摘要
轨道缺陷检测对列车的安全运行意义重大,传统依赖专业人员现场人工识别的方法存在效率低下、人力成本较高等诸多问题。但传统集中式训练方法要求共享参与方的私有数据,又带来隐私泄露、流量通信压力等问题,因此基于深度学习的检测方法应运而生。近年来提出的联邦学习方法是一种分布式机器学习训练策略,通过聚合多个联邦成员本地训练的参数更新,无需共享私有数据,即可有效建立全局模型。然而,在铁路等工业场景中使用的物联网边缘设备,其计算和网络资源难以承受复杂深度模型带来的计算和通信开销。因此,提出一种轻量级边缘联邦学习算法,通过对网络模型进行通道剪枝与权值量化,降低计算、存储和通信开销,加速边缘设备上进行的联邦学习过程,最终得到精度损失在一定的可接受范围内的全局模型。通过构建轨道缺陷检测数据集,并选取ResNet-18作为目标模型,进行实验验证。结果表明,与不进行剪枝量化的原始联邦学习模型效果相比,在数据集所包含的各类缺陷上达到最高93.49%的平均鉴别准确率,同时模型大小可压缩至10.75倍,全局加速2.33倍,验证了本方法的有效性。
Track defect detection is of great significance to the safe operation of trains.The traditional method relying on on-site manual identification by professionals has many problems,such as low efficiency and high labor cost.Therefore,the detection method based on deep learning arises at the historic moment.However,the traditional centralized training method that requires the sharing of private data of participants brings problems including privacy leakage and traffic communication pressure.The federated learning method proposed in recent years is a distributed machine learning training strategy,which can effectively build a global model by aggregating the parameter updates of local training of multiple federated members without sharing private data.However,the computing and network resources of IoT edge devices used in industrial scenarios such as railways cannot withstand the computing and communication overhead caused by complex depth models.Therefore,a lightweight edge federation learning algorithm was proposed to reduce the computation,storage and communication overhead through channel pruning and weight quantization of the network model,and to accelerate the federation learning process on the edge devices.Finally,the global model with accuracy loss within a certain acceptable range was obtained.By constructing the track defect detection data set,and selecting ResNet-18 as the target model,the experimental verification was carried out.The results show that compared with the effect of the original federated learning model without pruning quantification,the model size can be compressed up to 10.75 times,with the global acceleration up to 2.81 times,and the average identification accuracy rate of 93.34%on the various defects contained in the data set,which verifies the effectiveness of the method.
作者
任中伟
方维维
许文元
李中睿
胡一寒
REN Zhongwei;FANG Weiwei;XU Wenyuan;LI Zhongrui;HU Yihan(School of Computer and Information Technology,Beijing Jiaotong University,Beijing 100044,China;Beijing Key Lab of Traffic Data Analysis and Mining,Beijing Jiaotong University,Beijing 100044,China)
出处
《铁道学报》
EI
CAS
CSCD
北大核心
2023年第4期77-83,共7页
Journal of the China Railway Society
基金
国家自然科学基金(62172031)
北京市自然科学基金(L191019)。
关键词
联邦学习
缺陷检测
模型压缩
物联网
federated learning
defect detection
model compression
Internet of Things
