LAUSR

Abstract Grinding plays a vital role in modern gear manufacturing industry while the need for high quality products is continuously increasing. A methodology for gear grinding monitoring, exploiting the power of Deep Learning architectures and 2D representations, is presented in this paper. Vibration signals, measured during the grinding process under healthy and faulty conditions, are classified with high accuracy. Three types of faults i.e., a high profile form error, a high lead error, and a high profile slope variation, have been emulated. The Short-Time Fourier Transform (STFT) of each vibration signal is calculated, and the 2D time-frequency representations are input to a Deep Convolutional Neural Network (DCNN) for classification. Different filter sizes are tested, and the classification accuracy of 95.0% has been achieved, demonstrating the efficiency of the methodology for gear grinding monitoring.