摘要详情

ID / 提交时间

63 / 2025-03-30 00:39:15

标题

Research on Differentiated Training of Data for Intelligent Safety-critical Software

关键字

Differentiated Training,Safety-critical Software,Data Construction,Data Splitting,Artificial Intelligence

主题及专题

T12 安全关键系统

状态

全文待审

作者

健堃徐 / 29所

立金吴 / 船综院

泊江刘 / 船综院

游李 / 29所

摘要

With the advancement of technology and scientific development, intelligent safety-critical software plays an increasingly critical role in modern battlefields. Traditional approaches for intelligent safety-critical software are constrained by manually designed feature dimensions, often leading to feature loss. Integrating the data characteristics of intelligent safety-critical software, this paper proposes a differentiated training method tailored to control modules, processing modules, input-output modules, and health management modules – the four core data modules. First, this work analyzes the unique attributes of intelligent safety-critical software, clarifying its high reliability, real-time performance, reusability, and attack-resistant capabilities. Second, we select appropriate training data construction methods based on varying scenarios and mission requirements, specifying training scenarios and optimizing training approaches for control data, processing data, input-output data, and health management data. Third, we implement training methods for heterogeneous safety-critical data, including stratified sampling, adversarial validation, time-series segmentation, and cross-validation, establishing a high-reliability intelligent training framework for smart safety-critical software. Finally, the method is validated through case studies, demonstrating improvements in alarm accuracy and data effectiveness, reducing fault rates and control errors, thereby verifying the proposed methodology.