TL;DR
This paper introduces a novel, non-contact framework for accurately measuring fatigue crack lengths in engineering structures. By combining a robust convolutional neural network (CNN) for crack detection with a precise image processing algorithm for crack tip localization, the method achieves submillimeter accuracy, outperforming traditional manual inspection.
What Problem Does This Paper Address?
Fatigue failure is a critical concern for the safety of engineering structures like bridges and aircraft. Accurately monitoring crack propagation is essential for predicting remaining life and preventing catastrophic failures. Traditional methods, such as human inspection, are time-consuming and subjective. Existing vision-based methods often struggle to distinguish true cracks from noise (scratches, edges) and accurately pinpoint crack tips, which is crucial for length measurement. This paper aims to bridge this gap by developing an automated, accurate, and robust system for fatigue crack length measurement.
Methods
The proposed framework consists of two main stages:
- Crack Area Detection (CNN-based):
- A Convolutional Neural Network (CNN) is employed for robust crack detection. The architecture is based on the Single Shot Detector (SSD) meta-architecture, utilizing MobileNets as the feature extractor.
- This approach leverages the CNN's ability to learn complex features and distinguish true cracks from crack-like noises and background elements.
- The CNN is trained to detect the entire crack region, enhancing its reliability.
- Crack Tip Location and Length Measurement (Image Processing):
- Once the crack area is identified by the CNN, a multi-step image processing algorithm is applied to precisely locate the crack tip.
- Threshold Segmentation: A global threshold is used to separate the crack (darker pixels) from the background.
- Morphological Operations: Dilation and erosion are used to connect broken parts of the crack and remove small noise regions, resulting in a cleaner crack image.
- Crack Tip Identification: The rightmost pixel with a specific grayscale value (255) in the processed binary image is identified as the crack tip.
- Length Calculation: The crack length is calculated by measuring the pixel distance between the identified crack tip and a reference point, then converting this to physical distance using a calibration scale.
Datasets
- Training Data: A dataset of 4928 crack images was collected using a measuring magnifier under fatigue experiments. These images, with resolutions of 640 × 480 pixels, were manually labeled using LabelImg. The dataset was divided into training (80%) and validation (20%) sets based on a five-fold cross-validation principle.
- New Benchmark: While not explicitly stated as a new benchmark, the paper describes the creation and use of a specific dataset for training and validating their CNN model, focusing on fatigue cracks.
Key Insights and Ideas
- CNN for Robust Crack Detection: Using CNNs, particularly with architectures like SSD and MobileNets, significantly improves the ability to detect cracks amidst complex visual noise (scratches, edges, boundaries) that traditional methods struggle with.
- Two-Stage Approach for Accuracy: Combining a powerful deep learning model for initial detection with precise image processing for fine-grained localization (crack tip) is key to achieving high measurement accuracy.
- Morphological Operations for Crack Connectivity: Techniques like dilation and erosion are crucial for cleaning up segmented crack images, filling gaps, and ensuring the crack structure is continuous for accurate tip identification.
- Submillimeter Accuracy Achieved: The paper demonstrates that this combined approach can achieve measurement accuracy within submillimeters, a significant improvement over less precise methods.
Key Results
- Crack Detection Precision: The trained CNN model achieved mean average precision (mAP) of 0.86 at 0.5IOU and 0.35 at 0.75IOU after training.
- Crack Length Measurement Accuracy: The proposed method demonstrated a maximum absolute error of less than 0.15 mm and a maximum relative error of less than 2% when compared to human inspection.
- Robustness: The method effectively distinguished cracks from crack-like noises and performed well across different crack sizes and resolutions.
Limitations and Future Work
- Manual Threshold Selection: A limitation acknowledged is the manual selection of the image segmentation threshold, which is crucial for preserving crack tip information. The authors suggest developing an automatic image segmentation method.
- Mobile Platform Implementation: The authors propose further work on implementing this method on mobile platforms for automated crack inspection.
- Crack Width Measurement: The paper suggests modifying the method to measure crack width, which is important for reinforced concrete structures.
- Error Source: Figure 21 indicates that errors can arise from the morphological operation, which might connect noise with the crack, leading to inaccuracies in length calculation.
Summary updated