Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Fusion of Wave Simulation with Neural Networks for Structural Identification
Ultrasonic guided waves offer great opportunities in Structural Health Monitoring (SHM). Due to their short wavelength and their high frequency, however, computational costs for Ultrasonic guided wave simulations are numerically expensive. These simulations show a necessity towards unleashing the full potential of this method. Therefore, a mechanical informed surrogate will be developed based on Neural Networks.
Keywords: Neural Networks, Ultrasonic Guided Waves, Structural Health Monitoring, Fusion of data and models
This Master thesis is part of research towards Structural Health Monitoring (SHM) using Ultrasonic Guided Waves (UGWs). UGWs are particularly suited for structural defect detection for mainly two reasons. Guided Waves are propagated within thin structures (e.g. plates), and suffer less from attenuation effects, which render these more suitable for performing SHM in a large range. Moreover, due to their small wavelength, they are suited for detection of small defects.
To unleash the full potential of UGW testing, measurement data should be coupled with models of the monitored structure. These models might be based on the finite element method or other modelling approaches, e.g. the spectral element approach. A drawback of all these simulations is the high computational effort. This comes from the property that UGWs have a short wavelength and a high frequency. Both need fine discretisation in space and time, respectively.
To overcome this, reduced order models (ROMs) or surrogate models can be developed for reducing the required evaluation time of a full order model by orders of magnitude. For effective use within the context of SHM for inference of local defects, such models need to be parametric with respect to the location (and other properties) of the defect. This thesis will focus on methods of characterizing delamination defects within a composite plate. A Neural Network architecture will be used for constructing a suitable ROM. It will be shown that owing to the nature of the problem (Fourier transformation of shifted waves), the network should be able to recover a periodic behaviour in terms of parameter dependence, albeit on the basis of few training points.
This Master thesis is part of research towards Structural Health Monitoring (SHM) using Ultrasonic Guided Waves (UGWs). UGWs are particularly suited for structural defect detection for mainly two reasons. Guided Waves are propagated within thin structures (e.g. plates), and suffer less from attenuation effects, which render these more suitable for performing SHM in a large range. Moreover, due to their small wavelength, they are suited for detection of small defects. To unleash the full potential of UGW testing, measurement data should be coupled with models of the monitored structure. These models might be based on the finite element method or other modelling approaches, e.g. the spectral element approach. A drawback of all these simulations is the high computational effort. This comes from the property that UGWs have a short wavelength and a high frequency. Both need fine discretisation in space and time, respectively. To overcome this, reduced order models (ROMs) or surrogate models can be developed for reducing the required evaluation time of a full order model by orders of magnitude. For effective use within the context of SHM for inference of local defects, such models need to be parametric with respect to the location (and other properties) of the defect. This thesis will focus on methods of characterizing delamination defects within a composite plate. A Neural Network architecture will be used for constructing a suitable ROM. It will be shown that owing to the nature of the problem (Fourier transformation of shifted waves), the network should be able to recover a periodic behaviour in terms of parameter dependence, albeit on the basis of few training points.
1)Introduction to UGWs and neural networks
2)Selecting a suitable neural network architecture for the problem at hand
3)Apply the neural network for detecting and characterizing delamination-type defects within a composite plate
4)Optional: Apply the derived neural network architecture to a different type of defect, other than delamination
5)Writing the thesis report and offering a final presentation
1)Introduction to UGWs and neural networks 2)Selecting a suitable neural network architecture for the problem at hand 3)Apply the neural network for detecting and characterizing delamination-type defects within a composite plate 4)Optional: Apply the derived neural network architecture to a different type of defect, other than delamination 5)Writing the thesis report and offering a final presentation