计算机科学
一般化
鉴定(生物学)
干扰(通信)
特征提取
适应性
特征(语言学)
功能(生物学)
结构健康监测
人工智能
模式识别(心理学)
系统标识
传递函数
频道(广播)
数据挖掘
机器学习
稳健性(进化)
可分离空间
特征向量
工程类
电流(流体)
结构工程
信号(编程语言)
标识
DOI:10.1088/2631-8695/ae5ecd
摘要
Abstract Automated structural health monitoring is critical for ensuring the safety of civil infrastructure. However, current deep-learning methods frequently suffer from high false positive and false negatives rates, primarily due to multi-scale crack variations and complex background interference. To address these limitations, this study proposes YOLO-MFD, a novel crack detection model built upon the YOLOv11n architecture. The model integrates a C3k2_MSCB module with parallel separable convolutions for enhanced multi-scale feature extraction and fusion. Additionally, the model employs a Focused Linear Attention mechanism to amplify feature signals while suppressing environmental noise. Furthermore, a cascaded attention DynamicHead improves adaptability to complex targets, while the Inner-CIoU loss function refines localization accuracy. Ablation studies on a custom, real-world dataset confirm the efficacy of these contributions. Comparative evaluations demonstrate that YOLO-MFD outperforms mainstream architectures, achieving 98.5% precision, 89.3% recall, 93.7% F1-score, 94.8% mAP 50 , and 78.8% mAP 50-95 , all while maintaining a high computational efficiency of 110.8 FPS. Experiments on three independent datasets further validate the model’s robust generalization and transfer learning capabilities, confirming that the lightweight YOLO-MFD delivers real-time intelligent crack detection in complex environments.
科研通智能强力驱动
Strongly Powered by AbleSci AI