Figure 3.

(A) Overall architecture of the TCN-based model. The modified TCN is formed by five stacked residual blocks with filter size k = 3, number of filters n = 256, and dilation factors d = 1, 2, 4, 8, 16. The output channels of the first FC layer and second FC layer are 128 and 5, respectively. The weights in the network are initialized with no bias. (B) Architecture of a residual block. A 1×1 convolution is implemented if the input and output of the residual block have different dimensions. Z^l denotes the output of block l. (C) Dilated causal convolution network with two hidden layers. As each layer has the same length as the input layer, the convolution stride equals one. The receptive field of the convolution is computed as (k − 1)d, which means that higher layers can cover longer current segments.