A method, system and PC host for building and running intelligent AI

Through parallel training and hierarchical aggregation technology of distributed computing nodes, the problems of low efficiency and complexity of traditional AI construction methods are solved, and efficient and accurate construction and operation of intelligent AI models are achieved.

CN118485163BActive Publication Date: 2024-11-15DONGGUAN HUAMING TENG TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410734371.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-06
Publication Date
2024-11-15
Estimated Expiration
2044-06-06

AI Technical Summary

Technical Problem

Traditional AI construction methods rely on professionals, are time-consuming and labor-intensive, have low model training efficiency, and are complex in operation and management, which limits the popularization and application of AI technology.

Method used

By obtaining and cleaning the training resource data of the intelligent AI model, deploying distributed computing nodes for parallel training and hierarchical aggregation of models, optimizing model parameters, realizing distributed deployment and online operation, reducing the complexity of model construction and management.

Benefits of technology

It improves the parallelism and efficiency of model training, shortens training time, reduces labor costs, simplifies the construction and operation process of AI systems, and improves the accuracy and reliability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118485163B_ABST
    Figure CN118485163B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of intelligent AI technology, and in particular to a method, system and PC host for constructing and operating an intelligent AI. The method comprises the following steps: obtaining the training resource data required for the intelligent AI model and performing cleaning processing to obtain the training standard data required for the intelligent AI model; by deploying distributed computing nodes, and performing training resource partitioning processing and constructing and initializing the intelligent AI sub-model for the training standard data required for the intelligent AI model, the intelligent AI sub-models on each distributed computing node are obtained; the intelligent AI sub-models on each distributed computing node are subjected to model parallel training and model hierarchical aggregation to generate an intelligent AI global model; the intelligent AI global model is subjected to model optimization to obtain an intelligent AI global optimization model; and the intelligent AI global optimization model is subjected to online operation to obtain an intelligent AI online operation model. The present invention can realize the efficient construction and operation of the intelligent AI model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent AI technology, and in particular to a method, system and PC host for building and running intelligent AI. Background Art

[0002] In the current digital age, the application of artificial intelligence (AI) technology has penetrated into various fields, bringing huge potential and opportunities for social development. Traditional AI construction methods often rely on professional data scientists and engineering teams, and require manual data collection, feature extraction, model selection and tuning and other tedious steps, which consume a lot of time and manpower costs. Moreover, in the process of model training and optimization, parameters and algorithms need to be constantly adjusted, which often leads to high trial and error costs and low efficiency. In addition, once the model is built, its operation and management also require professional technicians to monitor and maintain, which increases the complexity and cost of intelligent AI systems. However, there are still some challenges in the construction and operation of existing intelligent AI systems, such as long model training time and high algorithm complexity, which limits the popularization and application of intelligent AI technology. Summary of the invention

[0003] Based on this, it is necessary for the present invention to provide a method, system and PC host for building and operating an intelligent AI to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above purpose, a method for building and running an intelligent AI includes the following steps:

[0005] Step S1: Acquire the training resource data required for the intelligent AI model, and clean the training resource data required for the intelligent AI model to obtain the training standard data required for the intelligent AI model; deploy distributed computing nodes in the network, and divide the training resources for the training standard data required for the intelligent AI model based on the distributed computing nodes to obtain the intelligent AI model training resource sub-data of each distributed computing node;

[0006] Step S2: constructing and initializing intelligent AI sub-models on each distributed computing node to obtain intelligent AI sub-models on each distributed computing node; performing model parallel training on the intelligent AI sub-models on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node to obtain intelligent AI model training parameters of each distributed computing node;

[0007] Step S3: Based on the intelligent AI model training parameters of each distributed computing node, the intelligent AI sub-models on each distributed computing node are aggregated in a hierarchical manner to generate an intelligent AI global model; the intelligent AI global model is optimized to obtain an intelligent AI global optimization model;

[0008] Step S4: The intelligent AI global optimization model is deployed to each distributed computing node for online operation to obtain an intelligent AI online operation model.

[0009] The present invention first obtains the training resource data required for the intelligent AI model, which is the first step in building an efficient model. This step can ensure that the model has enough information during the training process by collecting a large amount of resource data required for model training, thereby improving the accuracy and reliability of its predictions and decisions, and providing a sufficient data basis for subsequent data processing and model training, thereby laying a solid foundation for building an intelligent AI model with good generalization capabilities. By performing data cleaning processing on the training resource data required for the intelligent AI model, outliers and interference information in the data can be effectively removed to improve data quality. This processing can reduce the interference of noise data on the model during training, thereby making the model more robust and reliable. At the same time, standardization processing can make the data have similar scales in feature dimensions, avoiding model training instability and convergence difficulties caused by data scale differences, and can also improve the training efficiency and convergence speed of the model, thereby accelerating the model training process. At the same time, by deploying distributed computing nodes in the network and dividing the training resource data for processing, it is possible to fully utilize distributed computing resources and accelerate the model training process. The key to this step is to improve the parallelism and efficiency of model training, so that larger data sets can be processed and model training tasks can be completed in a shorter time, thereby shortening the model training time. Secondly, by building and initializing intelligent AI sub-models on each distributed computing node, the distributed deployment and initialization of the model can be achieved to prepare for the subsequent training process. The key to this step is to be able to establish an independent intelligent AI sub-model on each node, so that each node can independently perform model training, thereby improving the parallelism and overall training speed of the AI ​​model. By performing model synchronization and parallel training on the intelligent AI sub-models on each node based on the intelligent AI model training resource sub-data of each distributed computing node, the model parameters between nodes can be updated synchronously, which improves the efficiency and speed of model training. In this way, the distributed computing resources can be fully utilized, and the synchronous training of model parameters can be realized, thereby speeding up the overall training process of the intelligent AI model. By performing asynchronous parallel training on the model, the model parameters between nodes can be updated asynchronously, which improves the flexibility and stability of model training. The key to this step is to give full play to the computing power of each node and realize asynchronous training of model parameters, so that model training can better adapt to the computing resource conditions of different nodes. In addition, by performing arithmetic average analysis on the model synchronous parallel training parameters and the model asynchronous parallel training parameters of each distributed computing node, the training results of each node can be comprehensively considered to obtain the final intelligent AI model training parameters. The key to this step is to reasonably integrate the training results of each node to obtain unified training parameters, which provides a reliable parameter basis for the final model construction and application, thereby reducing the complexity of the algorithm of the intelligent AI model training process.Finally, by deploying the intelligent AI global optimization model to various distributed computing nodes for online operation, the real-time application and service of the intelligent AI model can be realized. The key to this step is to be able to apply the trained model to actual scenarios to provide users with intelligent services and decision support. By running the model online, the model performance can be continuously optimized, thereby improving the intelligence level of the intelligent AI system and user experience.

[0010] Preferably, step S1 comprises the following steps:

[0011] Step S11: Obtaining training resource data required for the intelligent AI model;

[0012] Step S12: performing noise data removal processing on the training resource data required for the intelligent AI model to obtain the intelligent AI model training resource denoised data;

[0013] Step S13: performing a missing value test on the intelligent AI model training resource denoising data to obtain the missing status of the intelligent AI model training resource data; performing interpolation compensation processing on the intelligent AI model training resource denoising data based on the missing status of the intelligent AI model training resource data to obtain the intelligent AI model training resource missing compensation data;

[0014] Step S14: standardizing the intelligent AI model training resource loss compensation data to obtain the training standard data required for the intelligent AI model;

[0015] Step S15: By deploying distributed computing nodes in the network, and dividing the training resources of the standard training data required for the intelligent AI model based on the distributed computing nodes, the intelligent AI model training resource sub-data of each distributed computing node is obtained.

[0016] The present invention first obtains the training resource data required by the intelligent AI model, which is the first step in building an efficient model. This step can ensure that the model has enough information during the training process by collecting a large amount of resource data required for model training, thereby improving the accuracy and reliability of its prediction and decision-making, and providing a sufficient data basis for subsequent data processing and model training, thereby laying a solid foundation for building an intelligent AI model with good generalization ability. Secondly, by performing noise data cleaning on the training resource data required for the intelligent AI model, outliers and interference information in the data can be effectively removed to improve data quality. This processing can reduce the interference of noise data on the model during the training process, thereby making the model more robust and reliable. Then, by performing missing value inspection on the training resource data that has been processed by noise cleaning and performing interpolation compensation processing, the missing information in the data can be effectively processed to ensure the integrity and availability of the data. Through this step, it can be ensured that the model can make full use of all available data information during the training process, thereby improving the training effect and performance of the model. Next, by standardizing the training resource data that has been processed for missing compensation, the data can have a similar scale in the feature dimension, avoiding the instability and convergence difficulties of the model training caused by the difference in data scale. This step can improve the training efficiency and convergence speed of the model, thereby accelerating the model training process. Finally, by deploying distributed computing nodes in the network and dividing the training resource data, the distributed computing resources can be fully utilized to accelerate the model training process. The key to this step is to improve the parallelism and efficiency of model training, so that larger data sets can be processed and model training tasks can be completed in a shorter time.

[0017] Preferably, step S15 comprises the following steps:

[0018] Step S151: Deploy distributed computing nodes in the network and perform computing resource statistics analysis on each distributed computing node to obtain computing resource status data of each distributed computing node;

[0019] Step S152: Performing statistical analysis on the data storage space of each distributed computing node to obtain the size of the data storage space range of each distributed computing node;

[0020] Step S153: performing data resource accommodation evaluation and analysis on each distributed computing node based on the computing resource status data of each distributed computing node and the size of the data storage space range, and obtaining the data resource accommodation limit range on each distributed computing node;

[0021] Step S154: Perform training resource division processing on the standard training data required for the intelligent AI model according to the data resource accommodation limit range on each distributed computing node to obtain the intelligent AI model training resource sub-data of each distributed computing node.

[0022] The present invention first deploys distributed computing nodes in the network and performs statistical analysis of computing resources on each node. The key to this step is to fully understand the computing power and load of each node, and provide data support for subsequent resource allocation and task allocation to ensure that each node can fully utilize its computing resources, thereby improving the performance and efficiency of the overall model training process. Secondly, by performing statistical analysis of data storage space on each distributed computing node, the key to this step is to evaluate the storage capacity of each node to determine the range of data that the node can accommodate, and provide a basis for subsequent data storage and management, thereby ensuring the balanced distribution of data between nodes and the rationality of storage. Then, by performing data resource accommodation evaluation and analysis on the nodes based on the computing resource status data and data storage space range size of each distributed computing node, the key to this step is to determine the range of data resources that each node can carry, and provide guidance for resource allocation and task scheduling, so as to ensure that model training can effectively process data and tasks under high load. Finally, the training standard data required by the intelligent AI model is divided and processed according to the data resource capacity limit range on each distributed computing node. The key to this step is to be able to reasonably distribute the training resource data to each node, make full use of the computing and storage resources of each node, realize parallel training, improve training efficiency and speed, thereby speeding up the model training process and improving the overall performance of the system.

[0023] Preferably, step S153 includes the following steps:

[0024] Performing CPU operation utilization statistics analysis on the computing resource status data of each distributed computing node to obtain the CPU operation utilization of each distributed computing node;

[0025] Performing network bandwidth utilization statistics analysis on the computing resource status data of each distributed computing node to obtain the network bandwidth utilization of each distributed computing node;

[0026] Based on the CPU operation utilization rate and network bandwidth utilization rate of each distributed computing node, the computing resource utilization capacity of each distributed computing node is evaluated and analyzed to obtain the computing resource utilization capacity data of each distributed computing node;

[0027] Based on the size of the data storage space of each distributed computing node, the data storage capacity of each distributed computing node is evaluated and analyzed to obtain the spatial storage capacity data of each distributed computing node;

[0028] The resource accommodation limit interval of each distributed computing node is determined according to the computing resource utilization capacity data and the space storage capacity data of each distributed computing node, and the data resource accommodation limit interval on each distributed computing node is obtained.

[0029] The present invention firstly performs CPU operation utilization statistics analysis on the computing resource status data of each distributed computing node. The key of this step is to be able to comprehensively evaluate the CPU resource utilization of each node to help determine the computing power and load of the node, thereby providing an important reference for subsequent resource allocation and task scheduling. Secondly, the network bandwidth utilization statistics analysis is performed on the computing resource status data of each distributed computing node to evaluate the network communication status between nodes and determine the utilization of network bandwidth, thereby providing basic data guarantee for subsequent processing. Then, the computing resource utilization capacity of the node is evaluated and analyzed based on the CPU operation utilization and network bandwidth utilization of each distributed computing node. The key of this step is to comprehensively consider the utilization of CPU and network resources to evaluate the overall computing capacity of the node, thereby providing comprehensive guidance for resource allocation and task scheduling. Next, the data storage capacity of each distributed computing node is evaluated and analyzed based on the data storage space range size of each distributed computing node. The key of this step is to be able to evaluate the storage capacity of the node to determine the amount of data that the node can carry, provide guidance for data storage and management, thereby ensuring the balanced distribution of data between nodes and storage rationality. Finally, the resource capacity limit range of each node is determined based on the computing resource utilization capacity data and storage capacity data of each distributed computing node. The key to this step is to comprehensively consider the utilization of computing and storage resources to determine the data capacity range of the node, providing an accurate reference for resource allocation and task scheduling, thereby ensuring the efficient operation and data processing capabilities of the subsequent model parallel training process.

[0030] Preferably, step S2 comprises the following steps:

[0031] Step S21: constructing and initializing an intelligent AI sub-model on each distributed computing node to obtain an intelligent AI sub-model on each distributed computing node;

[0032] Step S22: Distribute the intelligent AI model training resource sub-data of each distributed computing node to each distributed computing node by building a secure network communication channel;

[0033] Step S23: Based on the intelligent AI model training resource sub-data of each distributed computing node, the intelligent AI sub-model on each distributed computing node is trained synchronously and in parallel to obtain the intelligent AI sub-model synchronous and parallel training parameters of each distributed computing node;

[0034] Step S24: performing asynchronous parallel training on the intelligent AI sub-models on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node, and obtaining asynchronous parallel training parameters of the intelligent AI sub-models of each distributed computing node;

[0035] Step S25: Perform arithmetic mean analysis on the synchronous parallel training parameters of the intelligent AI sub-models of each distributed computing node and the asynchronous parallel training parameters of the intelligent AI sub-models to obtain the intelligent AI model training parameters of each distributed computing node.

[0036] The present invention first constructs and initializes the intelligent AI sub-model on each distributed computing node, which can realize the distributed deployment and initialization of the model and prepare for the subsequent training process. The key to this step is to establish an independent intelligent AI sub-model on each node, so that each node can independently perform model training, thereby improving the parallelism and overall training speed of the AI ​​model. Secondly, by constructing a secure network communication channel to distribute the intelligent AI model training resource sub-data of each distributed computing node to each distributed computing node, the security and integrity of the data transmission process can be ensured. The key to this step is to achieve the secure transmission of data, ensure the integrity and confidentiality of the training data, and provide reliable data support for subsequent model training. Then, by performing model synchronization and parallel training on the intelligent AI sub-models on each node based on the intelligent AI model training resource sub-data of each distributed computing node, the synchronous update of model parameters between nodes can be realized, and the efficiency and speed of model training can be improved. In this way, the distributed computing resources can be fully utilized, and the synchronous training of model parameters can be realized, thereby speeding up the overall training process of the intelligent AI model. Next, by performing asynchronous parallel training on the intelligent AI sub-models on each node based on the intelligent AI model training resource sub-data of each distributed computing node, the model parameters between nodes can be asynchronously updated, which improves the flexibility and stability of model training. The key to this step is to give full play to the computing power of each node and realize asynchronous training of model parameters, so that model training can better adapt to the computing resource conditions of different nodes. Finally, by performing arithmetic mean analysis on the synchronous parallel training parameters of the intelligent AI sub-models of each distributed computing node and the asynchronous parallel training parameters of the intelligent AI sub-models, the training results of each node can be comprehensively considered to obtain the final intelligent AI model training parameters. The key to this step is to reasonably integrate the training results of each node to obtain unified training parameters, thereby providing a reliable parameter basis for the final model construction and application.

[0037] Preferably, step S23 includes the following steps:

[0038] Step S231: performing synchronous training task allocation on the intelligent AI sub-models on each distributed computing node to obtain the intelligent AI model synchronous training task on each distributed computing node;

[0039] Step S232: Based on the intelligent AI model training resource sub-data of each distributed computing node, the intelligent AI model synchronization training task on each distributed computing node is trained in parallel to obtain an intelligent AI synchronization parallel training model on each distributed computing node;

[0040] Step S233: Perform model parameter synchronization update processing on the intelligent AI synchronous parallel training model on each distributed computing node to obtain the intelligent AI sub-model synchronous parallel training parameters of each distributed computing node.

[0041] The present invention first distributes the training tasks to the intelligent AI sub-models on each distributed computing node synchronously, so as to effectively and reasonably distribute the training tasks to each node to ensure that each node participates in the training process. The key to this step is that it can optimize the distribution and scheduling of training tasks, make full use of distributed computing resources, and improve the training efficiency and overall performance of the intelligent AI model. Then, by performing model synchronization and parallel training on the intelligent AI model synchronous training tasks on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node, the synchronous update of model parameters between nodes can be achieved, and the efficiency and speed of model training can be improved. The key to this step is that the training tasks of each node can be coordinated, the synchronous update of model parameters can be achieved, and the overall training process can be accelerated. Finally, by performing model parameter synchronization update processing on the intelligent AI synchronous parallel training model on each distributed computing node, the consistency and synchronization of model parameters between nodes can be guaranteed, and a reliable parameter basis is provided for the construction of the final model. The effect of this step is to ensure the uniformity of model parameters, improve the training stability and performance of the model, and provide a reliable basis for subsequent model parameter analysis.

[0042] Preferably, step S24 comprises the following steps:

[0043] Step S241: performing asynchronous training task allocation on the intelligent AI sub-model on each distributed computing node to obtain the asynchronous training task of the intelligent AI model on each distributed computing node;

[0044] Step S242: Perform task load analysis on the intelligent AI model asynchronous training tasks on each distributed computing node to obtain model training task load data on each distributed computing node;

[0045] Step S243: asynchronously load the intelligent AI model training tasks on each distributed computing node according to the model training task load data on each distributed computing node, and obtain the model asynchronous loading training tasks on each distributed computing node;

[0046] Step S244: performing asynchronous parallel model training on the model asynchronous loading training tasks on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node, to obtain the intelligent AI asynchronous parallel training model on each distributed computing node;

[0047] Step S245: Asynchronously update the model parameters of the intelligent AI asynchronous parallel training model on each distributed computing node to obtain the asynchronous parallel training parameters of the intelligent AI sub-model of each distributed computing node.

[0048] The present invention firstly distributes the training tasks asynchronously to the intelligent AI sub-models on each distributed computing node, and can flexibly distribute the training tasks according to the actual operation status and resource utilization of each node, so as to make full use of the computing resources and improve the overall training efficiency. Through asynchronous task distribution, the waiting time between nodes can be avoided, making the training process more coherent and efficient. Secondly, by performing task load analysis on the asynchronous training tasks of the intelligent AI model, the training task conditions on each node can be fully understood, including information such as task type and resource consumption, which helps to optimize the task scheduling strategy, avoid node load imbalance or overload, and thus improve the stability and performance of the overall model training. Then, by asynchronously loading the intelligent AI model training tasks on each distributed computing node according to the model training task load data on each distributed computing node, the loading order and priority of the tasks can be dynamically adjusted according to the actual situation of the node to ensure that the resources of each node are fully utilized and maximize the overall training speed and efficiency. Next, by asynchronously loading the model training tasks on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node, the model asynchronous parallel training can be fully utilized to accelerate the training process. This step can process multiple training tasks at the same time through asynchronous parallel training, thereby improving the throughput and training efficiency of the intelligent AI model. Finally, by asynchronously updating the model parameters of the intelligent AI asynchronous parallel training model, the model parameters can be dynamically updated during the training process, allowing the model to adapt to new data and situations more quickly, which helps to improve the generalization ability and training effect of the model, making the final training results more accurate and reliable.

[0049] Preferably, step S3 comprises the following steps:

[0050] Step S31: Dynamically adjust the weight of the intelligent AI model training parameters of each distributed computing node to obtain the model training contribution balance parameters of each distributed computing node;

[0051] Step S32: performing model hierarchical aggregation on the intelligent AI sub-models on each distributed computing node based on the model training contribution balance parameters of each distributed computing node to generate an intelligent AI global model;

[0052] Step S33: performing model compression processing on the intelligent AI global model to obtain an intelligent AI global compression model;

[0053] Step S34: Perform model pruning optimization on the intelligent AI global compression model to obtain an intelligent AI global optimization model.

[0054] The present invention firstly adjusts the weight of the intelligent AI model training parameters of each distributed computing node dynamically, and can adjust the parameter weight according to the contribution degree of the node and the resource utilization to achieve the contribution balance of each node in the training process. The effect of this step is to optimize the model training contribution between nodes, avoid excessive load or resource waste of some nodes, and improve the overall efficiency and performance of model training. Secondly, by performing model hierarchical aggregation on the intelligent AI sub-models on each distributed computing node based on the model training contribution balance parameters of each distributed computing node, the local model information of each node can be effectively integrated to generate an intelligent AI global model. The effect of this step is to achieve the cooperation and fusion of the models between nodes, improve the performance and generalization ability of the global model, and lay the foundation for subsequent model optimization and application. Then, by performing model compression processing on the intelligent AI global model, the parameter amount and computational complexity of the model can be reduced, and the lightweight and deployment efficiency of the model can be improved. The effect of this step is to optimize the storage and computing resource consumption of the model, making the deployment of the model on the mobile terminal or edge device more efficient and feasible. Finally, by performing model pruning optimization on the intelligent AI global compression model, the sparsity and efficiency of the model can be further improved by pruning redundant parameters and connections. The effect of this step is that it refines the structure of the model, reduces the complexity and computational complexity of the model, and enables the model to perform better in a resource-constrained environment, thereby improving the model’s reasoning speed and performance.

[0055] Preferably, the present invention provides a construction and operation system of an intelligent AI, which is used to execute the construction and operation method of the intelligent AI as described above, and the construction and operation system of the intelligent AI includes:

[0056] The training resource division module required by the model is used to obtain the training resource data required by the intelligent AI model, and clean the training resource data required by the intelligent AI model to obtain the training standard data required by the intelligent AI model; by deploying distributed computing nodes in the network, and dividing the training resource for the standard data required by the intelligent AI model based on the distributed computing nodes, the intelligent AI model training resource sub-data of each distributed computing node is obtained;

[0057] A distributed node model parallel training module is used to construct and initialize an intelligent AI sub-model on each distributed computing node to obtain an intelligent AI sub-model on each distributed computing node; based on the intelligent AI model training resource sub-data of each distributed computing node, the intelligent AI sub-model on each distributed computing node is trained in model parallel to obtain intelligent AI model training parameters of each distributed computing node;

[0058] The intelligent AI model hierarchical aggregation optimization module is used to perform hierarchical aggregation of the intelligent AI sub-models on each distributed computing node based on the intelligent AI model training parameters of each distributed computing node to generate an intelligent AI global model; and to optimize the intelligent AI global model to obtain an intelligent AI global optimization model;

[0059] The intelligent AI model online operation module is used to deploy the intelligent AI global optimization model to various distributed computing nodes for online operation to obtain an intelligent AI online operation model.

[0060] In summary, the present invention provides a construction and operation system for intelligent AI, which consists of a training resource partitioning module required for the model, a distributed node model parallel training module, an intelligent AI model hierarchical aggregation optimization module, and an intelligent AI model online operation module. It can realize any intelligent AI construction and operation method described in the present invention, and is used to combine the operations between computer programs running on each module to realize any intelligent AI construction and operation method. The internal structures of the system cooperate with each other, which can greatly reduce repetitive work and manpower investment, and can quickly and effectively provide a more accurate and efficient intelligent AI construction and operation process, thereby simplifying the operation process of the intelligent AI construction and operation system.

[0061] Preferably, the present invention also provides a PC host, comprising a processor, a memory, and a computer control program stored in the memory and executable on the processor, for executing the method for constructing and operating the intelligent AI as described above. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0063] Figure 1 A schematic diagram of the steps of the method for constructing and operating the intelligent AI of the present invention;

[0064] Figure 2 for Figure 1 Detailed step flow diagram of step S1;

[0065] Figure 3for Figure 2 Detailed step flow chart of step S15 in FIG. DETAILED DESCRIPTION

[0066] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0067] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0068] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0069] To achieve this, please refer to Figures 1 to 3 The present invention provides a method for constructing and operating an intelligent AI, the method comprising the following steps:

[0070] Step S1: Acquire the training resource data required for the intelligent AI model, and clean the training resource data required for the intelligent AI model to obtain the training standard data required for the intelligent AI model; deploy distributed computing nodes in the network, and divide the training resources for the training standard data required for the intelligent AI model based on the distributed computing nodes to obtain the intelligent AI model training resource sub-data of each distributed computing node;

[0071] Step S2: constructing and initializing intelligent AI sub-models on each distributed computing node to obtain intelligent AI sub-models on each distributed computing node; performing model parallel training on the intelligent AI sub-models on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node to obtain intelligent AI model training parameters of each distributed computing node;

[0072] Step S3: Based on the intelligent AI model training parameters of each distributed computing node, the intelligent AI sub-models on each distributed computing node are aggregated in a hierarchical manner to generate an intelligent AI global model; the intelligent AI global model is optimized to obtain an intelligent AI global optimization model;

[0073] Step S4: The intelligent AI global optimization model is deployed to each distributed computing node for online operation to obtain an intelligent AI online operation model.

[0074] In the embodiment of the present invention, please refer to Figure 1 As shown, it is a schematic diagram of the steps of the construction and operation method of the intelligent AI of the present invention. In this example, the construction and operation method of the intelligent AI includes the following steps:

[0075] Step S1: Acquire the training resource data required for the intelligent AI model, and clean the training resource data required for the intelligent AI model to obtain the training standard data required for the intelligent AI model; deploy distributed computing nodes in the network, and divide the training resources for the training standard data required for the intelligent AI model based on the distributed computing nodes to obtain the intelligent AI model training resource sub-data of each distributed computing node;

[0076] The embodiment of the present invention first determines the type and source of training resource data required by the intelligent AI model, and obtains data from the determined data source (including sensors, databases, API interfaces, file systems, etc.), including data in various forms such as images, text, and voice, so as to obtain the training resource data required for the intelligent AI model. Secondly, the collected training resource data required for the intelligent AI model is processed by using the corresponding data cleaning method to effectively remove outliers and noise interference information in the data, and interpolate to compensate for missing information in the data, and standardize the data so that the data has a similar scale in the feature dimension, avoiding model training instability and convergence difficulties caused by data scale differences, thereby obtaining the standard training data required for the intelligent AI model. Then, distributed computing nodes are deployed by selecting appropriate locations and quantities in the network according to the requirements of the intelligent AI system and the nature of the computing tasks, and the upper and lower limits of the data capacity limit are determined on each distributed computing node. The training standard data required for the intelligent AI model is divided into appropriate sub-data sets, taking into account the uniformity of data distribution and the parallelism of training tasks on each node, so as to allocate the divided sub-data sets to each node and ensure that each node can effectively use its resources for training, and finally obtain the intelligent AI model training resource sub-data of each distributed computing node.

[0077] Step S2: constructing and initializing intelligent AI sub-models on each distributed computing node to obtain intelligent AI sub-models on each distributed computing node; performing model parallel training on the intelligent AI sub-models on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node to obtain intelligent AI model training parameters of each distributed computing node;

[0078] The embodiment of the present invention constructs a network structure of an intelligent AI sub-model on each distributed computing node according to the model architecture and requirements, and initializes the parameters of the AI ​​model network structure on the corresponding node by using standard initialization methods (such as Xavier initialization, He initialization, etc.) to achieve distributed deployment and initialization of the model, thereby obtaining the intelligent AI sub-model on each distributed computing node. Secondly, by using encryption technology or authentication mechanism to design and establish a secure network communication channel, the allocated intelligent AI model training resource sub-data of each distributed computing node is accurately distributed to each distributed computing node, and by combining the intelligent AI model training resource sub-data allocated on each distributed computing node, the synchronous training tasks of the intelligent AI models on the corresponding distributed computing nodes are synchronously trained in parallel, so that each node can execute the training tasks at the same time, and synchronously update the model parameters on each node according to periodic or preset conditions within the same time step. At the same time, by combining the intelligent AI model training resource sub-data allocated on each distributed computing node, the intelligent AI sub-models on the corresponding distributed computing nodes are asynchronously trained in parallel, so as to make full use of the parallel computing power of the distributed computing environment on the node to independently execute the corresponding model training tasks, and independently update their own model parameters according to the asynchronously loaded training data and the computing power of the node, thereby obtaining the corresponding synchronous parallel training parameters and asynchronous parallel training parameters. Then, by collecting the synchronous parallel training parameters and asynchronous parallel training parameters of each distributed computing node and performing arithmetic average analysis on them, the intelligent AI model training parameters of each node are reasonably integrated to obtain unified model training parameters, and finally the intelligent AI model training parameters of each distributed computing node are obtained.

[0079] Step S3: Based on the intelligent AI model training parameters of each distributed computing node, the intelligent AI sub-models on each distributed computing node are aggregated in a hierarchical manner to generate an intelligent AI global model; the intelligent AI global model is optimized to obtain an intelligent AI global optimization model;

[0080] The embodiment of the present invention monitors the model training status and performance indicators of each distributed computing node in real time, such as loss function, accuracy, etc., and uses a distributed optimization algorithm or an adaptive learning rate algorithm to dynamically adjust the model training parameter weights of the nodes according to the model training status contribution and performance performance of each node obtained by monitoring, and integrates and aggregates the hierarchical structure of the intelligent AI sub-models on the corresponding distributed computing nodes by combining the model training parameters of each distributed computing node obtained after weight adjustment, using weighted average, model distillation and other technologies to form a global intelligent AI model, and retain the important information contributed by each node, thereby generating an intelligent AI global model. Secondly, by using corresponding compression technologies (such as weight sharing, low-rank approximation and other methods) to compress the intelligent AI global model obtained after aggregation, the balance between the performance loss and compression ratio of the model is considered to reduce the number of parameters and the computational complexity of the model, and by using corresponding pruning algorithms (such as structured pruning, parameter sensitivity pruning and other methods) to prune the intelligent AI global compression model obtained after compression, so as to further reduce the complexity of the model and the consumption of computing resources, and remove redundant connections and parameters in the model aggregation, and finally obtain an intelligent AI global optimization model.

[0081] Step S4: The intelligent AI global optimization model is deployed to each distributed computing node for online operation to obtain an intelligent AI online operation model.

[0082] The embodiment of the present invention deploys the optimized intelligent AI global optimization model to each distributed computing node for online operation to realize real-time application and service of the intelligent AI model, and monitors the model performance and operation status in real time to adjust and optimize the model in time, and finally obtains the intelligent AI online operation model.

[0083] The present invention first obtains the training resource data required for the intelligent AI model, which is the first step in building an efficient model. This step can ensure that the model has enough information during the training process by collecting a large amount of resource data required for model training, thereby improving the accuracy and reliability of its predictions and decisions, and providing a sufficient data basis for subsequent data processing and model training, thereby laying a solid foundation for building an intelligent AI model with good generalization capabilities. By performing data cleaning processing on the training resource data required for the intelligent AI model, outliers and interference information in the data can be effectively removed to improve data quality. This processing can reduce the interference of noise data on the model during training, thereby making the model more robust and reliable. At the same time, standardization processing can make the data have similar scales in feature dimensions, avoiding model training instability and convergence difficulties caused by data scale differences, and can also improve the training efficiency and convergence speed of the model, thereby accelerating the model training process. At the same time, by deploying distributed computing nodes in the network and dividing the training resource data for processing, it is possible to fully utilize distributed computing resources and accelerate the model training process. The key to this step is to improve the parallelism and efficiency of model training, so that larger data sets can be processed and model training tasks can be completed in a shorter time, thereby shortening the model training time. Secondly, by building and initializing intelligent AI sub-models on each distributed computing node, the distributed deployment and initialization of the model can be achieved to prepare for the subsequent training process. The key to this step is to be able to establish an independent intelligent AI sub-model on each node, so that each node can independently perform model training, thereby improving the parallelism and overall training speed of the AI ​​model. By performing model synchronization and parallel training on the intelligent AI sub-models on each node based on the intelligent AI model training resource sub-data of each distributed computing node, the model parameters between nodes can be updated synchronously, which improves the efficiency and speed of model training. In this way, the distributed computing resources can be fully utilized, and the synchronous training of model parameters can be realized, thereby speeding up the overall training process of the intelligent AI model. By performing asynchronous parallel training on the model, the model parameters between nodes can be updated asynchronously, which improves the flexibility and stability of model training. The key to this step is to give full play to the computing power of each node and realize asynchronous training of model parameters, so that model training can better adapt to the computing resource conditions of different nodes. In addition, by performing arithmetic average analysis on the model synchronous parallel training parameters and the model asynchronous parallel training parameters of each distributed computing node, the training results of each node can be comprehensively considered to obtain the final intelligent AI model training parameters. The key to this step is to reasonably integrate the training results of each node to obtain unified training parameters, which provides a reliable parameter basis for the final model construction and application, thereby reducing the complexity of the algorithm of the intelligent AI model training process.Finally, by deploying the intelligent AI global optimization model to various distributed computing nodes for online operation, the real-time application and service of the intelligent AI model can be realized. The key to this step is to be able to apply the trained model to actual scenarios to provide users with intelligent services and decision support. By running the model online, the model performance can be continuously optimized, thereby improving the intelligence level of the intelligent AI system and user experience.

[0084] Preferably, step S1 comprises the following steps:

[0085] Step S11: Obtaining training resource data required for the intelligent AI model;

[0086] Step S12: performing noise data removal processing on the training resource data required for the intelligent AI model to obtain the intelligent AI model training resource denoised data;

[0087] Step S13: performing a missing value test on the intelligent AI model training resource denoising data to obtain the missing status of the intelligent AI model training resource data; performing interpolation compensation processing on the intelligent AI model training resource denoising data based on the missing status of the intelligent AI model training resource data to obtain the intelligent AI model training resource missing compensation data;

[0088] Step S14: standardizing the intelligent AI model training resource loss compensation data to obtain the training standard data required for the intelligent AI model;

[0089] Step S15: By deploying distributed computing nodes in the network, and dividing the training resources of the standard training data required for the intelligent AI model based on the distributed computing nodes, the intelligent AI model training resource sub-data of each distributed computing node is obtained.

[0090] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Detailed step flow diagram of step S1 in FIG. 1 , in this embodiment, step S1 includes the following steps:

[0091] Step S11: Obtaining training resource data required for the intelligent AI model;

[0092] The embodiment of the present invention first determines the type and source of training resource data required by the intelligent AI model, and obtains data from the determined data source (including sensors, databases, API interfaces, file systems, etc.), including data in various forms such as images, text, and voice, to ensure that the model has sufficient information during the training process, and finally obtains the training resource data required by the intelligent AI model.

[0093] Step S12: performing noise data removal processing on the training resource data required for the intelligent AI model to obtain the intelligent AI model training resource denoised data;

[0094] The embodiment of the present invention eliminates the noise of the collected training resource data required for the intelligent AI model by using corresponding noise filtering methods (such as Gaussian filtering, smoothing filtering, etc.), so as to effectively remove outliers and noise interference information in the data, and finally obtain the denoised data of the intelligent AI model training resources.

[0095] Step S13: performing a missing value test on the intelligent AI model training resource denoising data to obtain the missing status of the intelligent AI model training resource data; performing interpolation compensation processing on the intelligent AI model training resource denoising data based on the missing status of the intelligent AI model training resource data to obtain the intelligent AI model training resource missing compensation data;

[0096] The embodiment of the present invention uses a missing value test method to test the denoised data of the intelligent AI model training resources obtained after denoising, so as to check whether there are missing values ​​in the denoised data and determine the distribution of the missing values, thereby obtaining the missing status of the intelligent AI model training resource data. Then, by combining the missing status of the intelligent AI model training resource data obtained through analysis, a suitable interpolation method (such as linear interpolation, polynomial interpolation, KNN interpolation, etc.) is used to interpolate and compensate the corresponding intelligent AI model training resource denoised data, so as to effectively interpolate the missing information in the compensation data, and finally obtain the missing compensation data of the intelligent AI model training resources.

[0097] Step S14: standardizing the intelligent AI model training resource loss compensation data to obtain the training standard data required for the intelligent AI model;

[0098] The embodiment of the present invention uses corresponding standardization methods (such as Z-score standardization, MinMax standardization, etc.) to standardize the intelligent AI model training resource missing compensation data obtained after missing compensation, so that the data has a similar scale in the feature dimension, and avoids model training instability and convergence difficulties caused by data scale differences, and finally obtains the training standard data required for the intelligent AI model.

[0099] Step S15: By deploying distributed computing nodes in the network, and dividing the training resources of the standard training data required for the intelligent AI model based on the distributed computing nodes, the intelligent AI model training resource sub-data of each distributed computing node is obtained.

[0100] The embodiment of the present invention deploys distributed computing nodes by selecting appropriate locations and quantities in the network according to the requirements of the intelligent AI system and the nature of the computing tasks, and deploys corresponding network monitoring tools on each distributed computing node or uses the system's own performance monitoring function to monitor the computing resources on each node in real time, including CPU, memory, network bandwidth, disk IO, etc. At the same time, by checking and recording the capacity information of the corresponding storage device on each distributed computing node, including the size of the hard disk, solid-state hard disk or other storage medium, the data storage space size of each node is evaluated and determined, and the network resource capacity of each corresponding distributed computing node is determined according to the analyzed computing resource situation and the storage space size, so as to determine the upper and lower limits of the data capacity limit on each node according to the computing resource capacity status and the actual storage space situation. Then, the training standard data required by the intelligent AI model is divided into appropriate sub-data sets according to the upper and lower limits of the data capacity on each distributed computing node, so as to take into account the uniformity of data distribution and the parallelism of training tasks on each node, and the divided sub-data sets are allocated to each node, and it is ensured that each node can effectively use its resources for training, and finally the intelligent AI model training resource sub-data of each distributed computing node is obtained.

[0101] The present invention first obtains the training resource data required by the intelligent AI model, which is the first step in building an efficient model. This step can ensure that the model has enough information during the training process by collecting a large amount of resource data required for model training, thereby improving the accuracy and reliability of its prediction and decision-making, and providing a sufficient data basis for subsequent data processing and model training, thereby laying a solid foundation for building an intelligent AI model with good generalization ability. Secondly, by performing noise data cleaning on the training resource data required for the intelligent AI model, outliers and interference information in the data can be effectively removed to improve data quality. This processing can reduce the interference of noise data on the model during the training process, thereby making the model more robust and reliable. Then, by performing missing value inspection on the training resource data that has been processed by noise cleaning and performing interpolation compensation processing, the missing information in the data can be effectively processed to ensure the integrity and availability of the data. Through this step, it can be ensured that the model can make full use of all available data information during the training process, thereby improving the training effect and performance of the model. Next, by standardizing the training resource data that has been processed for missing compensation, the data can have a similar scale in the feature dimension, avoiding the instability and convergence difficulties of the model training caused by the difference in data scale. This step can improve the training efficiency and convergence speed of the model, thereby accelerating the model training process. Finally, by deploying distributed computing nodes in the network and dividing the training resource data, the distributed computing resources can be fully utilized to accelerate the model training process. The key to this step is to improve the parallelism and efficiency of model training, so that larger data sets can be processed and model training tasks can be completed in a shorter time.

[0102] Preferably, step S15 comprises the following steps:

[0103] Step S151: Deploy distributed computing nodes in the network and perform computing resource statistics analysis on each distributed computing node to obtain computing resource status data of each distributed computing node;

[0104] Step S152: Performing statistical analysis on the data storage space of each distributed computing node to obtain the size of the data storage space range of each distributed computing node;

[0105] Step S153: performing data resource accommodation evaluation and analysis on each distributed computing node based on the computing resource status data of each distributed computing node and the size of the data storage space range, and obtaining the data resource accommodation limit range on each distributed computing node;

[0106] Step S154: Perform training resource division processing on the standard training data required for the intelligent AI model according to the data resource accommodation limit range on each distributed computing node to obtain the intelligent AI model training resource sub-data of each distributed computing node.

[0107] As an embodiment of the present invention, refer to Figure 3 As shown, Figure 2 Detailed step flow diagram of step S15 in FIG. 1 , in this embodiment, step S15 includes the following steps:

[0108] Step S151: Deploy distributed computing nodes in the network and perform computing resource statistics analysis on each distributed computing node to obtain computing resource status data of each distributed computing node;

[0109] The embodiments of the present invention deploy distributed computing nodes by selecting appropriate locations and quantities in the network according to the requirements of the intelligent AI system and the nature of the computing tasks, and deploy corresponding network monitoring tools on each distributed computing node or use the system's own performance monitoring function to monitor the computing resources on each node in real time, including CPU, memory, network bandwidth, disk IO, etc., and finally obtain the computing resource status data of each distributed computing node.

[0110] Step S152: Performing statistical analysis on the data storage space of each distributed computing node to obtain the size of the data storage space range of each distributed computing node;

[0111] The embodiment of the present invention evaluates the storage capacity of each node and determines the size of the data storage space of each node by checking and recording the capacity information of the corresponding storage device on each distributed computing node, including the size of the hard disk, solid-state drive or other storage media, and finally obtains the range size of the data storage space of each distributed computing node.

[0112] Step S153: performing data resource accommodation evaluation and analysis on each distributed computing node based on the computing resource status data of each distributed computing node and the size of the data storage space range, and obtaining the data resource accommodation limit range on each distributed computing node;

[0113] The embodiment of the present invention determines the network resource capacity of each corresponding distributed computing node by combining the computing resource status data of each distributed computing node obtained by analysis and the size of the data storage space range, so as to determine the upper and lower limits of the data capacity limit on each node according to the computing resource capacity status and the actual storage space situation, and finally obtain the data resource capacity limit range on each distributed computing node.

[0114] Step S154: Perform training resource division processing on the standard training data required for the intelligent AI model according to the data resource accommodation limit range on each distributed computing node to obtain the intelligent AI model training resource sub-data of each distributed computing node.

[0115] The embodiment of the present invention divides the standard training data required for the intelligent AI model into appropriate sub-data sets according to the data resource accommodation limit range on each distributed computing node, takes into account the uniformity of data distribution and the parallelism of training tasks on each node, distributes the divided sub-data sets to each node, and ensures that each node can effectively use its resources for training, and finally obtains the intelligent AI model training resource sub-data of each distributed computing node.

[0116] The present invention first deploys distributed computing nodes in the network and performs statistical analysis of computing resources on each node. The key to this step is to fully understand the computing power and load of each node, and provide data support for subsequent resource allocation and task allocation to ensure that each node can fully utilize its computing resources, thereby improving the performance and efficiency of the overall model training process. Secondly, by performing statistical analysis of data storage space on each distributed computing node, the key to this step is to evaluate the storage capacity of each node to determine the range of data that the node can accommodate, and provide a basis for subsequent data storage and management, thereby ensuring the balanced distribution of data between nodes and the rationality of storage. Then, by performing data resource accommodation evaluation and analysis on the nodes based on the computing resource status data and data storage space range size of each distributed computing node, the key to this step is to determine the range of data resources that each node can carry, and provide guidance for resource allocation and task scheduling, so as to ensure that model training can effectively process data and tasks under high load. Finally, the training standard data required by the intelligent AI model is divided and processed according to the data resource capacity limit range on each distributed computing node. The key to this step is to be able to reasonably distribute the training resource data to each node, make full use of the computing and storage resources of each node, realize parallel training, improve training efficiency and speed, thereby speeding up the model training process and improving the overall performance of the system.

[0117] Preferably, step S153 includes the following steps:

[0118] Performing CPU operation utilization statistics analysis on the computing resource status data of each distributed computing node to obtain the CPU operation utilization of each distributed computing node;

[0119] The embodiment of the present invention installs or uses a monitoring tool on each distributed computing node to monitor and evaluate the CPU resource utilization of each node in real time, so as to help determine the computing power and load of each node, and perform statistical analysis on the collected data to calculate the CPU utilization on each node, and finally obtain the CPU operation utilization of each distributed computing node.

[0120] Preferably, a network bandwidth utilization statistical analysis is performed on the computing resource status data of each distributed computing node to obtain the network bandwidth utilization of each distributed computing node;

[0121] The embodiment of the present invention monitors the usage of network bandwidth on each distributed computing node by using a network monitoring tool or a built-in performance monitoring function, and calculates the utilization rate of the network bandwidth on each node by performing statistical analysis on the collected network bandwidth utilization, and finally obtains the network bandwidth utilization rate of each distributed computing node.

[0122] Preferably, the computing resource utilization capacity of each distributed computing node is evaluated and analyzed based on the CPU operation utilization rate and network bandwidth utilization rate of each distributed computing node to obtain computing resource utilization capacity data of each distributed computing node;

[0123] The embodiment of the present invention uses a computing resource utilization capacity evaluation method to evaluate and analyze the corresponding distributed computing nodes by combining the CPU operation utilization rate and network bandwidth utilization rate of each distributed computing node obtained by analysis, so as to comprehensively evaluate the computing resource utilization capacity on each distributed computing node, and finally obtain the computing resource utilization capacity data of each distributed computing node.

[0124] Preferably, the data storage capacity of each distributed computing node is evaluated and analyzed based on the data storage space range size of each distributed computing node to obtain the space storage capacity data of each distributed computing node;

[0125] The embodiment of the present invention evaluates and analyzes each corresponding distributed computing node by combining the data storage space range size of each distributed computing node obtained by analysis, so as to analyze the storage requirements and storage space utilization on each node to evaluate the data storage capacity of each node, and finally obtain the space storage capacity data of each distributed computing node.

[0126] Preferably, the resource accommodation limit interval of each distributed computing node is determined according to the computing resource utilization capacity data and the space storage capacity data of each distributed computing node, so as to obtain the data resource accommodation limit interval on each distributed computing node.

[0127] The embodiment of the present invention determines the network resource capacity of each corresponding distributed computing node based on the computing resource utilization capacity data and space storage capacity data of each distributed computing node obtained through evaluation and analysis, so as to determine the upper and lower limits of the data capacity limit on each node according to the computing resource capacity requirements and the actual storage space situation, and ensure the effective utilization and stable operation of the node resources, and finally obtain the data resource capacity limit range on each distributed computing node.

[0128] The present invention firstly performs CPU operation utilization statistics analysis on the computing resource status data of each distributed computing node. The key of this step is to be able to comprehensively evaluate the CPU resource utilization of each node to help determine the computing power and load of the node, thereby providing an important reference for subsequent resource allocation and task scheduling. Secondly, the network bandwidth utilization statistics analysis is performed on the computing resource status data of each distributed computing node to evaluate the network communication status between nodes and determine the utilization of network bandwidth, thereby providing basic data guarantee for subsequent processing. Then, the computing resource utilization capacity of the node is evaluated and analyzed based on the CPU operation utilization and network bandwidth utilization of each distributed computing node. The key of this step is to comprehensively consider the utilization of CPU and network resources to evaluate the overall computing capacity of the node, thereby providing comprehensive guidance for resource allocation and task scheduling. Next, the data storage capacity of each distributed computing node is evaluated and analyzed based on the data storage space range size of each distributed computing node. The key of this step is to be able to evaluate the storage capacity of the node to determine the amount of data that the node can carry, provide guidance for data storage and management, thereby ensuring the balanced distribution of data between nodes and storage rationality. Finally, the resource capacity limit range of each node is determined based on the computing resource utilization capacity data and storage capacity data of each distributed computing node. The key to this step is to comprehensively consider the utilization of computing and storage resources to determine the data capacity range of the node, providing an accurate reference for resource allocation and task scheduling, thereby ensuring the efficient operation and data processing capabilities of the subsequent model parallel training process.

[0129] Preferably, step S2 comprises the following steps:

[0130] Step S21: constructing and initializing an intelligent AI sub-model on each distributed computing node to obtain an intelligent AI sub-model on each distributed computing node;

[0131] The embodiment of the present invention constructs a network structure of an intelligent AI sub-model on each distributed computing node according to the model architecture and requirements, and ensures the consistency of the construction of the sub-model with the overall model to facilitate subsequent model synchronization and parameter updates. At the same time, by using standard initialization methods (such as Xavier initialization, He initialization, etc.), the parameters of the AI ​​model network structure on the corresponding node are initialized to achieve distributed deployment and initialization of the model, and ensure that an independent intelligent AI sub-model is established on each node, and finally the intelligent AI sub-model on each distributed computing node is obtained.

[0132] Step S22: Distribute the intelligent AI model training resource sub-data of each distributed computing node to each distributed computing node by building a secure network communication channel;

[0133] The embodiment of the present invention designs and establishes a secure network communication channel by using encryption technology or authentication mechanism to accurately distribute the intelligent AI model training resource sub-data of each distributed computing node to each distributed computing node, so as to ensure the security and integrity of data transmission and avoid data loss or transmission delay.

[0134] Step S23: Based on the intelligent AI model training resource sub-data of each distributed computing node, the intelligent AI sub-model on each distributed computing node is trained synchronously and in parallel to obtain the intelligent AI sub-model synchronous and parallel training parameters of each distributed computing node;

[0135] The embodiment of the present invention combines the intelligent AI model training resource sub-data allocated on each distributed computing node to perform synchronous and parallel training of the model on the intelligent AI model synchronization training tasks corresponding to each distributed computing node, so that each node can execute the training task at the same time, and synchronously update the model parameters on each node according to periodic or preset conditions within the same time step, to ensure the consistency of the intelligent AI sub-model and the effectiveness of the training, thereby synchronizing the model parameters on each node to other nodes, and finally obtaining the synchronous and parallel training parameters of the intelligent AI sub-models of each distributed computing node.

[0136] Step S24: performing asynchronous parallel training on the intelligent AI sub-models on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node, and obtaining asynchronous parallel training parameters of the intelligent AI sub-models of each distributed computing node;

[0137] The embodiment of the present invention combines the intelligent AI model training resource sub-data allocated on each distributed computing node to perform asynchronous parallel training on the intelligent AI sub-models on the corresponding distributed computing nodes, so as to make full use of the parallel computing capability of the distributed computing environment on the node to independently execute the corresponding model training tasks, and independently update its own model parameters according to the asynchronously loaded training data and the computing power of the node, ensuring that the model parameters between each node are consistent, and finally obtaining the asynchronous parallel training parameters of the intelligent AI sub-models of each distributed computing node.

[0138] Step S25: Perform arithmetic mean analysis on the synchronous parallel training parameters of the intelligent AI sub-models of each distributed computing node and the asynchronous parallel training parameters of the intelligent AI sub-models to obtain the intelligent AI model training parameters of each distributed computing node.

[0139] The embodiment of the present invention collects the synchronous parallel training parameters of the intelligent AI sub-models and the asynchronous parallel training parameters of the intelligent AI sub-models of each distributed computing node and performs arithmetic average analysis on them to reasonably integrate the intelligent AI model training parameters of each node, thereby obtaining unified model training parameters, and finally obtaining the intelligent AI model training parameters of each distributed computing node.

[0140] The present invention first constructs and initializes the intelligent AI sub-model on each distributed computing node, which can realize the distributed deployment and initialization of the model and prepare for the subsequent training process. The key to this step is to establish an independent intelligent AI sub-model on each node, so that each node can independently perform model training, thereby improving the parallelism and overall training speed of the AI ​​model. Secondly, by constructing a secure network communication channel to distribute the intelligent AI model training resource sub-data of each distributed computing node to each distributed computing node, the security and integrity of the data transmission process can be ensured. The key to this step is to achieve the secure transmission of data, ensure the integrity and confidentiality of the training data, and provide reliable data support for subsequent model training. Then, by performing model synchronization and parallel training on the intelligent AI sub-models on each node based on the intelligent AI model training resource sub-data of each distributed computing node, the synchronous update of model parameters between nodes can be realized, and the efficiency and speed of model training can be improved. In this way, the distributed computing resources can be fully utilized, and the synchronous training of model parameters can be realized, thereby speeding up the overall training process of the intelligent AI model. Next, by performing asynchronous parallel training on the intelligent AI sub-models on each node based on the intelligent AI model training resource sub-data of each distributed computing node, the model parameters between nodes can be asynchronously updated, which improves the flexibility and stability of model training. The key to this step is to give full play to the computing power of each node and realize asynchronous training of model parameters, so that model training can better adapt to the computing resource conditions of different nodes. Finally, by performing arithmetic mean analysis on the synchronous parallel training parameters of the intelligent AI sub-models of each distributed computing node and the asynchronous parallel training parameters of the intelligent AI sub-models, the training results of each node can be comprehensively considered to obtain the final intelligent AI model training parameters. The key to this step is to reasonably integrate the training results of each node to obtain unified training parameters, thereby providing a reliable parameter basis for the final model construction and application.

[0141] Preferably, step S23 includes the following steps:

[0142] Step S231: performing synchronous training task allocation on the intelligent AI sub-models on each distributed computing node to obtain the intelligent AI model synchronous training task on each distributed computing node;

[0143] The embodiment of the present invention designs a corresponding synchronous training task allocation algorithm by considering factors such as communication delays, computing resources, and network bandwidth between each node, and synchronously allocates training tasks to each distributed computing node to ensure that each node is responsible for processing a part of the training tasks and avoid excessive concentration or uneven distribution of model training tasks, ultimately obtaining synchronous training tasks for intelligent AI models on each distributed computing node.

[0144] Step S232: Based on the intelligent AI model training resource sub-data of each distributed computing node, the intelligent AI model synchronization training task on each distributed computing node is trained in parallel to obtain an intelligent AI synchronization parallel training model on each distributed computing node;

[0145] The embodiment of the present invention uses the intelligent AI model training resource sub-data allocated on each distributed computing node to perform synchronous and parallel training of the model for the intelligent AI model synchronous training tasks on the corresponding distributed computing nodes, so that each node can execute the training tasks at the same time, and the model parameters and gradient information need to be synchronized between each node within the same time step, and finally the intelligent AI synchronous and parallel training model on each distributed computing node is obtained.

[0146] Step S233: Perform model parameter synchronization update processing on the intelligent AI synchronous parallel training model on each distributed computing node to obtain the intelligent AI sub-model synchronous parallel training parameters of each distributed computing node.

[0147] The embodiment of the present invention designs a parameter synchronization strategy to synchronously update the model parameters of the intelligent AI synchronous parallel training model on each distributed computing node by considering the communication cost and delay between nodes and the training requirements of each sub-model, so as to synchronously update the model parameters on each node according to periodic or preset conditions, and ensure the consistency of the model and the effectiveness of the training, thereby synchronizing the model parameters on each node to other nodes, and finally obtaining the intelligent AI sub-model synchronous parallel training parameters of each distributed computing node.

[0148] The present invention first distributes the training tasks to the intelligent AI sub-models on each distributed computing node synchronously, so as to effectively and reasonably distribute the training tasks to each node to ensure that each node participates in the training process. The key to this step is that it can optimize the distribution and scheduling of training tasks, make full use of distributed computing resources, and improve the training efficiency and overall performance of the intelligent AI model. Then, by performing model synchronization and parallel training on the intelligent AI model synchronous training tasks on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node, the synchronous update of model parameters between nodes can be achieved, and the efficiency and speed of model training can be improved. The key to this step is that the training tasks of each node can be coordinated, the synchronous update of model parameters can be achieved, and the overall training process can be accelerated. Finally, by performing model parameter synchronization update processing on the intelligent AI synchronous parallel training model on each distributed computing node, the consistency and synchronization of model parameters between nodes can be guaranteed, and a reliable parameter basis is provided for the construction of the final model. The effect of this step is to ensure the uniformity of model parameters, improve the training stability and performance of the model, and provide a reliable basis for subsequent model parameter analysis.

[0149] Preferably, step S24 comprises the following steps:

[0150] Step S241: performing asynchronous training task allocation on the intelligent AI sub-model on each distributed computing node to obtain the asynchronous training task of the intelligent AI model on each distributed computing node;

[0151] The embodiment of the present invention also designs an asynchronous training task allocation algorithm by considering factors such as communication delays, computing resources, and network bandwidth between each distributed computing node, and allocates tasks to the intelligent AI sub-models on each distributed computing node, so as to flexibly allocate asynchronous training tasks according to the actual operating conditions and resource utilization of each node, including training operations such as data loading and model updating, while avoiding excessive concentration or uneven distribution of tasks, and ultimately obtaining asynchronous training tasks for intelligent AI models on each distributed computing node.

[0152] Step S242: Perform task load analysis on the intelligent AI model asynchronous training tasks on each distributed computing node to obtain model training task load data on each distributed computing node;

[0153] The embodiment of the present invention monitors the execution status of training tasks on each node in real time, including the task start time, end time, computing resource utilization, etc., and collects node load information therefrom, including the utilization status of resources such as CPU, memory, GPU, and the speed and efficiency of task execution, and at the same time determines the processing capacity and load level of each node, and finally obtains the model training task load data on each distributed computing node.

[0154] Step S243: asynchronously load the intelligent AI model training tasks on each distributed computing node according to the model training task load data on each distributed computing node, and obtain the model asynchronous loading training tasks on each distributed computing node;

[0155] The embodiment of the present invention designs an asynchronous loading strategy by combining the model training task load data on each distributed computing node obtained by analysis and considering factors such as the node load level and network bandwidth to determine the loading order and speed of the training tasks on each node to reasonably arrange the loading order and priority of the tasks, and asynchronously loads the training tasks to each distributed computing node according to the designed asynchronous loading strategy. At the same time, during the task loading process, by considering the dependencies between tasks and the efficiency of data transmission, it is ensured that the loading process will not affect the normal operation of the node, and finally the model asynchronous loading training tasks on each distributed computing node are obtained.

[0156] Step S244: performing asynchronous parallel model training on the model asynchronous loading training tasks on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node, to obtain the intelligent AI asynchronous parallel training model on each distributed computing node;

[0157] The embodiment of the present invention combines the use of intelligent AI model training resource sub-data allocated on each distributed computing node to perform asynchronous parallel training of the model on the corresponding model asynchronous loading training tasks on each distributed computing node, so as to make full use of the parallel computing capabilities of the distributed computing environment on the node to independently execute the corresponding model training tasks, and finally obtain the intelligent AI asynchronous parallel training model on each distributed computing node.

[0158] Step S245: Asynchronously update the model parameters of the intelligent AI asynchronous parallel training model on each distributed computing node to obtain the asynchronous parallel training parameters of the intelligent AI sub-model of each distributed computing node.

[0159] The embodiments of the present invention asynchronously update the model parameters of the intelligent AI asynchronous parallel training model on each distributed computing node by using a periodic transmission of model parameters or an asynchronous parameter transmission mechanism, so as to independently update its own model parameters according to the asynchronously loaded training data and the computing power of the node, while the update operations between different nodes are independent of each other and ensure that the model parameters between each node remain consistent, and finally obtain the asynchronous parallel training parameters of the intelligent AI sub-model of each distributed computing node.

[0160] The present invention firstly distributes the training tasks asynchronously to the intelligent AI sub-models on each distributed computing node, and can flexibly distribute the training tasks according to the actual operation status and resource utilization of each node, so as to make full use of the computing resources and improve the overall training efficiency. Through asynchronous task distribution, the waiting time between nodes can be avoided, making the training process more coherent and efficient. Secondly, by performing task load analysis on the asynchronous training tasks of the intelligent AI model, the training task conditions on each node can be fully understood, including information such as task type and resource consumption, which helps to optimize the task scheduling strategy, avoid node load imbalance or overload, and thus improve the stability and performance of the overall model training. Then, by asynchronously loading the intelligent AI model training tasks on each distributed computing node according to the model training task load data on each distributed computing node, the loading order and priority of the tasks can be dynamically adjusted according to the actual situation of the node to ensure that the resources of each node are fully utilized and maximize the overall training speed and efficiency. Next, by asynchronously loading the model training tasks on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node, the model asynchronous parallel training can be fully utilized to accelerate the training process. This step can process multiple training tasks at the same time through asynchronous parallel training, thereby improving the throughput and training efficiency of the intelligent AI model. Finally, by asynchronously updating the model parameters of the intelligent AI asynchronous parallel training model, the model parameters can be dynamically updated during the training process, allowing the model to adapt to new data and situations more quickly, which helps to improve the generalization ability and training effect of the model, making the final training results more accurate and reliable.

[0161] Preferably, step S3 comprises the following steps:

[0162] Step S31: Dynamically adjust the weight of the intelligent AI model training parameters of each distributed computing node to obtain the model training contribution balance parameters of each distributed computing node;

[0163] The embodiment of the present invention monitors the model training status and performance indicators of each distributed computing node in real time, such as loss function, accuracy, etc., and uses a distributed optimization algorithm or an adaptive learning rate algorithm to dynamically adjust the model training parameter weights of the nodes according to the model training status contribution and performance performance on each node obtained by monitoring, so as to achieve model training load balancing and optimization of model convergence speed, and finally obtain the model training contribution balance parameters of each distributed computing node.

[0164] Step S32: performing model hierarchical aggregation on the intelligent AI sub-models on each distributed computing node based on the model training contribution balance parameters of each distributed computing node to generate an intelligent AI global model;

[0165] The embodiment of the present invention combines the model training contribution balance parameters of each distributed computing node obtained after weight adjustment, and uses weighted averaging, model distillation and other technologies to hierarchically integrate and aggregate the intelligent AI sub-models on the corresponding distributed computing nodes to form a global intelligent AI model, and retain the important information contributed by each node, thereby making full use of the training results of each node and finally generating an intelligent AI global model.

[0166] Step S33: performing model compression processing on the intelligent AI global model to obtain an intelligent AI global compression model;

[0167] The embodiment of the present invention uses corresponding compression technologies (such as weight sharing, low-rank approximation, etc.) to compress the intelligent AI global model obtained after aggregation, so as to consider the balance between the performance loss and compression ratio of the model, reduce the number of parameters and computational complexity of the model, and ensure that the predictive ability and accuracy of the model are maintained as much as possible while reducing the scale of the model, and finally obtain an intelligent AI global compression model.

[0168] Step S34: Perform model pruning optimization on the intelligent AI global compression model to obtain an intelligent AI global optimization model.

[0169] The embodiment of the present invention uses corresponding pruning algorithms (such as structured pruning, parameter sensitivity pruning, etc.) to prune and optimize the intelligent AI global compression model obtained after compression, so as to further reduce the complexity of the model and the consumption of computing resources, and remove redundant connections and parameters in the model aggregation, and finally obtain the intelligent AI global optimization model.

[0170] The present invention firstly adjusts the weight of the intelligent AI model training parameters of each distributed computing node dynamically, and can adjust the parameter weight according to the contribution degree of the node and the resource utilization to achieve the contribution balance of each node in the training process. The effect of this step is to optimize the model training contribution between nodes, avoid excessive load or resource waste of some nodes, and improve the overall efficiency and performance of model training. Secondly, by performing model hierarchical aggregation on the intelligent AI sub-models on each distributed computing node based on the model training contribution balance parameters of each distributed computing node, the local model information of each node can be effectively integrated to generate an intelligent AI global model. The effect of this step is to achieve the cooperation and fusion of the models between nodes, improve the performance and generalization ability of the global model, and lay the foundation for subsequent model optimization and application. Then, by performing model compression processing on the intelligent AI global model, the parameter amount and computational complexity of the model can be reduced, and the lightweight and deployment efficiency of the model can be improved. The effect of this step is to optimize the storage and computing resource consumption of the model, making the deployment of the model on the mobile terminal or edge device more efficient and feasible. Finally, by performing model pruning optimization on the intelligent AI global compression model, the sparsity and efficiency of the model can be further improved by pruning redundant parameters and connections. The effect of this step is that it refines the structure of the model, reduces the complexity and computational complexity of the model, and enables the model to perform better in a resource-constrained environment, thereby improving the model’s reasoning speed and performance.

[0171] Preferably, the present invention provides a construction and operation system of an intelligent AI, which is used to execute the construction and operation method of the intelligent AI as described above, and the construction and operation system of the intelligent AI includes:

[0172] The training resource division module required by the model is used to obtain the training resource data required by the intelligent AI model, and clean the training resource data required by the intelligent AI model to obtain the training standard data required by the intelligent AI model; by deploying distributed computing nodes in the network, and dividing the training resource for the standard data required by the intelligent AI model based on the distributed computing nodes, the intelligent AI model training resource sub-data of each distributed computing node is obtained;

[0173] A distributed node model parallel training module is used to construct and initialize an intelligent AI sub-model on each distributed computing node to obtain an intelligent AI sub-model on each distributed computing node; based on the intelligent AI model training resource sub-data of each distributed computing node, the intelligent AI sub-model on each distributed computing node is trained in model parallel to obtain intelligent AI model training parameters of each distributed computing node;

[0174] The intelligent AI model hierarchical aggregation optimization module is used to perform hierarchical aggregation of the intelligent AI sub-models on each distributed computing node based on the intelligent AI model training parameters of each distributed computing node to generate an intelligent AI global model; and to optimize the intelligent AI global model to obtain an intelligent AI global optimization model;

[0175] The intelligent AI model online operation module is used to deploy the intelligent AI global optimization model to various distributed computing nodes for online operation to obtain an intelligent AI online operation model.

[0176] In summary, the present invention provides a construction and operation system for intelligent AI, which consists of a training resource partitioning module required for the model, a distributed node model parallel training module, an intelligent AI model hierarchical aggregation optimization module, and an intelligent AI model online operation module. It can realize any intelligent AI construction and operation method described in the present invention, and is used to combine the operations between computer programs running on each module to realize any intelligent AI construction and operation method. The internal structures of the system cooperate with each other, which can greatly reduce repetitive work and manpower investment, and can quickly and effectively provide a more accurate and efficient intelligent AI construction and operation process, thereby simplifying the operation process of the intelligent AI construction and operation system.

[0177] Preferably, the present invention also provides a PC host, comprising a processor, a memory, and a computer control program stored in the memory and executable on the processor, for executing the method for constructing and operating the intelligent AI as described above.

[0178] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0179] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for constructing and operating an intelligent AI, characterized in that: The following steps are involved: Step S1: Acquire the training resource data required by the intelligent AI model, and clean the training resource data required by the intelligent AI model to obtain the training standard data required by the intelligent AI model; By deploying distributed computing nodes in the network, and performing training resource division processing on the training standard data required for the intelligent AI model based on the distributed computing nodes, the intelligent AI model training resource sub-data of each distributed computing node is obtained; wherein step S1 includes the following steps: Step S11: Obtaining training resource data required for the intelligent AI model; Step S12: performing noise data removal processing on the training resource data required for the intelligent AI model to obtain the intelligent AI model training resource denoised data; Step S13: performing a missing value test on the intelligent AI model training resource denoising data to obtain the missing status of the intelligent AI model training resource data; performing interpolation compensation processing on the intelligent AI model training resource denoising data based on the missing status of the intelligent AI model training resource data to obtain the intelligent AI model training resource missing compensation data; Step S14: standardizing the intelligent AI model training resource loss compensation data to obtain the training standard data required for the intelligent AI model; Step S15: By deploying distributed computing nodes in the network, and performing training resource division processing on the training standard data required for the intelligent AI model based on the distributed computing nodes, the intelligent AI model training resource sub-data of each distributed computing node is obtained; wherein step S15 includes the following steps: Step S151: Deploy distributed computing nodes in the network and perform computing resource statistics analysis on each distributed computing node to obtain computing resource status data of each distributed computing node; Step S152: Performing statistical analysis on the data storage space of each distributed computing node to obtain the size of the data storage space range of each distributed computing node; Step S153: performing data resource accommodation evaluation and analysis on each distributed computing node based on the computing resource status data of each distributed computing node and the size of the data storage space range, and obtaining the data resource accommodation limit range on each distributed computing node; Step S154: performing training resource division processing on the training standard data required for the intelligent AI model according to the data resource accommodation limit range on each distributed computing node, and obtaining the intelligent AI model training resource sub-data of each distributed computing node; Step S2: constructing and initializing intelligent AI sub-models on each distributed computing node to obtain intelligent AI sub-models on each distributed computing node; performing model parallel training on the intelligent AI sub-models on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node to obtain intelligent AI model training parameters of each distributed computing node; Step S3: Based on the intelligent AI model training parameters of each distributed computing node, the intelligent AI sub-models on each distributed computing node are aggregated in a hierarchical manner to generate an intelligent AI global model; the intelligent AI global model is optimized to obtain an intelligent AI global optimization model; Step S4: The intelligent AI global optimization model is deployed to each distributed computing node for online operation to obtain an intelligent AI online operation model.

2. The method for constructing and operating intelligent AI according to claim 1, characterized in that: Step S153 includes the following steps: Performing CPU operation utilization statistics analysis on the computing resource status data of each distributed computing node to obtain the CPU operation utilization of each distributed computing node; Performing network bandwidth utilization statistics analysis on the computing resource status data of each distributed computing node to obtain the network bandwidth utilization of each distributed computing node; Based on the CPU operation utilization rate and network bandwidth utilization rate of each distributed computing node, the computing resource utilization capacity of each distributed computing node is evaluated and analyzed to obtain the computing resource utilization capacity data of each distributed computing node; Based on the size of the data storage space of each distributed computing node, the data storage capacity of each distributed computing node is evaluated and analyzed to obtain the spatial storage capacity data of each distributed computing node; The resource accommodation limit interval of each distributed computing node is determined according to the computing resource utilization capacity data and the space storage capacity data of each distributed computing node, and the data resource accommodation limit interval on each distributed computing node is obtained.

3. The method for constructing and operating intelligent AI according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: constructing and initializing an intelligent AI sub-model on each distributed computing node to obtain an intelligent AI sub-model on each distributed computing node; Step S22: Distribute the intelligent AI model training resource sub-data of each distributed computing node to each distributed computing node by building a secure network communication channel; Step S23: Based on the intelligent AI model training resource sub-data of each distributed computing node, the intelligent AI sub-model on each distributed computing node is trained synchronously and in parallel to obtain the intelligent AI sub-model synchronous and parallel training parameters of each distributed computing node; Step S24: performing asynchronous parallel training on the intelligent AI sub-models on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node, and obtaining asynchronous parallel training parameters of the intelligent AI sub-models of each distributed computing node; Step S25: Perform arithmetic mean analysis on the synchronous parallel training parameters of the intelligent AI sub-models of each distributed computing node and the asynchronous parallel training parameters of the intelligent AI sub-models to obtain the intelligent AI model training parameters of each distributed computing node.

4. The method for constructing and operating intelligent AI according to claim 3, characterized in that: Step S23 includes the following steps: Step S231: performing synchronous training task allocation on the intelligent AI sub-models on each distributed computing node to obtain the intelligent AI model synchronous training task on each distributed computing node; Step S232: Based on the intelligent AI model training resource sub-data of each distributed computing node, the intelligent AI model synchronization training task on each distributed computing node is trained in parallel to obtain an intelligent AI synchronization parallel training model on each distributed computing node; Step S233: Perform model parameter synchronization update processing on the intelligent AI synchronous parallel training model on each distributed computing node to obtain the intelligent AI sub-model synchronous parallel training parameters of each distributed computing node.

5. The method for constructing and operating intelligent AI according to claim 3, characterized in that: Step S24 includes the following steps: Step S241: performing asynchronous training task allocation on the intelligent AI sub-model on each distributed computing node to obtain the asynchronous training task of the intelligent AI model on each distributed computing node; Step S242: Perform task load analysis on the intelligent AI model asynchronous training tasks on each distributed computing node to obtain model training task load data on each distributed computing node; Step S243: asynchronously load the intelligent AI model training tasks on each distributed computing node according to the model training task load data on each distributed computing node, and obtain the model asynchronous loading training tasks on each distributed computing node; Step S244: performing asynchronous parallel model training on the model asynchronous loading training tasks on each distributed computing node based on the intelligent AI model training resource sub-data of each distributed computing node, to obtain the intelligent AI asynchronous parallel training model on each distributed computing node; Step S245: Perform asynchronous update processing on the model parameters of the intelligent AI asynchronous parallel training model on each distributed computing node to obtain the asynchronous parallel training parameters of the intelligent AI sub-model of each distributed computing node.

6. The method for constructing and operating intelligent AI according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: Dynamically adjust the weight of the intelligent AI model training parameters of each distributed computing node to obtain the model training contribution balance parameters of each distributed computing node; Step S32: performing model hierarchical aggregation on the intelligent AI sub-models on each distributed computing node based on the model training contribution balance parameters of each distributed computing node to generate an intelligent AI global model; Step S33: performing model compression processing on the intelligent AI global model to obtain an intelligent AI global compression model; Step S34: Perform model pruning optimization on the intelligent AI global compression model to obtain an intelligent AI global optimization model.

7. An intelligent AI construction and operation system, characterized in that: For executing the construction and operation method of the intelligent AI as claimed in claim 1, the construction and operation system of the intelligent AI comprises: The training resource division module required by the model is used to obtain the training resource data required by the intelligent AI model, and clean the training resource data required by the intelligent AI model to obtain the training standard data required by the intelligent AI model; by deploying distributed computing nodes in the network, and dividing the training resource for the standard data required by the intelligent AI model based on the distributed computing nodes, the intelligent AI model training resource sub-data of each distributed computing node is obtained; A distributed node model parallel training module is used to construct and initialize an intelligent AI sub-model on each distributed computing node to obtain an intelligent AI sub-model on each distributed computing node; based on the intelligent AI model training resource sub-data of each distributed computing node, the intelligent AI sub-model on each distributed computing node is trained in model parallel to obtain intelligent AI model training parameters of each distributed computing node; The intelligent AI model hierarchical aggregation optimization module is used to perform hierarchical aggregation of the intelligent AI sub-models on each distributed computing node based on the intelligent AI model training parameters of each distributed computing node to generate an intelligent AI global model; and to optimize the intelligent AI global model to obtain an intelligent AI global optimization model; The intelligent AI model online operation module is used to deploy the intelligent AI global optimization model to various distributed computing nodes for online operation to obtain an intelligent AI online operation model.

8. A PC host, characterized in that: It includes a processor, a memory, and a computer control program stored in the memory and executable on the processor, and is used to execute the construction and operation method of the intelligent AI as described in any one of claims 1 to 6.