There needs to be an indication that the mechanism of each AoA sensor is reliable, independent of any other sensor reading.
For example, one could use electromagnetic coils to modulate a small variation of torque on the AoA vane to see if it moves as expected. If + and - torque yields no + and - change in the AoA sensor reading then the sensor is flagged as unreliable. Subsystem testing could continue to be used on the off chance the sensor frees up, and then return it to the voting pool. This could also be used on the ground before departure and runs less risk than using the entire plane as a roller coaster to check the sensors or hoping for a situation that would change the AoA a meaningful amount over digitization noise.
It does raise the question - since all three sensors aren't identical in reading, what does the flight control software do with the readings? Does it average them, take the two that are closest and average them, take the middle one? What is the fallback if one is eliminated? Average or flip a coin and pick one? They are within some margin of each other, but then what if one more is eliminated? It cannot be disagreement on reading, right?