Diagnosis of different degree of blockage levels in the Centrifugal Pump Employing Vibration, Pressure and Current Data through Artificial Neural Network

The appearance of flow instabilities like the blockage severity, impeller cut flaws, pitted cover plate flaws can cause to diminish the efficiency of centrifugal pump (CP), and may result in excessive vibration and noise, and their failure may lead to the system imploding. To bridge the gap of downfall in the efficiency of CP, it is crucial that a system can be created to monitor the condition of the CP and must be maintained. The present work proposes at identifying and determining the severity of various blockage levels in the inlet pipe with three different kinds of pumps using three distinct sensors. One pump works faultlessly (healthy pump), another has cuts artificially made on the impeller blade, and the third has pits artificially created on the cover plate. The inlet pipe blockage mimics pump blockage which is made more severe step by step. As the blockage gets worse and the flow slows down, recirculation starts, causing vapor bubbles to form. Utilizing a mechanical modulating valve, the inlet flow area of the pipe is partitioned into six intervals (0%, 16.7%, 33.3%, 50%, 66.6%, and 83.33%) to replicate pump blockage. This obstruction directly influences vibrations, current line signals, and fluid dynamic pressure. To gather data across a spectrum of blockage levels and operational frequencies (30 Hz, 35 Hz, 40 Hz, 45 Hz, 50 Hz, 55 Hz, and 60 Hz), a combination of a pressure transducer, accelerometer, and current probes were strategically employed in this investigation. Multiple sets of statistical features were extracted from the data, and through various algorithms, the most effective combined statistical feature set was determined. In this domain, the combination of standard deviation, mean, and entropy demonstrates superior performance compared to other features. This feature set was input into an ANN model, which is developed by optimizing parameters like hidden layer count, neurons, epochs and then the results of this investigation are then compared with existing literature. It has been noted that employing combinations of multiple sets of statistical features significantly improves the accuracy in identifying obstruction levels, often achieving near-perfect accuracy for various feature sets (nearly 100% across various combinations). In comparison to other SOTA methods, this approach achieves higher accuracy, ranging from 2.41% to 15.69% across different metrics. This study presents a method to classify inlet pipe blockages into various levels, enhancing maintenance prioritization and reducing downtime and repair costs, ensuring long-term equipment health and operational efficiency. The fault prediction methodology proves highly robust across various CP operating conditions. Conflict of Interest Statement The authors declare no conflicts of interest.