My friend and I wrote an algorithm that uses linear algebra to reliably diagnose 3D print fails, before they get bad. The algorithm and data are all open and linked.
Our most successful algorithm correctly detected all prints and failures that we've tested so far. It uses a Fourier transform of accelerometer data, which means that it converts from a raw vector acceleration signal to the power levels carried at different frequencies of oscillation. Ultimately, we look at the overall trend in the frequencies of oscillation to calculate whether the print is beginning to fail.
We designed and analyzed three techniques for testing, and ultimately chose a regression on the frequency domain of accelerometer data. A commodity cellphone accelerometer was more than accurate enough, sampling at <100Hz.
Our most successful algorithm correctly detected all prints and failures that we've tested so far. It uses a Fourier transform of accelerometer data, which means that it converts from a raw vector acceleration signal to the power levels carried at different frequencies of oscillation. Ultimately, we look at the overall trend in the frequencies of oscillation to calculate whether the print is beginning to fail.
We designed and analyzed three techniques for testing, and ultimately chose a regression on the frequency domain of accelerometer data. A commodity cellphone accelerometer was more than accurate enough, sampling at <100Hz.