Faults are common hazards in distribution networks and should be diagnosed immediately to prevent harm and restore power supply. This article proposes a hybrid approach for fault diagnosis that combines deep learning models and advanced signal-processing approaches. It creates a standard four-node test distribution grid in MATLAB/Simulink, extracts features using discrete wavelet transform and short-time Fourier transform, and trains deep learning models to identify, categorize, and pinpoint faults. Results show that while both STFT and DWT perform well in fault detection, STFT demonstrates superior robustness in fault localization and categorization under noisy conditions, owing to its uniform time–frequency resolution. The proposed framework uniquely integrates load demand uncertainty, measurement noise, and a hybrid MATLAB–Python pipeline, providing more realistic evaluation than previous works. The proposed models are not affected by variations in pre-fault loading conditions, fault inception angle, or fault resistance.