Anomaly detection in multi variant/column/input data

I need suggestion from forum on best approach to find Anomalies on multi variant data or multi column data

I have data in the below format
timestamp , metric1, metric2 , metrics3 … metric 10

Currently I am using MultiEncoder approach to detect anomalies . Is it right approach to find anomalies ? or calculate anomaly on individual metric apply Mathematical formulas suggestion the the below paper (section :Supplementary Material)

Thanks & Regards

The problem with using anomaly score/likelihood on one model with many metrics is that you can’t know which of the metric(s) is acting abnormally, just that the multi-metric steam as a whole is. I think the better way to go would be as suggested in the paper you cite. As I understand it they have separate models for each metric (each outputting anomaly scores/likelihoods at all time steps), and if the sum of all models’s anomaly likelihoods over a certain time window exceeds a threshold then an overall anomaly for the whole system is raised.

It’d be interesting to see the differences in anomalies raised between this recommended method and the simpler approach of cramming all metrics into one model.