Identification of People in a Household Using Ballistocardiography Signals Through Deep Learning

Figures

• 

  Figure 1. Comparison of resistive pressure sensor and piezoelectric sensor signals.

• 

  Figure 2. Resistive pressure sensor (top) and piezoelectric sensor (bottom) signals. (A) denotes the time of pressure increase, while (B) denotes the time of pressure decrease.

• 

  Figure 3. Data acquisition experiment configuration.

• 

  Figure 4. Original ballistocardial signals for M0 to M9 (left) and typical FFT spectrum of training data that correspond to each other (right). Training data are represented by the light blue shaded areas, and the test data used are represented by the yellow shaded areas.

• 

  Figure 5. Overview of the training phases for the neural network. CE: cross-entropy. a–e represent subject IDs.

• 

  Figure 6. Overview of the test phases for the neural network. a–e represent subject IDs.

• 

  Figure 7. The typical loss value variation for each epoch.

• 

  Figure 8. The classification accuracy rate for all 252 patterns. The horizontal axis represents the IDs, arranged in ascending order in terms of their mean classification accuracy across all patterns. The ID with the lowest mean classification accuracy is positioned on the left, whereas the ID with the highest mean accuracy is positioned on the right.

• 

  Figure 9. Average classification probability for 50 sets of test data in regard to each category. In regard to the classification probabilities, cases where the test data were correctly classified into its actual category were counted as positive, while cases misclassified into other categories were counted as negative. The graphs: (a) for IDs M0, M2, M6, M7, and M8 (“02678”) cases (best case); (b) for IDs M0, M1, M4, M6, and M7 (“01467”) cases (second best case); and (c) for IDs M0, M2, M3, M6, M7 (“02367”) cases (third best case).

• 

  Figure 10. Average classification probability for 50 sets of test data in regard to each category. In regard to the classification probabilities, cases where the test data were correctly classified into its actual category were counted as positive, while cases misclassified into other categories were counted as negative. The graphs: (a) for IDs M0, M5, M6, M8, and M9 (“05689”) cases (worst case); (b) for IDs M3, M5, M6, M7, and M9 (“35679”) cases (second worst case); and (c) for IDs M1, M5, M6, M7, M9 (“15679”) cases (third worst case).

• 

  Figure 11. Average classification probability for 50 sets of test data in regard to each category. In regard to the classification probabilities, cases where the test data were correctly classified into its actual category were counted as positive, while cases misclassified into other categories were counted as negative. The graphs: (a) for IDs M0, M1, M4, M5, and M8 (“01458”) cases (average case); (b) for IDs M1, M2, M7, M8, and M9 (“12789”) cases (second average case); and (c) for IDs M1, M3, M5, M6, and M7 (“13567”) cases (third average case).

• 

  Figure 12. Classification probabilities for re-inference (lower graph) and a typical spectrum (upper graph) for ID M0 as part of the “02367” classification case: (a) masking from the low-frequency (LF) side; (b) masking from the high-frequency (HF) side; and (c) masking a single specific peak. The arrow indicates the expansion of the spectral mask.