Data augmentation using improved conditional GAN under extremely limited fault samples and its application in fault diagnosis of electric submersible pump

Abstract

Electric submersible pump (ESP) in offshore oilfields is one of the important artificial lifting methods to achieve high and stable production. The complexity of the ESP system and the long pumping cycle result in data having the typical characteristics of "a large amount of data and a small amount of information". Therefore, the scarcity of valid samples causes a major challenge for ESP fault diagnosis. To address these practical problems, we propose an intelligent virtual sample generation method that introduces the idea of multi-distribution mega trend diffusion (MD-MTD) into conditional generative adversarial networks (MCGAN-VSG). In the MCGAN-VSG method, the acceptable diffusion range of the sample attributes is first obtained by estimating the samples using the triangular probability distribution model constructed in MD-MTD. Secondly, the Borderline-SMOTE and uniform distribution were added to describe the small sample properties, and suitable output samples are generated to fill the information gap between samples for resampling with Bootstrap. Thirdly, CGAN is used to generate the input samples corresponding to the output samples. Finally, the accuracy of the classification model is improved by generating a large number of virtual samples with an extremely limited number of fault samples. In order to verify the advantages of the proposed MCGAN-VSG, the quality of the input and output virtual samples generated via the method is investigated through a two-dimensional standard function. The proposed method is further applied to the fault diagnosis of ESP in an offshore oil field, and the effectiveness of MCGAN-VSG is verified with actual industrial data. The MCGAN-VSG was compared with most advanced methods such as MTD, TTD, Bootstrap and MD-MTD, and the experimental results show that the proposed method is superior to all other methods.