| 摘要: |
| 由于航空电子设备性能退化趋势与工作环境（温度、振动、负载等）存在强耦合关系，历史数据和实时数据分布存在难以量化的差异，因此航空电子设备的故障预测一直是业内难题。针对工程应用中的故障预测需求，提出了一种基于多深度置信网络（Multi-deep Belief Network，DBN）模型融合的故障预测方法，基于历史数据和实时数据对多个DBN模型进行含Dropout的迁移训练，有效解决历史域和目标域数据分布差异带来的预测偏差；采用改进遗传算法对DBN模型组进行融合，在迁移学习的基础上进一步提升故障预测的精准度。实验显示，所提方法预测的均方根误差为0.008，相对误差均值为0.9〖WT《Times New Roman》〗%〖WTBZ〗，相关度为0.964 7，预测精度高于单一DBN模型和支持向量机，在航空电子设备的故障预测领域有一定的应用价值。 |
| 关键词: 航空电子设备 故障预测 深度置信网络 迁移学习 |
| DOI： |
|
| 基金项目: |
|
| Fault Prognostics of Avionics Equipment Based on Multi-deep Belief Network Fusion |
| LIANG Tianchen |
| (Southwest China Institute of Electronic Technology,Chengdu 610036，China) |
| Abstract: |
| Due to the strong coupling relationship between the performance degradation trend of avionics and changeable environment(temperature,vibration,load and other),it is difficult to quantify the difference between historical and real-time data distribution,which brings the fault prognostics for avionics an industry challenge.For the prognostics requirements in engineering,a prognostics method based on multi-deep belief network(DBN) model fusion is proposed.Firstly,according to historical and real-time data,transfer learning train with Dropout for DBN models is performed to solve prognostics bias caused by data distribution differences between historical and target domains.Secondly,the enhanced genetic algorithm is used for fusing DBN model group to further improve the accuracy of fault prognostics based on transfer learning.After experimental verification,the root mean square error value of the proposed method is 0.008,the relative error is 0.9%,the correlation is 0.964 7,and the prognostics accuracy is higher than those of a single DBN model and support vector machine,which has certain application value in the field of fault prognostics for avionics. |
| Key words: avionics equipment fault prognostics deep belief network transfer learning |
