On-edge acoustic anomaly detection compared on ESP32-S3 and Nicla Vision
Original title: On-Edge Acoustic Continual Anomaly Detection: A Cross-Platform Comparative Study
A shared ShallowNet TFLM encoder scores 79.41% to 95% across five DCASE machine types on both boards.
The paper compares continual acoustic anomaly detection on an ESP32-S3 with an external I2S microphone and an Arduino Nicla Vision with an onboard PDM microphone. Both run the same TensorFlow Lite for Microcontrollers encoder, a lightweight ShallowNet that maps spectrograms to 32-dimensional embeddings and updates prototypes with an exponential moving average. Firebase and MQTT handle remote collection and retraining. Under a speaker-replay protocol with DCASE Task 2 sounds, ESP32-S3 accuracy is 82.35% to 95.00% and Nicla Vision is 79.41% to 95.00% across five machine types.
- Both boards use identical on-device prototype learning.
- ESP32-S3 accuracy spans 82.35% to 95.00% across five machine types.
- Nicla Vision accuracy spans 79.41% to 95.00% on the same protocol.
- Evaluation uses controlled speaker replay of standardized DCASE Task 2 sounds.
From the original article
1 min read at arxiv.orgAcoustic Anomaly Detection (AAD) serves as a non-contact approach for machine condition monitoring, yet much of the existing research is evaluated offline and rarely addresses embedded hardware constraints or microphone variations under operational drift.
By Rami Zewail · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.