| The design,manufacturing,and maintenance of automotive products generate a large amount of data.Using machine learning technology to mine the value of automotive big data is of great significance for advancing the digital transformation of the automotive industry.In tasks such as fault prediction,automotive big data is often class-imbalanced: the number of samples in different categories in the dataset is significantly different,and the cost of misclassification of the minority class is higher than that of the majority class.The class-imbalanced characteristic makes the machine learning model not enough to learn the minority class,and it is difficult to identify the existence of the minority class during prediction,which brings high cost and even risks to tasks such as fault prediction.Aimed at the above problems,a machine learning modeling process for classimbalanced datasets is constructed.At the feature level,a combined feature selection method based on Relief F and IG algorithms is studied.At the model level,an class-imbalanced classification model based on the Light GBM model is proposed.The case study and verification of the classimbalanced automotive big data provide a reference solution for efficient mining of the value of automotive big data.The main contributions and conclusions of this dissertation are summarized as:(1)The machine learning modeling process for class-imbalanced datasetsBy analyzing the basic framework of machine learning modeling for class-imbalanced datasets,a machine learning modeling process including data preprocessing,feature selection,model training and tuning,model prediction and evaluation is constructed.(2)Research on combined feature selection method based on Relief F and IG algorithmsAimed at the problem that the single feature selection algorithm does not fully evaluate the feature importance,the advantages of Relief F and IG algorithms are analyzed,and the combined strategy of two feature selection algorithms is studied: the operation of Relief F and IG algorithms are accelerated by random undersampling,normalization is used to deal with the feature importance acquired by Relief F and IG algorithms,and a new feature importance index is defined.Thus,a combined feature selection method is proposed,which enhances classification performance.(3)Research on class-imbalanced classification model based on the Light GBM modelAimed at the problem that the Light GBM model is not enough to deal with class-imbalanced datasets,based on cost-sensitive learning technology,class weights and L1 regularization are used to modify the loss function during model training,and the threshold shift method is used to reduce the classification threshold during model prediction.Based on the Bayesian optimization method,TPE algorithm is used to tune the hyperparameters of the model.Thus,a class-imbalanced classification model based on the Light GBM model is constructed,which enhances the modeling ability of class-imbalanced datasets.(4)Engineering application of machine learning methods for classimbalanced automotive big dataFor the APS Failure at Scania Trucks engineering case with classimbalanced characteristics,experiments based on the above two researches were carried out separately and serially.The experimental results show that the two researched in this dissertation effectively enhance the predictive ability of the engineering case with class-imbalances characteristics,and are superior to the research results of related literatures,which verifies the validity and engineering value of this dissertation. |
