Torque estimation has an increasing importance in the field of automotive control, as most engine controllers rely on torque estimation today. It seems appropriate to look for a physically based model structure which can be easily identified from data. The full path between accelerator pedal angle and torque is considered, assuming that rotational speed is known, which is usually the case. The nonlinearity of the engine prevents a simple description of the engine by classical linear methods, like the transfer function technique and state space method. Instead, observing that the time scale of the engine is fixed by the rotational speed, and that most parameters of the engine depends on it, a LPV (Linear Parameter Varying) model seems a sensible compromise between the too large complexity of general nonlinear systems and the information loss associated with a linear model. Nonlinear subspace identification method is used which exploits the natural dependency of many engine processes on the rotational speed, but does not use a physical model, which is very time consuming and difficult to obtain. First a LPV state space method is introduced and then experimental results are presented which show that the model exhibits high accuracy. The diesel torque model consists of a static nonlinear gain element cascaded with a LPV dynamic system, similar to a Hammerstein model. The results obtained from the identification experiments are promising. However, the models are only identified in their stable operation points with a limited range of input and output. This model is valid only at these operation points and a global engine torque model is still not obtained because its dynamics is very complex. This will be another research target and further theory research and experiments are needed. These will include the global model identification and application to torque control and fault detection of diesel engine systems.