Artificial Neural Network-Based Fault Detection System with Residual Analysis Approach on Centrifugal Pump: A Case Study

Abstract

Centrifugal pump is an instrument that is widely used in industry and has become the main driving component. A detection system is often needed to prevent damage to these pumps because they can interfere with the overall system performance. Therefore, this study discussed the development of a fault detection system for two centrifugal pump units, namely the Medium Pressure Oil Pump (MPOP) and the Water Injection Pump (WIP). In detecting the operating conditions of the pump, it was used a residual feature extraction technique in the time domain with a statistical approach. Residual was generated by using three sub-systems of a pumping system. Each sub-system was modeled using an artificial neural network with feedforward-back propagation architecture. Based on the feature values, the classifier was designed to classify pump conditions. Then the proposed fault detection system was applied in a condition monitoring system scheme. The test results (using data from the field) show that the fault detection system has an accuracy of 91.67% for MPOP and 94.8% for WIP cases. Meanwhile, the fault detection system has an accuracy above 99% during online monitoring simulations.