neuro-sat

This is an application for simulating the rotation speed controll of any spacecraft either through PID or DNN controller using reaction wheels. The application is written in C++ and shipped with a QT graphical user interface:

Features:

Simulate any 3D spacecraft given its inertia and reaction wheels
Generate sample data, train deep neural networks and save / load them for later use
Add simple precession-disturbance to see if your controller can handle that
Compatible with Windows, MacOS and Linux
Free, light-weight and not heavy in computation power

The application is built on top of the Eigen math library and the MiniDNN library for deep neural networks.
However, you wont have to install those header-only libraries for yourself since they're shipped together with the neuro-sat application.

Installation

The installation requires you to have qmake installed.
You can either have a version of QT-creator installed, or get qmake independantly.

Once you got that, installation is rather easy. In your neuro-sat folder:

Linux and MacOS

mkdir build
cd build/
qmake ..
make

Problems on Linux

If you run into
QT: nothing to be done for ''
try

sudo apt-get install qt5-default

Problems on MacOS

You might find youself not to be able to perform make.
If that's the case, try installing XCode (from the appstore) & XCode Command Line Tools, as described in this post.

Windows

Install QT-creator (free & open source)
Import the project using the SatSimApp.pro file
Specify an output path and create the .exe with building the project

First steps

Important note: You can not expect a beautifull behaivior for every setup of satellite + wheels + controller.
You will most likely run into a situation where some parameters have to be tuned in order for the controller to run perfectly smooth - or you will encounter a situation where the setup of satellite + wheels will not allow the controller to produce good results at all.

For example: Speed saturation.
You might have a massive satellite but only give it tiny wheels. Probably this setup will not be able to spin any faster than 0.001 radians per second, so you might run into some strange behaivior if trying to speed up to more than that.

Your first DNN Satellite attitude controller

In the menubar you can find some predefined Wheel Samples.

Start off with the 4 Wheels setup, which is a pyramid-like arrangement of the RSI68-170/60 reaction wheel
Construct a Cubesat with mass 400kg and length 1m
Select Neural network, then Train a new one
In the Sample generator you have to create some training samples for the DNN to learn from. Create 1000 samples and save them wherever you like
The next step may vary and is highly dependable on which setup you're using. You may find youself playing around with those values to get a good result. For the current setup, batchsize 50, 5000 epoches, 1 hidden layer, 50 neurons per hidden layer, RMSProp optimizer and 0.0007 learning rate works perfectly well
Start the training process and wait. The more power your machine has, the fewer time you will have to wait
Now you might want to save your trained network, so you might load it later
Add a new command to look at the 3D step-response that your network is able to perform. Keep speed saturation in mind! For this setup, speed values between -0.04 to 0.04 rad/s should work fine
Start the simulation. Congrats! You (probably) just trained your first spacecraft attitude-controlling neural network

Keep in mind

If you find yourself in that situation where you get great oszilations (or similar unexcpected behaivior) and can't get the controller to work properly, then most likely the satellite+wheel setup you're using is just poorly chosen.
Adjusting the control setup is all about fine-tuning. A PID controller might perform superbad with some values but without any uncertainty with other ones. Also, getting a DNN to work just as good (or even better) than a PID controller is not an easy task. Keep your eyes open for overfitting or too small sample sizes and fiddle around with the other parameters until you find a good solution.

About the author

This application is part of my bachelors thesis about learning attitude control and simulation for spacecrafts.
If you want to find out more about how this simulator works (or how to get a neural network to replace a PID controller in a closed loop), get involved into the project or just ask something else:
Feel free to create an issue!

fallow24/neuro-sat