Structural health monitoring of building structures using novel acceleration-based signal-to-image technique and 2D convolutional neural networks

Structural Health Monitoring (SHM) is essential for detecting damage in structures like buildings, bridges, aircraft, etc., as they can deteriorate over time or suffer sudden failure during extreme events. In this realm, the application of deep learning (DL) techniques has been observed to become quite an integral part of global-level SHM using acceleration-based measurements by automatic monitoring in a timely and precise manner. In the present work, a novel strategy has been proposed by considering acceleration-based measurements in the form of computer vision (through the encoding of time series into images) and proposes two novel strategies (employing DL models). Specifically, two advanced encoding strategies have been introduced: (i) compressed time series recurrence plot (CTSRP), which captures temporal dependencies and patterns in vibration data, and (ii) time delay encoding (dotTDE), which encodes time-delay features to enhance the representation of subtle damage characteristics. Transforming a time series to an image can bring contrasting features that may not be present in the raw time series. Such images are then utilized as input to the 2D convolutional neural network (CNN) for determining the possible damages in structures. The proposed strategies are finally validated using two building structures (numerically simulated) i.e., a shear frame structure and the IASC-ASCE benchmark structure (incorporating multiple damage scenarios). During the comparison of classification accuracy, it has been observed that the CTSRP method has achieved 96.1% accuracy and the dotTDE method has achieved 100% accuracy for the ten-story shear frame structure. These results have shown an improvement ranging from 0.9% to 92.9% for CTSRP and 5.0% to 100% for dotTDE over the state-of-the-art approaches. Additionally, for the IASC-ASCE benchmark structure, the CTSRP method has achieved 94.8% accuracy, with improvement ranging from 0.4% to 72.1%, while the dotTDE method has achieved 100% accuracy with improvement ranging from 5.9% to 77.3%. Moreover, the analysis of the robustness and scalability of the investigated DL models has also been reported. In the future, the investigation may be conducted by focusing on enhancing the proposed DL models to improve their ability to detect unseen damage classes and adapt to structural variations. Furthermore, the models can be applied to different real-world structures for broader validation and applicability.