NARXNET with input delay

Hi,

I am solving a problem incorporating measurement delay in a sensor. I have a set of measured acceleration based on the sensor reading (in time series). However, the measurements that I obtained have some time delay from the 'true' value of it. I want to design a neural network that can 'predict' the true acceleration value using the measured value that I obtained from the sensor measurement.

I am using NARXNET to model the neural network of it. To accommodate the delay in the input, 'inputDelays' is adjusted according to the delay in the sensor measurement (and, technically, no delay is needed for the 'feedbackDelays'). Am I doing it correctly?

Thank you,

Ghazi

Answer 1

[Greg Heath]

No.

TIMEDELAYNET is used to predict an output series, y, using an exogeneous external input series x.

NARXNET is used to predict an output series, y, using BOTH an exogeneous external input series x, along with feedback signals from the previous estimated values of y.

HOWEVER, you do not have an exogeneous external input series!

Therefore, you should use NARNET. First read the documentation

 help narnet
 doc narnet

Then search for some of my posts in BOTH the NEWSGROUP AND ANSWERS using

 narnet greg

and

 narnet greg tutorial

Begin by finding the statistically significant nonzero lags at the highest local peaks of the measurement autocorrelation function. Search

 nncorr narnet

Then use as small a continuous subset of lags as needed for your predictions.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Ghazi Binarandi]

Thank you, Greg! You are always helpful.

[Ghazi Binarandi]

Hi Greg,

Now I think that NARNET may not be suitable for my problem. (Please correct me if I'm wrong!) Let me repeat the problem so I can explain it in a better way:

In my neural network, - my inputs are going to be 'measured accelerations' from the sensor reading (which is delayed from the 'true' value) - my targets are going to be the 'true accelerations'

So, I think I may use TIMEDELAYNET for this particular problem since I have delay in my input. Is the way of my thinking right?

Thanks,

Ghazi

[Greg Heath]

Hmmm, you are trying to predict the past !!!

I think you are correct in using timedelaynet.

Is there a constant delay between the true signal and the measurement?

Greg

[Ghazi Binarandi]

Yes, I am trying to predict the past! I have tried using timedelaynet and got a pretty nice result.

The delay can be approximated as a constant in the modeling, although it is not a constant in the real problem.

Thanks so much for the help!

[Greg Heath]

How do you obtain the true values you use for design targets?

[Ghazi Binarandi]

I have another type of sensor which has better performance in terms of delay. So, I could model the measurements from this sensor as the true/ideal values (these are going to be my target). Then, by training it with neural nets, I could use the sensors with delay on it, but still obtain a performance that is close to the better sensor.