Publications

For suppressing the crosstalk problem due to wire resistances and contacted resistances of the long flexible cables in tactile sensing systems, we present a novel two-wire fast readout approach for the two-dimensional resistive sensor array in shared row-column fashion. In the approach, two wires are used for every driving electrode and every sampling electrode in the resistive sensor array. The approach with a high readout rate, though it requires a large number of wires and many sampling channels, solves the cable crosstalk problem. We also verified the approach's performance with Multisim simulations and actual experiments.

Aiming at the problem of gear fault diagnosis, in order to effectively extract the features and improve the accuracy of gear fault diagnosis, the method based on wavelet-packet independent component analysis and support vector machine with kernel function fusion is proposed in this research. The proposed wavelet-packet independent component analysis feature extraction method can effectively combine the advantages of wavelet packet and independent component analysis methods and acquire more comprehensive feature information. Besides, the proposed kernel-function-fusion support vector machine can well integrate the advantage characteristics of each kernel function. The energy features of wavelet packet coefficients are acquired with four-layer wavelet packet decomposition and then the extracted energy features are further optimized by the independent component analysis method. The kernel-function-fusion support vector machine method is adopted to realize the gear fault diagnosis. Two kernel function models with the best self-classification accuracy are employed to serve the gear fault diagnosis corporately. The test samples are primarily classified by the main kernel function model, and then some samples are selected to be reclassified with the other kernel function model. Finally, the two kernel function models cooperate to determine the type of test samples. The comparison investigations demonstrate that the proposed method based on wavelet-packet independent component analysis and support vector machine with kernel function fusion achieves very high diagnosis accuracy.

Inspired by recent success of foundation models in vision and language domains, time series foundation models (TSFMs) have garnered increasing attention in general time series analysis tasks like finance, weather, healthcare, and power. However, given high heterogeneity and severe annotation scarcity in time series sensor data, how to unlock the potential of large-scale general-purpose TSFMs for downstream activity recognition tasks remains yet unexplored? This paper makes the first attempt to address this timely challenge by adapting the self-supervised pre-trained TSFM (i.e., MOMENT) to few-shot activity recognition. We introduce a simple and efficient Sensor-Prompt Tuning (SPT) strategy, which employs multiple convolution-based sensor-friendly filters with a gating mechanism to act as learnable soft prompts, which can dynamically adapt sensor input space to the frozen TSFM backbone, effectively bridging domain gap between pre-training general time series data with wearable sensor stream. Extensive experiments across three public activity recognition benchmarks demonstrate that our SPT achieves up to 15.5% performance gains over existing state-of-the-art baselines under few-shot scenarios, while considerably outperforming other mainstream fine-tuning strategies with smaller than 1% of backbone parameters. Practical cloud-edge inference latencies are measured. This work offers a new prompt-tuning perspective on how to adapt pre-trained TSFMs for wearable activity recognition tasks. Code will be released.

Despite the fact that robotic platforms can assist user in training tasks requiring eye-hand coordination (EHC) motor skills, there are few instances where the robot-assisted training paradigm is more effective than unassisted practice for skill acquisition. This may be largely due to the reason that current studies on robotic EHC training have elucidated several challenges, such as large-delay gaze-based feedforward predictions of participant's focused-object and unintuitive augmented feedback for delivering skill characteristics. To this end, we develop a novel robotic training paradigm with gaze-informed haptic guidance for enhancing EHC learning in a simulated spatiotemporally critical interception task. Its gaze interface accurately captures the participant's visual attention on virtual moving object with only ~200ms latency (much shorter than ~2s of current studies), and then the robot immediately activates the kinesthetic feedback for teaching fine motions (when and how to move) facilitating successful interceptions. In this way, the proposed paradigm exhibits features previously shown to promote successful training: It avoids disruptive delays between the user's attention before hand movement and the task-specific robotic assistance, assisting trainees in completing the interception more frequently when it is in use; It encourages user engagement, since the participant's intentional focus has to be detected explicitly during training; It provides meaningful haptic guidance to the less-skilled learner for spatiotemporally critical tasks. Through user studies, we showed that the proposed robotic training paradigm with attention-triggered task-specific haptic feedback led to increased skill acquisition compared with unassisted practice.

To date, convolutional neural networks have played a dominant role in sensor-based human activity recognition (HAR) scenarios. In 2021, researchers from four institutions almost simultaneously released their newest work to arXiv.org, where each of them independently presented new network architectures mainly consisting of linear layers. This arouses a heated debate whether the current research hotspot in deep learning architectures is returning to MLPs. Inspired by the recent success achieved by MLPs, in this paper, we first propose a lightweight network architecture called all-MLP for HAR, which is entirely built on MLP layers with a gating unit. By dividing multi-channel sensor time series into nonoverlapping patches, all linear layers directly process sensor patches to automatically extract local features, which is able to effectively reduce computational cost. Compared with convolutional architectures, it takes fewer FLOPs and parameters but achieves comparable classification score on WISDM, OPPORTUNITY, PAMAP2 and USC-HAD HAR benchmarks. The additional benefit is that all involved computations are matrix multiplication, which can be readily optimized with popular deep learning libraries. This advantage can promote practical HAR deployment in wearable devices. Finally, we evaluate the actual operation of all-MLP model on a Raspberry Pi platform for real-world human activity recognition simulation. We conclude that the new architecture is not a simple reuse of traditional MLPs in HAR scenario, but is a significant advance over them.

During recent years, dynamic early-exit has provided a promising paradigm to improve the computational efficiency of deep neural networks by constructing multiple classifiers to let easy samples exit at shallow layers while avoiding redundant computations at deep exits, which has been seldom explored in the context of latency-aware human activity recognition (HAR) deployed on wearable devices. Particularly, most existing early-exit strategies have always treated all activity samples equally at each exit during training, which ignore such dynamic early-exit behavior at test-time, causing a potential mismatch between training and test. Intuitively, easy activity samples that often exit earlier at test-time should place more emphasis on the training loss of shallow classifiers, while hard activity samples should contribute more to the training loss of deep classifiers. To bridge this gap, this paper introduces a sample-reweighting approach for efficient activity inference, which employs a weight-predicting network to reweight the training loss of different activity samples at every exit. From a perspective of meta learning, a new optimization objective function is designed to jointly optimize both weight-predicting network and backbone network. We perform extensive experiments on three popular HAR benchmarks including UCI-HAR, WISDM, and UniMiB-SHAR, which demonstrate that while incorporating such test-time early-exit behavior into conventional training pipeline, it can consistently improve the accuracy-efficiency trade-offs under budgeted batch classification and anytime prediction patterns. Moreover, our approach has a natural advantage in handing class-imbalance HAR problem. Detailed ablation studies, visualized illustrations, and real hardware deployment are provided to support our statement. Code will be available at https://github.com/FuZenan/LtSRW-for-EEAR.