Tapping process fault identification by LSTM neural network based on torque signal singularity feature

Aiming at the problem in which the tap easily breaks during the tapping process of small-diameter internal thread, this study proposed a fault monitoring method for the tapping process based on the singularity characteristics of the torque signal in accordance with the phenomenon where the torque signal continuously and violently singularly fluctuates when the tap breaks. The wavelet coefficient maps were obtained by the continuous wavelet transform of the torque signal; the modulus maxima of the wavelet transform coefficients were extracted to reflect the fluctuation of the torque signal singularity; the maximum attenuation of the modulus maxima in the time scale plane was calculated as the Holder exponent (HE); and the distribution characteristics of the mean, standard deviation, and number of singular points of the HE in different fault states are investigated. Results showed that these three features have strong correlation with tapping faults. Thus, a categorical feature vector set was constructed. A fault identification model based on the long short-term memory tapping process was established, and the model training and fault identification test were completed by using the constructed feature vector set. The testing accuracy of the constructed fault identification model reached 96.11 %, which is remarkably better than other network models. Results verified the validity of the analysis of the singularity method, providing a possible solution for fault recognition in the tapping process, which lays the foundation for realizing the intelligence and automation of the tapping process.