By Mike Mohseni
Due to the broad application of welding and the mission-critical effect of weld quality, there are several diverse technologies in the market to monitor and enhance the quality of certain processes.
Welding cameras monitor processes in real-time. Its most prominent benefits include recording the process track and increasing work safety due to remote monitoring. An operator with domain knowledge then interprets the welding process quality based on the raw data collected. With the increased availability of hardware components, the price point for welding cameras is continuously decreasing, making them more accessible.
Laser systems in automated detection solutions provide geometrical feature assessments such as the welding location to ensure fitting before the process or the weld line quality post-process. Despite the enhanced automation and in-line application, welding inspection via laser-based systems remains a serial operation relative to the manufacturing process. This adds to the lead time while risking finding defects after the part is manufactured. For new defects or part geometries, the laser system should be re-calibrated, which can be challenging in welding operations as changeover rates are high in spaces like the automotive supplier industry.
X-ray and ultrasound are reliable methods of weld inspection as they reveal the internal structure of the weld. These methods are commonly used for in-service inspections. However, the equipment is expensive, sensitive, and requires expertise for operation and interpretation. The speed of measurements and analysis does not match requirements in an industrial operation setup or in-line applications.
Machine learning methods are experiencing growth in interest within the industry thanks to the awareness raised due to industrial automation programs and the accessibility of software development tools. Automated defect detection is an application of machine learning models. Initially, an operator with domain knowledge marks the process data such as images of the weld line as normal or defective. A computer program receives the marked data as input to an iterative process of model training. The trained model is then able to distinguish between normal and defective welds given new data.
The procedure above is an intuitive application of machine learning. Recently, industrial patents were issued to similar approaches in welding inspection. However, scalability remains to be an issue for these methods. Characteristics of post-mortem defects vary with different welding scenarios, part geometry, and weld setup. The software trained to detect specific defects in a given welding setup does not apply to other defects or if the welding setup changes. Therefore, machine-learning-powered solutions have not been used during the complex process of welding despite successful applications in other fields such as smartphone packaging inspection.
Table 1: Exhibiting the different welding inspection methods on the measures of type of defects defected, expertise required to operate, application in high changeover rate, and in-line inspection.