Frugal Heart Rate Correction Method for Scalable Health and Safety Monitoring in Construction Sites

Figures

• 

  Figure 1. The PineTime smartwatch measuring the user’s heart rate.

• 

  Figure 2. Test volunteer walking on a treadmill.

• 

  Figure 3. Test volunteer during upper lift exercises.

• 

  Figure 4. A fragment of the collected heart rate and acceleration data (study participant 1).

• 

  Figure 5. (a) Hidden layers size tuning results. (b) Heart rate and acceleration windows size tuning results.

• 

  Figure 6. The final neural network selected during neural architecture tuning.

• 

  Figure 7. Overall hardware architecture of the solution. The network connections are represented in a simplified manner for clarity. The number of devices varies depending on the specific deployment.

• 

  Figure 8. Overall software architecture of the solution in the context of the use case.

• 

  Figure 9. Histograms for train and validation MAE for models explored during the tuning phase. Final model’s results are indicated with dashed vertical lines.

• 

  Figure 10. The errors of the raw measurement, the moving average, and the proposed method, in relation to the ground truth, on the test dataset. The boxes show the quartiles of the error distribution. The whiskers show the extent of the distribution, except the outliers, indicated with diamonds.

• 

  Figure 11. Predictions of the model compared with the ground truth (data from the Equivital device) and the raw measurements from the smartwatch’s sensor. Data from the validation and test datasets was used for this example (study participant 4).

• 

  Figure 12. The frugality curve of the final model.