Hi everyone,
I’m facing the same issue and need some help (I’m new to ML, so please explain steps simply).
My setup
- Board: STM32 NUCLEO-G491RE (Cortex-M4F)
- Project: Fan Anomaly Detector (using Edge Impulse)
- Dataset: ~1000 normal audio files (train) + 100 normal and 100 anomaly (test)
- Total training data: 2h 46m 40s, about 19 000 windows
Impulse settings
Window size: 1000 ms
Window increase: 500 ms
Frequency: 16000 Hz
Pipeline: Audio → Audio (MFE) → Anomaly Detection (K-Means)
MFE params: frame length 0.02 s, frame stride 0.01 s, filters 40, FFT 256, high freq 8000, noise floor –52.
Problem
After clicking Generate features, everything finishes fine (19 000 windows created).
But in Feature explorer I only see dots for each audio file (hover shows file names).
No MFE_1…MFE_40 axes appear anywhere.
In Anomaly Detection, it shows audio feature 0–3959 instead, and training gives the error
“Requires at least two axes.”
Tried so far
- Deleted and re-added both MFE and Anomaly blocks
- Re-saved impulse and regenerated features
- Confirmed all parameters and window sizes
Still the same issue.
What I think
Maybe the Anomaly block is still linked to raw audio instead of MFE output, or this is a known bug where feature metadata doesn’t refresh.
Please help
Can someone explain (step-by-step if possible):
- How to make the Anomaly block read the MFE_# features instead of “audio feature #”?
- Any way to force Edge Impulse to refresh / rebuild the feature metadata?
- Or how to export the MFE features to check them manually?
Thanks a lot!
(Again, I’m new to ML — a simple explanation would really help.)