Machine learning assisted early anomaly detection of LEDs with spectral power distribution modeling

Conference paper (2023)

Authors

Minne Liu Fudan University
Mesfin S. Ibrahim New Territories
Minzhen Wen Fudan University
Sheng Li Shanhai Yaming Lighting Co.Ltd
Aijun Wang Shanhai Yaming Lighting Co.Ltd
Kouchi Zhang Electronic Components, Technology and Materials - EEMCS
J. Fan Electronic Components, Technology and Materials - EEMCS , Fudan Zhangjiang Institute,, Fudan University, Chinese Academy of Sciences
Research Group
Electronic Components, Technology and Materials (EEMCS) (TU Delft)
Copyright: © 2023 Minne Liu, Mesfin S. Ibrahim, Minzhen Wen, Sheng Li, An Wang, Kouchi Zhang, J. Fan
DOI
help_outline
https://doi.org/10.1109/SSLChinaIFWS57942.2023.10071010
Anomaly detection Principal component analysis K-nearest neighbor White LEDs Spectral power distribution
More Info
expand_more

Published Date

2023

Language

English

Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Faculty

Electrical Engineering, Mathematics and Computer Science

Department

Microelectronics

Research Group

Electronic Components, Technology and Materials

Abstract

Spectral power distribution (SPD) is the radiation power intensity at different wavelengths, containing the most basic photometric and colorimetric performance of the illuminant, which is able to predict the lifetime of LEDs. This paper proposes an SPD model assisted by machine learning algorithms to detect the early failure of white LEDs. The SPD features of 3W high-power white LEDs were firstly extracted by the statistical models of Gaussian, Lorentz, and Asym2sig functions. An unsupervised learning method, principal component analysis (PCA), was then used to reduce the extracted features parameters’ dimensions. Next a K-nearest neighbor (KNN)-based method was used to detect LEDs’ anomalies by dividing the main cluster into groups, and estimating the distance from the center of mass of each cluster to the test point. The results showed the following: (1) for selected white LEDs, the Asym2sig function has a better fitting result than Gaussian and Lorentz functions; (2) machine learning methods can significantly assist in LED anomaly detection and can decrease the amount of anomaly detection time to 789.6 h, compared to the 1311 h when lumen maintenance degradation reaches 70% as required by IES TM21.