PSI - Issue 64

Eshwar Kumar Ramasetti et al. / Procedia Structural Integrity 64 (2024) 557–564 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

558

2

SHM systems can address the following physical quantities: corrosion, cracking, displacement, strain, temperature, and vibrations. Similarly, the most utilized sensors in SHM systems are accelerometers and strain sensors. The SHM process commences with the acquisition of data through the sensors from the excitation and material properties of the structures. In the first step, data cleaning is done to remove inconsistent data and normalize values from the dataset. The next step, feature selection in the SHM process is very essential as it converts the data to be clearly understandable by any machine learning algorithm in the pattern recognition stage (Onur et al., 2021). In the past few decades, SHM systems have been installed on several bridges to monitor the health of the bridge and to increase its life span. The Norwegian University of Science installed a SHM system along the main girder of the Hardanger Bridge in Norway with twenty accelerometers and nine anemometers to monitor the wind speeds and accelerations. The description of the monitoring system is explained by Fenerci et al. (2021), and this measured open access data has been the subject in many various research papers. The University of Sheffield installed a SHM system on the Tamar suspension bridge in England to monitor wind speeds and temperatures (Cross et al., 2013). In Germany, SHM systems were installed on several bridges in the Scherkonde valley between the cities of Erfurt and Leipzig/Halle. For a period of twelve years, the bridge temperature and displacement were monitored (Herbers et al., 2023). A SHM system was installed at Maintal-Bridge in Gemünden am Main by Simon et al. (2022) to monitor the traffic impacts and apply stochastic subspace damage detection to the structure. The machine learning algorithms have become popular to predict patterns using the data generated from the SHM system. The commonly used deep learning algorithms are artificial neural networks (ANNs) and convolutional neural networks (CNNs), which extract features from the raw data to assess the structural integrity of the structure. Lawal et al. (2023) proposed an ANN architecture to classify the vehicle movement on a railway bridge from the acceleration dataset obtained from the SHM system. The dataset consists of 210 events, with train crossings and minor impacts having 105 events each. The dataset was split into 70 % training and 30 % testing data. The predicted results achieved a classification accuracy of 98.67 %, and the model did not show overfitting. Yang Yu et al. (2019) developed a deep convolutional neural network architecture (DCNN) to identify the damages at building structures using the acceleration data from the sensors. The results show the proposed model achieved high accuracy for damage detection on raw noisy signals and outperformed the other ML-based methods in terms of root mean square error (RMSE). Khodabandehlou et al. (2019) also developed a similar CNN-based technique to classify if damage occurred or not on a concrete box-girder bridge. The accuracy of the proposed model was almost 100 % in predicting the different damage states from acceleration signals. Sitton et al. (2019) proposed a simple neural network to classify if there was any train passing on the railroad bridges. The data was recorded using SENSR CX1 sensors on eight low-clearance bridges. The CX1 consists of 3-axis micro-electro-mechanical system (MEMS) accelerometers, and samples were taken at a rate of 150 Hz. The model detected and classified impacts with an accuracy range of 91 - 100 % with false rates limited to 0.0 - 0.75 %. Arnold and Keller (2024) developed a deep learning model to classify the vehicle type on the bridge with the vibration dataset. The results for the differentiation between cars and trucks showed an accuracy of 100 % with all the models. In the current project, a SHM system was installed by the Federal Institute of Materials Research and Testing (BAM) on the Nibelungen Bridge located in the city of Worms, Germany. The SHM system setup is tailored to access the individual structural behavior of the bridge and is therefore specific. Based on these SHM systems and the unique datasets it is vital to develop classifiers with quick adoption capabilities between similar structures and different SHM hardware. Different hierarchy levels of classifiers can be used to identify local load states, and based on lower-level classifiers, extended classifiers can predict individual structural behavior and overall load states. The convolutional neural network (CNN) algorithm was developed to classify whether a vehicle is passing or not on the bridge based on the acceleration data. Furthermore, when a vehicle passed over the bridge, the model was used to classify whether it was a car, truck, or large truck. In continuation of this work, the goal will be to transfer this model to the other sensor datasets located at different areas of the bridge and to see if the model can predict similar results. This developed deep transfer learning model aims to reduce the training data to produce accurate results and transfer it to other bridges in the future. Section 2 begins by discussing the installation of the SHM system and the analysis of the dataset. Section 3 discusses the proposed classification neural network approach. Section 4 presents the results of the proposed approach, and Section 5 concludes it. The primary goal of this project is to develop efficient and accurate models to study the vehicle movement on the bridge. The results describe the accuracy of the proposed model.