MLP Networks for Variable Estimation in Vehicle Diagnostic Context Elie R. Accari and Chaiban Nasr Faculty of Engineering Lebanese University Tripoli, Lebanon elie.accari@ieee.org Denis Hamad Laboratoire d’Analyse des Systèmes du Littoral Université du Littoral Côte d’Opale 62228 Calais, France Denis.Hamad@lasl.univ-littoral.fr Adrian Ciocan and Ahmed Hajjaji Centre de Robotique d'Electrotechnique et d'Automatique Université de Picardie Jules Verne 80000 Amiens, France Ahmed.Hajjaji@u-picardie.fr Keywords: Diagnosis, MLP, Estimation, Vehicle, Vertical Reaction, Wheel Slip, Yaw rate. ABSTRACT This paper deals with a neural network approach to estimate functional vehicle variables from measurable variables. Advance in technology is rendering the integration of this intelligence in vehicles a practice more and more applicable. Modelization is the alternative of an otherwise very expensive and dangerous test-bed and it is used for tool professional simulation software like veDyna to validate our results. We setup slalom, velocity profile and double lane change maneuvers on dry and wet roads and identified the forces acting on the wheels and the vehicle’s center of gravity from the accelerations using multi-layer perceptron neural networks. Finally we note that even though the vehicle can be subject to all situations during production, it is important to provide the system with the functionality to be trained online for the final owner of the vehicle. 1 - INTRODUCTION The public adoption of a security or comfort system is an obligation to all manufacturers to provide their clients with an exact replica of the performance they’ve become to expect. Embedding all the intelligence on a single chip, as used in the cutting edge technology sedans, has become more and more popular and security systems such as Anti-lock Breaking System (ABS), Vehicle Stability Assist (VSA), Traction Control System (TCS) and Electronic Stability Program (ESP) are now vastly used all over the world, despite the lack of standards in this trade. The ultimate goal is to detect any abnormal dynamic behavior and to inform the pilot in order to return to a normal security state; and to do that several hard-to-measure variables must be estimated. This work comes as a contribution to the estimation of such variables in the dynamic behavior of automotive vehicles by a neural network approach. Neural networks have reached a methodological and technological maturity ensuring great confidence and large demand by the industry to solve modeling, control and security issues (Rivals et al. 1993, Johnson and Calise 2002). Furthermore, security systems must be fast and have reduced false and non detection error rates (Persson et al. 2002), and feed-forward multilayer perceptrons fall within this category. The vehicle state is obtained thanks to an online estimation of some important variables such as vertical reactions, yaw rate and wheel slip using MLPs. In order to generate the training samples data-base essential for the training phase of the networks, we create slalom and velocity profile maneuvers on dry and wet roads. We validate the results on a double lanechange profile on various adherence road types. Vehicle simulation software, like the veDyna simulator used in this work, offer unmatched flexibility allowing them to reproduce at a cheap cost a large variety of normal and abnormal driving situations. This way, they spare fastidious and dangerous maneuvers on real test vehicles. The organization of this paper is as follows. The second section presents vehicle variables estimation equations based on simplifications of the vehicle’s physical model. Section 3 introduces the Multilayer Perceptron MLP system used as black-box estimator. Experimental results are presented in section 4 and we conclude in section 5. 2 - VEHICLE VARIABLES ESTIMATION Physical Vehicle models are complex, having a large number of variables many of which cannot be measured. Moreover, several parameters must be known with precision. These models are useful for the simulation of the vehicle behavior as they try to create situations close to real life. The limitations of the simplified models used to write the equations in this paragraph include, but are not limited to: decoupling the front and rear axles which means that what happens on one axle does not affect the other, part of the difference in wheel speed is to be interpreted as a result of the difference in tire pressures instead of change in the road’s radius of curvature and vehicle yaw rate; and the assumption that the point of application of the vertical force Fz on the tread is not variable. Rectifying these assumptions requires the use of variables that are hard to measure and the purpose here is in fact to evaluate the error caused by these assumptions. The more the variables that can be considered as given or known and included among the inputs, the more precise the model will be, but since this is not practical we choose to use estimators for this task. 2.1 - Vertical Reaction Force Each wheel is subject to a vertical reaction force which is the resultant of the sum of all the small forces applied over the surfaces of contact between the tread and the road. This force is not applied necessarily at the center of this surface and various .....