基于注意力多通道卷积ON-LSTM的APU剩余寿命预测

APU Remaining Useful Life Prediction Based on Attention Multi-Channel ConvolutionalON-LSTM Method

针对飞机辅助动力装置状态监测数据纬度高、数据量大等特点,提出一种基于注意力机制和多通道卷积ON-LSTM的剩余寿命预测方法。首先利用一维卷积神经网络对传感器参数进行局部特征提取;其次利用ON-LSTM能够学习序列长期依赖的优势,对传感器参数进行时序特征提取;再次通过注意力机制确定各参数的权重,准确预测APU的剩余使用寿命;最后为验证方法的有效性,利用APU实际在翼监测数据开展测试。研究结果表明,所提出的模型预测精度优于现有的支持向量机、深度置信网络和长短期记忆神经网络等方法,为APU的健康管理提供重要支撑。

Aiming at the characteristics of high dimension and large scale of condition monitoring data of aircraft auxiliary power unit,a remaining useful life prediction method based on attention mechanism and multi-channel convolutional ordered neuron long short-term memory(ON-LSTM)was proposed.Firstly,one dimension convolutional neural network was used to extract local feature of parameters.Secondly,the temporal feature of sensor data was extracted by ON-LSTM,in view of its advantage of learning long-range dependencies.Thirdly,attention mechanism could determine the weight of each monitoring sensor,thus predicting APU remaining useful life accurately.In addition,to validate the effectiveness of the proposed method,a test was carried out on a real APU on-wing monitoring dataset.The result demonstrated that compared with SVM,DBN and LSTM,the proposed method could achieve better performance,providing strong support for health management of APU.