Abstarct: Fast and selective protection systems play a crucial role in preventing fault propagation, reducing outages, and protecting power grid equipment, and fault detection and classification are of particular importance for achieving these systems. Distance protection is considered one of the most important protection systems for protecting transmission lines as primary and backup protection due to its proper performance, but it has challenges when faced with phenomena such as power swings. Considering the logic of distance protection operation, detecting and classifying the type of fault and faulty phases has a significant impact on its correct and fast operation. In this study, a deep learning-baxsed approach is proposed to improve the accuracy and speed of detecting stable and unstable power swings from faults, classifying types of power swings, and classifying types of faults under normal and power swings conditions. In the proposed method for signal preprocessing, the combination of Hilbert transform and Wigner-Weil distribution on a quarter cycle of the current signal is used to produce suitable features for the input of the residual network-baxsed deep learning network (ResNet-50). To evaluate the proposed method, the standard test systems of 9-bus WSCC and 39-bus New England are used in different conditions of noise change, sampling frequency, training data set, fault type, fault occurrence moment, load angle, fault resistance and fault location. The simulation results show that the detection of power swing types from faults, classification of stable and unstable power swings and classification of fault types in the various conditions mentioned are performed with an accuracy of more than 99 percent and the average detection and classification time of the protection system is one and five milliseconds, respectively.