Hybrid Driven Based Online Detection Of High Power Laser Welding Process

With the fast development of high-power solid-state laser machine in recent years, laser welding gradually replaces traditional welding technology and becomes a primary technology that guarantees high weld quality and efficiency. Especially, in thick-plate material welding, laser welding technology has conspicuous advantages such as deep penetration, small thermal deformation and large depth-to-width ratio. Laser welding process generally involves dramatic conversion of thermal energy. This poses high demand on technical parameters and workpiece fixation precision. Even a minor change is likely to cause serious weld defects. With the widespread use of laser welding technology in various fields, it has become an essential concern to conduct quality monitoring over welding process. This thesis firstly compares both the advantages and weakness of current detection methods. Taking actual industrial application into consideration, it then investigates the effective integration of photoelectric sensing detection technology, visual imaging detection technology and spectrum analysis technology. Based on the multiple non-linear-status coupling variation features during welding process, a multiple-sensor detection experimental platform was set up. Characterization of welding status was analyzed, and the interaction patterns among different feature parameters with the occurrence of typical weld defects have been specified. Attempts were made to explore a data-driven detection approach based on multi-sensor information fusion.On the basis of multi-sensor information fusion detection technology, the following detection platforms were devised and setup:laser welding stainless steel single-signal detection platform (Chapter Two); laser welding stainless-steel two-signal synchronous detection platform (Chapter Two); scanning laser welding stainless-steel four-signal synchronous detection platform (Chapter Three); laser welding high-strength steel three-signal synchronous detection platform (Chapter Four) and laser welding stainless-steel six-signal synchronous detection platform (Chapter Five and Chapter Six). Sensors used in the current research include visible light sensing photodiode, laser reflection sensing photodiode, spectrometer, near-infrared imaging, visible imaging, auxiliary light imaging and X-ray imaging high-speed camera.Firstly, a preliminary research was made on the detection efficiency of mono-sensor in laser welding. The respective applications of photodiode, near-infrared visual sensor and visible range visual sensor to optical radiation feature extraction, welding seam position deviation and molten pool surface defect detection were studied. Experimental results show that the medium-frequency component and high-frequency component of the optical radiation signals were rather sensitive to the changes in welding parameters such as laser power, weld speed and defocusing position. By performing difference operation on the near-infrared images, reliable position parameters can be obtained and were used as the basis for adaptive Kalman filtering prediction on weld seam deviation. The Elman neural network was used to compensate the overall error of Kalman filter. Besides, using delayed recognition algorithm helped to accurately extract the features of plume and spatters on molten pool surface, which provided effective referential basis for defect recognition during laser welding process.Secondly, signal-based data driven detection and knowledge-based data driven detection were performed to realize the effective integration of multi-sensor information. On the signal level, experimental results suggested that the time-domain signals captured by photodiode can be used for detecting the penetration condition. Besides, using the low-frequency component correlation of the signals captured by photodiode and visual sensor was effective in detecting keyhole status when weld penetration occurred. It has also been proved that the low-frequency component positive correlation of the integrated signals is corresponding to the welding instability. On the knowledge level, by adding more sensors and extracting various features of optical intensity radiation, keyhole, molten pool and metallic vapor, classification precision of the SVM training model was greatly improved. This helped to realize accurate recognition of complicated defects (such as partial underfill, entire underfill) on molten pool surface.Finally, a six-signal synchronous detection experimental system was established to realize accurate measurement of the multi-model parameters during laser welding process. How focal energy density inside the keyhole affects the status of typical weld defects was analyzed, and non-linear features during the transition of different statuses were also specified. The correlation analytical results suggest that optical sensor measurement features are closely related to physical geometric features. Based on the above-mentioned analysis, a status detection technology based on model-driven, signal-driven and experience-driven welding was proposed. By replacing the high-cost experimental sensors with low-cost industrial sensors, accurate prediction on laser welding status parameters and effective recognition of welded defects were realized. A multi-input and multi-output non-linear model was devised and the keyhole geometrical parameters measured by the high-speech imaging system were used as identification reference in model based driven. The high-frequency signals collected by photodiode and spectrometer were used as input variable in signal based driven. Multi-scale features extraction of optical signal was performed by using wavelet packet decomposition technology. The extracted features, together with the feature vector output of the identified non-linear model, served as the input reference of feature selection. A low dimensional feature space was eventually devised and used for feature input during the knowledge based driven. Spatters and penetration status measured by the high-speech imaging system were used as the basis for the estimation output of knowledge based driven. The surface quality of welded seam was also considered as the basis for the classification output of knowledge based driven.Experimental results show that after the completion of model recognition and training, the hybrid-driven detection technology proposed by the current thesis can guarantee high prediction and classification accuracy with a few principal components by using industrial sensors (such as photodiode and spectrometer). The research findings provide new theoretical approach and reliable verification data for the comprehensive detection of laser welding process.