Product quality prediction in pulsed laser cutting of silicon steel sheet using vibration signals and deep neural network

This study investigated the feasibility of utilizing vibration signals, measured from the pulsed laser cutting machine, to predict the product quality in terms of average kerf width of a straight slot cutting by using the Deep Neural Network (DNN) predictive models. Non-oriented silicon steel sheet with a thickness of 0.1 mm was chosen as the workpiece. Two ways of preprocessing input features in the DNN models were considered. There are the statistical features of time-domain raw vibration signals in 3-axis directions and the extracted features from five levels of wavelet decomposition signals. The output of the DNN was chosen as the average kerf width of the workpiece. Then, the relation between input features and kerf width was examined using the Pearson correlation coefficients and Select K-best method. This process provides the most significant features for two preprocessing input data. The performance comparisons of the developed DNN model associated with the selected features were discussed. Furthermore, the combination of selected features with laser machining parameters as the input features was also addressed. In general, integrating statistical features extracted from raw vibration and wavelet decomposition signals as inputs in the proposed DNN model is effective for predicting the kerf width of a straight slot laser cut.