Abstract Scope |
Weld defects detection usually requires post-weld inspection. Therefore, defects formation could not be known until the process was finished. Some defects may be avoided if initiation of defect formation could be detected and consequent parameter adjustment could be made. As a result, real-time defect detection and process control become critical. In this research, effect of welding parameter on weld anomaly formation is studied. More than 100 features in time domain, frequency domain and time-frequency domain are extracted from the voltage, current and audio data in welding process. The importance of features is sorted and analyzed by correlation analysis method, and some important features are selected to establish machine learning model. Both supervised learning and unsupervised learning machine learning models are established for abnormal detection in welding process, including support vector machine, random forest, gradient boosting decision tree, isolation forest, local outlier factors, etc. The results show that the welding anomaly detection model established in this paper can accurately detect welding defects and provide feedback for welding process control. |