Using System Identification and Central Pattern Generators to Create Synthetic Gait Data

One way to accelerate fall prevention research is through the generation of synthetic gait data. The objective of this paper is to explore a proof-of-concept method using a central pattern generator (CPG) and system identification techniques to generate synthetic human gait data. The two neuron CPG model used in this paper actuated a single point mass that represented the height of the human hip. This continuous model was converted into a discrete system with four state variables - hip height, descending drive, and two CPG inhibition variables. To capture the step-to-step variation in the experimental dataset, we performed linear and nonlinear system identification on the discrete system. Compared to the experimental dataset, the linear model yielded a 12% error, while the nonlinear method yielded an 11% error, indicating that this system with the data used was predominantly linear. Using the identified linear model, we generated synthetic data by repeatedly applying the linear model to the current step's state to generate the next step's state. The resulting synthetic data had some similarity to the experimental data but clearly did not fully replicate the experimental signal properties, indicating that the methods used here have promise but are currently insufficient.