A Study on Health Indicator for Pump Degradation Diagnosis Using Autoencoder with Bayesian Neural Network

Abstract

The Pump is the main mechanical system for power and transferring fluids. Failure of the pump results in financial and temporal costs, and failure diagnosis is essential. Recently, data driven condition monitoring has been used to detect abnormalities. In addition, there is a need to evaluate the level of degradation as well as to determine whether or not there is a pump defect. Therefore, this paper proposes a robust health diagnostic method based on a data-based approach for pump state-based maintenance decisions. Data-based approaches require condition data for learning artificial intelligence model. Abnormal detection and degradation level evaluation are performed using pump vibration data acquired under normal operating conditions. The latent vector that is more robust to the uncertainty of the deviation that exists between the training data and test data distributions is extracted using Bayesian neural network based Autoencoder model. The latent vector is converted into a single factor using the Mahalanobis distance. The probability distribution of the factor value in the normal condition and the probability distribution of the factor value of the system of interest are compared through the Fisher Discriminant Ratio to evaluate the anomaly and the level of degradation. The validity of the artificial intelligence model-based diagnostic method proposed in this paper is verified through simulation cases of hydraulic piston pump and centrifugal pump cavitation. This study is significant in that it developed a robust health diagnosis method considering the inherent variability of normally operating pumps and the uncertainty of field data.