After my first blog post “Topic Modeling” , I would also like to shed light on another extremely exciting topic: “Anomaly Detection”. which was another highlight of the “Practical conference about ML, AI & Deep Learning applications” – Machine Learning Prague 2019 , which we attended from February 22nd – 24th, 2019. I would like to briefly explain what Anomaly Detection is and then give you some examples! Here we go: 

Using Machine Learning for Anomaly Detection 

Anomaly Detection is a topic that I think is becoming more widely used.  There was a very interesting presentation by Vítězslav Vít Vlček from Broadcom called “Data-driven System health determination in Monitoring Softwares for Operational Intelligence”.

The general idea is to use past data to decide if new data are conforming or anomalous.

Here are a few examples where Anomaly Detection would be useful:

Anomaly detection for Systems protection 

One common use of Anomaly Detection is to protect systems (server infrastructures, networks, etc.).  This was the example on which Vítězslav based his presentation.  His program uses a pattern matching algorithm inspired by the wave function collapse in Quantum Mechanics.

This basically means, that his matching algorithm compares chunks of past data to new data to determine if it fits – somewhat like a puzzle.

When a hacker attempts to enter the system, or it becomes unhealthy for some reason, the program detects an anomaly.

The subsequent actions then depend on the needs of the customer, it could alert a systems administrator or take more drastic action such as logging out all sever users. Very useful!

Anomaly detection to identify fraud 

Another important application of Anomaly Detection is fraud detection, which is probably used by your bank (at least I really hope it is).

If you have ever received an email or call from your bank after making a strange purchase, it is probably because an AI algorithm decided that it was anomalous.

Such a system collects data from purchases made through the bank and compares new purchases to the past.  Such algorithms are becoming more widely used to protect people and systems from malign behavior.

Predictive Maintenance applications

What the presentation also made me think about is “Predictive Maintenance”.  Here the algorithm records data from a machine to determine if it will break in the near future.  For example, vibrations, mechanical tensions, and the sound it produces are continuously measured.

The past data are used to determine if the behavior of the machine has changed and if maintenance should be undertaken.

Perhaps the methods described by Vítězslav could be used for predictive maintenance as well?  By intervening to fix a mechanical problem early, you can prevent costly down-times that are incurred when a part completely breaks.

We are excited to keep exploring the many possibilities in this field of Machine Learning in the future!


Interested in finding out more or have a chat about Machine Learning and how Anomaly Detection in particular could benefit your business?

We are happy to have a talk about possible solutions that meet your needs.

Let’s talk: