RCA on top of existing anomalies

Question

When we work with timeseries data containing multiple features eg. a sensor data. we can detect anomalies using cluster, supervised and semi supervised based approaches (Eg. Isolation, Autoencoder Etc), but those approaches only detect anomalies and not their root causes. Example certain event which has occurred at time 1, might cause ripple effect which can in turn lead to anomalies at time 2.

How can we perform such operations, in general how can we extend our anomalies models to detect root causes or are their any algorithms available given anomalies can pip point possible root causes.

Answer 1

Root Cause Analysis can be approached with a ML classifier trained with supervised learning, potentially using specialized methods for causal modelling.

The primary challenge in developing such a model tends to be access to a sufficiently large and diverse dataset. One basically needs a dataset that contains many examples of each and every anomaly of interest, annotated with their root cause. It also must cover other sources of variance in the data, like background noise, sensor characteristics, machine characteristics, operation conditions, et.c. Depending on the exact problem, it may require 10-10'000 examples per failure mode.

When such a dataset exits, it can become a straightforward supervised learning problem. If such a dataset does not exist, then the effort should go into the processes that will build the dataset.