Reliability Prediction of TFT-LCD Modules in Harsh Environments Using Physics-Guided Machine Learning

Accurate Remaining Useful Life (RUL) prediction of TFT-LCD modules is critical for industrial predictive maintenance, yet it remains heavily challenged by complex degradation mechanisms in different climates. Traditional purely data-driven models (SVR, LSTM) often lack physical interpretability, struggling to filter out environmental noise or predict irreversible failures. To address this, we propose a highly reliable prognostic tool based on a Physics-Informed Gaussian Process Regression (PI-GPR) framework, by embedding cumulative thermal load and thermo-mechanical stress into the model’s prior function. Evaluated using one-year field exposure data, the physical constraints empower the model to accurately predict device lifetime under highly variable environments, including luminance fluctuations in tropical hygrothermal conditions and device failures in cold environments. Quantitative results demonstrate that the unified PI-GPR framework achieves an outstanding coefficient of determination (R2 = 0.93) and reduces the RUL prediction error to merely 7.5 days, significantly outperforming conventional shallow learning, deep sequence, and standard probabilistic baselines. Ultimately, this study provides a robust, physically grounded methodology for the health monitoring and life cycle management of display modules in practical industrial applications.