An automated welding system is essential to ensure a stable and good welding quality and improve productivity in the gas metal arc welding (GMAW) process. Therefore, various studies have been conducted on the establishment of smart factories and the demand for good weldability in the fields of production and manufacturing. In shipbuilding welding and pipe welding, the uniformly generated back-bead is an important criterion for judging the mechanical properties and weldability of the welded structure, and is also an important factor that enables the realization of an automated welding system. Therefore, in this study, the welding current signal measured in real-time in the GMAW process was pre-processed by a short time Fourier transform (STFT) to obtain a time-frequency domain feature image (spectrogram). Based on this, a back-bead generation detection algorithm was developed. To accelerate the training speed of the proposed convolution neural network (CNN) model, we used non-saturating neurons and a highly efficient GPU implementation of the convolution operation. As a result of applying the proposed detection model to actual welding process, the detection accuracies with and without the back-bead regions were 95.8% and 94.2%, respectively, which confirmed the excellent classification performance for back-bead generation.