论文部分内容阅读
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。
A new fast multi-classification SVM algorithm is proposed to solve the multi-class fault diagnosis of aeroengines in the case of large samples. Firstly, Hierarchical Support Vector Machine (H-SVM) is used to implement multi-class classification. Data is represented by various data centers. Clustering is carried out by using SOFM (Self-organizing Feature Mapping Neural Network) Data into the same sub-class for training, get H-SVM hierarchy. Secondly, when training the binary classifier in H-SVM, the relative boundary vector (RBV) is used instead of all the training samples, the number of training samples is greatly reduced while keeping the classification accuracy almost unchanged, and the training time is significantly shortened At the same time, the classification time is shortened due to the reduced number of support vectors. In the case of more severe aliasing, the new algorithm eliminates aliasing heterogeneous data, then calculates RBV, and selects the samples whose distance to the calculated RBV is less than a certain value to train SVM, which ensures the rationality of RBV. Prevent the loss of key data, effectively improve the classification accuracy. The classification of a Category 5 composite fault in an aero-turbojet engine is simulated. The correct rate of the total fault classification reaches 91.2%. The training time of binary SVM is only 16.20% of the original. When the total number of training samples reaches 7500, Under this algorithm, the number of samples after reduction is only 3.05%. Simulation results show that the proposed algorithm is effective, reliable and easy to implement.