How to Fix Inference Timeout Errors with Edge Impulse Model on Arduino Nano 33 BLE Sense?

Hello everyone, I’m currently developing a vibration anomaly detection system using an Arduino Nano 33 BLE Sense. The machine learning model was trained with Edge Impulse, and it worked well in the testing phase. However, I’m facing an issue when running the model on the Arduino. After deploying the model, I receive the following error on the serial monitor:
Edge Impulse model failed - Inference Timeout on Layer 5: Model exceeded allocated inference time
Edge Impulse model failed - Inference Timeout on Layer 5: Model exceeded allocated inference time
The error appears consistently during inference, specifically at Layer 5, which leads me to believe it could be related to the complexity of the model or the processing limitations of the Arduino Nano 33 BLE Sense. My qst is - Could this be caused by the model being too computationally intensive for the microcontroller? If so, what steps can I take to simplify the model without losing accuracy? - Are there specific configurations or adjustments in the Edge Impulse export settings that might help resolve this timeout issue? Thanks in advance!
Solution:
@wafa_ath I have no experience with Edge Impulse but maybe you can try to reduce your input size of your sample. Are you working with raw data as an input or do you use features? In both cases you can try to reduce the sampling rate and/or the time interval. This can have an effect on the accuracy but you need to find an optimum.
Jump to solution
2 Replies
Solution
anniekeerdekens
anniekeerdekens3mo ago
@wafa_ath I have no experience with Edge Impulse but maybe you can try to reduce your input size of your sample. Are you working with raw data as an input or do you use features? In both cases you can try to reduce the sampling rate and/or the time interval. This can have an effect on the accuracy but you need to find an optimum.
wafa_ath
wafa_ath3mo ago
Thanks for the suggestion! I'm currently using raw data. I’ll try reducing the sampling rate and time interval to see if that helps. If I switch to feature extraction, are there specific features you’d recommend that are less demanding but still effective for anomaly detection?
Want results from more Discord servers?
Add your server