Electrical Power Data Based Fault Diagnosis Of Oil Pumping Units

To adapt to the development toward digitalization and intelligentization,electrical power signal based oil-pumping-units fault diagnosis becomes increasingly popular and is widely used in real-world application.In this paper,we aim to establish a fault diagnosis system for oil pumping units using electrical power signal.To realize this,novel machine learning methods need to be proposed to adapt to specific characteristics of target data.There are two main problems,i.e.class imbalance and large data collection,need to be addressed when using electrical power data to establish a diagnosis system.In our work,the imbalanced classification is employed to solve the problems caused by class imbalance,and the Active Learning framework is used alleviate the learning burden and manual annotation cost caused by large data collection.The paper also covers active learning methods for imbalanced data sets,so that these two problems can be solved simultaneously.In the first place,we introduce the electrical power signal and its feature extraction method.Specifically,this chapter introduces the source of electrical power signal,sampling and storage method,and electrical power signals of typical working patterns.The feature extraction for the electrical power signal is also included;we use wavelet transform to decompose the electrical power signal and calculate energy of sub-signals as features.In addition,we analyze the electrical power signals according to the mechanism of oil-pumping units to get the signal energy features,power features and AUC features.To solve the imbalanced problem of target data,imbalanced classification is employed.In our work,the Support Vector Machine is used as the basic classification algorithm,and it is combined with undersampling method(SVM-RANDU),oversampling method(SVM-SMOTE)and cost-sensitive method(SVM-WEIGHT).Moreover,the binary imbalanced classification is extended to multi-class imbalanced classification via one-against-one strategy.This method is applied to electrical power data.The experimental results show that imbalanced classification is beneficial for learning on the electrical power data.As to the problem of large data collection,we use active learning strategy to improve the learning efficiency and alleviate the labeling cost of human annotators.Since the electrical power data is also imbalanced,our research focuses on active learning strategies for large imbalanced data set.We propose clustering based active learning to solve the malfunction of traditional active learning methods on imbalanced datasets.The proposed method incorporates clustering analysis during the instance selection process during active learning and the main purpose is to efficiently select minority instances during instance selection.The proposed method is evaluated on benchmarks and obtains satisfactory results.Further,the clustering based active learning is used to analyze real-world data recorded in the oil-field production.The proposed method also shows its advantage over tradition methods in establishing fault diagnosis system for electrical power data.