Motivated by the common-seen model uncertainty of real-world systems, we propose a likelihood ratio-based approach to statistical detection for a rich class of partially observed systems. Here, the system state is modeled by some jump-diffusion process while the observation is of additive white noise. Our approach can be implemented recursively based on some Markov chain approximation method to compare the competing stochastic models by fitting the observed historical data. Our method is superior to the traditional hypothesis test in both theoretical and computational aspects. In particular, a wide range of different models can be nested and compared in a unified framework with the help of Bayes factor. An illustrating numerical example is also given to show the application of our method.