The Imbalanced Learning Method Of Optimal Margin Distribution Machine And Its Industrial Application

The imbalanced data and samples are widely used in pattern recognition applications in the fields of fault diagnosis,fraud detection,cyber attack supervision,and disease diagnosis.They always have large negative impact on the pattern recognition accuracy and the generalization performance of learning models.The optimal margin distribution machine has a rigorous basic theory and good generalization performance,based on various theories and methods such as kernel modification of classifier,feature engineering and deep learning,this paper proposes several optimal margin distribution machine for unbalanced sample learning.Combining the extraction architecture of fusion features from industrial imbalanced data,aiming at the problem of poor recognition accuracy of abnormal sintering conditions in rotary kilns,a high-precision,robust and low-risk sintering condition recongnition framework was developed based on the imbalanced thermal signals collected from industrial site.In this way,the proposed imbalance learning method has verified by the industrial application.The main work and innovations of this paper are as follows:(1)Under the condition of imbalanced samples,the separator learned by conventional classification algorithms is usually skew toward the minority area,resulting in a small margin of minority samples and poor generalization ability of the classifier.In this regard,based on the mechanism of the optimal margin distribution machine(ODM)and the kernel modification theory,a novel conformal transformation function based on the margin of training samples is constructed to modify the kernel function of ODM,and a new type of imbalanced classification model,which is named margin based kernel modified ODM(mb KMODM),is proposed.The performance of the mb KMODM is verified on two-dimensional visualization data and UCI standard data sets.The experimental results show that the proposed mb KMODM can alleviate the skewness of the learned separator caused by imbalanced samples,and balance the classification accuracy of different classes.While ensuring the generalization ability of the recognition model.(2)For the construction of a conformal transformation function based on the margin,the standard classifier needs to be pre-trained.However,pre-training the classifier on imbalanced samples may bring about problems such as low model calculation efficiency,and the margin cannot reliably reflect the spatial distribution of the samples.Aim to this,this paper proposes the methods for extracting boundary samples in kernel space.Using the relative position between samples and the obtained boundary samples to replace the margin,the space distribution of samples is represented more reasonably.Based on this,a new type of conformal transformation functions is constructed and the imbalanced classification models bs KMODM is obtained.By analysising the classification results of proposed model on the UCI data set,it can be proved that the the proposed imbalance learning models have excellent imbalance classification performance.(3)In view of the complicated process of extracting boundary samples in kernel space and calculating the relative position,this paper proposes a method for calculating the average distance of heterogeneous samples in kernel space.The average distance of heterogeneous samples in kernel space can be directly calculated by the kernel matrix,which has high calculation efficiency and can reflect the spatial distribution of training samples more accurate.Based on these,this paper redesigned the construction method of conformal transformation function,and proposed an imbalanced classification model dd KMODM.By analyzing the spatial expansion coefficients of different regions in kernel space,and comparing the classification results with multiple latest models on the UCI dataset,the ability of dd KMODM to balance the effects of different class of samples by adjusting the kernel spatial resolution is verified.It also further confirms that dd KMODM has the advantage of lower computational complexity.(4)The existence of imbalanced raw data in complex industrial sites usually causes the problems such as low separability of the extracted features and unclear decision boundaries of sample.In view of this,this paper proposes an imbalanced feature extraction framework that deeply integrates the prior knowledge of experts and hidden information of raw data.The framework fully considers the common characteristics,such as large time delay and imbalance data volume of different modes,of complex industrial process data.It mainly includes three modules: delay analysis,feature extraction that integrates prior knowledge and hidden information and feature reduction,and the application verification was carried out in the feature extraction task of thermal signal of rotary kiln.By comparing the separability of the single feature and the fusion feature after dimensionality reduction,it proves the superiority of the proposed imbalanced feature extraction framework that combines prior knowledge and depth information.(5)Due to the different frequency of different sintering conditions at the production site of the rotary kiln,the occurrence probability of normal condition is much higher than that of abnormal conditions,the samples of different sintering conditions are not balanced,and the recognition accuracy of abnormal conditions by traditional classification algorithms is not high.In view of this,by using the imbalanced classification model and feature extraction framework proposed in this paper,this paper constructs a high-precision,robust and low-risk sintering condition recognition framework for rotary kiln based on imbalanced thermal signals.Through real thermal data experiments and industrial field applications,the excellent performance of the framework proposed in this paper is verified.