A COMPARATIVE STUDY OF AI MODELS FOR SAW WELD QUALITY ASSESSMENT WITH AN IOT-BASED HYBRID MONITORING SYSTEM

The assessment of weld quality is a fundamental aspect of contemporary manufacturing operations, specifically in industries where the safety and performance of their product is intricately related to its structural integrity. Submerged Arc Welding (SAW) is a welding process that is commonly implemented in shipbuilding, pipe manufacturing, and the production of large-scale structural components due to its high deposition rates, deep penetration, and overall efficient nature. Unfortunately, some defects, such as porosity, slag inclusions, undercut, and crack patterns can arise during the SAW process, posing a risk to the strength and reliability of the welded joints. Common quality control methods such as visual inspection, ultrasonic testing, and radiography often do not achieve their intended outcome of detecting defects in a timely manner, largely because the aforementioned defects are frequently invisible to the naked eye. Moreover, frequently these methods do not offer a statistically significant quality control tool for complex manufactured products ¹, ³. With recent advances in Artificial Intelligence (AI), the weld inspection process is now stronger than it has ever been before. AI offers exciting ways to automate defect detection and improve the performance of weld quality assessments and controls. Whether through statistical evaluations or the utilization of deep learning algorithms, like Convolutional Neural Networks (CNN), considerable confidence has emerged from the ability of AI to uncover patterns with weld data, and a strong ability to capture visual anomalies ², ⁴. The analytical models developed with the process parameters of the welding operation, such as arc voltage, welding current, and torch speed, predict defect incidences by recognizing latent relationships, or patterns not achievable through traditional evaluation approaches ⁶. CNNs are acclaimed for their exceptional image processing abilities, which can extract complex features from weld images allowing for precise, non-destructive identification of minor defects ¹¹. Nevertheless, despite the rising use of these techniques, there remains a comparative gap in knowledge with conventional statistical models and deep learning approaches when it comes to SAW quality assessments one trades. Most studies tend to only explore these techniques independently leaving a void in analyzing comprehensive performance in real world industrial conditions ⁵, ⁸. The complication of SAW processes with various thermal and metallurgical interactions only further complicates the ability to detect welding imperfections through traditional post-process inspection approaches. There is a need for a real-time, intelligent detection systems which can identify and classify weld defects during the execution stage of the process. This study aims to fill this gap by studying and combining statistical modeling approaches and CNN-based architectures for real-time quality monitoring in SAW applications. By examining defect-labeled SAW datasets and imagery, the study aims to evaluate the deliverables of these AI techniques under actual industrial processes. The driver behind this research is an increase in the popularity of intelligent quality assurance programs as a means to reduce manual inspection time, lower costs of products and prevent defective products. The AI-IoT framework fits under the umbrella of Industry 4.0 and evolves smart manufacturing by utilizing real-time data gathering, big data analysis, and autonomous decision making ¹⁰. The comparative evaluation provides valuable direction for manufacturers that want to either implement or optimize AI-enabled quality assurance systems specific to their operations ⁷, ⁹. This research has the following objectives: - To develop a statistical model for assessing weld quality in SAW. - To deploy a CNN framework for real-time defect detection utilizing weld images. - To assess and perform a comparative study on the statistical and CNN framework in terms of accuracy, reliability and real-time capabilities. Achieving these objectives will provide a basis for a robust hybrid quality assessment system in order to not only identify defects with a high level of accuracy, but also predict defects before they jeopardize product quality. We expect the results to provide both practical solutions and theoretical understanding to intelligent welding systems for high-reliability manufacturing environments.