Figure 4
The diagram presents a pipeline for extracting spatial features from time-series data using graph convolutional networks. The diagram is divided into several labeled sections: “Data”, “G C N 1”, “G C N 2”, “G C N 3”, and the final output labeled “Spatial Features”, each connected by a right-pointing arrow. On the left side, a dashed box labeled “Data” contains a waveform plot representing a time-series signal. Below the waveform, a legend shows colored circular nodes labeled “Sample point 1”, “Sample point 2”, “Sample point 3”, followed by an ellipsis, and “Sample point n”. These colored nodes represent individual data samples that will be treated as nodes in a graph. An arrow points from the data section toward the first graph convolution block. The first processing block is labeled “G C N 1”. At the top of the block, a diagram labeled “Graph Convolution” shows a small network of connected nodes representing the graph structure. Below it, two sequential layers are labeled “Batch Norm” and “R e L U”, indicating batch normalization followed by a rectified linear unit activation. The second block labeled “G C N 2” repeats the same structure. A graph convolution layer processes node relationships, followed by a “Batch Norm” layer and a “R e L U” activation layer. The third block labeled “G C N 3” again contains a “Graph Convolution” diagram followed by “BatchNorm” and “R e L U”. Arrows between the blocks show the flow of information from one layer to the next. After the third graph convolution block, the output is passed to a node labeled “G A P”, which stands for “Global Average Pooling”. This operation aggregates node-level information into a single feature representation. The final output appears as a vertical column of circular nodes labeled “Spatial Features”, representing the extracted spatial feature vector derived from the graph-based processing of the data.Spatial dimension model. Source(s): Figure created by authors