Intelligent Machining Monitoring Using Sound Signal Processed With the Wavelet Method and a Self-Organizing Neural Network

Abstract

A methodology was developed using the sound signal based on wavelet analysis and self-organizing neural network (NN) to monitor cutting accuracy in an extremely noisy machining process. Sawing experiments were conducted under different levels of feed speed, depth of cut, and rotation speed to monitor the longitudinal waviness of sawn samples using laser displacement sensors as an index of cutting accuracy and sawing deviation. The sound of the cutting and idling processes was recorded using a microphone. The acquired acoustic signals were pre-processed using the wavelet de-noising method for background noise elimination. As the signal still encompasses the low-frequency components corresponding to the idling process (machine motor, saw vibration, etc.), a systematic wavelet thresholding method was applied to the coefficients of the decomposed signal at different levels to discard the sound signal components associated with the idling process. Inverse wavelet transform was then applied to make a synthesized signal from the original one. Different features were extracted from the original and synthesized signals and used to train a self-organizing NN. Group method of data handling (GMDH) NN was utilized for predicting the waviness from the sensory features. The GMDH model trained with features extracted from the synthesized signal outperformed the one trained with the original signal features. The results suggested that employing the proposed wavelet-based methodology enables the acoustic signal to be used in monitoring the manufacturing processes in an extremely noisy environment. Self-organizing NN was shown to have a promising performance without facing the difficulties in finding the network optimal architecture, which is a typical challenge in the conventional backpropagation NNs.