ESP32 Time-Drift Issue and Its Impact on Model Test Accuracy

[s1614149076]

I am working on a gesture recognition project using Wi-Fi CSI data. When I train a model using a batch of data, the accuracy is good. However, after some time, when I collect a new test set under the same hardware setup, the model fails to recognize the gestures.

I conducted an experiment where I collected static CSI data for 30 minutes and looked at the signal's behavior over time. In the first plot, I analyzed if there were any drifts over 30 minutes.

My question is related to the standard deviation of the data. In the first plot, while the mean seems to change gradually, the standard deviation exhibits stepwise changes. Could these stepwise changes indicate something about the signal's stability or possible external factors affecting the signal? Could this behavior be contributing to the instability of my model's performance when the training and testing data distributions shift over time?

[mike2872]

One thing to investigate is Automatic Gain Control (AGC) - I've found this can significantly affect CSI signal characteristics over time.

The ESP32's Wi-Fi radio automatically adjusts receiver gain based on signal conditions. This could explain both the gradual baseline shifts and the stepwise changes you're seeing in your plots. When AGC adjusts the gain settings, it changes the scaling of your CSI measurements, so the same gesture produces different CSI values depending on the current AGC state.

You can try fixing the AGC values to a constant setting and then test if your model stability improves.

[Franco]

Have you looked at ESP-CSI?
https://github.com/espressif/esp-csi