Blackbox system identification with neural network

This research trains a multilayer perceptron (MLP) to identify an N-port electrical network given S-parameter files of the system. Since the rational expansion of the transfer function is known, the problem is simplified to finding poles and residues of a fixed number of terms to approximate the transfer function. We take S-parameters at different frequencies as input, preprocess the data and pass it into a simple MLP to output poles and residues. Once the poles and residues arrive, we feed them back into the rational expansion to calculate the approximated transfer functions’ values given different frequencies. To obtain the loss of the network, we compute the difference between the calculated transfer functions output and the given S-parameters, since S-parameters are values of the real transfer function. Finally, we update the weights of the MLP by minimizing the loss. Our code is written in TensorFlow. It gives users the flexibility to choose the number of ports, poles and residues to describe the electrical system. It also allows users to define the batch size, weight initializers, number of layers, size of layers and activation functions for the neural network.Ope