MS17.18 - Structural Health Monitoring

Structural Health Monitoring of civil engineering structures is experiencing an increasing progress in the last decades. The present work focuses mainly on static behavior of a highway bridge due to environmental temperature effects or ground settlements. The first goal is to compare the results of the finite element simulations to the classical geodesy surveying measurements for deformation monitoring of a large, curved highway bridge. The second goal is to test the applicability of artificial intelligence methods to predict such deformations online without comprehensive computer simulations if a certain amount of measurement data is available for training purposes. This study is based on the work within the LEVANGO project funded by the German Federal Ministry of Research and Education in cooperation with the Airbus Defence and Space and the AllTerra Deutschland companies. The Wehretal bridge as a part the federal highway A44 was completed in 2019 and should be opened in 2022. This highway bridge is a prestressed concrete structure with a total length of approximately 670 m and a monolithic deck laying on a series of column pairs. A specific feature of the bridge is a relatively large curvature of the longitudinal axis with a constant radius of about 480 m and special supports with sliding bearings that enable a longitudinal and transversal displacement of the deck with a certain friction. A safe operation of the bridge requires a trouble-free behavior of sliding bearings under thermal deformations. The obtained results show a good correlation between simulation and measurement results as well as a good potential applicability of artificial neural networks to predict bridge deformation even in presence of several monitoring difficulties and data limitations. The advantages and problems of the applied approaches will be discussed in detail.

New wind power installations surpassed 90 GW in 2020 – a 53% growth compared to 2019. This magnitude of growth is necessary to meet the ambitious COP26 global climate targets whilst maintaining the supply stability for an ever increasing environmentally conscious consumer population. To increase energy capture and cost efficiency, turbines are being located in remote offshore locations, with their blades being constructed from lighter, more flexible materials. In general, monitoring of turbines remains a manual process with inspections carried out at pre defined intervals driving operation and maintenance costs prohibitively high. This research will develop a vibration based structural health monitoring (VBSHM) methodology for remote monitoring and damage severity assessment of a laboratory scale wind turbine blade under simulated wind like excitation. The methodology will exploit the fact that structural degradation will manifest itself through a notable shift in pre defined damage sensitive features and use this to predict damage accrued on the structure. The finite element model updating (FEMU) procedure adopted involves the creation of a “digital twin” by minimising a fitness function containing the discrepancy between model responses and observed dynamic responses. The application of deterministic FEMU can be considered idealistic as uncertainty can have a non negligible influence on the accuracy of the final solution. To this end, the authors incorporated non probabilistic fuzzy theory, modelling membership functions of output parameters to build membership functions associated with input parameters. This accounts for limitations associated with determinism and enables modelling and measurement errors to be accounted for in a meaningful way. The method was demonstrated on a 2.36m blade from a 5kW domestic wind turbine subject to wind like excitation. Operational modal analysis techniques were used to obtain dynamic responses of the structure with metaheuristic optimization algorithms implemented to calibrate the numerical models using a modified version of the Abaqus2matlab toolbox. Through this process, a digital twin of the baseline structure was successfully constructed, with longitudinal modulus and shear modulus calibrated to reduce the maximum percentage deviation in natural frequencies from 19.4% to 1.4%. This calibrated model was then used as a baseline for further damage detection studies. To facilitate damage severity assessment non destructively, two typical field observed damages were considered. Localized stiffness reduction, comparable to transverse cracking, was replicated by adding small masses to the blade whilst gradual boundary degradation was simulated through addition of neoprene sheet to increase joint flexibility. The VBSHM developed was able to detect with sufficient accuracy all five damage scenarios (0.05kg, 0.10kg, 0.20kg and 0.40kg on the blade’s trailing edge only and 0.20kg on both trailing and leading edges). Benefits of considering uncertainty were demonstrated through creation of membership functions for each scenario to prevent false alarms and provide confidence in the results. Boundary degradation was successfully identified experimentally however the analytical sensitivity of responses to variation in rotational and translational springs was insufficient to facilitate updating using the analytical model created. This contribution highlights the ability to account for uncertainties in a non computationally expensive and intuitive way and

Recently, with the advancement of computers and computational resources, the use of deep learning models is becoming a common practice in Structural Health Monitoring. This can facilitate the procedure by extracting proper features from data as well as detecting patterns in the data. However, there are several challenges in this effort among which two cases are very crucial: data shortage and long training time. Researchers address the former by different approaches, such as fusion at the data or decision level, incorporating the physics of the problem either by defining a hybrid cost function or by generating synthetic data from a finite element model. But these approaches may increase the overall training time. With the advancement of quantum computers and the associated algorithms, different fields have started to exploit inherent quantum mechanical features, e.g. superposition and entanglement, and investigate if these effects can be beneficial for them. Among different quantum applications, quantum machine learning (QML) methods are known for their faster convergence with smaller datasets than those of classic ML. This is one of the main reasons for their recent widespread applications. However, since limited numbers of qubits can remain stable at the same time, implementing QMLs in different fields is not feasible yet. To overcome this issue one could employ Hybrid Quantum-Classic Machine Learning approaches. In this study, a preliminary investigation on the application of Hybrid Quantum-Classical ML to structural health monitoring has been done. In this regard, a deep neural network (DNN) model with a quantum layer is selected as the hybrid model and the training dataset is obtained from ultrasonic inspection of a wind turbine blade. The effect of the quantum layer on the performance of the DNN for damage detection is then investigated by performing convergence and accuracy analyses of the model with and without the quantum layer. The promising results indicate the benefit of employing hybrid models in this field.