Tech Stack
Description
Built a CNN-based autoencoder for the MIMII dataset (malfunctioning industrial machine investigation and inspection), labeled as normal or abnormal. Trained the encoder-decoder only on normal data, so reconstruction loss stays low on normal samples and spikes on abnormal ones.
Preprocessed audio with Torchaudio mel spectrograms before feeding them into the CNN. Reached 0.98 AUC on the fan id_02 subset, close to the 0.99 AUC baseline reported in the original MIMII paper.
Reached 0.98 AUC on fan id_02, versus a 0.40 AUC trivial loudness baseline and a 0.99 AUC research-paper baseline.
W&B: tracked training runs and hyperparameter sweeps, selected the 90th-percentile anomaly threshold.
ONNX: exported the trained model for framework-agnostic inference.
Deployed a Gradio demo to Hugging Face Spaces and a Dockerized FastAPI REST endpoint.
Prometheus + Grafana: instrumented and visualized live inference metrics.