摘要
六足机器人在实际运维工作中,很难采集到足够的故障数据,这导致故障数据存在不平衡,并严重影响故障诊断的效果。针对六足机器人实际应用中故障数据不平衡影响深度学习模型训练导致故障诊断精度低的问题,提出一种多阶段生成对抗网络数据增强算法。将多传感器的时序数据作为时间序列生成对抗网络(TimeGAN)的输入,生成具有时间动态特征的序列数据;将时间序列生成对抗网络合成的数据代替带梯度惩罚WGAN生成对抗网络(WGAN-GP)中的噪声输入数据,且判别器将图卷积网络、长短时记忆网络和注意力机制结合,更有效地挖掘多源异构传感数据的时空相关性,增强对时序数据的判别能力。实验结果表明:所提算法能有效生成具有高质量、多样性的时序特征数据;与同类型的算法相比,能显著提高故障诊断的准确率。
In the actual operation and maintenance work,it is difficult to collect enough fault data,which leads to the imbalance of fault data and seriously affects the effect of fault diagnosis.In order to solve the problem that fault data imbalance affects deep learning model training in practical applications of hexapod robots and leads to low accuracy of fault diagnosis method,a multi-stage generation adversarial network data enhancement algorithm was proposed.The time series data of multiple sensors was used as the input of time-series generative adversarial networks(TimeGAN)to generate the time-dynamic data.The noise input data in WGAN-GP with gradient penalty was replaced by the data synthesized by the time series generating adversarial network,and the discriminator combined the graph convolutional network,long and short time memory network and attention mechanism to more effectively mine the spatio-temporal correlation of multi-source heterogeneous sensor data and enhance the discriminant ability of time series data.Experimental results show that the proposed algorithm can effectively generate time series feature data with high quality and diversity.Compared with the algorithms of the same type,it can significantly improve the accuracy of fault diagnosis.
作者
张志强
斯帅
佃松宜
赵涛
郭斌
刘佳
ZHANG Zhiqiang;SI Shuai;DIAN Songyi;ZHAO Tao;GUO Bin;LIU Jia(College of Electrical Engineering,Sichuan University,Chengdu Sichuan 610065,China;High End Talent Service Center in Da′an District,Zigong Sichuan 643000,China)
出处
《机床与液压》
北大核心
2025年第11期1-9,共9页
Machine Tool & Hydraulics
作者简介
张志强(1998-),男,硕士研究生,研究方向为机器人故障诊断。E-mail:2020382189@qq.com;通信作者:佃松宜(1972-),男,博士,研究方向为先进控制理论、感知与人工智能算法。E-mail:scudiansy@scu.edu.cn。
