Design and experimental validation of intention recognition algorithms for a powered knee orthosis

Exoskeletons are becoming powerful tools to support therapists during rehabilitation and assist patients with locomotion difficulties. Most of the exoskeletons in the literature (Sanz-Morere et al. 2019), (Giovacchini et al. 2015) are based on a hierarchical control structure: (i) a high-level controller that estimates the user’s locomotion intent, (ii) a mid-level controller that implements an adaptive assistive strategy, and (iii) a low-level controller that manages the actuation units. This work aims at designing and validating a high-level classification algorithm for a portable active knee orthosis (AKO), previously developed by the team of engineers of the Wearable Robotics Lab, at The BioRobotics Institute (Scuola Superiore Sant’Anna, Pisa). For this purpose, an analysis of the scientific literature on control algorithms for wearable robots was initially performed. As a starting point of the work, the classification algorithm designed for a bilateral hip exoskeleton developed by the Wearable Robotics Lab was taken as a reference (Parri et al., 2017). The work presented an algorithm based on fuzzy logic for classifying walking, stair ascent and stair descent for assistance delivery with a powered hip orthosis. A dataset for offline analysis was acquired on six healthy volunteers. The experimental protocol consisted of a dynamic circuit that included ground-level walking (GLW), stair ascent (SA), stair descent (SD), ramp ascent (RA), and ramp descent (RD). The participants were requested to wear: (i) the AKO controlled in zero-impedance mode to measure the knee joint angle, (ii) two sensorized insoles to measure the vertical ground reaction force (vGRF) and center of pressure (CoP), and (iii) seven inertial measurement units (IMUs) to measure the acceleration and angular velocity of the trunk and the main lower-limb segments (thighs, shanks, feet). The development of the intention recognition algorithm followed a step-wise approach, leading to the training and testing of four different versions of the algorithm. Each time a new version was trained and tested, a procedure for the automatic feature selection was performed (the method consisted of the selection of the features that minimized the percentage of the overlapped areas of the membership functions across various locomotion modes). Lastly, the performance of the intention recognition algorithms was evaluated offline with the Leave-One-Out-Cross-Validation (LOOCV) method. The final version of the algorithm consists of a two-phase logic. First, at the heel strike, the algorithm recognizes the three main modes of locomotion, (ground walking GW, SA, and SD). Then, for the GW steps, the algorithm refines the classification at the mid-stance and classifies the GLW, RA, and RD modes. The classification at the heel strike resulted on average, in 100% accuracy for the three locomotion modes. For the mid-stance classification, the GLW, RA, and RD are then classified, respectively, with 84.3%, 96.8 %, and 100% accuracy. In the future, this classification algorithm will be implemented in real-time on the AKO and verified in terms of accuracy when knee assistance is provided according to the output of the classification.