Automated Recognition of Weld Pool Characteristics from Active Vision Sensing

While penetration occurs underneath the workpiece, the raw information used to detect it during welding must be measurable to a sensor attached to the torch. Challenges are apparent because it is difficult to find such measurable raw information that fundamentally correlates with the phenomena occurring underneath. Additional challenges arise because the welding process is extremely complex such that analytically correlating any raw information to the underneath phenomena is practically impossible; therefore, handcrafted methods to propose features from raw information are human dependent and labor extensive. In this paper, the profile of the weld pool surface was proposed as the raw information. An innovative method was proposed to acquire it by projecting a single laser stripe on the weld pool surface transversely and intercepting its reflection from the mirror-like weld pool surface. To minimize human intervention, which can affect success, a deep-learning-based method was proposed to automatically recognize features from the single-stripe active vision images by fitting a convolutional neural network (CNN). To train the CNN, spot gas tungsten arc welding experiments were designed and conducted to collect the active vision images in pairs with their actual penetration states measured by a camera that views the backside surface of the workpiece. The CNN architecture was optimized by trying different hyperparameters, including kernel number, kernel size, and node number. The accuracy of the optimized model is about 98% and the cycle time in the personal computer is ~ 0.1 s, which fully meets the required engineering application.