Diagnosing Power Module Degradation with High-Resolution, Data-Driven Methods

This work proposes a data-driven approach to monitor lifetime-varying thermal impedance frequency response data from power modules and to diagnose different degradation modes with high resolution. To demonstrate the monitoring and diagnosis approach, the paper develops a piece-wise linear thermal model that describes the variable effects of convection, thermal interface material, and die-attach degradation. To identify critical thermal impedance frequencies, the model is investigated using a comprehensive degradation sensitivity trend analysis. An introduced sensitivity metric reveals degradation sensitivity extrema at high resolution and, ultimately, allows for optimal design of degradation monitoring systems. Next, it is shown how in-situ thermal impedance spectroscopy can compute the thermal impedance data at identified special excitation frequencies. Finally, to perform the estimation of lifetime-varying parameters on the basis of thermal impedance data, artificial neural networks are designed and trained, and their overall capability to diagnose degradation is quantified. The investigation shows that a neural network can resolve different sources of degradation with 1% errors using thermal impedance phase information. Overall, the developed technology synthesizes robust physical degradation data to be used for realizing predictive maintenance schemes in highly-reliable power electronics.

Automated summary

This work presents a data-driven approach to monitor lifetime-varying thermal impedance frequency response data from power modules and diagnose different degradation modes with high resolution. By developing a piece-wise linear thermal model and conducting degradation sensitivity trend analysis, critical thermal impedance frequencies are identified and thermal impedance data is computed at special excitation frequencies. Through the design and training of artificial neural networks, different sources of degradation can be resolved with 1% errors.