About: This document contains summary of a real-time anomaly detection system developed by me (and team) for Anadarko's crude oil extraction facilities during my time at Quantiphi.
- Our client has facilities of 2 types. For both, we want to have models that can do anomaly detection in real time.
- For type I we have some samples marked as anomalies, so we can either train a supervised ML model (with the challenge of an extreme bias dataset) or we can use the model trained for type II dataset but with some finetuning using the annotation. We go with the 2nd route.
- For type II we do not have any marked samples and we have to train a model in unsupervised setting
- We want high recall with decent precision since false negatives incur higher costs (replacing entire components) whereas false positives incur marginal costs (extra component checks)