Heat-Flux-Based Condition Monitoring of Multichip Power Modules Using a Two-Stage Neural Network
Borong Hu, Zedong Hu, Li Ran, Chong Ng, Chunjiang Jia, P. McKeever, Peter Tavner, Chi Zhang, Huaping Jiang, Philip Mawby
Abstract
Power semiconductor chips are paralleled in modules to increase current rating. Under thermo-mechanical stresses in service, the die-attach solder layers will gradually develop into different levels of degradation. Early fault detection requires to monitor the occurrence of such an uneven degradation pattern. Internal temperature differences only slightly affect the current sharing between chips, and some temperature sensitive electric parameters are considerably weakened at module terminals. This article presents an external heat-flux-based condition monitoring method, implemented in a two-stage neural network. The first stage consists of a set of subnetworks to represent the mapping between the electrical operating point of the module and its external temperature distribution, for a range of solder degradation patterns and severities. The respective levels of matching are then extracted and applied to the second stage to diagnose the health condition of the power module. This condition monitoring method, validated by experiment, can sensitively detect individual solder degradation in multichip power modules or multimodule power electronic systems.