Temperature Calibration Using Machine Learning Algorithms for Flexible Temperature Sensors.

Area of Science:

• Materials Science
• Sensor Technology
• Structural Health Monitoring

Background:

• Thermal fluctuations pose risks to large structures like bridges and buildings.
• Flexible temperature sensors are crucial for structural health monitoring (SHM) systems.
• Printed sensors exhibit complex dynamic hysteresis, challenging their reliability.

Purpose of the Study:

• To develop a reliable calibration method for EHD inkjet-printed flexible temperature sensors.
• To address the challenge of dynamic hysteresis in sensor readings.
• To enhance the accuracy of SHM systems under rapid thermal changes.

Main Methods:

• Fabrication of a flexible temperature sensor using EHD inkjet printing.
• Development and training of a Long Short-Term Memory (LSTM) calibration model.
• Model training exclusively on quasi-static temperature data (20-70 °C).
• Validation using unseen dynamic temperature data.

Main Results:

• The LSTM model achieved a 33.563% improvement in prediction error compared to polynomial regression on quasi-static data.
• When tested on dynamic data, the LSTM model's RMSE decreased from 12.451 °C to 4.899 °C.
• The statically trained LSTM model demonstrated superior generalization capabilities for dynamic conditions.

Conclusions:

• Data-driven models, specifically LSTM, offer a robust solution for calibrating flexible printed sensors.
• The developed LSTM approach significantly enhances sensor reliability in real-world SHM applications.
• This method improves the accuracy and trustworthiness of thermal monitoring in large structures.