A CNN-Based Method for Intent Recognition Using Inertial Measurement Units and Intelligent Lower Limb Prosthesis

No prior work had fully resolved the challenge of achieving seamless locomotion transitions for individuals using powered artificial limbs. Existing systems often required numerous sensors mounted directly onto the prosthetic device itself. Researchers previously relied on machine learning models that depended heavily on manually selected features from time-series data. That uncertainty drove the need for more efficient, automated approaches to intent recognition. It was already known that motion intent precedes physical transitions between different walking environments. However, current methods frequently struggled to maintain natural movement patterns during complex activities like stair climbing. This gap motivated the development of strategies that utilize biological signals from the unaffected limb. The field required a more robust framework to bridge the divide between human intention and mechanical actuation.

The aim of this study is to develop a new method for training an intent recognition system for powered lower limb prostheses. Researchers sought to enable seamless transitions between different locomotion states for transfemoral amputees. The project addresses the limitations of existing systems that rely on multiple sensors mounted directly on the artificial limb. This work explores the mapping between the healthy leg's motion and the user's intended movement. By focusing on the unaffected limb, the authors intend to predict upcoming transitions before they happen. The team specifically investigates the early swing phase as a source of predictive data for the recognition model. This motivation stems from the need to improve the naturalness of walking on stairs, ramps, and flat surfaces. The study ultimately seeks to provide a more intuitive control strategy for advanced prosthetic devices.

The review approach focuses on a novel deep learning framework designed for predictive motion classification. Investigators utilized inertial measurement units to collect kinematic data from the unaffected limb of participants. This design captures movement patterns specifically during the initial swing phase of the gait cycle. The team implemented a convolutional neural network to process these inputs without manual feature engineering. Researchers evaluated the model using thirteen different categories of locomotion, covering both steady and transitional states. The study included testing on both able-bodied individuals and unilateral amputees to ensure broad applicability. Data processing involved mapping healthy leg motion to the intended prosthetic response. This systematic approach allows for the automated learning of complex movement signatures across various terrains.

Key findings from the literature indicate that the system achieves a recognition accuracy of 94.15% for able-bodied subjects. For amputee participants, the model reached an accuracy of 89.23% across all tested classes. The framework successfully classified five steady states and eight transitional states involving stairs, ramps, and level ground. These results demonstrate that early swing phase data provides sufficient information for accurate intent prediction. The convolutional neural network effectively learned feature mappings without the need for expert-guided input selection. The performance metrics confirm that the method remains robust across different user groups. This high level of accuracy supports the feasibility of using healthy limb kinematics for prosthetic control. The findings reveal that the system can reliably predict transitions before they physically occur.

The authors demonstrate that their approach achieves high recognition accuracy across thirteen distinct motion classes. Synthesis and implications suggest that leveraging healthy leg kinematics provides a reliable signal for prosthetic control. This study confirms that convolutional neural networks can successfully learn feature mappings without requiring human expert intervention. The findings imply that predicting intent before a transition occurs significantly improves the adaptability of powered devices. Researchers conclude that this method functions effectively for both unilateral amputees and able-bodied individuals. The evidence supports the integration of these sensors into future prosthetic control strategies. This work highlights the potential for automated feature extraction to replace traditional, labor-intensive data processing techniques. The results indicate that this framework offers a viable path toward more intuitive and responsive limb replacement technology.