**1. Introduction**

In the 21st century, the aged population has increased dramatically. Among elders, a considerable number of people suffer from stroke and related complications such as hemiplegia, disability, etc., which lead to problems in daily caring [1]. To restore self-care capabilities, stroke patients usually require a long rehabilitation period after surgery [2,3]. Patients' needs at different rehabilitation stages vary, thus rehabilitation therapy should also be changed accordingly. To address this issue, human–machine interaction (HMI) technology is developed for rehabilitation exoskeletons [4–6]. In brief, all HMI technologies serve three purposes, which are intention capture, intention recognition, and physical response [7].

Capturing exoskeleton user intention traditionally relies on feedback from sensors, such as force transducers [8–10], cameras [11], strain gauges [12], and lasers [13], each of which possesses inadequate sensor–machine synergy in dealing with complex gesture/action and leads to low intention-recognition accuracy. Recently, electromyography

**Citation:** Xia, K.; Chen, X.; Chang, X.; Liu, C.; Guo, L.; Xu, X.; Lv, F.; Wang, Y.; Sun, H.; Zhou, J. Hand Exoskeleton Design and Human–Machine Interaction Strategies for Rehabilitation. *Bioengineering* **2022**, *9*, 682. https://doi.org/10.3390/ bioengineering9110682

Academic Editors: Christina Zong-Hao Ma and Ramana Pidaparti

Received: 21 October 2022 Accepted: 9 November 2022 Published: 11 November 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

(EMG) and electroencephalogram (EEG) have been extensively studied for HMI due to their high intention-detection accuracy potential, which benefits from multiple signal channels [14–16]. However, EMG and EEG usually require huge data manipulation efforts, which lead to a significant delay in real-time control [17]. To balance sensor–machine synergy and real-time control performance, sensor matrices have been developed in many studies. Upon distributing a flexible skin tactile sensor array on the 'Baxter' robotic forearm, the real-time human touching detection accuracy reached 96% [18]. Moreover, by utilizing a piezoelectric force sensor matrix, the gesture recognition accuracy of a 'smart glove' could reach ~98%. In addition to the tactile sensor matrix, Inertial Measurement Units (IMU) are also renowned in wearable devices due to their compact size, high resolution, fast response, low cost, and compatibility with different systems [19]. The synergy of multiple IMUs led to successful applications in gesture recognition [20], dance mimics [21], gait analysis [22], tumble detection [23], daily life activity classifications [24], etc.

Intention recognition is another aspect of HMI, which refers to the prediction of human activities based on sensor output data [25]. In practice, the intention-recognition accuracy is affected by factors such as the resolution of the sensor, the install location of the sensor, the complexity of the gesture/action, and the types of sensors synergized for prediction [19]. In addition to sensor selection and setup, a data processing and intention prediction model is also crucial for intention-recognition accuracy. In recent years, the research on intention prediction has mainly focused on the following approaches: Statistics [26,27], machine learning [10,28], and deep learning [29–32]. Representative statistic approaches such as the least-squares method and the Kalman filtering algorithm possess advantages such as low computational complexity and good real-time control performance. However, to achieve high prediction accuracy, a linear correlation is required between data captured by the sensor and the demanded action trajectory [33]. In other words, the statistical approach is only applicable to simple motion prediction. To address this issue, machine learning and deep learning methods have been extensively studied. Representative machine learning approaches such as the Maximum Entropy Markov Model (MEMM) and the Support Vector Machine (SVM) usually require heavy data pre-processing such as Wavelet Transform (WT) or Principal Component Analysis (PCA) to optimize eigenvalues of the data sets [10,28]. Although the popular SVM model can make reasonable predictions on data sets with non-linear correlations, compared to deep learning methods such as the Convolutional Neural Network (CNN), more computational time is usually required for large sample sizes [34], and the prediction accuracy of SVM is more sample-size-dependent, due to its inferior feature-extracting capability [34].

To assist users in rehabilitation and daily life activities, a reliable mechanical structure design of hand exoskeletons is indispensable. Based on the force transmission mechanism, the hand exoskeleton can be classified as pneumatic [35], cable/tendon-driven [36,37], smart-material-based artificial muscle-driven [38–40], and linkage-driven [41,42] technology. 'Stiff hand' is usually observed in stroke patients, and significant torque force is required to perform successful rehabilitation. Artificial muscles based on smart materials such as dielectric elastomers [39] and electroactive polymers [40] are not applicable as they are usually insufficient in the generation of power, force, and deformation. Due to the compressible and temperature-sensitive nature of gas, the bending angle and bending speed of each finger joint cannot be precisely controlled by a pneumatic 'muscle' [43]. The cable-driven design reduces the weight of the exoskeleton. In practice, the cables and artificial tendons usually experience elastic deformation in operation, which may require constant calibration to avoid misalignment with the rotation center of finger joints [44,45]. Most existing cable-driven designs only drive the fingers through the stretch or bend phase, and the complete bend–stretch process cannot be repeated without the intervention of additional complex mechanisms [46]. Overall, soft design, which involves pneumatic 'muscle', artificial tendon, or smart material, provides a comfortable wear experience; however, most exoskeletons with a low-rigidity design are heavily underactuated and one active Degree of Freedom (DoF) is usually considered for each digit, which limits its applications [47].

Compared with soft exoskeletons, this design involves linkages that are bulky and rigid, which potentially provides an uncomfortable wear experience and a heavy burden for the user [48–50]. Furthermore, misalignment of the finger joint (axis) and exoskeleton joint (axis) is commonly found in current designs, which potentially leads to discomfort and skin abrasion [36,51,52]. However, the linkage-driven mechanism is still widely adopted in hand exoskeletons due to the large force transmission efficiency, precise joint trajectory control potential, and reliability of the mechanism [53].

To facilitate post-stroke rehabilitation and provide assistance for complex daily life activities, a complete smart hand exoskeleton rehabilitation system, which covers accurate digit joints' motion control, adjustable dimensions, a reliable intention-detection approach, and high intention-recognition accuracy, is proposed in this study. Based on the physiological structure of a human hand, a compact linkage-driven design with 10 active DoFs and 3 passive DoFs is proposed, which enables accurate control of a wide range of postures. Adopting the dimension-adjustable design, the device can be equipped by the majority of the population in the world. Based on the preferences of the user, the hand exoskeleton can be mounted on the existing up-limb exoskeleton system via a link module, which greatly diminishes the weight burden for the user. Three rehabilitation/daily life assistance modes are developed for various personal needs, namely, robot-in-charge, therapist-in-charge, and patient-in-charge modes. Considering HMI, a thin-film force sensor matrix and IMUs are installed in the exoskeleton, and the corresponding controller aims to capture/detect user intentions by tracing the force on the exoskeleton and the rotation angle of finger joints. The reliability of the sensor composition synergized with this device is assessed by the trigger rate, Kernel Density Estimation (KDE), and a confusion matrix. To recognize user intention, a genetic algorithm (GA) is applied to search for the optimal hyperparameters of CNN aiming for high intention-recognition accuracy.

#### **2. Design of Hand Exoskeleton**

#### *2.1. Hand Skeleton Model Construction*

The physiological structure of the hand can be revealed by analyzing the existing model of the hand skeleton in the OpenSim library. The skeleton of the hand capitates near the wrist, metacarpals, and phalanges segments. In the hand skeleton, all digits contain 1 metacarpal segment. The 4 fingers have 3 segments, namely, proximal, intermediate, and distal phalanxes. The thumb possesses 2 phalanx segments, which are proximal and distal phalanxes. The joints of the hand are named according to the bones to which they connect. Consequently, there is 1 metacarpophalangeal joint (MCP), 1 distal interphalangeal joint (DIP), and 1 proximal interphalangeal joint (PIP) for the 4 fingers, while the thumb contains only 1 MCP and 1 DIP joint (Figure 1). In addition, there is a carpometacarpal joint (CMC) for each digit near the wrist.

**Figure 1.** Constructed 3D model of hand skeleton based on anatomy.

Parameter for subjects' fingers (units: mm).

— 28

–

–

The joint between each phalanx can be treated as a 1 DoF hinge joint, as 2 phalanges can only bend and extend along the vector direction shown in Figure 1. The MCP joint is equivalent to 2 DoFs, a ball-and-socket model that can rotate along the two directions. The CMC joints can be regarded as a 2 DoF saddle joint [54]. All digits in one hand have a total of 29 DoFs, where the thumb contains 5 DoFs and each of the four fingers has 6 DoFs. If all 29 DoFs are adopted as active DoFs, the weight of the hand exoskeleton device would be a huge burden and the reliability of the device in both motion transmission and motion control would be low. To carry out a successful grasp, each digit acts independently for flexion–extension, and the trajectory of each joint is constrained in a single plane.

Notably, the four fingers and the thumb do not share the same physiological structure. The intermediate phalanx is absent for the thumb. Moreover, the DIP joint of the thumb possesses a significantly larger active rotation range (compared with the DIP joints of fingers). However, the CMC joint (especially the CMC of the thumb) plays an essential role in grasping in terms of flexibility and force transmission. The simple grasp action can be performed with all metacarpals fixed, and rotation of the CMC joint is not mandatory. Therefore, the hand exoskeleton designed in this paper only considers the DoFs required by flexion–extension, which mainly involves PIP and MCP joints for the four fingers, while DIP and MCP joints are considered for the thumb.

#### *2.2. Finger Kinematics*

In order to conduct a finger kinematics analysis, the measurement of a volunteer is necessary. Measurements are conducted on phalanges and metacarpals with the aid of a vernier caliper. These measurements are recorded in Table 1. Note that while the exoskeleton is developed based on a single subject, the fitness for a larger population is considered, which is thoroughly discussed in Section 2.3.1. An experiment on subjects with different hand sizes is presented in Section 3.2.3.

**Table 1.** Parameter for subjects' fingers (units: mm).


Considering all the possible gestures/actions performed by the hand, the skeleton of the hand plays an essential role in posture support, and the length of each digit stays approximately the same during the rotation process. Taking the index finger as an example, we treat the metacarpal bone as a fixed base frame and the metacarpal bone, proximal, middle, and distal phalanxes form an open-chain four-linkage mechanism. For a grasping action, the 3 DoFs in the four-linkage mechanism are all rotational, and the rotation angle ranges are 0–90◦ , 0–110◦ , and 0–70◦ for MCP, PIP, and DIP, respectively. Based on a modified hand skeleton model (Figure 2), D-H parameters (Table 2) of the equivalent four-linkage mechanism are established to study the kinematics of the hand, where O<sup>0</sup> is the coordinate system fixed at one end of the metacarpal close to carpals (CMC joint) and the O1, O2, O3, and O<sup>4</sup> coordinate system is located at the geometric center of the MCP, PIP, DIP, and fingertip, respectively.

– – –

**Figure 2.** Coordination system of the 3D hand model.



In this research, the hand exoskeleton is designed to carry out rehabilitation training and aid patients in daily life activities such as object grasping. To implement the grasp action, muscles and tendons drive the MCP joint first, followed by PIP and DIP joints. In this study, workspace refers to the collection of spatial positions that a joint can reach under constraints.

Based on the Monte Carlo method [55], workspace studies on the fingertip and (the geometric center of) the DIP joint are carried out first to build and validate the initial design of the hand exoskeleton. Random valid rotation angles of each joint are substituted into a kinematics matrix based on the D-H setup to obtain the workspace cloud map of the index fingertip and DIP joint (Figure 3). As can be seen, the workspace of the DIP joint lays inside the workspace of the fingertip; however, the relatively smaller DIP joint workspace is enough for grasping large objects such as a bottle.
