Multitask fatigue crack recognition network based on task similarity analysis

For fatigue cracks in steel box girder bridges, comprehensive perception is urgent and practical since it provides diverse damage information. Multitask learning networks, which learn multiple tasks jointly, could realize various detection tasks simultaneously. However, multitask networks are hard to avoid task conflicts, thus limiting recognition performance. In this study, the representational similarity analysis method was introduced to quantify task similarity, which helps decouple the parameters of unrelated tasks. Considering task correlation, a multitask network called CrackFormer was developed to recognize the initiation location, repair scene, and region morphology of fatigue cracks. From the perspective of learning difficulty, multiple tasks were trained and coordinated dynamically. Finally, the trained model was evaluated by comparative tests. Results showed that task similarity analysis was verified to be reasonable by transfer learning test. Compared with isolated tasks, CrackFormer demonstrated superior accuracy, efficiency, and robustness by mutual complementation of task knowledge. After optimizing the parameter sharing design, detection performance was further improved.