of claims 1-3 to complete collaborative reasoning and decision control in multi-industry intelligent monitoring and control scenarios.

Images