基于改进Xgboost算法的多联机空调系统故障诊断策略

Fault Diagnosis Strategy for Variable Refrigerant Flow Air-conditioning System Based on Improved Xgboost Algorithm

采用Xgboost算法处理多联机故障诊断问题时,存在算法参数设置难、特征选择稳定性低等缺陷,严重影响诊断精度与效率。本文提出一种改进Xgboost算法,以获得预测性能好、稳定性高的特征变量。本文选择Xgboost算法作为基特征选择器获得5组特征排序,使用交叉验证与网格搜索确定算法参数,采用贪心搜索集成策略对特征子集进行集成,得到集成特征子集。最后,分别选择30,000与40,000组样本数据,对比集成的特征选择法与单一特征选择法在Adaboost模型和Xgboost模型上的预测准确率和泛化性能,结果表明,改进Xgboost算法可以获得最高的预测准确率,在两种模型选择的特征数量较少时,准确率提升明显,最好时分别提升了15.28%和4.68%。

When using the Xgboost algorithm to deal with multi-line fault diagnosis problems, there are defects such as difficult algorithm parameter settings and low stability of feature selection, which seriously affects the accuracy and efficiency of diagnosis. An improved Xgboost algorithm to obtain feature variables with good prediction performance and high stability is proposed in this paper. Xgboost algorithm is selected as the base feature selector to obtain five sets of feature rankings, cross-validation and grid search are used to determine the algorithm parameters, and the greedy search integration strategy is used to integrate the feature subsets to obtain the integrated feature subsets. Finally, 30,000 and 40,000 sets of sample data are selected respectively, and the prediction accuracy and generalization performance of the integrated feature selection method and the single feature selection method on the Adaboost model and the Xgboost model are compared. The results show that, the improved Xgboost algorithm can obtain the highest prediction accuracy. When the number of features selected by the two models is small, the accuracy is improved, and even up to 15.28% and 4.68%, respectively.