A hybrid inverse dynamic-neural network approach to lower limb exoskeleton control
MetadataShow full item record
A powered machine that is wearable over all or part of the human body can be referred to as a powered exoskeleton. The role of powered exoskeletons is usually to provide ergonomic structural support while using motor power to synchronize to and assist with intended movements. One specific category of exoskeletons is the lower limb exoskeleton. There is a variety of applications for lower limb exoskeletons, including assistance, rehabilitation, and augmentation. A challenge in developing any of these forms of exoskeletons is the design of controllers which are able to perform well under a variety of scenarios, such as change in speed while walking or stair climbing, as well as with a variety of users. There are many different controllers that have been developed accordingly. One of these approaches involves estimating joint torques and applying these directly as control torques. This can be done in one of two ways: estimating the torques based on a few subjects and applying these prescribed torques to everyone, or estimating and applying joint torques in real-time. Many existing controllers which estimate and apply joint torques in real-time, only do so with a portion of the joint torques. For instance, this can be done by computing and applying joint torques which result from gravity only. The challenge with the estimating and applying joint torques in real-time is developing an accurate model to represent the dynamics of the system and accurately measuring all required state signals. The signals which are most problematic for measurement is ground contact force measurements. As forceplates are not useful for continuous overground measurements and instrumented insoles can be unreliable, an alternative approach is required. To fill this gap and generate a robust real-time joint torque estimator, a hybrid inverse dynamic-neural network model is proposed. In addition, a data-driven solution is proposed and comprises of an end-to-end neural network for direct joint torque estimation. The hybrid model computes joint torques with the use of kinematic information only. Eliminating the need for kinetic measurements allows ease with implementation in scenarios where forceplates are not available; this is done with a neural network for ground contact force estimation. The hybrid model was validated with 11 subjects during treadmill walking, including several different gait patterns. In comparison to the end-to-end direct torque estimator, the hybrid model has slightly worse performance at the knee and hip joints during treadmill walking which includes speed changes, asymmetrical walking, and start-stops. However, when testing these two approaches with a participant wearing an exoskeleton, the hybrid model outperforms the end-to-end network. This validates the versatility of the hybrid model to generalize to many different conditions and subjects. The hybrid model was then implemented as a controller in a lower limb exoskeleton. A second pre-defined direct torque controller was also developed. The pre-defined torques are recorded from the response of a feedback controller used on one participant. These torques are then applied as the direct torque control as a function of walking speed and gait phase. Both these controllers can be considered to be feedforward control approaches as the applied torques are not explicitly encoded by feedback errors. These controllers were tested individually and in combination for treadmill and overground walking with nine participants. A combination of the two controllers, with more contributions from the hybrid control, produces the overall best results in terms of spatiotemporal metrics. At a joint-level, all the tested controllers have similar performance in terms of range of motion and joint angle correlation to natural walking. The controller consisting of a combination of both hybrid and direct torque control, with more weight on the hybrid model, was also able to decrease the activation of four out of six muscles measured in the lower limbs, which includes knee flexors and extensors, and ankle dorsi- and plantarflexors, on average when compared to walking with the exoskeleton in passive mode. The decrease in muscle activity indicates that this control approach is able to provide assistance as well as improve the spatiotemporal performance. As the joint-level performance was not meaningfully improved by this controller consisting of a combination of both approaches, this control alone would be insufficient for users who require assistive as well as corrective torques from the exoskeleton. For example, those who have suffered from an incomplete spinal cord injury or post-stroke hemiparesis do not have the ability to walk with a natural gait, therefore can benefit from corrections from the exoskeleton to achieve a natural gait. The addition of corrective torques in the form of a position feedback control (FB) to the previously defined feedforward control (FF) is designed to provide both assistive and corrective torques to the user. In a pilot study with two participants for both treadmill and overground walking, the feedforward control alone has the best spatiotemporal performance while the feedback alone has the best joint-level performance. A combination of the two controllers will produce a balance of these two characteristics. All three of these controllers (FF, FB, FF-FB) were able to produce some reduction in muscle activation of the knee extensor and ankle dorsiflexor muscles, compared to passive exoskeleton walking. This indicates that all the controllers provide some level of assistance. However further testing is required to validate this hypothesis as well as optimize the method for combining these two control approaches. This thesis demonstrated that the application of biological joint torques as an exoskeleton controller can be further improved with the addition of other control strategies. It is possible that combining biological torques with other control approaches, including those not explored in this thesis, will be more suitable for those suffering from physical impairments such as hemiparesis or severe muscle weakness.
Cite this version of the work
Hannah Dinovitzer (2022). A hybrid inverse dynamic-neural network approach to lower limb exoskeleton control. UWSpace. http://hdl.handle.net/10012/18600