基于多模型融合的航空电子产品故障预测方法

针对复杂机载环境应力条件下航空电子产品故障预测所面临的退化趋势差异大、训练数据样本量小等问题，提出了一种改进长短期记忆(Long Short-Term Memory, LSTM)神经网络模型与集成学习框架相结合的故障预测方法，以满足现代综合航空电子系统智能调度管理与自主维护保障的需求。该方法在LSTM模型中引入Dropout机制，构建基于不同历史数据集的差异性LSTM模型组，以解决故障预测时序信息记忆问题与小样本条件下数据驱动模型训练过拟合问题；采用Adaboosting算法计算模型权重，并基于实时数据动态调整，以滤除复杂机载环境应力引入的预测误差，解决多模型融合的性能差异问题。最后，采用NASA公开的锂电池退化数据集进行仿真验证，实验结果表明，相较于传统BP神经网络、经典LSTM和LSTM基模型，该方法具有更高的趋势拟合度和预测精度。

An avionics fault prognostic method based on multi-model fusion

Under the complex airborne environment stress conditions,the fault prognostic of avionics products faces such problem as large differences in degradation trends and small sample size of training data.To resolve these problems,and meet the needs of intelligent scheduling and autonomous maintenance of modern integrated avionics systems,a fault prognostic method combining the Long Short-Term Memory(LSTM) neural network model and the integrated learning framework is proposed.To solve the problem of fault prognostic time series information memory,the Dropout mechanism is introduced into the LSTM model,and a differential LSTM model group is built based on different historical data sets to solve the problem of data-driven model training over fitting under the condition of small samples.In the process of model fusion,the Adaboosting algorithm is used to calculate the weight of a single model,and it is dynamically adjusted based on real-time data to filter out the prognostic error caused by the complex airborne environmental stress and solve the performance difference problem of different models.Finally,the lithium battery degradation data set published by NASA is used for simulation verification.The experimental result shows that compared with the traditional BP neural network,the classic LSTM and the LSTM base model,the proposed method has higher trend fitting degree and prognostic accuracy.