Data Augmentation Considering Distribution Discrepancy for Fault Diagnosis of Drilling Process With Limited Samples

The fault diagnosis during drilling is necessary to prevent the accidents develop to more serious status. Data-driven diagnosis methods have great advantages in nonlinear industrial process, however, the problem of limited samples restricts its further application. This article proposes a data augmentation method based on synthetic data generation and updating for drilling fault diagnosis with limited samples. First, the generator is trained with generative adversarial nets (GAN), and the GAN is improved by the design of parameter selection module, and loss function in generative model. Then, sufficient samples are obtained, and a balanced dataset is constructed for modeling. Meanwhile, by considering the distribution discrepancy, the self-organizing incremental neural network-based synthetic data updating is realized to track the changes of data distribution when the data drift appears. Finally, the actual data acquired from two wells are employed for the method validation. The experimental results illustrate that the proposed method is helpful for improving the performance of diagnosis model with limited samples, and the negative impact to diagnosis model due to the distribution discrepancy also can be overcome.