Damage Classification of a Three-Story Aluminum Building Model by Convolutional Neural Networks and the Effect of Scarce Accelerometers

Abstract

Structural health monitoring (SHM) plays a crucial role in extending the service life of engineering structures. Effective monitoring not only provides insights into the health and functionality of a structure but also serves as an early warning system for potential damages and their propagation. Structural damages may arise from various factors, including natural phenomena and human activities. To address this, diverse applications have been developed to enable timely detection of such damages. Among these, vibration-based methods have received considerable attention in recent years. By leveraging advancements in computer processing capabilities, machine learning and deep learning algorithms have emerged as promising tools for enhancing the efficiency and accuracy of vibration-based SHM. This study focuses on the application of convolutional neural networks (CNNs) for the classification and detection of structural damage within a steel-aluminum building model. An experimental platform was devised and constructed to generate data representative of building damage scenarios induced by bolt loosening. Both the typical placement of sensors on each floor and the utilization of only one accelerometer were employed to understand the effect of scarcity of accelerometers. By subjecting the building model to controlled vibrations and environmental conditions, the response data from both sensor configurations were collected and analyzed to evaluate the effectiveness of the CNN approach in detecting structural damage under varying sensor deployment strategies. The findings demonstrate that the CNNs exhibited high accuracy in both damage classification and detection, even under scenarios with limited sensor coverage. Moreover, the proposed method proved effective in identifying structural damage within building structures.

Tiny-Machine-Learning-Based Supply Canal Surface Condition Monitoring

“…The evaluation results for the model are presented in Table 3. ShuffleNetV2 highlights efficient resource utilization, featuring a low number of parameters (1,194,515) and FLOPs (20,652,799), making it one of the models with the lowest complexity and number of parameters. With an accuracy of 94.95 ± 1.84% and an F1-score of 94.75 ± 1.70%, ShuffleNetV2 achieves a balance between performance and size, making it ideal for deployment on resource-constrained devices.…”

“…In [19], a vision-based method using deep CNNs is proposed for detecting concrete cracks in infrastructure; the CNN is trained on 40,000 images and achieved 98% accuracy, demonstrating the potential of deep learning for damage detection in building infrastructure inspection, particularly for automatic detection and classification. Study [20] explored the capacity of CNNs to classify and detect structural damage in steel and aluminum building models. It found that CNNs can efficiently classify damage even with limited sensor coverage, demonstrating their capacity to enhance the accuracy and efficiency of structural health detection.…”