In this paper, the authors present a new transfer-learning-based robust technique for detection and classification of multi-type faults in low-voltage power distribution grids. Three-phase current and voltage signals were initially measured and sampled at a medium-voltage/low-voltage substation upon occurrence of a certain type of fault. Subsequently, a Hilbert–Huang transform was applied to the corresponding sampled fault signals to construct a time–frequency energy matrix as a \(35 \times 35\) pixel matrix of the digital image after passing through a band-pass filter. Additionally, and for the sake of comprehensiveness, a residual network (ResNet) was designed as a fault detector and classifier to accurately identify the location and type of faults, while guaranteeing the robustness of the proposed technique against both white and connection noises. For the sake of evaluation, a three-phase 10 kV test system was implemented in the MATLAB/Simulink environment under different faulty operating conditions designs. The performance of the introduced technique was also examined in the Python environment. Based on simulation results, several conclusions can be drawn: (i) The ResNet outshines the neural architecture search network, Inception, and Xception by elevating the average accuracy by 2.23%; (ii) the ResNet-101 model is robust against white Gaussian noise and electromagnetic interference with values of 20, 30, and 40 dB; (iii) the proposed technique enhances functional coverage up to a radius of 30 km for fault detection along 14 points in the test system; and (iv) in the presence of distributed energy resources, the proposed technique is superior to other network architectures with an average of 5.44%.
