ANN Based EMI Filter Optimization with Limited Data in Vehicle Power Systems

In an in-vehicle power system, three-phase AC, 100 V AC, and 12 V DC are generated from a high-voltage battery, requiring efficient EMI filter design. Time-domain simulations for IGBT and MOSFET switching are costly, limiting data availability. This study compares orthogonal arrays (OAs) and Latin hypercube sampling (LHS) for optimal learning of an artificial neural network (ANN) with limited data. Three ANN models were trained for different noise terminal voltages, using seven filter parameters as inputs and 300 -point frequency spectra in the range of 0.1 MHz to 100 MHz as outputs. Results show OAs achieving higher accuracy than LHS.