基于决策树的涡轴发动机气路健康评估方法

Gas Path Health Assessment Method of Turboshaft Engine Based on Decision Tree

关键词:  涡轴发动机  健康评估  梯度提升决策树  特征筛选  分类器

Key words:  Turboshaft engine  Health assessment  Gradient boosting decision tree  Feature selection  Classifier

摘要:

为解决涡轴发动机退化模式多、故障样本大、分类困难的问题，将梯度提升决策树（GBDT）算法应用于涡轴发动机气路健康评估。首先对涡轴发动机故障标签进行了拆解，针对每一类标签设计了一个分类器。通过特征重要度排序筛选了6个分类标签的输入特征，有效降低模型复杂度。通过遍历确定树的最优数量和深度，确保了分类的准确性。仿真结果表明，特征筛选减少测试时间16%～35%；相比广泛使用的支持向量机（SVM）算法，在特征数量相同条件下，GBDT测试时间缩短了31.88%～65.28%，相比极限学习机（ELM），误诊样本数量低于其千分之一。可见GBDT算法在涡轴发动机大样本故障评估中表现出更高的分类精度，验证了其在发动机气路健康评估上的有效性。

Abstract:

In order to solve the problem of turbofan engine health assessment which is characterized by multi-degraded modes， tremendous fault samples and hard to classify， the gradient boosting decision tree （GBDT） algorithm was applied to the gas path health assessment of turboshaft engine. Firstly， the fault labels of turboshaft engine were disassembled， and a classifier was designed for each tag. The input features of six classification tags were screened out according to the feature importance ranking， which effectively reduced the complexity of the model. The traversal method was adopted to determine the optimal number and depth of the tree to ensure the accuracy of the classification. The simulation results show that feature selection reduces test time by 16% to 35%. Compared with the widely used support vector machine （SVM） algorithm， under the condition of the same number of features， the test time of GBDT is reduced by 31.88% to 65.28%， and compared with extreme learning machine （ELM）， the number of misdiagnosed samples is less than 1/1000 that of ELM. It can be seen that the GBDT algorithm has higher classification accuracy in the health assessment of turboshaft engine with a large amount of fault samples， which verifies its effectiveness in engine gas path health assessment.