Optimizing Locomotor Task Sets in Biological Joint Moment Estimation for Hip Exoskeleton Applications

TL;DR

Optimizing Locomotor Task Sets in Biological Joint Moment Estimation for Hip Exoskeleton Applications addresses the data burden in training deep-learning-based hip moment estimators. The authors propose a task-set optimization pipeline using PCA for dimensionality reduction, k-means clustering with silhouette-based cluster selection, and a task-weight analysis to select eight representative tasks, trained with a fully connected neural network under leave-one-subject-out cross-validation. The optimized task set achieves RMSE $0.30±0.05$ Nm/kg and $R^2=0.57±0.10$, comparable to the full-task approach ($RMSE$ $0.28±0.05$, $R^2=0.64±0.07$) and significantly better than cyclic-only data ($RMSE$ $0.38±0.08$, $R^2=0.32±0.09$). The results validate that reduced, biomechanically diverse task sets can preserve estimator accuracy, potentially lowering data collection costs for hip exoskeleton designers.

Abstract

Accurate estimation of a user's biological joint moment from wearable sensor data is vital for improving exoskeleton control during real-world locomotor tasks. However, most state-of-the-art methods rely on deep learning techniques that necessitate extensive in-lab data collection, posing challenges in acquiring sufficient data to develop robust models. To address this challenge, we introduce a locomotor task set optimization strategy designed to identify a minimal, yet representative, set of tasks that preserves model performance while significantly reducing the data collection burden. In this optimization, we performed a cluster analysis on the dimensionally reduced biomechanical features of various cyclic and non-cyclic tasks. We identified the minimal viable clusters (i.e., tasks) to train a neural network for estimating hip joint moments and evaluated its performance. Our cross-validation analysis across subjects showed that the optimized task set-based model achieved a root mean squared error of 0.30$\pm$0.05 Nm/kg. This performance was significantly better than using only cyclic tasks (p<0.05) and was comparable to using the full set of tasks. Our results demonstrate the ability to maintain model accuracy while significantly reducing the cost associated with data collection and model training. This highlights the potential for future exoskeleton designers to leverage this strategy to minimize the data requirements for deep learning-based models in wearable robot control.

Figures (5)

• Figure 1: Task Selection and Model Training. Average joint moment, angle, and velocity profiles were used for principal component analysis (PCA), k-means clustering, and task-weight analysis to determine a smaller optimized set of tasks that were representative of the full set. The time-series data for each training dataset condition (all tasks, optimized tasks, and cyclic tasks) were used for the fully connected neural network (FCNN) model training with leave-one-subject-out cross-fold validation.
• Figure 2: PCA and K-means Clustering Parameter Selection. (a) We used the first three principal components since they explained over 70% of the variance. (b) The Silhouette method revealed that $K$ = 8 yielded the best silhouette score, indicating it as the optimal number of clusters for our dataset.
• Figure 3: PCA and K-means Clustering Results. (a) All biomechanical feature data represented in the reduced dimensions. Each task mode is colored differently and each cluster is marked in a different marker shape. (b) Separate plots for each cluster, with the ellipsoid representing the space of 99% confidence for each cluster.
• Figure 4: Model Performance. (a) The all-task and optimized-task models outperformed the cyclic-task model, with significantly lower root mean squared errors (RMSEs). (b) Similarly, the optimized-task model showed higher R$^{2}$ compared to the cyclic task model, but similar performance compared to the all-task model. Gray dots represent values from 11 models per condition. * indicates a statistical difference from the cyclic tasks model (p<0.05).
• Figure 5: Qualitative Performance Evaluation. Ground truth and estimated hip joint moments for the optimized-task (blue) and all-task (red) models on two cyclic and two non-cyclic tasks using leave-out subject data with median RMSE performance. Ground truth is shown as a black dashed line. Both models follow the ground truth trend, but peak moment estimates show a ceiling effect.