GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.3 GNGTS 2024 trained the NN we can scale every signal to the trained SD and make the inference that will output a signal with half the frequency, for the Low-frequency model and one with a double central frequency, for the High-frequency model. Thanks to the constraints given by the custom loss functon with the frequency counterpart taken into account, we are able to further generalize the results to other frequencies just by applying a Scaling Factor (SF), i.e. a factor applied to the SD therm that multplies the number of samples in the output. This feature allows us to infer quite easily diferent frequency predictons. In Figure 1 we plot some of these predictons with varying SF: low-frequency model in Figure 1A where the input SD is divided by the SF and high-frequency model, in Figure 1B, where the SD is multplied by the SF. The input in both of these results is a frequency fltered version of the Viking Graben Line 12 (Keys and Foster, 1998), fltered with a band-pass flter at 6-20Hz. Figure 1: Low (A) and High (B) frequency inference of the Viking Graben Line 12 on the x-axis frequency, and on the y- axis the Scaling Factor. The solid line in A and B marks the chosen frequency that will be used in the next secton. Results and Discussion Low Frequency inference Results of the applicaton of the methodology are shown in Figure 2. We moved the central frequency from 14Hz to 8Hz, as shown in the amplitude-frequency plot in Figure 2C. If we compare the input (Figure 2A) and the predicton (Figure 2B), it is clear that, as expected from the low frequency counterpart, more importance is given to main refectors, e.g. horizontal refector at 2s in the data. We can furthermore appreciate that amplitude is preserved as expected and interference is properly predicted, e.g. in the wedge around 1s in the frst 750m .