Alpha Version of Anomaly Detection AI Reaches 92% Accuracy on Test Data

We are thrilled to announce a major technical milestone for the AIZoOM project. Our software development team has successfully trained the alpha version of our Deep Learning anomaly detection model.

Model Architecture

The model utilizes a Convolutional Autoencoder architecture, processing Mel-spectrograms generated from raw audio. To overcome the initial lack of "sick" animal audio, the model was trained exclusively on open-source datasets of healthy rodent vocalizations (primarily from the Sciuridae family) using a self-supervised learning approach.

Simulation Results

To test the model, we programmatically introduced "silence anomalies" and artificial noise profiles into the healthy audio streams to simulate a sudden colony die-off. The model successfully detected 92% of the simulated epizootic events with a false-positive rate of only 4.5%.

Looking forward: The team is now integrating Spatio-Temporal Attention mechanisms to help the model ignore background wind and insect noise, which will be critical when processing the live data arriving from our field sensors later this spring.