Improved vibration-based condition monitoring for Pelton turbines: experimental assessment on a prototype

To support grid stability, hydropower plants are increasingly required to operate flexibly, subjecting turbines to frequent start-stop cycles and off-design conditions. In this context, Pelton turbines face a higher risk of fatigue-related damage, highlighting the need for advanced condition monitoring (CM) tools to support early fault detection and effective maintenance. This article presents the experimental implementation and validation of a vibration-based CM system on a horizontal-axis Pelton prototype, combining a physics-based approach with data-driven techniques. The system collected vibration signals from various machine locations and operating conditions from supervisory control and data acquisition. Condition indicators (CI) were defined based on prior experimental and numerical analysis of the turbine’s dynamic behavior. Next, multidimensional hill charts of the machine’s vibration across its operating range were generated. Reference, alarm, and trip layers were defined using convolutional neural network-low short-term memory-attention models. Following a period of operation, a fatigue-induced runner bucket failure occurred. The turbine continued to operate until the next scheduled overhaul, allowing for a direct comparison of vibration data before and after damage. While overall vibration levels remained relatively stable, certain CIs showed a high sensitivity to the failure. These findings show that using a reduced set of dynamic-based indicators can enhance monitoring while keeping computation and storage demands low.