Research on Detection Optimization of the 3.0 Base Version of the Metering Automation System Based on Reinforcement Learning Algorithm

• Published in: IEEE access Vol. 13; pp. 155697 - 155713
• Main Authors: Zeng, Lukun, Lin, Guoying, Yang, Jingxu, Chen, Guanyu, Li, Sheng, Liu, Guobo, Bi, Jinghu
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation models; base version; Deep learning; Dynamic tests; feature engineering; Feature extraction; Heuristic algorithms; Intelligent agents; Machine learning; Markov processes; Measurement automation system; Optimization; Power system dynamics; Priority scheduling; Quality assessment; Quality control; quality inspection; Real time; Real-time systems; reinforcement learning; Resource management; Smart grid; Testing; Throughput
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3606062

In response to the problems of low detection efficiency and insufficient strategy optimization ability of the current base version detection tools in dynamic and complex scenarios, this paper proposes a detection optimization framework based on deep reinforcement learning. Firstly, by constructing a dual-mode dynamic testing model that integrates local detection and online detection, the base version quality assessment process is formalized as a Markov decision process, where the state space covers key indicators such as functional integrity, interface consistency, and storage throughput (refer to the GB/T 25000 standard). Secondly, an intelligent agent based on the Deep Q-Network (DQN) is designed to achieve adaptive optimization of detection dimension selection, feature index priority scheduling, and test case generation. A multi-objective hybrid reward mechanism (R =0.6 defect discovery rate +0.3 throughput gain - 0.1* time cost) is innovatively introduced to balance detection quality and resource consumption in dynamic interaction. Experiments show that this method improves detection efficiency by 37.2% compared to traditional schemes, with a feature index coverage rate of 98.5%, and its robustness in multi-unit detection scenarios is verified in third-party testing. The developed microservice architecture detection tool integrates a reinforcement learning decision interface, supporting real-time visualization of detection progress and automatic generation of electronic reports. This research provides an innovative solution for the full life cycle quality control of smart grid base versions.