Model Selection and Optimization for Fault Diagnosis in Two-Stage Switching Power Supplies of HTOL Testing Systems

In the context of High Temperature Operating Life (HTOL) aging tests for integrated circuits, the stability and precision of the power supply are of paramount importance. Even minute voltage ripple variations can lead to experimental failures or chip destruction. Existing diagnostic methodologies are often insufficient to meet the demand for rapid and precise fault localization in aging test bench power supplies, especially when dealing with high sampling rates and large volumes of data. To address these challenges, this paper presents a study on fault diagnosis for switching power supplies in aging test benches. Initially, a Simulink simulation model of the switching power supply was constructed, and faults were intentionally injected into the MOSFETs and capacitors. Subsequently, output voltage signals from four distinct operational states were sampled at a frequency of 1.5 MHz, forming the raw dataset for analysis. Both conventional fault diagnosis methods, specifically Support Vector Machines (SVM) and Random Forests, and the deep learning model, Convolutional Neural Network (CNN), were utilized for independent model training and testing. The diagnostic efficiency and accuracy of each model were evaluated to determine the optimal performer. Finally, an ensemble learning model, which determines model weights based on diagnostic accuracy, is proposed to further optimize the achieved results. This research establishes a technical architecture that will facilitate the future development of multi-channel secondary switching power supply fault diagnosis systems for aging test benches.