Gait State Estimation for a Powered Ankle Orthosis Using Modified Fractional Timing and Artificialc Neural Network1

摘要

An ankle foot orthosis (AFO) provides assistance to individuals with lower limb muscle impairment to correct gait deficiencies [1]. Powered AFOs and exoskeletons have been developed to explored robotic gait assistance. The portable pneumatically powered ankle foot orthosis (PPAFO) is capable of providing plantarflexor and dorsiflexor torque at the ankle [2]. The PPAFO has three sensors: force resistive sensors (FSR) under the heel and ball of foot to detect foot contact, and a Hall effect sensor to record ankle angle.Control of a powered AFO will be most effective, if we can successfully identify and actuate the AFO at specific times during the gait cycle, i.e., when to apply dorsiflexor or plantarflexor torque at the ankle. We define that one gait cycle has 101 gait states, which corresponds to 0–100% gait cycle (% GC). A simple approach for identifying gait events and controlling the AFO is by directly checking whether the sensor signals go above or below a predetermined threshold (direct event control) [2]. An alternative approach is the use of predefined models to estimate the gait state, or % GC, from sensor signals and then use these estimated states to identify specific events during the gait cycle (state estimation control).A controller based on estimating state proposed by Li et al. had been used to actuate the PPAFO [1]. They proposed a fractional time (FT) state estimation approach that depended only on data from the heel FSR and estimated all other gait states from only one gait event, heel strike [1]. However, if gait speed changes, the FT estimator must wait for the next heel strike to begin adapting the speed change and will need another four to five strides to catch up with the new speed. Two new estimators (modified fractional time (MFT) and artificial neural network (ANN)) are introduced here which addressed the limitations of speed changes and reliance on only the heel strike event of the FT estimator.The MFT and ANN estimators used data from three sensor signals (heel contact, toe contact, and ankle angle) to estimate multiple gait events during a gait cycle.The MFT estimator targeted eight gait events. MFT was hypothesized to adapt faster to changes in gait speed than the FT estimator since the estimated states were updated eight times per cycles, rather than only once at heel strike. For MFT, the gait cycle was divided into eight events (heel strike, middle of initial contact, end of loading response, midstance, terminal stance, toe off, preswing, and midswing). The eight gait events were detected by statistical parameters obtained from the sensor values. We assumed that the gait states increased linearly between two events according to the average period of the gait cycle.In our second approach, a feedforward ANN with one hidden layer was also designed to estimate gait states. Unlike MFT, which identified eight events and then estimated the intervening states via linear interpolation, the ANN estimator continuously estimated states across a gait cycle based on a moving window from the previous six data points. Therefore, the network had 18 inputs (three sensors and six tap delays). The hidden layer had ten neurons and used log-sigmoid activation functions. The output layer had a single neuron and used a linear activation function, which provided the estimated state value during walking. Our primary goal was to find each state by minimizing a cost function. Estimation of the state from collected training data was performed by the Levenberg–Marquardt algorithm with Bayesian regularized training described by Hagan et al. [3,4].An experimental study examined the performance of the three state estimation algorithms by using experimental data from the three sensor signals that were collected while a subject walked on a treadmill that recorded ground reaction forces (Bertec, Columbus, OH). Ankle angle was also recorded from a motion capture system (Vicon, Oxford, UK). For actuation control consistency in the st