fitnet Neural Network Temperature Compensation of Pressure Sensor Output

Project info

• Field of studyInstrumentation and Sensors
• LevelDoctoral dissertation (S3)
• Method typeMachine Learning
• Estimated build time± 7 working days
• Toolsfitnet, train, perform (Deep Learning Toolbox)
• PriceUS$38 - US$238

Research goal

Predict compensated pressure from raw sensor output and sensor temperature using Neural Network (fitnet) and assess model performance.

Variables

Example result

• 0.98Test R
• 0.184Test RMSE
• 0.016Best val. MSE
• 14Epochs
• 14Hidden neurons
• 300Samples

A neural network with 14 hidden neurons predicts Compensated pressure from Raw sensor output and Sensor temperature with a test correlation of R=0.980 and RMSE 0.184 after 14 epochs; the best validation MSE is 0.0155.

Computed from illustrative synthetic data, not from a real study.

About Neural Network (fitnet)

fitnet builds a feedforward neural network for regression or nonlinear function approximation. It suits modelling complex input-output relations when the equation form is unknown and enough data are available.

Steps

1. 1Prepare data with features as rows and samples as columns
2. 2Create the network with fitnet and set the hidden neuron count
3. 3Set train, validation, and test division and the training algorithm
4. 4Train with train and monitor performance
5. 5Evaluate RMSE and R on the test data
6. 6Plot predictions against actuals and training history

Assumptions to check

• There are enough data relative to the number of weights
• Features are normalised (fitnet does this automatically)
• Training is stochastic so results depend on the seed and initialisation

Reading and reporting the results

Report the architecture, algorithm, data division, and test RMSE and R. Compare with linear regression to show the need for a nonlinear model.

Example MATLAB code

A short starting script with this project's variable names. The delivered version is adapted to your data and fully commented.

rng(10);
n = 500;
raw_sensor_output = [10*rand(1, n); 5*rand(1, n)];            % features x samples
compensated_pressure = sin(raw_sensor_output(1, :)) .* raw_sensor_output(2, :) + 0.2*randn(1, n);

net = fitnet(10, 'trainlm');
net.divideParam.trainRatio = 0.7;
net.divideParam.valRatio = 0.15;
net.divideParam.testRatio = 0.15;
net.trainParam.showWindow = false;
[net, tr] = train(net, raw_sensor_output, compensated_pressure);

pred = net(raw_sensor_output);
te = tr.testInd;
rmse = sqrt(mean((pred(te) - compensated_pressure(te)).^2));
R = corrcoef(pred(te), compensated_pressure(te));
fprintf('Test RMSE = %.3f, R = %.3f, epochs = %d\n', rmse, R(1, 2), tr.num_epochs);
b = [ones(numel(te), 1) raw_sensor_output(:, te)'] \ compensated_pressure(te)';
lin = [ones(numel(te), 1) raw_sensor_output(:, te)'] * b;
fprintf('Linear baseline RMSE = %.3f\n', sqrt(mean((lin - compensated_pressure(te)').^2)));

figure;
subplot(1,2,1); plot(compensated_pressure(te), pred(te), '.'); hold on; refline(1, 0);
xlabel('Actual'); ylabel('Predicted');
subplot(1,2,2); semilogy(tr.perf); hold on; semilogy(tr.vperf); semilogy(tr.tperf);
legend('Train', 'Validation', 'Test'); xlabel('Epoch'); ylabel('MSE');

Packages and prices

Choose how much help you need. Each higher package includes everything in the lower ones. Prices exclude 11% VAT.

• Payment: 50% down payment before work starts, balance after the deliverables are handed over
• Tax: Prices exclude 11% VAT
• Data and academic integrity: Results are computed from your own research data; you write and stand behind your manuscript, we help with the analysis and your understanding of it
• Revisions: Up to 2 revisions are included in the price (Package 3: 3 rounds)

Frequently asked questions