基于孪生减元注意力网络的航空发动机故障诊断  

Aero-engine fault diagnosis based on Siamese reduced-neuron attention networks

作　　者：王月 赵明航 刘雪云 林琳[1] 钟诗胜[1,2] WANG Yue;ZHAO Minghang;LIU Xueyun;LIN Lin;ZHONG Shisheng(School of Mechatronics Engineering,Harbin Institute of Technology,Harbin 150001,China;School of Ocean Engineering,Harbin Institute of Technology at Weihai,Weihai Shandong 264209,China)

机构地区：[1]哈尔滨工业大学机电工程学院,哈尔滨150001 [2]哈尔滨工业大学(威海)海洋工程学院,山东威海264209

出　　处：《航空动力学报》2023年第7期1784-1792,共9页Journal of Aerospace Power

基　　金：国家自然科学基金联合基金(U1933202);山东省自然科学基金(ZR2020QE156)。

摘　　要：针对传统故障诊断方法在故障样本缺乏条件下容易遭遇过拟合,以及强噪声条件下微弱故障特征难以提取的问题,提出了一种基于孪生减元注意力网络的航空发动机故障诊断方法。根据孪生神经网络的原理,将训练样本集中的样本随机两两配对,使输入从样本变为样本对,实现样本量的扩增。在特征提取模块引入减元注意力机制。其中,注意力机制能够通过全局扫描,快速找到有用特征,并且抑制冗余特征,这与航空发动机微弱气路故障特征被噪声所淹没的情况吻合良好;减元操作可以降低模型的参数量,缓解过拟合现象。研究结果表明:该方法在某航空公司CFM56-5B/7B系列发动机的真实监测数据上取得了88.39%的平均准确率。In view of the problems that traditional fault diagnosis methods are prone to over-fitting under the condition of insufficient fault samples,and the weak fault features are difficult to be extracted under strong noise conditions,an aero-engine fault diagnosis method based on Siamese reduced-neuron attention networks was proposed.According to the principle of Siamese neural network,pairwise coupling of the samples in the training dataset was conducted,so that the input was changed from samples to sample pairs,and the diversity of input was improved.A reduced-neuron attention mechanism was integrated into the feature extraction module.Among them,the attention mechanism can quickly find useful features through global scanning,and suppress redundant features,which was in good agreement with the situation where the weak gas path fault features of aero-engines were submerged by noise;the reduced-neuron operation can reduce the amount of parameters and alleviate overfitting.The results show that this method achieves an average accuracy of 88.39%on the real monitoring data of CMF56-5B/7B series engines of an airline.

关 键 词：故障诊断 孪生神经网络 减元注意力 涡扇发动机 小样本 

分 类 号：V263.6[航空宇航科学与技术—航空宇航制造工程]