predictive maintenance pipeline

industrial equipment failures are expensive precisely because they’re rare — which also makes them a bad fit for standard supervised classification. there usually aren’t enough labelled failure examples to train a reliable “will this fail” classifier, so i went with an anomaly-detection framing instead: model what normal sensor behaviour looks like, and flag deviations from it.

the core is an autoencoder trained on time-series sensor telemetry from equipment operating normally — reconstruction error becomes the anomaly signal, with feature engineering on the raw time-series (rolling stats, frequency-domain features) doing a lot of the work before the model ever sees the data. airflow orchestrates the pipeline end to end — ingestion, feature computation, training, scoring — and mlflow tracks every experiment run and model version, which mattered once i started iterating on feature sets and needed to compare runs honestly instead of by memory.

the real lesson here was that the ml model was the easy part; getting the orchestration and versioning right so results were actually reproducible run to run took longer than the modeling itself.