This paper summarizes a novel approach to improved localization of fuel defects by fusing existing data sources and methods within a neural network model to make accurate and quantifiable identification earlier than existing processes. The approach is demonstrated through application to a CANDU reactor and utilizes a small, manually labeled set of delayed neutron data augmented with neutronic power data to train a neural network to estimate the probability of a fuel channel containing a defect. Results demonstrate that the model is often capable of identifying likely defects earlier than existing methods and could support earlier decision making to enable a reduction in cost and time required to localize defects. The approach described has broader application to other reactor types given the general difficulty of detecting fuel defects via fission product measurement and the large quantities of ancillary parameters normally already recorded that can be leveraged using machine learning techniques.