Dans le cadre du projet DETER-DRIVE II, notre étude vise à caractériser le processus de la conduite automobile ; l’objectif est de disposer de capteurs pour définir les ensembles de mesures caractéristiques d’un bon fonctionnement ou d’un dysfonctionnement et de déclencher automatiquement une alarme au franchissement entre ces deux états. Le problème devient complexe lorsqu’il s’agit de distinguer les caractéristiques de la conduite personnalisée. Nous proposons trois outils de diagnostic pour bien identifier les habitudes du conducteur : la réglementation routière, les procédures de conduite et les capacités réactionnelles du conducteur. Nous avons choisi une approche neuronale de fusion multisensorielle. Un système multicapteur effectue la fusion de 32 variables. Notre étude concerne trois types de carrefours et, pour chaque type, le conducteur effectue trois tâches différentes générant neuf niveaux de difficulté pour la conduite. Cette approche neuronale comporte deux phases : apprentissage et généralisation. We describe, in this paper, the general approach for the detection of changes in driving behaviour in the global framework of a Smart driver copilot System we have developped within the European Community DRIVE-DETER II research. The driver System is composed of three parts: the Comparator Component based on an error detection with respect to the driving rules, the Driver Impairment Monitor depending on a detection of changes in driving behaviour and the Tutoring part. The 3 principles we have used are: 1. Counting driving errors according to driving rules; 2. Studying procedural changes in specific reproducible driving situations; 3. Studying reactional capability of the driver when an dangerous situation happens. Only part 2 is present ed. As many industrial, aeronautical, driving processes get more complex, human tasks consist in controlling, monitoring them by means of terminal boards. The operators have to keep attention for a long period of time in that case. A lack of attention, a performance impairment may lead to a catastrophic situation for the process, the environment, even for the operator himself. The final aim of the driver copilot System will be to detect any changes in his behaviour before any fault occurs. As driving is a control task which consists in controlling the speed and the position of the vehicle on the road and is driver dependent and context dependent (i.e. depends on the habits of a driver and the near by environment of the road, the weather, the traffic) we use an unsupervised learning System to build up a probabilistic model of the driver’s behaviour with training data obtained from a same driver negotiating the m meters of his way before a T-junction or cross-roads. The current behaviour of the driver is then « acquired » for this specific situation. Without the need to monitor the driver himself by means of physiological and sophisticated sensors to measure his « wakefulness », the behaviour of the vehicle driver can be observed through a set of relevant parameters from sensors located on the car and beacons located on the road. A sequence of pattern vectors {Xt} t>0 of dimension k sampled at regular time interval t is provided for a same driver and an experiment from time t=0 to t-T. For this presentation, the k parameters used are: the distance of the vehicle recorded from the beacon (a T-junction or cross-roads), the speed of the vehicle, the positions of the foot on the brake pedal, the accelerator, the clutch pedal. During the learning phase, using multidimensional training data, a pattern recognition component (Ollero et al. 1984) models the instantaneous estimates of the posterior class probabilities given the features P (wll(tl,t2)/X(tl,t2)), P(wl2(tl,t2)/X(tl,t2)), P(w13(t1,t2)/X(tl,t2))... P(w21(t2,t3)/X(t2,t3)), P(w22(t2,t3)/X(t2,t3),p(w23(t2,t3)/X(t2,t3)... for different predefined lengths ofthe time intervals (ti.tj). where ∑. (ti, tj)-T and T is the total duration of the experiment, towards a distribution of the data into a certain number of classes, resulting in a reduction of the dimensionality of the initial multidimensional set of data for each time interval. During the generalisation phase, using multidimensional test data, we can follow the driver behaviour, for each predefined time interval (ti, tj), the pattern recognition component assigns the multidimensional set of data to a predefined class or an unexisting class, according to a decision rule based on maximum membership function of the observed test data. Some results will be given using experimental data.