*3.2. Model Training*

The velocity spectrum of CDP 50 is used as the tracking target of the training dataset. As shown in Figure 6b, there are four energy clusters. We take the energy cluster of 0.2–0.35 s (as shown in Figure 7a) as an example to explain the generation of a training dataset. Based on the energy cluster of 0.2–0.35 s, 100 training samples labeled as positive are generated by stretching, compression, and filtering. Figure 7b,c are examples of images after deformation and stretching, and Figure 7d–f are examples of images after filtering. The positive samples are set to 1.

**Figure 7.** Examples of positive sample. (**a**) Tracking target image; (**b**) image after 1.5 times stretching; (**c**) image after 3 times stretching; (**d**) image after low-pass filtering; (**e**) image after bandpass filtering; (**f**) image after high pass filtering.

The energy clusters of 0.1–0.2 s, 0.35–0.5 s, and 0.5–0.65 s in Figure 6b are stretched, compressed, and filtered to generate 300 training samples marked as negative, as shown in Figure 8. The negative samples are set to 0. The training process has 500 iterations, and the learning rate is set to 0.0001. After this training process is completed, set the tracking target as the energy cluster of 0.35–0.5 s, and set the energy clusters of 0.1–0.2 s, 0.2–0.35 s, and 0.5–0.65 s as the negative samples, and then repeat the same training process.

**Figure 8.** Examples of negative sample.
