I am working on a problem where a control method, backed by a Neural Network (NN), dictates the movement of a 1D actuator to influence a specific process. This actuator can move linearly within a set range (e.g., 10 units to the left and 10 to the right) at a predetermined speed (e.g., 10 units/second). The current position can be observed whilst the target location of the actuator is set by the control method, meaning that it will move towards that position at the set speed * time interval each iteration.

The primary objective of the control method is to optimize this process, and the actuator's movement is a critical variable (among others) it adjusts to achieve this. I'm interested in introducing various potential faults into the actuator's operation, such as:

  • The actuator getting stuck at a certain position.
  • The preset range might change.
  • The actuator can only move at set intervals.
  • It might get delayed.
  • The movement speed might depend on the current location.

There are a lot more faults that occur (this description is a simplified version). I aim for a method to handle these faults without (or with limited) prior knowledge of what they could be. However, I need to pass some form of information to the main control method so that its NN can handle it effectively.

A potential approach I am considering is using a Recurrent Autoencoder. This would take the current and target positions as inputs and provide the next predicted position as an output. I anticipate that the latent space from this might be an useful input for the main control methods. While this approach might handle unexpected faults (also ones that arise during the process), especially with adequate training data, it feels somewhat unconventional (and has some obvious downsides that retraining this encoder does not hold any consistency for the other network).

Are there any suggestions for methods or resources that could guide me in this challenge?

  • $\begingroup$ Do you want to just detect that a fault has happened, or to optimally compensate for it after it has happened? The actuator probably uses a PID controller, do you have access to its internal signals, e.g. whether it is saturating its actuator (pushing with max force) or how hard it has to push to maintain speed? $\endgroup$
    – maxy
    Commented Oct 12, 2023 at 17:46
  • $\begingroup$ Just the detection that a fault has happened. The only the controller just sets the target position and has no other control. It can observe the current position. $\endgroup$ Commented Oct 12, 2023 at 19:41


You must log in to answer this question.