基于最大均值差异迁移学习的飞机燃油泵故障诊断  

Fault Diagnosis of Aircraft Fuel Pump Based on Maximum Mean Discrepancy Transfer Learning

作　　者：刘振 张小军[1,2] 潘俊[3] 叶茂 苗扬[1,2] LIU Zhen;ZHANC Xiaojun;PAN Jun;YE Mao;MIAO Yang(Faculty of Materials and Manufacturering,Beijing University of Technology,Beijing 100124;Beijing Key Laboratory of Advanced Manufacturing Technology,Beijing University of Technology,Beijing 100124;AVIC Nanjing Electromechanical Hydraulic Engineering Center,Nanjing,Jiangsu 211102;Shanghai Institute of Mechanical and Electrical Engineering,Shanghai 200050)

机构地区：[1]北京工业大学材料与制造学部,北京100124 [2]北京工业大学先进制造北京市重点实验室,北京100124 [3]中国航空工业集团南京机电液压工程研究中心,江苏南京211102 [4]上海机电工程研究所,上海200050

出　　处：《液压与气动》2023年第10期62-69,共8页Chinese Hydraulics & Pneumatics

基　　金：国家自然科学基金(51975011);北京工业大学国际科研合作种子基金(2021b24);北京工业大学材料与制造学部“启航计划”(QH202206)。

摘　　要：燃油泵是飞机燃油系统的核心部件。燃油泵的工作环境恶劣,一旦发生故障,会对飞机的飞行安全造成严重的后果,所以对飞机燃油泵进行故障诊断至关重要。现阶段飞机燃油泵故障存在一些问题,如故障较为随机,故障数据稀缺等。针对这些问题,提出了基于深度迁移学习的飞机燃油泵故障诊断算法。首先,将飞机燃油泵和其他结构相似泵的数据分别作为目标数据和训练数据,利用小波包分解,对两种泵的原始数据进行特征提取。其次,对分解后的两种故障数据进行过采样处理,增加数据量。使用最大均值差异算法(MMD)作为衡量域损失的度量,并将其嵌入到1维卷积神经网络(CNN)结构中。使用该算法对数据进行训练,最终完成故障分类。实验结果表明,该算法相对于BP神经网络、LSTM以及CNN有更好的准确性。Fuel pump is the core component of aircraft fuel system.Fuel pumps operate in harsh environments.It will cause serious consequences to the aircraft flight safety when the fault occurs.Therefore,Timely fault diagnosis of aircraft fuel pump is necessary.At present,aircraft fuel pump faults have some problems,such as random faults and scarce fault data.To solve these problems,this paper proposed a fault diagnosis algorithm based on deep transfer learning.Firstly,the data of aircraft fuel pump and other pumps with similar structure are used as target data and training data respectively,and this method extracts features from the original data of the two pumps by wavelet packet decomposition.Secondly,the two kinds of decomposed fault data are oversampled to increase the amount of data.These data are trained by the algorithm which inserts Maximum Mean Discrepancy(MMD)in 1D-CNN network structure.Wherein,MMD is used to measure the domain loss.Experimental results show that the algorithm has better accuracy than BP neural network,LSTM and CNN.

关 键 词：飞机燃油泵 故障诊断 迁移学习 最大均值差异 

分 类 号：TH137[机械工程—机械制造及自动化]