Deep Learning for Motion Classification in Ankle Exoskeletons Using Surface EMG and IMU Signals

TL;DR

A motion classification framework that integrates three Inertial Measurement Units (IMUs) with eight surface Electromyography (sEMG) sensors fabricated as towel-based textile electrodes, which improve comfort, durability, and usability for long-term deployment compared to traditional gel electrodes is presented.

Abstract

Ankle exoskeletons have garnered considerable interest for their potential to enhance mobility and reduce fall risks, particularly among the aging population. The efficacy of these devices relies on accurate real-time prediction of the user's intended movements through sensor-based inputs. This paper presents a novel motion prediction framework that integrates three Inertial Measurement Units (IMUs) and eight surface Electromyography (sEMG) sensors to capture both kinematic and muscular activity data. A comprehensive set of activities, representative of everyday movements in barrier-free environments, was recorded for the purpose. Our findings reveal that Convolutional Neural Networks (CNNs) slightly outperform Long Short-Term Memory (LSTM) networks on a dataset of five motion tasks, achieving classification accuracies of $96.5 \pm 0.8 \%$ and $87.5 \pm 2.9 \%$, respectively. Furthermore, we demonstrate the system's proficiency in transfer learning, enabling accurate motion classification for new subjects using just ten samples per class for finetuning. The robustness of the model is demonstrated by its resilience to sensor failures resulting in absent signals, maintaining reliable performance in real-world scenarios. These results underscore the potential of deep learning algorithms to enhance the functionality and safety of ankle exoskeletons, ultimately improving their usability in daily life.

Figures (5)

• Figure 1: Data collection (a) The schematic illustrates the custom-made electrode assembly used for EMG recordings Tang2023. (b) Motion data was captured using commercial IMUs. (c) The dataset includes recordings from various activities: squatting, walking forwards and backwards, and turning left and right.
• Figure 2: Signal recordings Representative examples of the relationship between recorded signals and physiological processes (a-d). During forward walking, intermittent forward acceleration is detected by the IMU (a), which corresponds to regular activation of the soleus muscle (c). Similarly, during a squat movement, slight forward motion of the shank is observed by the IMUs, which is linked to activity in the tibialis anterior muscle (d). Orange arrows highlight key instances where the two signals align, emphasizing representative matches. The EMG signal leads the IMU data, as there is an inherent latency between the electrical signals indicating muscle contraction and the subsequent initiation of motion detected by the IMU.
• Figure 3: Classifier performance (a) Performance of LSTM and CNN trained on different datasets (IMU only, EMG only and combined EMG and IMU). As a control, the accuracy of randomly choosing a class for each test sample was also reported. (b) Confusion matrix for the CNN trained on both IMU and EMG data. (c) Architecture of the CNN used for motion classification.
• Figure 4: Transfer learning and model robustness. (a) For transfer leaning the model is trained using data from 2 subjects only. The resulting model is then optionally fine-tuned using 50 samples before being evaluated on samples from a new subject. (b) Performance of the different models. (c) For model robustness the model is trained as previously but test data channels are corrupted. (d) The effect of different sensor corruptions on the classification results.
• Figure 5: Signal processing. Representative examples of EMG and IMU signal processing recorded during a forward walking trial. (a) The raw IMU signals are bandpass filtered to reduce noise (b). (c) EMG signals undergo outlier removal using a Hampel filter before bandpass filtering (d).