CN-121996256-A - Model end side batch deployment and intelligent operation and maintenance method and system in autonomous controllable environment

CN121996256ACN 121996256 ACN121996256 ACN 121996256ACN-121996256-A

Abstract

The embodiment of the invention provides a model end side batch deployment and intelligent operation and maintenance method and system in an autonomous controllable environment, and belongs to the technical field of model deployment. The method for batch deployment and intelligent operation and maintenance of the model end sides in the autonomous controllable environment comprises the steps of obtaining a trained model file and domestic hardware, constructing a plurality of deployment packages according to the trained model file and the domestic hardware, and distributing the deployment packages to target equipment corresponding to target end sides. The method has the advantages that the mode that the model files and domestic hardware are combined to respectively construct deployment packages and are distributed to corresponding target devices in a targeted mode is adopted, the precision, reliability and efficiency of model batch deployment can be effectively improved, the key indexes of the target devices are monitored and uploaded through the lightweight agent, dynamic monitoring of the target devices can be achieved, deployment dynamic adjustment is conveniently achieved, and the deployment precision and operation and maintenance precision of the target devices are further improved.

Inventors

Xing Guoyong
ZHANG LINYU
LI XIAONING
HUANG XIAOGUANG
QIU ZHEN
WANG XIAODONG
QIN YU
LU DAWEI
ZHOU YIPING
BAI JINGPO
LIU YUANYUAN

Assignees

国网信息通信产业集团有限公司
福建亿榕信息技术有限公司
北京国网信通埃森哲信息技术有限公司

Dates

Publication Date: 20260508
Application Date: 20251219

Claims (10)

1. The method for batch deployment and intelligent operation and maintenance of the model end sides in the autonomous controllable environment is characterized by comprising the following steps: Obtaining a trained model file and domestic hardware; Constructing a plurality of deployment packages according to the trained model file and the domestic hardware; Distributing the deployment package to target equipment corresponding to a target end side; Running a lightweight agent on each of the target devices in the target end side; the lightweight agent installs and starts the deployment package; and the lightweight agent monitors and uploads key indexes of the target equipment.
2. The method for model end-side batch deployment and intelligent operation and maintenance in an autonomous controllable environment according to claim 1, wherein constructing a deployment package according to the trained model file and the domestic hardware comprises: Constructing a basic environment mirror image and a mirror image package of the domestic hardware; acquiring a hardware model of each target device in the target end side; Acquiring a corresponding basic environment image and an image package according to the hardware model of the target equipment; and packaging the basic environment mirror image and the mirror image package of the target equipment, and acquiring a deployment package.
3. The method for model end-side batch deployment and intelligent operation and maintenance in an autonomous controllable environment according to claim 2, wherein constructing a basic environment mirror image and a mirror image package of domestic hardware comprises: acquiring an identifier of the domestic hardware and a software ecological dependency list; constructing a domestic hardware feature library according to the identifier of the domestic hardware and the software ecological dependency list; Constructing a minimized runtime environment according to the software ecological dependency list; And packaging the model file, the minimized runtime environment and the startup script to obtain a mirror package, and adopting a configuration template to carry out management model and service configuration on the mirror package.
4. The method for model end-side batch deployment and intelligent operation and maintenance in an autonomous controllable environment according to claim 1, wherein the distributing the deployment package to the target device corresponding to the target end side comprises: acquiring static attribute and dynamic attribute of each target device in the target end side; grouping the target devices according to the static attribute and the dynamic attribute of each target device to obtain a deployment group; and distributing different deployment packages to the corresponding deployment groups.
5. The method for model end-side batch deployment and intelligent operation and maintenance in an autonomous controllable environment according to claim 4, wherein distributing different deployment packages to the corresponding deployment groups comprises: distributing the deployment package in a controllable concurrency mode; and controlling the global concurrent deployment task number by adopting a token bucket algorithm.
6. The method for model end-side batch deployment and intelligent operation and maintenance in an autonomous controllable environment according to claim 4, wherein distributing different deployment packages to the corresponding deployment groups further comprises: Judging whether the deployment package belongs to a large deployment package or not; under the condition that the deployment package is judged to belong to a large deployment package, the large deployment package is segmented to obtain a plurality of deployment package small blocks; Acquiring the ACK feedback speed of the target end side; acquiring the size of a sliding window according to the ACK feedback speed of the target end side; and transmitting a plurality of deployment packet small blocks by adopting the sliding window.
7. The method for model end-side batch deployment and intelligent operation and maintenance in an autonomously controllable environment according to claim 4, wherein the static attributes include hardware architecture, AI accelerator type and number, memory/storage capacity and operating system version.
8. The method for model end-side batch deployment and intelligent operation and maintenance in an autonomously controlled environment according to claim 4, wherein the dynamic attributes include current network bandwidth, real-time CPU/memory load and geographic area.
9. The utility model provides a model terminal side is deployment in batches and intelligent fortune dimension system under independently controllable environment which characterized in that includes: the target end side module comprises a plurality of target devices; The cloud management platform is connected with the target end side module and used for executing the model end side batch deployment and intelligent operation and maintenance method in the autonomous controllable environment according to any one of claims 1-8.
10. A computer readable storage medium storing instructions for being read by a machine to cause the machine to perform the model end-side batch deployment and intelligent operation and maintenance method in an autonomously controllable environment according to any one of claims 1-8.

Description

Model end side batch deployment and intelligent operation and maintenance method and system in autonomous controllable environment Technical Field The invention relates to the technical field of model deployment, in particular to a model end side batch deployment and intelligent operation and maintenance method and system in an autonomous controllable environment. Background Along with the large-scale application of the artificial intelligence technology in key industries such as electric power, industrial manufacturing and the like, the deployment of a trained lightweight AI model (such as a power transformation equipment fault identification model and a forest fire early warning model) to massive end-side equipment (such as a patrol terminal, a camera and an edge server) has become a key link for realizing service intelligence. However, in the context of emphasis on autonomous control of technology, environments based on domestic hardware (e.g., hua Sheng Teng, hei Guang DCU) and domestic operating systems (e.g., UOS) face significant challenges. Currently, the mainstream deployment methods mostly depend on manual operation or general software deployment tools, and have the following inherent defects: 1. the deployment efficiency is low, the quantity of the terminal equipment is huge, the geographic positions are distributed, the hardware is heterogeneous (domestic AI accelerators with different architectures), the environment, the installation model and the dependency library are manually configured for each equipment, the time and the labor are consumed, and the rapid large-scale popularization cannot be realized. 2. The domestic environment has poor compatibility, and a general deployment tool does not perform deep optimization on the driving and calculation library (such as the rising CANN) of the domestic chip, so that the model cannot fully utilize the hardware acceleration characteristic, and even has the problems of incompatibility and incapability of running. 3. The operation and maintenance management is not intelligent enough, and an effective monitoring means is lacking after the model is deployed. The running state (such as reasoning performance and hardware resource occupation) of the model at the end side cannot be perceived in real time, the performance attenuation, data drift or hardware faults of the model cannot be found in time, the operation and maintenance are completely dependent on manual inspection, the response is slow, and the cost is high. In the prior art, as in the chinese patent application number 202410761684.0, "an edge computing device deployment management method and system", a deployment scheme based on edge devices is disclosed. The scheme realizes the management and deployment of the device, but has the following problems: The environment adaptation capability is weak, the difference of domestic heterogeneous AI chips is not considered in the container mirror image, the specific driving and accelerating libraries required by different hardware cannot be dynamically adapted, and the deployment success rate in mixed environments such as rising, sea light and the like is low. The deployment strategy is single, the batch deployment process is 'cut-off', intelligent grouping and differentiated deployment strategy scheduling are not carried out according to the network condition and hardware performance of the equipment, and network congestion or low-performance equipment deployment failure is easy to cause. The operation and maintenance capability aiming at the AI model is lacking, namely the scheme only monitors the life cycle of the container, but cannot monitor the business indexes (such as accuracy and delay) inferred by the AI model in the container, cannot identify the specific faults (such as abnormal input data and performance drift) of the model, has single operation and maintenance dimension and has low intelligent level. Meanwhile, in the field of edge computing, in order to deploy AI models on resource-constrained devices, model compression (such as pruning) is one of the key technologies. Specifically, in the prior art, for example, a chinese patent application No. 202511127140.X "a model deployment method based on pruning compression in an edge device" discloses a dynamic adaptive pruning technique. According to the technology, the compression ratio of each layer of the model is determined by analyzing the hardware architecture, and the redundant weight blocks are reserved during pruning, so that the model can be subjected to online fine adjustment according to the operation environment after being deployed, and the self-adaption capability of a single model on single equipment is effectively improved. However, the prior art and the similar schemes thereof are mainly focused on microscopic optimization of a single model, and have obvious limitations when large-scale application scenes of a plurality of lightweight models are uniformly and efficiently deployed