which is the best network to predict vibration response( output) against time varying force (input) ?

will this kind of network work for it. in the attachement coulumn 2 is f(t) input and column 3 is x(t) i.e output? can anyone cross check what i have to modify to get the results correctly.

% Load and preprocess data

inputData = Predictors_train; % Time series force input data (100 values)

outputData = Responsers_train; % Corresponding time series displacement data (100 values)

% Split data into training and validation sets

cvp = 0.6; % 60% for training

idxTrain = 1:round(cvp*length(inputData));

idxVal = (round(cvp*length(inputData))+1):length(inputData);

XTrain = inputData(idxTrain, :);

YTrain = outputData(idxTrain, :);

XVal = inputData(idxVal, :);

YVal = outputData(idxVal, :);

% Define the number of features and responses

numFeatures = size(XTrain, 2); % Number of features in input data

numResponses = size(YTrain, 2); % Number of responses

% Define LSTM network architecture

layers = [ ...

sequenceInputLayer(numFeatures)

lstmLayer(12, 'OutputMode', 'sequence')

dropoutLayer(0.2)

% lstmLayer(32, 'OutputMode', 'sequence') % 'sequence' for sequence-to-sequence

% dropoutLayer(0.2)

fullyConnectedLayer(numResponses)

regressionLayer];

% Specify training options

maxEpochs = 10;

miniBatchSize = 16;

options = trainingOptions('adam', ...

'MaxEpochs', maxEpochs, ...

'MiniBatchSize', miniBatchSize, ...

'Shuffle', 'every-epoch', ...

'Plots', 'training-progress', ...

'Verbose', false);

% Convert training data to cell arrays

XTrain = num2cell(XTrain, 2); % Convert each sequence to a separate cell

YTrain = num2cell(YTrain, 2); % Convert each sequence to a separate cell

% Train LSTM network

net = trainNetwork(XTrain, YTrain, layers, options);

% Validate trained network

YPredVal = predict(net, num2cell(XVal, 2)); % Convert XVal to cell array

valRMSE = sqrt(mean((YVal - cell2mat(YPredVal)).^2)); % Convert YPredVal back to matrix

disp(['Validation RMSE: ', num2str(valRMSE)]);

figure;

plot(cell2mat(YPredVal));

new_input_data = Predictors_test;

% Make predictions on new data

XNew = num2cell(new_input_data, 2); % Convert new input data to cell array

YPred = predict(net, XNew); % Predict using the trained network

Ypred_mat=cell2mat(YPred);

Y_actual = Responsers_test;

figure;

plot(time_series,Ypred_mat,'r','LineWidth',1.2);

hold on;

plot( time_series, Y_actual, 'b','LineWidth',1);

Answer 1

[Shivansh]

Hi Thushar!

It looks like you have set up an LSTM network for predicting vibration response to time-varying forces.

The LSTM networks are indeed a good option for your problem due to their ability to capture temporal dependencies.

You can follow some common practices to find the issue with your current model and getting better results from them.

You can try to normalize your data for consistent scale, experiment with model complexity (adding or removing layers and neurons), adjust hyperparameters like epochs and batch size for better training, or use cross-validation for robust evaluation.

You can also explore different metrics such as MAE or MAPE for a comprehensive performance assessment.

The comparison of LSTM with other sequence models like GRUs or Transformers could also reveal a more effective solution for your specific problem.

You can refer to the following link to learn more about LSTMs in MATLAB: https://www.mathworks.com/discovery/lstm.html.

I hope it helps!

[Thushar]

the link is not working though

[Shivansh]

I don't know the issue with the previous link.

You can refer to this link: https://www.mathworks.com/discovery/lstm.html.