Pump Cavitation Detection with Machine Learning: A Comparative Study of SVM and Deep Learning

In the pursuit of enhancing industrial pump reliability and efficiency, this paper addresses the challenging issue of pump cavitation detection through the innovative application of machine learning techniques. Cavitation, a prevalent problem in pumps, significantly compromises their performance, causing damage and operational inefficiencies. Traditionally, cavitation detection has relied on numerical analysis and signal processing methods, which, despite their merit, often fall short in real-world applications due to their requirement for extensive domain knowledge and controlled operational conditions. This study diverges from conventional approaches by harnessing the power of machine learning to predict cavitation occurrences in pumps under varying real-world conditions with high accuracy. We present an analysis of a cavitation dataset compiled by the Danish pump manufacturer Grundfos, which includes vibration data from 297 experiments on seven different pumps, using both traditional machine learning models, specifically Support Vector Machine (SVM), and advanced deep learning techniques. Our methodology includes a detailed examination of the dataset, feature engineering, target definition, problem formulation, model design, and rigorous model testing on target hardware. Remarkably, our study not only demonstrates that machine learning models, particularly deep learning models, can adaptively and accurately predict cavitation but also emphasizes the importance of testing these models on target hardware to ensure their practical applicability. This work is accompanied by an open-source implementation.