摘要
退化预测是装备健康管理的重要技术途径,近年来,大量时间序列预测方法在退化预测中得到应用。然而,众多大型装备由于结构复杂,功能多样,在退化过程中存在明显的阶段性,采用单一的模型对不同阶段的退化进行预测将会出现明显的精度降低,针对不同阶段对模型重新训练也会带来时间和算力的损失。针对多阶段退化的问题,引入了迁移学习的思想,提出了一种退化阶段识别与LSTM-fine-tune结合的多阶段退化预测方法,采用退化数据对LSTM模型训练,之后对部分网络参数进行冻结,在识别到装备出现新的退化阶段后,利用新阶段的退化数据对模型进行微调,以快速匹配不同阶段的数据。为验证模型的有效性,本文以氧气浓缩器为例进行模型应用。结果表明,本文方法能够有效识别氧气浓缩器3个阶段的退化,每个阶段的预测均方差分别为0.507、8.976和0.375,远低于不分段直接预测的均方误差76.87,在训练时间上,对比于每个阶段重新训练时间大幅缩短,在训练精度上,明显优于维纳过程、Lstar等传统方法。
Degradation prediction is an important technical approach for equipment health management.In recent years,a large number of time series prediction methods have been applied in degradation prediction.However,due to the complex structure and diverse functions of many large equipment,there are obvious stages in the degradation process,and the application of a single model to predict the degradation at different stages will significantly reduce the accuracy,and the retraining of the model for different stages will also bring the loss of time and computing power.To solve the problem of multi-stage degradation,this paper introduced the idea of transfer learning and proposed a multi-stage degradation prediction method combining degradation pattern recognition and LSTM-fine-tune.The LSTM model was trained with degradation data,and then part of network parameters was frozen.After identifying the new degradation stage of equipment,the model is fine-tuned with the degraded data of the new stage to quickly match the data of different stages.In order to verify the validity of the model,this paper takes oxygen concentrator as an example to apply the model.The results show that the proposed method can effectively identify the degradation of oxygen concentrator at three stages,and the mean square error of prediction for each stage is 0.507,8.976 and 0.375 respectively,which is far lower than the mean square error of direct prediction without segmentation of 76.87.In terms of training time,compared with the retraining time of each stage,the training accuracy is obviously superior to the traditional methods such as Wiener process and Lstar.
作者
黄崧琳
景博
潘晋新
焦晓璇
王生龙
Huang Songlin;Jing Bo;Pan Jinxin;Jiao Xiaoxuan;Wang Shenglong(Aeronautics Engineering College,Air Force Engineering University,Xi′an 710038,China)
出处
《电子测量与仪器学报》
CSCD
北大核心
2022年第7期136-143,共8页
Journal of Electronic Measurement and Instrumentation
基金
十四五装备预研共用技术项目(50902060401)
基础研究项目群(514010504-304)资助。
