应用快速多分类SVM的航空发动机故障诊断方法

Aero-Engine Fault Diagnosis by a New Fast Multi-Class Support Vector Algorithm

关键词:  航空发动机  支持向量机  故障诊断  大规模训练集  样本约减  神经网络

Key words:  Aero-engine  Support vector machines  Fault diagnosis  Large-scale training set  Sample pruning  Neural networks

摘要:

提出了一种新的快速多分类SVM算法,用于解决大样本情况下航空发动机的多类故障诊断问题。首先,选用层次支持向量机(H-SVM)来实现多类分类,用各类数据中心代表该类数据,通过自组织特征映射神经网络(SOFM)进行聚类,把类中心之间距离较近的数据归为同一个子类进行训练,得到H-SVM层次结构。其次,在训练H-SVM中的二元分类器时,应用相对边界向量(RBV)代替全部训练样本,在保持分类精度几乎不变的条件下大幅度减少了训练样本数,使训练时间明显缩短；同时,由于支持向量的数量减小,分类时间也相应缩短。在分类数据混迭较为严重的情况下,新算法先剔除混迭的异类数据,再计算RBV,并且把与计算的RBV距离小于一定数值的样本都选择来训练SVM,保证了RBV的合理性,防止了关键数据的丢失,有效提高了分类精度。针对一个航空涡喷发动机5类复合故障的分类进行了实例仿真,总的故障分类正确率达到 91.2%,二元SVM的训练时间最多只有原来的16.20%；当训练样本总数达到7500的大规模情况下,根据本算法,约减后的样本数量只有原来的3.05%。仿真结果表明,提出的算法有效、可靠,容易实现。 

Abstract:

A new fast multi-class support vector algorithm was presented to solve the multi-class fault diagnosis problem for aero-engine at the condition of large-scale training set. According to the fast algorithm, a hierarchical support vector machine (H-SVM) was chosen for multi-class fault classification. Before SVM training, the training data are first clustered according to their class center Euclid distances with self-organizing feature mapping (SOFM) neural networks. The patterns which have close distances are divided into the same sub-classes to train. This ensures the H-SVM reasonable hierarchical construction and better generalization performance. The relative boundary vectors (RBVs) instead of all the original training samples are used for the training of the binary SVM fault classifiers in H-SVM. This pruning strategy decreases the number of final training sample significantly and can keep classification accuracy almost invariable. Accordingly, the training time is shortened greatly,compared with basic SVM classifier. Meanwhile, owing to the reduction of support vector number, the classification time is also reduced. When a serious sample aliasing exists, the aliasing sample points which are not the same class are eliminated before the relative boundary vectors are computed. Besides, the samples near the relative boundary vectors are selected for SVM training in order to prevent the loss of some key sample points resulted from aliasing. This can improve classification accuracy effectively. A simulation example to classify 5 classes of combination fault of a turbojet engine is finished:the total fault classification accuracy reaches 91.2% and the maximum training time of all the binary SVMs is only 16.20% of basic SVM classifier. When the original training sample number is increased to 7500, the final training sample number after pruning is reduced to 3.05% of basic SVM. Simulation results show that this fast multi-class support vector algorithm is effective, reliable and easy to be implemented for engineering application.