Hello, I am a master's student studying in Yonsei Univeristy, Korea.
I am trying to estimate the state of satellite, using Neural Network.
Below is a simple flow of my study.
1. Train (t0 ~ t1)
Train neural network using known observation & true state data
2. Validation (t1 ~ t2)
Using observation data starting from t1, validate the network
3. Test (t3 ~ t4)
With new observation data, estimate the true ECI coordinate at different time.
[For all steps]
Input : observation data ( RADAR SEZ coordinate data or Orbital Element data )
Output : true data ( ECI coordinate data)
I know that the validation is already done while training,
but the validation part is for checking whether the network is well-trained.
I used "narxnet" from the deep learning toolbox, and it worked well until the validation part.
However, in order to use the network made with "narxnet" for the test part,
I had to retrain using data from just before.
(to estimate t3~t4, need tx ~ t3 data trained network)
So all my work have failed, and I am going to restart on doing this.
Here is what I want to ask.
Thank you very much for reading my questions.
Jee Hoon, Kim.
Hi Kim,
I am not an expert in your applicaition domain.
This being said, I can answer your questions in a general context as:
I found that most of codes in MATLAB related to neural network is for image training. Is it better to use other program for this type of work? (e.g. Python, Tensorflow...)
>> I suggest you do not use MATLAB.
In my experience best combination is: Ubuntu + TensorFlow + Python
I found that is it better to use recurrent neural network, and time series input. Is MATLAB "train" code available for this?
>> Never looked into this, so I can't give you a definitive answer. However, recurrent NNs are not that different than regular ones. The underlying (temporally pipelined) architectures are the same anyway. So, this should be possible but I personally find MATLAB too limiting.
I cannot find much information on the documentations. I would like to know if there is good example I can refer to.
>> Once you switch to TensorFlow + Python combo you will find your chances of finding examples and code snippets to be much improved.
Finally, if you intend to keep working in this direction, you should invest in learning Python. You will find it much rewarding.
Thank you for your answer. I have been trying to find my way out using MATLAB. But same as you have said, MATLAB is too limiting. It would be much better to use Python and Tensorflow. Thank you very much.
Dear Kim Jee Hoon ,
Look the link, please.
Conference Paper Satellite orbits guidance using state space neural network
Regards,
Shafagat
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training