Fault identification of centrifugal pump using WGAN-GP method with unbalanced datasets based on kinematics simulation and experimental case

Abstract

Abstract In practical engineering applications, the accuracy and stability of fault identification for centrifugal pump will be significantly reduced due to unbalanced distribution between normal and fault datasets, i.e., the number of normal working samples is far more than the fault samples. To alleviate this bottleneck issue, this paper explores the fault identification of centrifugal pump based on Wasserstein generative adversarial network with gradient penalty (WGAN-GP) through combining kinematics simulation and experimental case. Specifically, ideal unbalanced vibration datasets from failure patterns such as damaged impeller of centrifugal pump are simulated and collected by prototype ADAMS software, then the unbalanced vibration signals are transformed into 2D grey-scale images. Furtherly, the generated grey-scale image datasets are feed into the original grey-scale image dataset as new datasets for training when the Nash equilibrium of the WGAN-GP model is reached. Eventually, the fault patterns of centrifugal pump are identified using confusion matrix graph. Meanwhile, another public dataset of centrifugal pump is employed for verifying the accuracy of the WGAN-GP model. Results indicate that fault identification accuracies with 95.07% and 98.0% of both kinematics simulation and experimental case are obtained, respectively, and the issues of unbalanced distribution and insufficient dataset can be overcome effectively.