I'm working on writing a regression test for a hardware device (based on FPGA) used for integrating signals. The goal of the test is to verify the integrator, and I can interface with the device through software for eg. Python to acquire the integrated result.
This device can also generate and read signals; so for this test, I generate a known signal and feed it back into the device, the integrator will then integrate this signal using weights which are also known to me.
So my test program simply simulates this generated signal and the weights, and then integrates (summing) them together to get the "expected" value. I compare this with the integrated value read from the device by calculating the error difference. However, when the integrated values tend are very small in the order of 2-3 decimal points or more, the error differences tend to be very high.
Now, this can be because the software simulated signal is perfect, in contrast, the signal acquired by the device will have noise added to it which is a potential cause affecting the integrated values at lower precision. I'd like to think not about the error due to quantization, bits, the architecture of the FPGA, but I want to try to use something other than a simple error difference check along the statistical route. I was thinking along the lines of adopting something like the t-test method or co-variance to best model the acceptance criteria of this test, and I was hoping if someone could advise how to do so.
Thanks! I'm still a novice at regression tests and statistics and would like to learn more about it so excuse if some terminologies don't make sense.