Flow Rate Estimation in Pipes using Flow-induced Vibration and Machine Learning

Abstract

Flow-induced vibration (FIV) has traditionally been considered a source of mechanical failure and noise in pipeline systems. However, this study introduces a novel approach that leverages FIV as a diagnostic tool for non-intrusive flow rate estimation using vibration signals and machine learning (ML). A custom-built experimental test rig was developed, comprising a 2-meter PVC pipeline, a controlled centrifugal pump, and both single-axis and triaxial vibration sensors strategically placed along the pipe. This setup enables high-resolution vibration data acquisition under varying flow conditions without penetrating the pipe wall. The novelty lies in a non-intrusive approach and the application of supervised ML algorithms to decode flowrate from structural vibrations---an area that remains inadequately explored in the current literature. Initial polynomial regression analysis of single-axis sensor data revealed strong flow-vibration correlation (R$^2> 0.96$), validating the physical basis for indirect flow measurement. Subsequently, six ML models---including Gradient Boosted Trees and Deep Learning---were trained on multiaxial vibration data, achieving high predictive performance with correlation coefficients up to 0.94 and RMSE as low as 1.81. In a two-class flowrate classification test (32 vs 25 mm$^3$/s), all models demonstrated near-perfect accuracy. This work provides the first integrated experimental-ML framework for real-time, low-cost, and non-intrusive flow monitoring using FIV, offering significant potential for industrial applications where conventional flow meters are impractical.