Figure 3
The detailed pipeline for feature extraction from time-series data is divided into three labeled sections: “Local Feature Extraction”, “Global Feature Extraction”, and “Key Features”. On the left side of the section “Local Feature Extraction”, an input signal is shown as a vertical waveform plot. Segmented portions of the signal are highlighted and fed into two parallel processing streams with an ellipsis between them. Each stream begins with a block labeled “CONV”, representing convolutional layers applied to extract local patterns. The output passes through blocks labeled “B N” (batch normalization) and “R e L U” activation. Circular nodes labeled “Max” indicate “Max pooling” operations that reduce dimensionality while preserving important features. This sequence—“CONV”, “B N”, “R e L U”, and “Max pooling”—is repeated twice in each stream, producing stacked feature maps. The outputs from multiple streams are then combined into vertical feature vectors, each shown by a rectangle containing stacked circular nodes, representing extracted local temporal features. In the center, a section labeled “Global Feature Extraction” models temporal dependencies using a bidirectional structure. Input nodes labeled “x subscript 1” and “x subscript t” represent features at different time steps. These connect to hidden nodes labeled “vector h subscript 1” and “vector h subscript t” through weighted connections labeled “w subscript 1”, “w subscript 2”, “w subscript 3”, “w subscript 4”, “w subscript 5”, and “w subscript 6”. Two pathways are shown: a “Forward layer” and a “Backward layer”, indicating bidirectional processing of temporal information. Arrows illustrate the flow of information across time steps in both directions. Output nodes labeled “y subscript 1” and “y subscript t”, with an ellipsis between them, represent globally extracted temporal features. On the right side, a section labeled “Key Features” applies a block labeled “Multi-head Attention”. This module takes the globally extracted features and computes attention weights to emphasize the most important temporal information. The output is a set of three stacked circular nodes, with an ellipsis, labeled “Temporal Features”, representing refined feature vectors after attention-based selection.Time dimension model. Source(s): Figure created by authors