Do I need to normalise input and target data when training an ANN using fitnet?

I am training an artificial neural network using MATLAB's `fitnet` function with 5 input variables (e.g., compression ratio, HES, SOI, injection duration, RPM) and 10 outputs (e.g., BTE, NOx, BSFC, Torque, CO, etc.).

I understand that `fitnet` automatically normalises input data using `mapminmax`. However, I am not sure whether I also need to manually normalise the **target (output)** data before training.

Do I need to normalise the **outputs** manually, especially when they vary widely in scale (e.g., NOx in ppm, BTE in %, BSFC in kg/kWh)?

Also, what is the correct way to **denormalise predictions** after training?

Any clarification would be appreciated.

Answer 1

[Manish]

Hi,

The 'mapminmax' function scales inputs and targets to fall within the range [–1, 1].

The following code demonstrates how to use it:

[pn,ps] = mapminmax(p);
[tn,ts] = mapminmax(t);
net = train(net,pn,tn);

No manual effort is required for normalization.

If needed, you can reverse the normalization using the 'reverse' option with 'mapminmax'.

Refer to the documentation link for better understanding:

• https://www.mathworks.com/help/deeplearning/ref/mapminmax.html

[Muhammad Usman Saeed]

Thank you so much for your response!

Could you please let me know should I denormalise the predections or it is also denormalised automatially?