Résumé
The present work develops an effective method to improve in situ welding monitoring and control. This work develops an in-process control chain to guarantee the quality of the welding operation by detecting weld pool deviations. First, sensors are chosen to measure different quantities (process parameters and dimensions) adapted to a harsh environment. The weld pool is observed using cameras and image processing algorithms to automatically detect the weld pool contour and obtained its shape. Due to the complexity of the physics, machine learning algorithm were used to detect deviations in weld pool behaviour. At the heart of machine learning is the database to train and run the algorithms. Given the variability of welding conditions from one weld to another during the assembly of thick components, the creation of a comprehensive database for use in multi-pass welding operations in industrial settings poses a significant challenge. In fact, it is not sure that the model obtained with a database can give good result in another configuration. To circumvent this fact, two artificial intelligence models are used. The first step is to use an unsupervised learning model to identify anomalies in the behaviour of the weld pool based on geometric characteristics for the first beads. The second step is to build a classification model that can predict subsequent runs based on the output of the previous model. The model has an accuracy score 0.94%. The database will be enriched during production in an efficient and robust way. The article demonstrated that this kind of method can be applied in real industrial configuration.