A low-cost non-intrusive thermal flow meter and fault detector enhanced by machine learning

Non-intrusive flow meters measure flow rate without direct interaction with the flowing fluid. This reduces the cost of flow measurement by eliminating production stoppage for installation and maintenance. However, most commercially available non-intrusive flow meters come at a high purchase cost and require calibration and careful installation for accurate measurements. This study presents a low-cost, non-intrusive flow meter and fault detector combined into a single device for steady-state water flow in a steel pipe. Instead of relying on traditional empirical correlations, various machine learning techniques are employed to establish relationships between temperature response and flow rates. Pipe surface temperature is measured for volumetric flow rates ranging from 5.99×10⁻⁴ m³/s to 2.39×10⁻³ m³/s while a band heater applies heat to the pipe for 60 s. Multiple regression learning techniques are used to correlate temperature measurements with volumetric flow rate, and classification learners are evaluated for fault detection. Three temperature-based parameters are used to train the machine learning models: temperature rise, average rate of temperature rise, and average rate of temperature drop after the heating period ends. The Fine Tree model demonstrated the highest accuracy in predicting flow rate, while the Bagged Trees model achieved the best performance for fault detection. Despite a narrower flow rate range and higher uncertainty compared to commercial flow meters, the proposed device costs <10 % of the average price of four commercially available alternatives with similar specifications.