AI/ML Success with RTTY
Here is a quick result from an #ai #deeplearning project for #amateurradio that we’ve been working on at Open Research Institute, Inc. Special thank you to MathWorks for supporting our educational, personal, and professional development through #opensource mission.
This result is from a deep learning model that recognizes RTTY messages. It’s given a RTTY transmission in audio format. Each transmission is short, only 10 characters long. The characters are randomly selected, which drives up the chance that the control characters that switch between letters and figures are involved. Another additional complication is that each message gets a random amount of AWGN from 0dB to 60dB. So, the signals aren’t clean.
The model is over 90% accurate under these conditions. The image is of a “confusion chart”, which is a compact way of communicating how well a classification system performed. To use this type of chart, look up what character was actually sent, and then see on the other axis how well the model performed in identifying it.
The model type used was a “long short term memory” model (LSTM), which is useful for finding relationships developed over time. This is a reasonable match to RTTY, where the control characters, once sent, determine which lookup table one uses for many characters in the future. In RTTY the symbols are “double booked”, in that any five bit digital representation of a character has two assignments. When you want to send figures (numbers and punctuation) then you send a figures control character. The receiver then uses the figures lookup table until a letters control character is sent. After that, the receiver uses the letters lookup table. And so on. This makes it not a simple “digits in characters out” scheme.
Note the blank space for “10”. There is no “10” character in RTTY. This is possibly a repercussion of having to give the model a one-character overhead based on Mathworks documentation about this specific deep learning technique. We are going to follow-up on this and “figure” it out.
If you would like to learn more about how to learn AI/ML using the Amateur Radio bands, then please visit ORI’s getting started page today, at openresearch.institute/getting-s…
