5 Generative Adversarial Networks for Labelled Vibration Data Generation 43 Fig. 5.1 Steel frame grandstand simulator [9] through the network, it helps the model to be more accurate and take less computation time. However, only a single operation is necessary for the GANs, which is only training, thus normalization is not needed unless the input signal has spikes, otherwise different scaled weight propagations can lower the model quality during the training. Additionally, the used 1-D W-DCGAN consists of batch and instance normalization layers which help the batches of data to be normalized during the training. After several trials with normalized input and raw input, the result did not noticeably change and in fact, it is believed that the 1-D W-DCGAN model can learn the spatial and temporal features better if the raw signal is fed in the model. Thus, the input is not normalized, and the raw 262,144 length of damage tensor is used as an input. 5.2.3 Model Architecture In the model, the generator takes the [1×256] dimensional noise tensor (z) and pass it through 5 1-D transpose convolutions, then followingly the [1 ×1024] vibration tensor is created. This is the point where both the created tensor from generator and the sampled batches of [1 ×1024] from the 1-D input data (×) enters the discriminator which is named as critic in the W-GAN paper. Then the critic takes the inputted 1-D tensor and after 5 1-D convolutions, it yields the decision score of the Critic to backpropagate and optimize the network afterwards. The readers are directed to the GAN [24], DCGAN [25], and W-GAN [26] papers for more details about the original models and the training process. After several trials, the best architecture is formed as shown in Fig. 5.2. Additionally, the generator starts the 1-D transpose convolutions with filter =64, stride =2, padding=0. Then, continues for the rest with filter =4, stride =2, padding=1. The critic takes the same filter, stride, and padding values in reverse, but the last 1-D convolution layer takes filter =64, stride =2, padding =0. Note that due to the limited specifications of the used PC, the maximum number of length of features in the generated dataset could be 1024. That means, after the training, the model is going to learn the 262,144-length of dataset and generate a 1024-length dataset of variation of the input. 5.2.4 Fine-Tuning for Training Training GANs and its variants are notoriously the most challenging ones among other DL models. The most common challenges are: powerful discriminator over the generator; high oscillation and unstable training thus no convergence; mode collapse; in other words, generator always creates the same output; diminishing gradients – losing gradients through the network with more epoch. A couple of approaches have been used to tackle these challenges and observed positive effects: Using W-GAN significantly helps in more stable and less sensitive training progress and leads to convergence with a trade-off with more time; Using dropout in the critic helped to reduce the capacity of the critic and overfitting the network; a decaying
RkJQdWJsaXNoZXIy MTMzNzEzMQ==