Fig. 6: The proposed neural architecture.
From: Flight trajectory prediction enabled by time-frequency wavelet transform
The network is cascaded by an input embedding network, an encoder, multiple decoders corresponding to all wavelet components, and an IDWT module. In each decoder, a wavelet attention module is devised to capture scale-oriented features of the trajectory sequence. The generated components by decoders illustrate global flight trends and local motion details. An orthodox inverse discrete wavelet procedure further transforms wavelet components into flight trajectory sequence of the past period and next instant. The mean squared error of generated wavelet components serves as the loss function to update the trainable parameters of the neural networks.
