چكيده به لاتين
Structural Health Monitoring (SHM) is a significant research area in structural engineering, aimed at detecting, locating, and assessing damage in structures to ensure their safety, reliability, and efficiency. Traditional SHM methods often face challenges due to the need for manual inspections, high costs, and the inability to detect hidden damages. This study presents an innovative approach that combines signal processing, image processing, and deep learning techniques for SHM.
In this method, raw vibration signals are first cleaned using preprocessing techniques such as noise removal and filtering. Then, the signals are converted into time-frequency images using Continuous Wavelet Transform (CWT), highlighting damage-related features. The CWT images are segmented into homogeneous regions using the watershed algorithm, and statistical and geometric features are extracted from each region. To generate initial labels without the need for labeled data, a self-supervised clustering algorithm (FSSC) is employed. Finally, a custom deep convolutional neural network with a self-attention mechanism is trained to classify the structural health status.
The proposed method was evaluated on three diverse datasets: real data from the Tianjin Yang Bridge, simulated data from the finite element model of the Central Florida University Bridge, and laboratory data from a steel structure at Qatar University. The results demonstrated that this method could detect damage with an accuracy of over 96% in all cases, indicating its generalizability and adaptability. A comparison with other common methods confirmed the significant superiority of the proposed approach.
The main innovations of this research include the use of Continuous Wavelet Transform to produce time-frequency images, the application of the watershed segmentation algorithm to extract damage-related features, the use of self-supervised clustering for learning from unlabeled data, and the design of a custom deep neural network architecture with an attention mechanism. Despite the promising results, limitations such as the need for larger datasets and computational complexity remain.
Future research areas include exploring more advanced neural network architectures, integrating multi-sensor data, developing transfer learning approaches, and applying explainability techniques. With advancements in these directions, machine learning-based SHM systems are expected to be increasingly adopted in the industry, leading to improved safety, reliability, and efficiency of engineering infrastructures.
Overall, this study makes a significant contribution to the field of SHM and the application of AI technologies in structural engineering. The proposed method demonstrates how combining advanced signal processing, image processing, and deep learning techniques can lead to the development of innovative and effective systems for automated damage detection. The findings of this research can inspire further development of AI-based SHM methods and ultimately contribute to the enhancement of engineering infrastructure management and maintenance.
