According to the modelling principle which consists in an iterative comparison between a knowledge model and a representation model, we are building a software which simulates the dynamics of a Diesel engine. Modelling a Diesel engine requires a good accuracy in combustion simulation; therefore Vibe's two-peak model was chosen. But determining the seven necessary parameters is quite difficult due to this non-linear representation. The “downhill gradient search method” solves this identification problem if weighting factors and initial filtering are chosen carefully. After presenting our principle, we shall apply this method to two different engine speeds. Lastly, we shall compare the data measured in the test-bed with the new computed ones using our model.