A study published in Buildings introduces a deep learning-based approach to structural health monitoring that leverages vibration data from two beam-ABAQUS models as input. Using a convolutional neural network (CNN), the method demonstrated near-perfect accuracy in locating damages as small as 3 and 6 cm in beams through laboratory experiments involving four beams with varying damage scenarios.
Study: Research on Concrete Beam Damage Detection Using Convolutional Neural Networks and Vibrations from ABAQUS Models and Computer Vision. Image Credit: Happy_stocker/Shutterstock.com
Background
Vibration data and deep learning techniques, such as CNNs, have become valuable tools for detecting structural damage. Structural vibration information is typically collected through image-based displacement sensors. However, challenges remain in determining whether these methods can reliably identify small internal damages, such as hollow areas in concrete beams.
Dynamic features like natural frequency, mode shape, and damping ratio are commonly used to assess structural health. These features enable the identification and localization of damage by correlating vibrations with structural conditions based on dynamic theory. However, minimal damages often introduce subtle changes to these characteristics, and environmental factors can further complicate their reliability.
To address this, the study explored an edge-detection-based displacement sensor for recording vibrations and evaluated its potential for training a CNN to accurately identify and classify damage.
Methods
The researchers used an ABAQUS model to simulate structural vibrations, generating data from several measurement points under dynamic loads. To make the dataset reflect real-world conditions, they added Gaussian white noise with varying signal-to-noise ratios (SNRs). This approach tested the model’s ability to handle noise while still detecting damage effectively.
The CNN model was built with three main layers, each serving a specific purpose:
- Convolutional layer: Identified key features in the vibration data, such as patterns that could signal damage.
- Pooling layer: Reduced the size of the data, keeping essential information while making the model faster and easier to train.
- Fully connected layer: Combined all the features to classify the type and location of the damage.
The input data consisted of raw vibration signals recorded from specific points on the beams. These signals were structured as a two-dimensional matrix, capturing both spatial and temporal details. As the data passed through the CNN, it was refined layer by layer, eventually producing a one-dimensional vector in the final layer. This output directly matched specific damage categories, providing clear and interpretable results.
To test the method, the team ran experiments on four beams with different types of internal damage. They collected 800 vibration samples (200 per beam) using an optimized computer vision-based displacement sensor. The dataset was divided into 70 % for training, 20 % for validation, and 10 % for testing. This setup allowed the researchers to train the CNN and evaluate how well it detected and classified damage.
Results and Discussion
The CNN model delivered solid results during training, validation, and prediction. For example, when the noise level was relatively high (SNR of 90 dB), the model achieved an accuracy of 74.36 %. As the noise level decreased, the accuracy improved significantly, eventually hitting 100 %.
Some errors cropped up, mostly around the three endpoints of the beams. This makes sense because damage near the free ends of the beams tends to have less of an effect on their overall behavior. Another issue was that overlapping damage points on the beams made it harder for the model to differentiate between specific cases.
Training the model with different SNR levels also showed some interesting patterns. While the training curves had a similar shape overall, lower SNRs meant the model needed more time to stabilize. As the noise increased, the accuracy dropped, and there was a bigger gap between the training and validation losses. Still, by the end of the training process, both losses leveled out. The final testing accuracy came in at an impressive 97 %, which speaks to the effectiveness of the approach.
The vibration data collected from the beams also confirmed that the image-based displacement sensors worked well. The sensors consistently captured clear and similar vibration patterns across all four beams. In total, researchers recorded 800 vibration samples—200 for each beam. The differences in these samples were due to the various forces applied during testing, as the team used a hammer to excite the beams with different levels of impact.
Conclusion
This study demonstrated how a CNN, paired with an image-based displacement sensor, could effectively detect damage in structural beams. The CNN was able to analyze raw vibration data and accurately classify damage, while the sensor reliably captured vibrations from several measurement points. Together, they provided a promising approach to structural health monitoring.
However, like any new method, there were some challenges. The sensor’s accuracy depended on factors like the camera quality and its distance from the structure, while the CNN required consistent and reliable data to perform at its best. These are areas that could be refined in future research.
Despite these limitations, the researchers are optimistic about the potential of this approach. They believe it can be scaled to larger datasets and applied to more complex structures, as the core principles of the method remain consistent across different scenarios. This makes it a valuable tool for improving how we monitor and maintain infrastructure in the future.
Journal Reference
Bai, X. & Zhang, Z. (2025). Research on Concrete Beam Damage Detection Using Convolutional Neural Networks and Vibrations from ABAQUS Models and Computer Vision. Buildings, 15(2), 220–220. DOI: 10.3390/buildings15020220, https://www.mdpi.com/2075-5309/15/2/220
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.