Kernel ridge regression for volume fraction prediction in electrical impedance tomography

We investigate using a kernel learning machine, specifically kernel ridge regression (KRR), to predict volume fractions in typical industrial electrical impedance tomography (EIT) applications. The 'curse of dimensionality' associated with applying such methods to physically captured EIT training data is overcome with a new training method, involving sampling of training data during rapid random repositioning of a set of physical objects in the measurement plane. We compare the performance to multi-layer perceptron (MLP) neural networks which appear to be the most common computational intelligence approach to the EIT reconstruction problem. We use empirically trained static situations so as to compare the results to previous research. Dynamic situations are also investigated, and KRR is shown to outperform MLP methods in both cases. Furthermore, KRR is shown to be a useful tool in EIT for extracting process information from industrial flows without first performing conventional image reconstruction.