深度置信网络在发动机气路部件性能衰退故障诊断中的应用研究

发布时间：2019-03-26 17:42

【摘要】：为提高发动机转动部件性能衰退故障诊断精度,针对传统的浅层网络和支持向量机(SVM)方法在诊断时存在泛化能力欠缺、易产生局部最优解等问题,引入近年来在模式识别领域取得巨大突破,模拟人脑多层结构的深度置信网络(DBN)进行发动机部件性能衰退故障的诊断。为改进深度置信网络性能,提出一种在无监督和有监督训练阶段都可自适应调整权值的改进算法(ad_DBN)。以涡扇发动机为对象,将两种DBN算法与BP,RBF和SVM方法从诊断精度、计算时间、抗噪能力三方面进行综合比较分析。结果表明DBN算法诊断精度明显优于反向传播(BP)神经网络,径向基(RBF)神经网络和支持向量机(SVM)方法,得益于权值的自适应调整,ad_DBN诊断的平均精度高达97.84%,其抗噪声能力也明显优于其他算法,能够提高故障诊断的有效性和可靠性。
[Abstract]:In order to improve the fault diagnosis accuracy of engine rotating component performance decline, the traditional shallow layer network and support vector machine (SVM) method have some problems, such as lack of generalization ability, easy to produce local optimal solution and so on. In recent years, a great breakthrough has been made in the field of pattern recognition, which simulates the deep confidence network (DBN) of the multi-layer structure of the human brain to diagnose the performance degradation of the engine components. In order to improve the performance of deep confidence networks, an improved algorithm (ad_DBN), which adaptively adjusts weights in both unsupervised and supervised training stages, is proposed. Taking turbofan engine as an object, two kinds of DBN algorithms and BP,RBF and SVM methods are compared and analyzed comprehensively from three aspects: diagnostic accuracy, calculation time and anti-noise ability. The results show that the diagnostic accuracy of DBN algorithm is obviously better than that of backpropagation (BP) neural network, radial basis (RBF) neural network and support vector machine (SVM) method, which benefit from the adaptive adjustment of weights. The average accuracy of ad_DBN diagnosis is as high as 97.84%, and its anti-noise ability is obviously superior to other algorithms. It can improve the validity and reliability of fault diagnosis.
【作者单位】：海军航空工程学院飞行器工程系;海军航空工程学院研究生管理大队;
【基金】：国家自然科学基金(51505492) “泰山学者”建设工程专项经费资助
【分类号】：V263.6

【相似文献】