Classifying and localizing anomalies in the performance of nuclear reactors is an issue to assure the safety of nuclear reactors and, consequently, guarantee their long-term operations. It is of particular interest to detect perturbations while are occurring without disturbing the operation of the nuclear reactor. This detection can be done with the analysis of neutron flux oscillations around a steady state value, called neutron noise. In this work. we propose a deep learning framework for the deconvolution of reactor transfer functions from perturbation-induced neutron noise sources. The main objective of this work is to develop tools based on deep learning techniques using convolutional neural networks, working with simulated data with different Gaussian noise levels, and to study the number and location of detectors that need to be active. In particular, the data used has been simulated using the neutron diffusion approximation and the first-order noise approximation for the BIBLIS 2D reactor and IAEA 3D reactor. High-accuracy results are obtained both when predicting the type of the perturbation and when locating the place of the perturbation, with a low error rate even when only four to eight detectors are available.