How to mix continuous and discrete categorical signals in LSTM-Autoencoder or other anomaly detection methods?

Solution for How to mix continuous and discrete categorical signals in LSTM-Autoencoder or other anomaly detection methods?
is Given Below:

I want to detect anomalies in a system that includes continuous data as well as some switches and other sources giving categorical input. Something like:

  • sensor signals (speed, angle,etc.)
  • controller outputs (forces etc)
  • input switches with additional logical information (like speed signal 123 has to be zero when switch 1 has state 5, else it is an anomaly, or controller xyz only should be active if switch 2 is ON)

Of course it would be possible to check all those logic-based conditions prior to feeding the continuous signals into an autoencoder. But there are multiple switches with up to 128 conditions each.

Given I have enough training data with normal operation so the network could learn those coherences and dependences, how would such a network look like?

The problem I am seeing is is that in a continuous signal, an input of 49 is not much difference to 48 or 50. But if each value represents a state, state 49 could be totally different from 48 or 50. So the network would have to learn that a whole different behaviour could be right depending on a minor change in the input value of one signal.

I’d appreciate any input from you, also papers or links.