Temperature Calibration Using Machine Learning Algorithms for Flexible Temperature Sensors.

Abstract

Thermal imbalance can cause significant stress in large-scale structures such as bridges and buildings, negatively impacting their structural health. To assist in the structural health monitoring systems that analyze these thermal effects, a flexible temperature sensor was fabricated using EHD inkjet printing. However, the reliability of such printed sensors is challenged by complex dynamic hysteresis under rapid thermal changes. To address this, an LSTM calibration model was developed and trained exclusively on quasi-static data across the 20-70 °C temperature range, where it achieved a low prediction error, a 33.563% improvement over a conventional polynomial regression. More importantly, when tested on unseen dynamic data, this statically trained model demonstrated superior generalization, reducing the RMSE from 12.451 °C for the polynomial model to 4.899 °C. These results suggest that data-driven approaches like LSTM can be a highly effective solution for ensuring the reliability of flexible sensors in real-world SHM applications.