Weldpool recognition using deep learning in the laser cladding process

Abstract

Artificial intelligence models have gained popularity in recent years due to their ability to extract information from different data sources including images, being able to perform activities that 50 years ago were unthinkable in the field of computing such as the recognition of different objects within the same image. Therefore, there is an opportunity in the use of these tools in different branches of engineering such as manufacturing and computerized control systems. A manufacturing process that has gained importance in the last years is the laser cladding process. In this process a laser is used to generate a molten pool called WeldPool, in which particles are projected adding material and generating a coating. Factors like the temperature and shape of the Weldpool are key parameters in the control of the process and these are the reasons than several investigations has been carried out on this subject. In the following work, deep learning models based on convolutional neuronal networks (CNN) are trained to characterize the WeldPool formed in the laser cladding process through object recognition, being these models the YOLO model, the SSD model and a simpler proposed model called OWN. To train these models, frames of laser cladding deposition videos are employed. The videos are preprocessed, labeled and artificially augmented to augment the database. The labels used focus on capturing the width and length of the WeldPool through a rectangle, as they correspond to the two most important features for visual recognition. There have been other attempts to perform WelPool characterization through image segmentation, but they have not been focused on generating models that are capable of making predictions with speeds approaching real time, nor occupying camera images in the visual spectrum. The study deals with the creation of machine learning models from three different data sources corresponding a videos recorded with different image setting like brightness and contrast. Two approaches are used: one involves training models on one data source and testing them on others, while the second approach involves training models with data from all sources and evaluating their generalization ability. The processing time of the models is also compared in terms of frames per second (FPS). The key takeaway from this research is that individual data sources cannot independently create predictive models for other sources. Instead, combining different data sources produces accurate predictions for WeldPool length and width. YOLO stands out as the model with the best speed-performance balance, making it a promising addition to a laser cladding process monitoring and control system.