Hello to everyone,
I’m a computer engineering student of University of Salerno, Italy. I’m studying for my thesis and I’ve read the paper “Real-Time Anomaly Detection for Streaming Analytics” about HTM.
For my research purpose I’d like to develop a system which execute on raspberry pi 3 real time anomaly detection on a continuous data flow from multiple sensors.
I’d like to know if in your community someone has already tried to perform HTM algorithm on an embedded system and, if yes, if he can give me some hints about the project, I’m a newbie in this study field.
I’ve successfully built nupic on a Raspberry Pi 3 and run through a couple of the examples, however I’m also a newbie to the HTM field. If you’re looking for help getting it up and running I might be able to help. However, if you’re looking for information on HTM, I’m definitely not going to be any help.
Hey @Andrea_Giordano, are you talking about writing code that implements HTM theory to run on the Pi? Or are you talking about trying to get an existing implementation of HTM compiled and running on the Pi? You can find HTM implementations here and here.
I have reached this community after read “Real-Time Anomaly Detection for Streaming Analytics”.
For my project I need an implementation of HTM to perform real time anomaly detection on a raspberry on a continuous flow of data from multiple sensors.
I’ve tried HTM studio on my pc and fit almost perfectly my requirements except for the fact I need to analyse a real time flow.
So I’m looking for something like implement HTM studio features in real time ways and performed on a raspberry pi.
Is it a requirement that the HTM system run on the Pi? It is pretty easy to push data off the Pi to a server on the same network.
You’re going to need to decide which HTM implementation to use. NuPIC is developed by Numenta. It is Python / C++. Others I linked in my last post. It should be easier to build HTM.Java on a Pi. Are you very familiar with programming? If so, what languages?
I know Java and C++ but I prefer Java.
It’s preferable HTM system run directly on Pi, but I can consider to run it also in a cluster of commodity hardware.
Let me explain my architecture:
I have some wearable sensors that pull data on rapsberry as they are generated.
Pi (maybe) performs real time anomaly detection, then it forward received sensor data to a Kafka broker. On the other side I have a SparkStreaming cluster which consumes data from Kafka.
Now, if run HTM on Pi is not possible can I consider to perform real time anomaly detection on spark’s cluster? It would be better?
I’ve never used Spark or Kafka, but there is an HTM.Java integration with Apache Flink, which would allow you to build a system where the HTM anomaly detection can run as a part of the stream cluster. @EronWright talked about this here: