Metal structures develop cracks under fatigue loading, which subsequently propagate. The size of the cracks directly affects the fatigue life of the structure. Accurate prediction of crack lengths under various loading conditions is crucial for the safe service of structures. And the crack length has a significant influence on the local strain of the structure. In this paper, finite element analysis (FEA) is used to extract strain data from various measurement points of compressive and tensile (CT) specimens under different loading conditions. The Deep Forest (DF) model is employed to optimize the training of the data. Compensation is applied to the measured dynamic strain data for predicting crack length. Experimental results show that multi-dimensional input signals in the XY plane can accurately predict crack length. Additionally, based on the Pearson correlation coefficient, this paper proposes a self-diagnostic coefficient for strain sensors. Combined with the DF model, it enables self-diagnosis of the strain sensor. The proposed crack length prediction and strain sensor self-diagnosis methods enhance the intelligence level of crack state monitoring to some extent.

Detecting metal surface crack defects is of great significance for the safe operation of industrial equipment. However, most existing mainstream deep-object detection models suffer from complex structures, large parameter sizes, and high training costs, which hinder their deployment and application in frontline construction sites. Therefore, this paper optimizes the existing YOLO series head structure and proposes a lightweight detection head structure, ZoomHead, with lower computational complexity and stronger detection performance. First, the GroupNorm2d module replaces the BatchNorm2d module to stabilize the model’s feature distribution and accelerate the training speed. Second, Detail Enhanced Convolution (DEConv) replaces traditional convolution kernels, and shared convolution is adopted to reduce redundant structures, which enhances the ability to capture details and improves the detection performance for small objects. Next, the Zoom scale factor is introduced to achieve proportional scaling of the convolution kernels in the regression branch, minimizing redundant computational complexity. Finally, using the YOLOv10 and YOLO11 series models as baseline models, ZoomHead was used to replace the head structure of the baseline models entirely, and a series of performance comparison experiments were conducted on the rail surface crack dataset and NEU surface defect database. The results showed that the integration of ZoomHead effectively improved the model’s detection accuracy, reduced the number of parameters and computations, and increased the FPS, achieving a good balance between detection accuracy and speed. In the comparative experiment of the SOTA model, the addition of ZoomHead resulted in the model having the smallest number of parameters and the highest FPS, while maintaining the same mAP value as the SOTA model, indicating that the ZoomHead structure proposed in this paper has better comprehensive detection performance.

Steel surface quality critically determines the service safety and structural reliability of industrial products. Defects such as cracks, inclusions, patches, pitting, rolled-in scale, and scratches severely compromise product safety, making accurate and efficient detection a key step in quality control. However, the native A2C2f module in YOLOv13 exhibits insufficient multi-scale feature extraction for tiny defects and weak robustness under complex industrial backgrounds, hindering the detection of these six defect types. To address these gaps, we propose a multi-scale denoising enhanced module, A2C2f-MSDE, which constructs a multi-scale multi-kernel fusion branch (MSKF) with learnable adaptive weights, integrates a lightweight SEL channel attention and a DE denoising module, and employs a dual learnable residual scaling structure, while preserving the original multi-scale fusion architecture. We embed A2C2f-MSDE into the YOLOv13 backbone, perform ablation studies to verify each component’s contribution, compare it with mainstream advanced detectors on the public NEU-DET dataset, and conduct generalization tests on the GC10-DET dataset. Experiments on NEU-DET show that the improved YOLOv13n achieves mAP50-95 of 0.454 (9.4% relative gain over baseline, absolute gain 0.039), with mAP50 and mAP75 reaching 0.774 and 0.466, at an inference speed of 555 FPS, respectively, outperforming the compared mainstream models. On GC10-DET, mAP50 reaches 0.704, comparable to the baseline, maintaining stable overall detection capability, while mAP75 and mAP50-95 improve by 0.033 and 0.019, verifying the module’s performance advantages under high localization accuracy requirements and its cross-dataset generalization ability. The proposed module effectively balances detection accuracy and lightweight characteristics, providing a high-performance solution for industrial steel defect detection.