At the beginning of 2022 there are 442 nuclear power reactors operating worldwide and other 51 power reactors are under construction. The operation of nuclear reactors leads to the production of spent fuel, which accounts for the majority of nuclear material placed under safeguards. Non-destructive assays (NDA) based on the radiation emitted from spent fuel, mainly neutrons and gamma-rays, are being used in-field for the safeguards verification. NDA techniques such as the Fork detector and the passive gamma-ray emission tomography are authorized for verification by the International Atomic Energy Agency, and many other NDA techniques are being developed. Among those, the Partial Defect Tester (PDET) is a NDA technique specifically designed to detect partial defects, i.e. missing or replaced fuel pins. The PDET consists of a set of neutron and gamma-ray detectors placed in several positions within and outside the spent fuel assembly. Over the past years a dataset containing more than 1000 virtual spent fuel assemblies has been developed with Monte Carlo simulations including complete fuel assemblies and assemblies with replaced pins. Given the complexity of spent fuel geometry and of the PDET detector response, machine learning models have been developed for the detection of partial defects. Decision trees and k-nearest neighbors were used as machine learning approaches and showed promising results. Previous work highlighted the capability to roughly classify the fuel assemblies in terms of percentage of fuel pins missing. This contribution builds on the previous work and improves the developed machine learning models. The objectives of the machine learning models presented in this work are to determine the number of replaced fuel pins and attempt to localize the replaced fuel pins within the fuel assembly. Results show that the developed models are able in more than 80% of the cases to exactly quantify the number of replaced pins, but the localization of pin replacement is still a challenging task.