摘要Monitoring concrete cracks for structural health in civil engineering presents a significant challenge.This is primarily due to the reliance on manual investigation methods,impacts of global climatic shifts stress,and geohazard threats to engineer-ing structures.To cope with this challenge,state-of-the-art Deep Learning(DL)models are utilized to predict concrete cracks and accurately identify subtle variations in crack patterns and sizes,which lighting conditions and surface textures can influence.Previous studies indicate that model accuracy may decrease when faced with obscured concrete cracks,irregular shapes,or limited datasets for real-world problem scenarios.Feature fusion enhances model performance by combining complementary information,resulting in more accurate predictions,but may increase complexity and potential information redundancy.The study presents the Fractur Encoder to Decoder(FractED)block,a novel architecture consist-ing of three sub-blocks:the inner block(Encoder),intermediate block(Intermediate block),and outer block(Decoder).This approach integrates fused features into the model without additional fine-tuning steps,allowing for comprehensive feature refinement and enhancement,ultimately optimizing model performance.The study investigates a DL methodology on three datasets,demonstrating its effectiveness in handling complex classification scenarios in civil engineering.The model achieved high accuracy rates,with 88.41％for multiclass(Deck,Pavement,and Walls)classification tasks,91.94％on the Pillow Dam Borehole image binary dataset,and 99.77％on the Surface Crack binary dataset.The FractED block integration ensures adaptability and scalability,making it valuable for various Artificial Intelligence(AI)applications in civil engineering.The research also provides a scientific foundation for automatizing civil engineering inspection instru-ments for the future.