Identification of NARX problem with death time between Inputs and Outputs

Dear All

I would like to identify a modeling problem and set up a right kind of (dynamic) neural netowrk.

We have an experimental MIMO combustion system with delayed time response between fuel flow (ANN input, x) and a water temperature (ANN output, y), i.e. when we change fuel input the response in the temperature is noticed 10s later. Sampling period is 1s so in terms of data the response starts to appear exactly after 10samples.

At time t we have both the fuel flow x(t) and the temperature y(t). If we would like to make e.g. NARX model that considers the delay in response (10-samples death time) and at the same time consideres history of 5samples, would the following NARX set-up be suitable/correct?

% inputs & outputs

     X = [x y]; % since at time t we have both x and y    
     T = y;     

% narxnet set-up 

     net = narxnet(9:13,1:5,e.g. 10); % In inputs: (0:8)-death time,(9:13)-history; In outputs: (1:5)-history of feedback

% prepare data

     [Xs,Xi,Ai,Ts] = preparets(net,X,{},T);

% train

     net = train(net,Xs,Ts,Xi,Ai);

Please give suggestions and comments!

Thank you kindly!

Answer 1

[Greg Heath]

% inputs & outputs

     X = [x y]; % since at time t we have both x and y

     T = y;     % narxnet set-up

1. NO. ONLY x

    net = narxnet(9:13,1:5,e.g. 10); % In inputs: (0:8)-death time,(9:13)-history; In outputs: (1:5)-history of feedback

2. NO.

NET = NARXNET(0:9,1:5,10); & USE LOWER CASE

Hope this helps.

Greg

Answer 2

[Rajiv Singh]

Have you looked at nlarx, idnlarx functions in system identification toolbox? The input argument "nk" denotes delay from input to output.

Answer 3

[Greg Heath]

The code looks It looks reasonable if your assumption of the effective signal length is correct.

Hope this helps.

Greg