It is of great significance to recognize hand gestures via measuring biological signals from forearm muscles for portable human-machine interaction (HMI). Decreasing the number of sensor nodes is imperative for practical HMI applications. However, it would be an enormous challenge to maintain gesture recognition performance for traditional myoelectric interface with sparse-site sensing. To overcome this drawback, we present a novel HMI predicting more than ten hand and wrist motions relying on only two hybrid mini-grid surface electromyography (sEMG), mechanomyography (MMG) and near-infrared spectroscopy (NIRS) sensor nodes. Beyond the time domain (TD) features of sEMG, additional information containing movement intention is measured from motor unit action potential trains (MUAPt) according to the decomposition of four-channel arrayed sEMG. Furthermore, low-frequency myofiber vibration and haemodynamics are extracted from MMG and NIRS, respectively. Experiments are performed on thirteen healthy subjects to recognize 12 hand and wrist gestures. The results indicate that combining motor unit discharge feature yields consistently higher ( p <0.01) classification accuracy (91.6%) than traditional TD features of sEMG (87.8%) by using linear discrimination analysis classifier. Additionally, both MMG and NIRS features are demonstrated effective supplementary to distinguish muscular activation patterns, producing significantly enhanced (4.2%~11.2%, p <0.05) recognition performance with the fused information. The outcomes of this study are promising for the HMI applications such as controlling prosthetic hand and wearable device.
The research on bidirectional human-machine interface (BHMI) is of great significance, which can realize the interaction between human and machine. As the traditional Human-Machine Interface (HMI) device is limited by its wired connection and unidirectional functionality, this paper intends to develop a wireless wearable solution for BHMI with gesture recognition and vibration feedback. Compact design and wireless connection improve the flexibility and wearability of bidirectional interface node (BIN). By establishing a Distributed Wireless Sensor Network (DWSN), the proposed system allows up to 8 BINs to detect muscle activities simultaneously and control each Linear Resonant Actuator (LRA) independently. The data transmission throughput of the proposed system can be reduced by transmitting signal features instead of raw signals. With a sample rate of 2000 Hz and a resolution of 12 bits, BIN can achieve 94.48%±4.56% classification accuracy of six gestures. To evaluate the performance of the system’s bidirectional functionality, virtual control experiment was carried out. The experimental results show that the completion rate of the virtual control experiment reached 87% within 3.5s with vibration feedback. Overall, the outcomes of this study have great potential to promote the practical applications of bidirectional human-machine interaction, including rehabilitation therapy, fatigue prevention, virtual reality/augmented reality (VR/AR), and sensorimotor control of artificial devices.
1 September 2023 in IEEE Transactions on Instrumentation and Measurement
Gesture recognition via the measurement of surface electromyography (sEMG) has been extensively applied in human-machine interface (HMI) scenarios. Although high-density (HD) electrodes are effective for gesture recognition, channel selection from HD electrodes contributes to computational efficiency and wearable feasibility in practical applications. Nonetheless, the majority of the existing methods prioritize classification accuracy over operational effectiveness. During the framework implementation, channel forms, the optimization-oriented number of gestures, and the number of trials for optimization and training (O&T) are related to the optimized layout and the duration of frameworks. To improve the operational effectiveness, we presented an electrode configuration method to simplify the framework process. The muscle-matched electrode configuration (MEC) method was evaluated on eleven subjects with eleven-gesture recognition, and compared with two traditional channel selection methods, i.e., uniform configuration (UNI) and sequential forward searching configuration (SFS). The experimental results showed that the bipolar channel outperformed the monopolar channel in terms of classification accuracy (CA). Further results demonstrated the superiority of MEC in operational effectiveness by reducing the optimization-oriented number of gestures and the O&T-oriented number of trials. Specifically in the six-channel configuration, SFS required nine-trial and five-gesture optimization to accomplish a comparable CA (90.25% ± 3.49%), whereas MEC only took one trial with the same number of gestures optimization to achieve reliable performance (92.61% ± 5.18%), which saved nearly 85 % of the training time and memory space. Overall, the outcomes of this study provide meaningful guidelines about the efficient configuration framework for sEMG-based gesture recognition in practical applications.
16 May 2024 in IEEE Transactions on Industrial Electronics
Haptic interfaces create information transmission channels from machine to human with a natural way. Robust haptic interfaces with compact forms that provide high information throughput are of particular interest in industrial applications such as teleoperation, soft displays, virtual, and augmented reality. Here, we propose a haptic interface with spatiotemporal displays to convey high-dimensional information during successive hand-to-object tasks, which is architectural and psychophysical lightweight. The wearable and flexible architecture incorporates seven independently addressable vibration units at a pitch of 16 mm to allow spatiotemporal modulation. To prove the superiority of the haptic interface, four-day perceptual experiments are performed on 12 subjects with pattern discrimination tasks. We compared recognition accuracy (RA) between space-dependent and spatiotemporal strategies across the forearm, which are encoded based on finger states and morphological deformations. Furthermore, the mixed biomimetic strategy derived from both is intuitive and showed strength in terms of RA (over 88%) and information transfer (IT) (over 2.4 bit/s). With prior knowledge, are significantly better (RA over 90%) at the discrimination of sequential patterns. Moreover, the results demonstrate robust performance of the biomimetic strategy on location shift, which prove that subjects perceive relative information among vibratory units rather than absolute locations on the skin.