Online Monitoring Of Laser Welding Penetration Status And Its Pattern Classification

Penetration status is a critical standard in judging welding quality of deep penetration laser welding process. For most welding joints, full penetration is a fundamental requirement; that is, partial penetration and excessive penetration should be avoided. Penetration status is affected by numerous factors, and proper penetration can not be obtained merely by setting welding parameters. Thus, it is real-time monitoring and control of welding process that serve to be vital in order to obtain satisfactory processing quality.In this study, a multi-sensor signal real-time monitoring system has been developed to pick up the audible sound signal (20~20k HZ) and the plasma-induced ultraviolet signal (400~440nm), both of which are closely related to the penetration status, as well as the infrared radiation signal (1200~1600nm) which relates to the state of welding pool. The signal sampling and processing programs have been developed based on LabVIEW.According to the features of the welds, there are four typical classes of penetration status: partial penetration, unstable penetration, full penetration and excessive penetration. To study different welding penetration status, the parameter-varified plane plate experiment, the parameter-set cuneiform plate and the scalariform plate experiments have been carried out. From the plane plate experiment, it can be found that laser energy density plays a more important role than linear energy input in affecting metal vapor pressure and plasma intenstiy. From the cuneiform plate and the scalariform plate experiments, energy distribution of audible sound and ultraviolet signals on frequency band of 2000~3500Hz properly reflects the typical penetration status, and can be used as discriminating referenceBased on feature fusion technology, 6 features of 3-channel signals have been fused utilizing the concept of information fusion. The simulated algorithm, which is composed to achieve feature fusion, has been achieved on basis of MATLAB."Feature-fusion coefficients"of 6 features have been figured out, thereby obtaining the optimized signal feature inputs. A pattern classifier has been designed based on BP network. Set the signal samples without feature fusion and fused signal samples as inputs of neural network respectively. Consequently, a better result has been achieved with the fused samples; that is, a classification of 88%~100% has been made for detection of the four distinct penetration states with feature fusion. Eliminating certain"abnormal"samples of unstable penetration process, 100% classification can be obtained.