Application of a convolutional neural network (CNN) for the classification of semiconductor diode junction temperature

The paper presents a method for analysing the junction temperature of semiconductor diodes using a convolutional neural network (CNN). The research was conducted using thermographic measurements and simulation studies. The Finite Element Method (FEM) was applied to determine the relationship between the junction temperature (Tj) and the package temperature (Tc) of a 6A6 diode. A detailed three-dimensional thermal model of the diode was developed in SolidWorks and experimentally verified. The differences between measured and simulated results did not exceed 3°C, confirming the model’s reliability. Thermograms were converted into numerical matrices and normalised for training the CNN model in the MATLAB environment. The developed neural network automatically classified thermograms into two categories: diode junctions operating at the correct temperature and diodes with an overheated junction. The obtained validation accuracy ranged from 92.17% to 100%, depending on the composition of the dataset. Unlike previous studies focused mainly on thermogram analysis and hot-spot detection, the proposed method combines FEM-based junction temperature estimation with CNN classification using radiometric temperature matrices. The proposed approach demonstrates that CNN-assisted thermographic analysis can effectively detect thermal anomalies in semiconductor components, providing a fast and non-invasive diagnostic tool for power electronics applications.