基于增量学习的工业软件:转炉图像异常识别系统  

Industrial Software Based on Incremental Learning:Anomaly RecognitionSystem of Converter Image

作　　者：武星[1] 殷浩宇 姚骏峰 金小礼[3] WU Xing;YIN Haoyu;YAO Junfeng;JIN Xiaoli(School of Computer Engineer and Science,Shanghai University,Shanghai 200444,China;CSSC Seago System Technology Co.,Ltd,Shanghai 200010,China;Central Research Institute of Baoshan Iron&Steel Co.LTD,Shanghai 201900,China)

机构地区：[1]上海大学计算机工程与科学学院,上海200444 [2]中国船舶集团海舟系统技术有限公司,上海200010 [3]宝山钢铁股份有限公司中央研究院,上海201900

出　　处：《武汉大学学报（理学版）》2024年第3期293-301,共9页Journal of Wuhan University:Natural Science Edition

基　　金：上海市青年科技启明星计划项目(21QB1401900);科技部重点研发计划项目(2022YFB3707800)。

摘　　要：为简化工业软件的研发过程,提高图像异常识别的准确性和效率,提出了一种基于增量学习的转炉图像异常识别系统。该系统采用机器视觉技术采集转炉图像,引入深度残差网络形成转炉图像异常识别模型,并利用采集到的图像训练该模型。系统采用低代码开发方法实现,并结合增量学习算法优化了模型的迭代更新。对比基于不同神经网络架构的转炉图像异常识别模型的识别准确率,并在低代码平台中对比了增量学习和全量学习在模型精度和时间消耗上的差异。实验结果表明,本系统在图像异常识别中展现出良好的精确性和稳定性;在低代码开发平台中,基于增量学习的系统软件在处理大规模数据和实时场景中表现出色,为转炉图像异常识别提供了一种高效、低成本的解决方法。A converter furnace image anomaly detection system based on incremental learning has been proposed to simplify the research and development process for industrial software and enhance the accuracy and efficiency of image anomaly detection.This system employs machine vision technology to capture images of the converter furnace.It introduces a deep residual network to form an image anomaly detection model,using the captured images to train this model.The system is implemented through low-code development methods and optimizes the iterative updates of the model by incorporating incremental learning algorithms.The recognition accuracy of the converter furnace image anomaly detection model is compared across various neural network archi⁃tectures.Additionally,the disparities in model accuracy and time consumption between incremental learning and full learning are assessed on a low-code platform.Experimental results demonstrate that this system exhibits good precision and stability in image anomaly detection.On the low⁃code development platform,the system software based on incremental learning performs excellently in handling large-scale data and real-time scenarios,providing an efficient and low⁃cost solution for converter furnace image anomaly detection.

关 键 词：增量学习 转炉图像识别 低代码平台 深度残差网络 异常检测 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]