基于改进LightGBM模型的汽车故障预测方法研究

Research on Vehicle Fault Prediction Scheme Based on Improved LightGBM Model

针对机器学习技术在汽车行业的应用中存在的汽车故障数据规模大和类别不平衡引起的模型训练速度慢、故障查全率低的问题,对LightGBM模型进行两方面的改进:模型训练时,设置类别权重和L1正则化项修正模型的损失函数,并通过贝叶斯优化获得修正项系数的取值;模型预测时,使用阈值移动法降低模型的分类阈值。在斯堪尼亚货车故障数据集上进行验证。结果表明,本文中所提出的改进LightGBM模型训练速度快,故障查全率高,具备工程应用价值。

In view of the defects of slow model training and low recall rate caused by the large scale and class imbalance of vehicle fault data in the application of machine learning technique to automotive industry,the improvements in two respects are conducted on LightGBM model:firstly,the setting of class weights and the loss function of correction model for L 1 normalization in model training,with the value of correction coefficient obtained by Bayesian optimization;secondly,the lowering of model classification threshold by threshold moving scheme in model prediction.The results of verification on Scania truck fault dataset show that the improved LightGBM model proposed has fast training speed,high recall rate and great engineering application significance.