Application of Machine Learning Models in Predictive Maintenance of Francis Hydraulic Turbines

Cavitation is a phenomenon that reduces the useful life of hydraulic machines, taking place in function of the variation of the pressure gradient at a constant temperature. In turbines, cavitation occurs when the turbine operates beyond nominal conditions, generating abnormal vibrations, erosion to blades and other key components, thus resulting in stoppage for maintenance. This article proposes a cavitation monitoring system based on the analysis of vibration spectra via two Machine Learning (ML) models: a Multilayer Perceptron (MLP) neural network and a Radial Basis Function (RBF) neural network. Drawing upon vibration analysis and pressure coefficient parameter standards, such models are capable of identifying the vibratory state of a given machine, distinguishing its cavitating and non-cavitating states. Moreover, it is proposed that these models may estimate real conditions for turbine functioning, thus enabling planning for the most opportune moment to carry out turbine maintenance. Both ML models were evaluated through a series of experiments with data from a Francis turbine, where vibration spectra and flow pressure coefficients were monitored; they identified cavitating and non-cavitating states with precision levels between 95% and 100%, thus demonstrating satisfactory performance and serving as an important step in the development of a system for monitoring hydropowers