Neural network-based software tool for predicting magnetic performance of strip-wound magnetic cores at medium to high frequency

Although magnetic wound cores have simple geometries, their magnetic properties vary in a complex manner depending on many factors such as core geometry, dimensions and interlaminar air flux between adjacent wound layers. These can set up internal stress which seriously affects internal flux distribution and hence permeability and losses. For core manufacturers to satisfy customer requirements, a need arises for a reliable tool for more accurate prediction of magnetic performance of strip wound cores. The magnetic performance of a range of strip wound cores has been measured over a wide frequency range (50 Hz-100 kHz). Using this information a neural network-based software package coined ´MagnetWolf´, has been developed for predicting core performance. Input parameters include core geometry, core dimension, core material and strip width. A subneural network approach employed reduces the amount of representative data required for network training. This provides rapid network development and enhances accuracy. A JAVA graphical user interface makes it possible for the tool to be accessible via the internet. A comparison of predicted and measured results shows this to be a reliable tool with potential industrial applications.