A study on prognostics of LED luminaires based on gradient-boosting decision tree classifiers

This paper presents the prognosis of LED luminaires based on supervised machine learning. Several gradient-boosting decision tree (GBDT) classifiers, viz. AdaBoost, GBM, XGBoost, CatBoost and LightGBM are selected for their high accuracy and speed. At first, a temperature-dependent linear model of LED is developed and validated. Using this model, complete LED systems comprising of LED module and constant-current flyback driver, having 50 W to 120 W rated power, four dimming levels within 25% to 100% range and various series–parallel combinations of the LED array are designed and implemented in NI-Multisim software. From the simulation results, datasets required for the training and testing of the GBDT classifiers are generated. The trained classifiers can efficiently predict five operational conditions, viz. healthy, overheating, faulty capacitor, open circuit and short circuit by taking time-domain electrical features as input. A comparative evaluation conducted among the GBDT classifiers indicates that LightGBM is the best performer as it exhibits the highest validation accuracy (97.78%) and test accuracy (95.83%) with smallest margin of error (±3.58%), the lowest generalization error (0.042) and the fastest speed of execution (0.43 s). Therefore, LightGBM can be recommended for the prognosis of LED luminaires for smart lighting applications.