A study on fatigue of welded structures: Predictive modeling based on automatic learning, numerical analysis, and experimental results

Welded joints in metallic structures are prone to fatigue failure, and thus are frequently the controller of service life. A study on fatigue damage of welded structures has been carried out, with the emphasis put on welded ship structures. This study utilized a new technical approach based on application of automatic learning technology. The purpose of the study was to build predictive modeling tools to make the best use of available information, and predict fatigue lives of the structures.; Data sets composed of experimental and analytical results were constructed using four fatigue experiments on welded cruciform and beam structures made of A36 steel and AL-6XN stainless steel. Potentially important parameters, including geometry, material properties, loading history, and local stress fields around likely cracking sites, were either obtained from experimental records, or generated using numerical analysis. Two-dimensional and three-dimensional finite element analysis and post-treatment procedures were developed to produce the data sets for the likely cracking zones. Optimized load combination was used to simulate the loading histories, and sub-modeling with refined local mesh was used for large structures to improve the accuracy. After the finite element analysis, a two-dimensional or three-dimensional window was positioned at each possible cracking area to obtain the local stress fields.; A hybrid automatic learning system was built to analyze the database, and extract the mathematical relations between the parameters and the recorded fatigue lives. The database was partitioned into the learning set and the testing set. The automatic learning tools, such as learning expert system, neural networks, principal components analysis, and statistical analysis software were trained on the learning set, and subsequently verified on the testing set. The obtained rules, weights, and regression functions produced accurate predictions of fatigue-prone areas, and projected fatigue lives as lower and upper bounds for fatigue failure when there was a substantial crack initiation period.; This study offers an effective way of estimating structural fatigue life, and opens a new way for design of fatigue resistant structures. This approach is open to further improvement as more data become available, and research efforts are being carried out to realize this potential.