Full-Waveform Echo Component Analysis Using Hybrid Attention Xception Networks

Determining the number of echo components is a critical and highly challenging step in full-waveform echo-signal decomposition. Traditional methods typically infer component counts from local extrema and inflection points, but these heuristics suffer substantial misclassification when confronted with overlapping or asymmetric echoes. To overcome these limitations, this study proposed a classification network named SESA-Xception. The model adopts Xception as its baseline and incorporates both Squeeze-and-Excitation (SE) blocks and self-attention mechanisms for structural optimization. This innovative integration significantly enhances the network's capability to comprehensively understand the morphological characteristics of echo signals. Experimental results show that, in complex echo scenarios, the overall classification accuracy of SESA-Xception reached 99.4%, which is 1% higher than the baseline model. When more than three complex waveforms are included, the classification accuracy is 3% higher than the baseline model. By learning the characteristics of waveforms, the SESA-Xception network can accurately identify the differences between asymmetric components and aliased waveforms and hidden components, make precise judgments on the number of echo components, and provide strong support for subsequent waveform decomposition work.