Fig. 3: The convolutional encoder-decoder (CED) model for predicting fault friction from scalograms, obtained from kinetic energy (simulation) and acoustic emission (experiment).
a CED model architecture. The encoder branch contains a b Preprocessing block and c four DownSampling 2D blocks that populate the d latent space. The decoder branch reverses the procedure using e four UpSampling 2D blocks and a f Postprocessing block. The encoder and decoder models are connected by skip connections (red dashed lines) between the downsampling and upsampling blocks as shown in a. The number of filters (f) for each block are shown in a. The image size after each layer block are provided in parentheses. The blue dashed lines indicate the sub-models used when computing the hierarchical components39 and the associated training loss function to obtain the total loss (Ltotal). The model layer notations are Conv (convolutional layer), ConvT (convolutional transpose layer), BatchNorm (batch normalization layer), linear (linear connected layer), and ReLU (rectified linear unit).
