Abstract Summary
This work focuses on the structural health monitoring of jacket-type foundations used by offshore wind turbines. A vibration-only response mechanism based on accelerometer data is specifically suggested. Classification models (supervised models) are one possible method for achieving the aim; however, it is difficult to collect data on structural damage to wind turbines, resulting in unbalanced data sets. The use of anomaly detection techniques to detect damage to wind turbines has been the subject of much research in an attempt to find a solution to this problem. This work, based on an anomaly detection model, has developed a methodology to detect crack bar deterioration at the wind turbine jacket consisting of two training phases with only healthy data: training of a generative adversarial network (GAN), and encoder training of an autoencoder based on the GAN model that has already been learned. A generator and a discriminator may be obtained through the GAN network training process. This model is used to train an encoder that permits the mapping of healthy pictures into a latent vector. Following encoder training, the encoder places the data at points in latent space that correspond to the input data's healthy state. The mapping of the image space to the latent space through the encoder and the subsequent mapping of the latent space to the image space through the generator should closely resemble the input image in the event of a healthy input image. However, when damage-state input images are used, the model output does not resemble the input. The image reconstruction error and a residual error comparison of the discriminator properties are the final two loss functions used for anomaly identification. The following are the work's key contributions: : i) the proposed strategy is based solely on healthy data; ii) a time-frequency feature extraction preprocess based on the Wigner-Ville (WV) transform is performed, as this transform gives accurate time-frequency representations for non-stationary signals; and iii) a signal-to-image conversion of WV features in multichannel grayscale images with as many channels as there are sensors in the structural health monitoring system. The proposed strategy has been tested through laboratory experiments on a scale model.
