深度学习识别光网络单元故障的设计与应用

为解决依赖装维上门鉴别光网络单元故障带来的不便,可以从机器视觉入手实现自动化故障识别。近年,ImageNet挑战赛的成功推动了物体识别技术的跨越式发展,特别是基于卷积的深度学习技术在视觉识别方面已经达到人类水平,为光网络单元故障的自动识别提供了技术基础。文章对识别光网络单元的工作状态进行了研究,将设备工作状态分为7个场景,提出了利用手机APP采集图片识别故障的解决方案并投入了实际生产;重点阐述了深度学习模块的设计与实现,提出一种通过算法整合的方式综合运用物体检测和图像分类算法,分3阶段逐步求精,解决了图片过滤,光网络单元型号和状态识别等问题,实现了基于计算机视觉自动识别光网络单元故障。从数据上看产品的端到端准确率超过84%,识别速度达到10 FPS,月均提供服务超过1万人次,在减少用户等待的同时节约了人力资源。

Design and Application of Identifying Malfunctions in Optical Network Units Based on Deep Learning

To solve the inconvenience caused by the failure of the manual identification optical network unit,automatic fault identification can be realized from machine vision.The success of the ImageNet Challenge has promoted the leap-forward development of object recognition technology,especially the convolution-based deep learning technology has reached the human level in visual recognition,providing a technical basis for automatic identification of optical network unit malfunctions.We study the working status of optical network units,divide the working status of equipment into seven scenarios,and put forward a solution of using the mobile phone APP to collect pictures to identify faults and put it into actual production. We focus on the design and implementation of the deep learning module and propose a comprehensive method of object detection and image classification by algorithm integration,which is in three stages,solving the problems of image filtering,optical network unit model and status recognition,and finally achieving the automatic identification of optical network unit faults based on computer vision.Saving human resources while reducing user waiting,the end-to-end accuracy rate of the product exceeds 84%,the recognition speed reaches 10 FPS,and the monthly service provides more than 10 000 times.