The paper deals with description, design, learning and inference process of a convolutional 2D neural network for detection of shortened winding turns of a three-phase permanent magnet synchronous motor. Input datasets for aforementioned procedures have been created by sensing vibration data on the real motor using accelerometers with a possibility of artificially induce short circuit in the motor winding. Only simple pre-processing of a time signal has been done – the time waveform was reshaped into 2D greyscale images with a size of 64 x 64 points and led directly into the neural network. No pretrained network has been used – internal parameters have been learned from scratch. Learning process as well as inference of the network have been performed on a standard personal computer with nVidia GeForce RTX 2080 Ti graphics card and implemented usign Python in Keras/TensorFlow environment. Datasets for different working states of the motor, such as speed, torque, error type and its severity have been used. Training procedure of the network has been done within lower tens of minutes and final validation accuracy was 100 % in the most cases, while classification accuracy during inference process has reached the value of more than 99 %. Obtained results confirmed a fact, that faults’ detection of the mechatronic system based on sensing of mechanical quantities and their evaluation is very reliable even in the case of electrical-based faults.