Laser Welding Quality Monitoring via Graph Support Vector Machine With Data Adaptive Kernel

Abstract

Laser welding is a rapidly developing technology that is of utmost importance in a number of industrial processes. The physics of the process has been investigated over the past 50 years and is mostly well understood. Nevertheless, online laser-quality monitoring remains an open issue until today due to its dynamic complexity. This paper is a supplement to existing approaches in the field of in situ and real-time laser-quality monitoring that presents a novel combination of state-of-the-art sensors and machine learning for data processing. The investigations were carried out using laser welding of titanium workpieces. The quality was estimated a posteriori by the visual inspection of cross-sections of the welded joints. Four quality categories were defined to cover the two main laser welding regimes: conduction and keyhole. The signals from the laser back reflection and optical and acoustic emissions were recorded during the laser welding process and were decomposed with the M-band wavelets. The relative energies of narrow frequency bands were taken as descriptive features. The correlation of the extracted features with the laser welding quality was carried out using the Laplacian graph support vector machine classifier. Also, an adaptive kernel for the classifier was developed to improve the analysis of the distributions of the complex features and was constructed from Gaussian mixtures. The presented laser welding setup and the developed adaptive kernel algorithm were able to classify the quality for every 2 µm of the welded joint with an accuracy ranged between 85.9% and 99.9%. Finally, the results of the developed adaptive kernel were compared with stateof-the-art machine learning methods.

Supervised deep learning for real-time quality monitoring of laser welding with X-ray radiographic guidance

“…The domain of wavelet spectrograms is also a 2D time-frequency domain, to which existing CNN can be applied directly. At the same time, in practical applications, the spectrograms preserve the special features of the original signals, and evidence of this was successfully demonstrated in earlier investigations 22,30,31,34,35 .…”

“…The built-in germanium (Ge) photodiode originally had a spectral sensitivity in the range of 800-1,800 nm. A narrow band-pass filter (FB1070-10, Thorlabs Inc., USA) with a center wavelength of 1070 ± 2 nm and a Full Width Half Max (FWHM) of 10 ± 2 nm was installed to the photodiode to provide a selective transmission of the LBR radiation from the process zone 22,[29][30][31] . The photodiode signal was acquired with an oscilloscope Teledyne LeCroy HDO6104 at sampling rates within the range of 50-2,000 kHz.…”

“…WPT was used to substitute the collected LBR and AE signals by wavelet spectrograms, formed as relative energies of the narrow frequency bands 32,33 . More details of this technique can be found in our previous works 22,30,34,35 . Such signal representation has four major advantages.…”

“…The choice for the optimal criterion was the minimum wavelet approximation error of the given original LBR and AE signals. Based on the search, Daubechies wavelet with ten vanishing moments was found to fit best the collected data 22 .…”

“…2B, and grouped according to the corresponding quality-significant events. Based on the results of our previous works 22,30 , 4096 frequency bands from each spectrogram were extracted and fed to the CNN. The wavelet transformation was tuned so that the spectrograms had the same size, regardless of the number of data points in the patterns.…”