The objective of the present invention is to introduce a two-dimensional (2D) convolutional multi-dimensional attention unit to be applied in the analysis of three-dimensional (3D) multivariate time series (MTS) data with cyclical properties, utilizing a recurrent neural network (RNN) architecture. This unit is capable of constructing an independent attention vector (α) for each variable of the MTS, using 2D convolutional operations to capture the relevance of a temporal step within segments and surrounding temporal steps.
To achieve this, the bidimensional attention unit is composed of a division block, an attention block, a concatenation block, and a scaling block.
Publications:
