1 June 2019 in Materials Chemistry and Physics
26 May 2020 in Smart Materials and Structures
Brillouin scattering (BS)–based distributed optical fiber sensors (DOFS) provide distributed sensing capabilities by monitoring the strain along entire segments of structures. Large cracks, such as those with large crack opening displacements (COD) can be detected by strain peaks or singularities along the measurement length of distributed sensors. Microcracks do not provide visible pronounced local peaks along the length of measured distributed strains. The peaks corresponding to microcracks are submerged within the measurement noise due to low signal-to-noise ratio (SNR) of BS systems. Deep learning (DL) methods have the potential to automatically extract feature representations from data exhibiting lower SNRs, and improve classification accuracy. Accordingly, a novel DL method is proposed in this study to improve the crack detection sensitivity of the BS-based DOFS. Development of the proposed DL method includes construction of model architecture, design of a training algorithm and the detection process. A 15 m-long wide-flange steel beam with artificial defects is built and employed in this study. A comprehensive experimental program is undertaken in order to train, validate and test the generality of the proposed DL method. Experimental results demonstrate that the DL method is capable of extracting highly discernable microcrack features from the distributed strains, and distinguish the crack-induced local peaks from the noise. Microcracks with CODs as small as 23 microns are accurately detected in the present work.
1 November 2023 in Expert Systems with Applications
RMRavie Chandren Muniyandi
3 December 2023 in Engineering Applications of Artificial Intelligence
10 October 2019 in Structural Health Monitoring
Accurate micro-crack detections on the whole surface of civil structures have great significance. Distributed optical fiber sensor based on Brillouin optical time-domain analysis technology exhibits great facility to measure strain distributions along the whole surface of structures with a high spatial resolution, thus providing a potential and competitive solution to the detection problem. However, mainly due to low signal-to-noise ratio in measurements, such sensor system is still limited in crack detection–based structural health monitoring applications. How to extract high-quality micro-crack feature representations from the low signal-to-noise ratio–distributed strain measurements is crucial to solve the problem. It has been demonstrated in field of pattern recognition that deep learning can automatically extract high-quality noise-robust feature representations from mass chaos data. Therefore, a micro-crack detection method is proposed herein based on deep learning to analyze the full-scale strain measurements. Each measurement is normalized and segmented into a set of equal-length subsequences. Autoencoders, a typical kind of building block of deep neural network, are stacked layer-wise into a deep network and then exploited to automatically extract feature representations from the subsequences. Each extracted feature representation is labeled as one of the two categories by a Softmax regression. One category originates in the subsequences acquired from structure sections with crack defects and another from sections without any cracks. The micro-crack detections are achieved by solving such a crack/non-crack binary classification problem. A 15-m-long steel I-beam with artifact crack defects is built up in laboratory to verify the proposed method. Experimental results demonstrate that the minimum size of detectable crack opening width reaches to 23 μm, and besides, the proposed method is significantly better than traditional Fisher linear discriminant analysis method and classical support vector machine on the detection accuracy.
8 February 2025 in Engineering Applications of Artificial Intelligence