小样本条件下元学习门控神经网络的齿轮退化趋势预测研究  

Gear Degradation Trend Prediction by Meta-learning Gated Recurrent Unit Networks under Few Samples

作　　者：余晓霞 邓蕾[1] 汤宝平[1] 夏乙[2] 李琪康 YU Xiaoxia;DENG Lei;TANG Baoping;XIA Yi;LI Qikang(The State Key Laboratory of Mechanical Transmission,Chongqing University,Chongqing 400044;Chongqing Aerospace Vocational and Technical College,Chongqing 400021)

机构地区：[1]重庆大学机械传动国家重点实验室,重庆400044 [2]重庆航天职业技术学院,重庆400021

出　　处：《机械工程学报》2022年第3期149-156,共8页Journal of Mechanical Engineering

基　　金：国家重点研发计划(2020YFB1709800);重庆市自然科学重点基金(cstc2019jcyj-zdxmX0026);国家自然科学基金(51775065)资助项目。

摘　　要：针对齿轮小样本条件下深度神经网络训练容易过拟合导致齿轮退化趋势预测不精准的难题,提出了元学习门控神经网络网络齿轮退化趋势预测模型(Meta-learning gated recurrent neural networks,MLGRU)。将齿轮原始振动信号经过征提取、特征筛选以及特征融合后的指标作为门控神经网络的输入;在元学习网络框架下搭建门控神经网络,在防止齿轮预测模型过拟合的同时通过多任务的训练方式优化元学习门控神经网络,使其在小样本条件下快速收敛;通过门控神经单元深度学习齿轮演化规律并进行退化趋势预测。试验结果表明所提元学习门控神经网络能够充分学习齿轮演化机理,实现齿轮退化趋势预测。To address the low accuracy of gear degradation trend prediction due to over-fitting in deep neural network training under small gear sample conditions,a meta-learning gated recurrent unit(MLGRU)network is proposed for gear degradation trend prediction.The gated recurrent neural network is built on the framework of meta-learning,and the network is optimized through multi-task training to reduce over-fitting of the gear prediction model,so that the MLGRU networks can converge quickly under small-sample conditions.Deep learning of gear evolution patterns and prediction of degradation trends by gated recurrent unit network.The experimental results show that the proposed meta-learning gated neural networks can fully learn the gear evolution mechanism and predict the gear degradation trend.

关 键 词：齿轮 退化趋势预测 元学习 门控神经网络 小样本 

分 类 号：TG156[金属学及工艺—热处理]