Part 2 ended with two directions: keep the single-electrode setup and try to build control around alpha waves, or change the hardware and move to the motor cortex.
The original target was a 3-state control: forward, left, right. In the previous part I showed that one frontal electrode located on Fp1 failed to find any motor imagery-related activity in the mu band. The signal that I could capture was the broader alpha band activity that was rising with mental effort. Pushing harder on motor imagery decoding would not have helped, because the signal I was looking for was never in the recording in the first place.
Mu and alpha overlap in frequency
The mu rhythm and the alpha rhythm both sit in 8-13 Hz, so a spectrum alone cannot separate them. They differ by location and generator. Mu comes from sensorimotor cortex (around C3 and C4) and desynchronizes during real or imagined movement. Frontal alpha rises with mental effort. An Fp1 electrode sits over frontal cortex, nowhere near the motor strip, so in that band it reads frontal alpha climbing with concentration, not mu.
As a result, I decided to drop motor imagery and try controlling my car with a binary system: drive forward on concentration and stop on relaxation. The controller I built opened with a 10-second calibration stage during which I stayed with my eyes closed and relaxed. The program then set the median alpha band power over that period as a baseline and set the activation threshold at 1.3x of the baseline. Every second, the controller processed a 1-second window (512 samples at 512 Hz) of data, estimated the power spectrum (PSD) with Welch’s method, and measured mean power in the 8-13 Hz band. When above threshold, it sent a signal to go forward, and when below to stop. From the concentration methods I tried - mental arithmetic, counting backwards, and focused attention - all of them raised the signal above the threshold. It ran, and it did occasionally move the car, but it felt slow and unpredictable.
Power Spectral Density (PSD)
The PSD describes how a signal’s power is distributed across frequency. For a 1-second EEG window it says how much power sits at each frequency, so the mean over 8-13 Hz gives the alpha band strength. Welch’s method estimates it by splitting the window into overlapping segments, taking a spectrum for each, and averaging them.
The problem was the signal, not the tuning. The 1-second window I used put a 1-second floor under every decision. Alpha added more delay on top: it did not switch cleanly and instead took a couple of seconds to build up above threshold, and then the same amount of time to decay. As a result, even after I stopped concentrating, the car would continue moving for a couple of seconds before the signal fell back below the threshold. I tried to work around it. Shorter windows (128 samples) with bandpass energy instead of PSD. Hybrid control - jaw clench for movement, alpha for “turbo.” None of it fixed the core issue: alpha was too slow. Fine for meditation apps, but not for real-time control.
Both paths on Fp1 were now dead. Concentration was too slow, and motor imagery was never visible on a frontal electrode. A car that steers has to tell left-hand imagery from right-hand imagery, and one electrode carries no information that separates them, so the hardware had to change.
In the next blog post I will explain how I decided what my future hardware should look like and what methods I could use to analyze my new signal.