LAUSR

A hybrid dynamics modeling method based on expectation‐maximization (EM) is proposed in this article. The dynamics include continuous dynamic equations that depend on the combination of continuous and discrete states, so it is often used to describe the dynamic behavior of cyberphysical system. Modeling such a system is a nonconvex problem, and if symbolic regression or mixed integer programs are used to solve it, the computational complexity will be tricky in many practical applications. This article proposes a two‐step‐based approach. In the first step, we use the method of EM and repeatedly solve the convex optimizer problem generated by the null space until all subsystems are identified. The second step is to identify the switching logic between the subsystems by sparse logistic regression with sample weights. Experiments show that the method is robust to noise and achieves satisfactory identification results on piecewise autoregressive models with exogenous input, ordinary differential equations, partial differential equations, and other nonlinear examples.