Published in Elsevier's Engineering Applications of Artificial Intelligence
This study was carried out to improve the initial experiments investigated for the undamaged-to-damaged acceleration response domain translation, which was previously published in Elsevier's MSSP
The primary difference between this and that initial study is that the model used in this study is enforced with frequency loss, the model architecture is improved, and the model is trained with more datasets that are strategically chosen from the structure to perform the undamaged-to-damaged application better.
Motivation and Problem:
Unpaired image-to-image translation is a popular research topic in computer vision and graphics. Recently, the authors of this paper took a similar approach and translated the domain of acceleration responses collected from a steel grandstand structure. In doing so, the undamaged response is translated to damaged, and the damaged response to undamaged. For that, a variant of the CycleGAN model is trained with undamaged (bolt tightened at the joint) and damaged (bolt loosened at the joint) responses from a single joint in the structure. However, the success of the domain translation on the test joints was very limited.
Objective and Scope:
Thus, this study investigates improvements to the model and the training procedure for further accuracy. First, the model in this study gets a more extensive training procedure to increase the model’s domain knowledge. During the training, a novel signal coherence-based index is considered to account for the similarity of frequency domains of the original and the translated data. Second, the Gated Linear Units, skip-connections, and Mish activation function are used to minimize the gradient loss and to learn the broader features in the data. Third, the total loss function of the generator is supplemented with a new frequency domain-based loss to better capture the frequency content of the data. Fourth, random decaying noise is added to the inputs for better generalization in the test data. Last, the model is evaluated using modal parameters such as natural frequencies, damping ratios, and singular value decomposition of the estimated spectral densities. The improvements presented in this study demonstrate a successful domain translation of acceleration responses for the tested joints compared to past study. The findings of this paper show that domain translation can be advantageous in Structural Health Monitoring applications, such as having access to the damaged or undamaged response of the structure while it is in pristine or unhealthy condition.
config.py provides the configurations used in the training
critic.py provides the critic model
generator.py provides the generator model and the other blocks used in both critic and generator
metrics.py provides some of the metrics used in the training
train.py is the file for training the generators and critics
utils.py is only used for gradient penalty used for the critics during the training
The dataset used in the paper can be obtained here
Domain translation applications are typically applied for images, e.g., zebra to horse conversions etc. Signal type of data instances can be a little different depending on the length on the signal. What I did in this paper is the division of the 1024-second signals from undamaged and damaged domains into 16-second tensors (each tensor is a 16 second vibration signal csv file). Then, used those 16-second tensors from both undamaged and damaged domain as if they are image data to train my model. More details about how this is achieved is available in the paper.
Training data flow for (a) undamaged-to-damaged domain translation and (b) damaged-to-undamaged domain translation
The frequency domains of the concatenated response signals are plotted and shown in Fig. 8, Fig. 9, Fig. 10, Fig. 11, Fig. 12, Fig. 13. For instance, Fig. 8(a) shows the frequency domain plotting of undamaged and synthetic undamaged acceleration response signals at Joint 5; on the other hand, Fig. 8(b) shows the frequency domain plotting of damaged and synthetic damaged acceleration response signals at Joint 5. For modal identification results, please see the original paper https://www.sciencedirect.com/science/article/pii/S0952197623003305
