CN-121982722-A - Method and device for data annotation and model training scheduling based on intelligent agent cooperation

CN121982722ACN 121982722 ACN121982722 ACN 121982722ACN-121982722-A

Abstract

The application discloses a method and a device for data annotation and model training scheduling based on agent cooperation, wherein the method for data annotation and model training scheduling based on agent cooperation comprises the steps of constructing an initial network architecture of an agent by utilizing image data and a knowledge graph, conducting parallel training on a neural network in the initial network architecture by utilizing first marked image data to obtain the network architecture of the agent, enabling the agent to conduct target detection on the image data by utilizing a convolution neural network to obtain a target in the image data, conducting abnormal classification on the target by utilizing a deep learning network to obtain an annotation result, enabling the agent to further be used for determining confidence of the annotation result according to the type of the neural network in the network architecture, the type of the image data, the target and the annotation result, inputting second image data into the agent to obtain the initial annotation result and the confidence output by the agent, and obtaining the annotation result of the second image data according to the initial annotation result and the confidence.

Inventors

CHEN LE
LV YAN
WANG ZHAO
WANG MINGLIANG
LI JINBU

Assignees

中国移动紫金(江苏)创新研究院有限公司
中国移动通信集团江苏有限公司
中国移动通信集团有限公司

Dates

Publication Date: 20260505
Application Date: 20260204

Claims (11)

1. A method for data annotation and model training scheduling based on agent coordination, which is applied to a first device, the method comprising: constructing an initial network architecture of the intelligent agent by utilizing the image data and the knowledge graph, wherein the initial network architecture comprises at least one neural network, namely a convolutional neural network for carrying out target detection on the image data and a deep learning network for carrying out abnormal classification on the image data; The method comprises the steps of obtaining a network architecture of an intelligent agent by training a neural network in an initial network architecture in parallel by using first marked image data, wherein the intelligent agent is used for detecting an image data by using a convolution neural network to obtain a target in the image data, and performing abnormal classification on the target by using a deep learning network to obtain a marking result; Acquiring second image data to be marked; Inputting the second image data into an agent of the first device to obtain an initial labeling result of the second image data output by the agent and a confidence coefficient of the initial labeling result; and obtaining the labeling result of the second image data according to the initial labeling result and the confidence coefficient.
2. The method of claim 1, wherein the training the neural network in the initial network architecture in parallel using the annotated first image data results in a network architecture of the agent, comprising: Training the neural network in the initial network architecture in parallel by using the first image data with a first batch size to obtain the utilization rate of the graphic processor of the first equipment in the training process; Adjusting the first batch size according to the utilization rate of the graphic processor by utilizing a near-end strategy optimization algorithm to obtain a second batch size; And training the neural network in the initial network architecture in parallel by using the first image data with the second batch size to obtain the network architecture of the intelligent agent.
3. The method according to claim 1, wherein the obtaining the labeling result of the second image data according to the initial labeling result and the confidence level includes: Under the condition that the confidence coefficient is smaller than or equal to a confidence coefficient threshold value, acquiring a manual annotation result of the second image data corresponding to the initial annotation result, and acquiring an annotation result of the second image data according to the manual annotation result; And under the condition that the confidence coefficient is larger than a confidence coefficient threshold value, taking the initial labeling result as a labeling result of the second image data.
4. A method according to claim 3, characterized in that the method further comprises: updating and training the intelligent agent by using a semi-supervised learning algorithm and first target image data; the first target image data is second image data with the confidence coefficient smaller than or equal to a confidence coefficient threshold value.
5. The method according to claim 1, wherein the method further comprises: and under the condition that the initial labeling result of the second image data indicates the defect type appearing for the first time, updating the network architecture of the intelligent agent according to the second image data and the initial labeling result of the second image data.
6. The method according to claim 1, wherein the method further comprises: obtaining a target deployment mode of the intelligent agent according to the hardware configuration mode of the first equipment, the network bandwidth supported by the first equipment and the deployment mode supported by the first equipment; and deploying the agent into the first device by using the target deployment mode.
7. The method according to claim 1 or 6, characterized in that the method further comprises: Acquiring the equipment performance of the first equipment in the operation process of the intelligent agent; Under the condition that the intelligent body operates abnormally, determining a repairing mode of the first equipment according to the equipment performance; Wherein the device capabilities include at least one of: The utilization rate of the central processing unit; The utilization rate of the graphics processor; Response delay.
8. An apparatus for data annotation and model training scheduling based on agent coordination, the apparatus comprising: the first processing module is used for constructing an initial network architecture of the intelligent agent by utilizing the image data and the knowledge graph, wherein the initial network architecture comprises at least one neural network, namely a convolutional neural network for carrying out target detection on the image data and a deep learning network for carrying out abnormal classification on the image data; the system comprises a first processing module, a second processing module, an intelligent agent, a deep learning network, a first processing module and a second processing module, wherein the first processing module is used for performing parallel training on a neural network in an initial network architecture by using marked first image data to obtain a network architecture of the intelligent agent; The first acquisition module is used for acquiring second image data to be marked; The third processing module is used for inputting the second image data into an agent of the first device to obtain an initial labeling result of the second image data output by the agent and the confidence coefficient of the initial labeling result; And the fourth processing module is used for obtaining the labeling result of the second image data according to the initial labeling result and the confidence coefficient.
9. An electronic device comprising a processor, a memory and a program stored on the memory and executable on the processor, the program when executed by the processor implementing the steps in the method of agent-collaborative based data labeling and model training scheduling of any one of claims 1-7.
10. A readable storage medium, wherein a program is stored on the readable storage medium, which program, when executed by a processor, implements the steps in the method for agent-collaborative based data labeling and model training scheduling of any one of claims 1-7.
11. A computer program product comprising computer instructions which, when executed by a processor, implement the steps in the method of agent-collaborative based data labeling and model training scheduling of any one of claims 1-7.

Description

Method and device for data annotation and model training scheduling based on intelligent agent cooperation Technical Field The application belongs to the technical field of artificial intelligence, and particularly relates to a method and a device for data annotation and model training scheduling based on intelligent agent cooperation. Background In recent years, with rapid development of artificial intelligence, computer vision and big data technology, an industrial vision artificial intelligence (ARTIFICIAL INTELLIGENCE, AI) platform in the industrial field gradually becomes an important component in intelligent manufacturing and automatic production. The platforms can perform tasks such as efficient defect detection, product identification, production line monitoring and the like in a complex industrial environment by integrating advanced image processing, machine learning and deep learning algorithms. With the rise of industrial Internet of things (Industrial Internet of Things, IIoT), various intelligent sensors and cameras are popularized, and industrial visual artificial intelligent platforms are widely applied. The platform obtains visual data of machine equipment, production lines, products and the like in real time, and performs marking processing on the data by utilizing technologies such as image recognition, feature extraction, mode recognition and the like, so as to provide intelligent decision support in the production process. However, the traditional industrial visual artificial intelligent platform still depends on manual labeling on data labeling, and particularly in a complex production environment, the high cost and high time consumption of the data labeling task become main barriers for platform application. Disclosure of Invention The embodiment of the application provides a method and a device for data annotation and model training scheduling based on agent cooperation, which are used for solving the problems that the existing industrial vision artificial intelligent platform still depends on artificial annotation on data annotation and has high cost and long time consumption. In a first aspect, an embodiment of the present application provides a method for data annotation and model training scheduling based on agent coordination, applied to a first device, where the method includes: constructing an initial network architecture of the intelligent agent by utilizing the image data and the knowledge graph, wherein the initial network architecture comprises at least one neural network, namely a convolutional neural network for carrying out target detection on the image data and a deep learning network for carrying out abnormal classification on the image data; The method comprises the steps of obtaining a network architecture of an intelligent agent by training a neural network in an initial network architecture in parallel by using first marked image data, wherein the intelligent agent is used for detecting an image data by using a convolution neural network to obtain a target in the image data, and performing abnormal classification on the target by using a deep learning network to obtain a marking result; Acquiring second image data to be marked; Inputting the second image data into an agent of the first device to obtain an initial labeling result of the second image data output by the agent and a confidence coefficient of the initial labeling result; and obtaining the labeling result of the second image data according to the initial labeling result and the confidence coefficient. Optionally, the training the neural network in the initial network architecture in parallel by using the noted first image data to obtain a network architecture of the agent includes: Training the neural network in the initial network architecture in parallel by using the first image data with a first batch size to obtain the utilization rate of the graphic processor of the first equipment in the training process; Adjusting the first batch size according to the utilization rate of the graphic processor by utilizing a near-end strategy optimization algorithm to obtain a second batch size; And training the neural network in the initial network architecture in parallel by using the first image data with the second batch size to obtain the network architecture of the intelligent agent. Optionally, the obtaining the labeling result of the second image data according to the initial labeling result and the confidence coefficient includes: Under the condition that the confidence coefficient is smaller than or equal to a confidence coefficient threshold value, acquiring a manual annotation result of the second image data corresponding to the initial annotation result, and acquiring an annotation result of the second image data according to the manual annotation result; And under the condition that the confidence coefficient is larger than a confidence coefficient threshold value, taking the initial labeling result a