Changes between Version 114 and Version 115 of Other/Summer/2025/mlCoexist

Timestamp:

Jul 28, 2025, 5:06:26 PM (8 days ago)

Author:

aw1086

Comment:

—

Other/Summer/2025/mlCoexist

@@ -155 +155 @@
 
 **Progress:** 
+- Problem: found impracticality in using binary classification (with/without RFI) on transmitting signal spectrograms. As they are pre-RFI and contain no features that reveal the impact of signal interference, Tx graphs are not suitable as input for training CNN, since the model would become a simple color detector (hue of spectrograms based on intensity of power, e.g., orange = RFI, blue = no RFI), which reduces the entire purpose of using ML for RFI detection
+
+- Solution: pivoted to using cropped Tx spectrograms (only signals leaked into L-band) to estimate brightness temperature, and changed the model from classification to regression
+ - Force the model to extract more graphical features than just color (ex. using the area of the brighter color in graphs to approximate total Tb when leaked signal reaches the receiving site) 
+
 - MATLAB code modification: 
  - Added a variability margin to the signal gains to simulate the uncertainty in physical transmissions and obtain multiple samples from a single scenario
  - Edited the y-axis (frequency range: 1400 MHz–1427 MHz) to illustrate only the section of signals that leaked into the L-band, in order to prevent the CNN model from picking up insignificant details during training
-- Problem: found impracticality in using binary classification (with/without RFI) on transmitting signal spectrograms. As they are pre-RFI and contain no features that reveal the impact of signal interference, Tx graphs are not suitable as input for training CNN, since the model would become a simple color detector (hue of spectrograms based on intensity of power, e.g., orange = RFI, blue = no RFI), which reduces the entire purpose of using ML for RFI detection
- - Solution: pivoted to using cropped Tx spectrograms (only signals leaked into L-band) to estimate brightness temperature, and changed the model from classification to regression
- - Force the model to extract more graphical features than just color (ex. using the area of the brighter color in graphs to approximate total Tb when leaked signal reaches the receiving site) 
 - New data table for model training: 
-
 \\
 \\