Fig 2. DeepGANnel network architecture.

On the left is a generator network of DeepGANnel in which the input is a random noise and output is a 2D image by 2x1280 that is converted from 1-D time series. The generator model in this architecture uses three deconvolution layers (upsampling) to produce an image from seed (random noise shape by 2x1280). Several Dense layers are used to take the noise input and transform into a desired image by up-sampling (Deconv or transposed convolutional layer) steps size of 2x1280. At each layer, Leaky Rectified Linear Unit (LeakyReLU) is used as activation function with batch normalisation, except output layer which uses tanh activation function. The discriminator model comprises of several convolutional layers with the same activation function (LeakyReLU) at each layer to classify whether an image is real or fake by comparing the sampled data to real data. Dropout layers were also appended to all convolutional layers except input layer with the value of 0.3 to reduce overfitting. Finally, a Dense output layer is used after flattening the network structure with classified labels.