Due to the highly dangerous nature of electrical failures, and more specifically electric arc faults, the detection of such problems has become absolutely necessary. In contrast to the majority of methods proposed in scientific literature that are based on frequency analysis, we propose a method of detection based on CNN models (LeNet5 - $28^{\ast}28$ and $64^{\ast}64$ images). For this method, the line current must first be recorded (the dataset is composed of about 11000 signatures with and without arc faults). Series-arc faults are produced in circuits that comprise a 270-Volt DC supply voltage and loads that are mainly resistive. The selected sections of the current signals are then transformed into a 2D matrix (images). Then, the network must then be trained, tested and validated using the dataset. The performance of this method must also be also studied and discussed. In fact, the results of the detection process must then be presented using a confusion matrix in order to provide more precise information. Experimental results show that the method that we are proposing can effectively detect arcing faults.