Sequence oder Featureinput layer ?

Hello dear Community,

i am currently trying to use a Feed-Forward Neural Network for a regression problem.

The input data i want to feed in are "rotational speed", "torque" and "time".

The target data is the temperature of an electrical drive.

And then i will have different operating points dependend on speed:

2000 rpm, 4000 rpm ....

The torque and time however stays the same for each point.

Right now, i am using a sequenceInputLayer and arranged the Data in 4 cells, where each cell contains 3 columns of data. One cell represents one OP.

The results are not good, and my gut tells me, that i might have a misunderstanding the way i treat the input data.

I am using MATLAB r24a and am training the network with the "trainnet"-function.

If you have any Suggestions to do this in a different way, go ahead.

Thank you very much in advance.

Answer 1

[Abhas]

Hi Tobias,

Instead of using a "sequenceInputLayer", you can use a standard "feedforwardnet" in MATLAB. You can follow the below steps to restructure your approach:

• Combine your input data ("rotational speed", torque, time) into a single matrix where each row represents a different operating point and ensure your target data (temperature) is in a corresponding vector.
• Use a standard feedforward neural network with input and output layers designed for regression and use the "train" function instead of "trainnet".

Here's is a sample example MATLAB code to demonstrate the same:

% Example Data
% Assuming you have 10 operating points
rotational_speed = [2000, 4000, 6000, 8000, 10000, 12000, 14000, 16000, 18000, 20000]; % Your actual data here
torque = repmat(50, 1, 10); % Example constant torque
time = repmat(10, 1, 10); % Example constant time
temperature = [70, 75, 80, 85, 90, 95, 100, 105, 110, 115]; % Your target temperature data

% Combine inputs into a matrix (each row is a feature, each column is a sample)
inputs = [rotational_speed; torque; time];

% Target data
targets = temperature;

% Create a Feed-Forward Neural Network
hiddenLayerSize = 10; % Adjust as needed
net = feedforwardnet(hiddenLayerSize);

% Train the Network
[net, tr] = train(net, inputs, targets);

% View the Network
view(net);

% Test the Network
outputs = net(inputs);
errors = gsubtract(targets, outputs);
performance = perform(net, targets, outputs);

% Display Results
disp('Performance:');
disp(performance);

You may refer to the attached output images and also the following MathWorks documentation links to have a better understanding on "feedforwardnet" and "train" functions:

1. feedforwardnet: https://www.mathworks.com/help/deeplearning/ref/feedforwardnet.html
2. train: https://www.mathworks.com/help/deeplearning/ref/train.html

[tobias]

Thank you for the fast reply.

Unfortunately, i can not use the feedforwardnet function, since i do have a custom layer in my architecture.

[Abhas]

In that case, "dlnetwork" might be helpful for you. You may refer to the below documentation links and try to incorporate them:

1. https://www.mathworks.com/help/deeplearning/ref/dlnetwork.html
2. https://www.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layers.html