Correct knowledge of noise statistics is essential for an estimator or controller tohave reliable performance. In practice

however

they are usually not known perfectly. Early effortson noise identification are mostly limited to stationary noise or noise with slowly-varying statisticsonly. We present a new noise identification algorithm for hybrid system. The new algorithm combines the innovation filtering and the interacting multiple model algorithm. The former decreasesthe sensitivity of probabilitic weightings to measurement noise and the later implement dynamic interaction and dynamic changing of multiple modes based on the Markovian chain process of thesystem models. Simulation results show the effectiveness of the new algorithm.