*4.2. Intention Detection*

The reliability of the sensor–device synergy is assessed by the trigger rate, Kernel Density Estimation (KDE), and the confusion matrix. The trigger rate is defined by assessing the data in each data set. For a data set correlated with action 6 (grasp a cylinder with a radius of 38 mm), all five columns of data need to be considered. If all pressure sensor readings exceed 0.2 N and all angle variations exceed 2◦ , a successful trigger is concluded. Assessing all 10,000 data sets, a 100% trigger rate is observed for each gesture/action (Figure 16a). Moreover, good training and estimation are more likely to be achieved based on similar testing and training data set patterns. Therefore, the KDE method is applied to illustrate the probability density distribution of a random training and testing data set. As can be seen, the two distribution patterns agree well with each other (Figure 16b). The difference in distribution patterns is mainly due to the desired diversity of data sets (i.e., random motion of the arm, shaking of hands, different forces applied to the pressure sensor, different finger-bending angles, and digit length of volunteers). In addition, the high prediction accuracy for both CNN and SVM also suggests a reliable sensor–machine synergy and a good data acquisition process.

**Figure 15.** Exoskeleton trajectory validation for DIP and PIP joints. (**a**) Opensim simulation setup; (**b**) trajectory of DIP joints; (**c**) trajectory of PIP joints. –

– **Figure 16.** Reliability of the sensor-device synergy. (**a**) Trigger rate of all sensors necessary for gestures/actions; (**b**) probability density distribution for random training and testing data sets using KDE.

In the data-acquisition phase, each repeat is performed slowly, and 200 data points are collected, utilizing a low data collection frequency of 40 Hz. The data collection frequency of the system can be adjusted to a much higher level, which improves the overall system response time significantly. The low data collection frequency adopted in this study aims to guarantee data set diversity for training purposes. After thousands of repeats, the joint rotation in fingers cannot be controlled precisely with a high data collection frequency (due to the fatigue of the human hand), which may jeopardize the diversity of the data set. In a real rehabilitation scenario, the user may perform actions at a different pace; however, more training data sets with high diversity may help with intention-recognition accuracy.
