Gait Prediction And Recognition Based On Extreme Learning Machine

Exoskeleton allows for the integration of the complex mechanisms of human locomotion with the locomotor and load-bearing capability of robots,in order to aid the human body in the accomplishment of high-intensity tasks that involve heavy load carriage.In regard to the lower extremity assistive exoskeleton,the ability to accurately recognise the gait phases involved in the daily activities of human body is a key element in the exoskeleton control strategies.However,there is a delay between the recognition of gait phases and the implementation of controlling behaviour by the exoskeleton.In order to prevent the delayed controlling behaviour from affecting the locomotion of the body,data regarding human gait phases should be predicted in advance,to ensure that the behaviour of exoskeleton synchronizes with that of the human body.First of all,a data collection system based on plantar pressure is designed.The distribution mode of sensors,performance parameters of the system,data transmission mode,communication protocol and program flow between the systems are introduced in detail through hardware and software.The system was used to collect the data of the foot pressure of walking,jogging,walking up and down stairs.Extreme Learning Machine(ELM)is used to predict the plantar pressure during four activities:walking,jogging,walking upstairs and walking downstairs.Considering the fact that the accuracy of the predictions may be affected by the weights and biases that are randomly generated online,the Firefly Algorithm(FA)has been used to optimize the ELM As shown by the experimental results,the FA-ELM solves the problem that a normal ELM would have regarding the choice of parameters,thereby enhancing the predicting accuracy of the prediction model.It provides data support for gait recognition.The gait phases have been differentiated and key gait events for each phase have been determined.A gait event recognition algorithm has been proposed based on the characteristics of the plantar pressure predicted for each of the four activities mentioned above.Such algorithm solves the problem of the threshold value of feet striking or leaving the ground being inapplicable to the experiment after changing the experimental conditions.Additionally,many limiting conditions have been taken into account to avoid inaccurate recognition caused by statistical fluctuations.As a result,an average recognition rate of around 96.14%and an average recognition accuracy of around 93.80%has been achieved.