The use of Artificial Neural Networks for Sensor Fault Detection, Isolation, and Accommodation in Automotive Engines Domenico Capriglione (1) , Consolatina Liguori (1) , Cesare Pianese (2) , Antonio Pietrosanto (3) (1) DAEIMI, University of Cassino, via G. Di Biasio 47, 07047, Cassino (FR), Italy, E_mail liguori@unicas.it (2) DIMEC, University of Salerno, via ponte Don Melillo, 84084, Fisciano (SA), Italy, E_mail pianese@unisa.it (3) DIIIE, University of Salerno, via ponte Don Melillo, 84084, Fisciano (SA), Italy, E_mail apietrosanto@unisa.it Abstract – The paper describes the hybrid solution, based on Artificial Neural Networks, ANNs, and production rule adopted in the realization of an Instrument Fault Detection, Isolation, and Accommodation scheme for automotive applications. Details on the ANN architectures and training are given together with diagnostic and dynamic performance of the scheme. Keywords – Automotive engine, sensor fault location, fault tolerant systems, automatic measurement station, artificial neural network I. INTRODUCTION In the l ast decade an increasing number of sensors and actuators have been employed in automotive systems [1]. The first reason is that the correct operation of modern electronically controlled engines is based on a high number of measurements (e.g. crankshaft speed, crankshaft position, inlet manifold pressure, inlet manifold air and coolant temperatures, battery voltage, throttle position, air fuel ratio at the exhaust). Moreover, the growing demand on security and comfort pushes toward an increasing use of suitable sensors and actuators. The antilock braking system and anti-spin traction control are now used in every new vehicle. Environmental sensors are used both for measuring and optimizing in- vehicle conditions, and for evaluating the outside conditions (presence of ice, intensity of rain, road characteristics, visibility). Anti -crash sensors and GPS are further sensor systems set up to improve the security and comfort of modern automotive systems. The driver status is also planned to be on- line monitored by more complex measurement systems that are aimed to reduce the number of crashes due to human errors. This means that like nuclear plants, aircrafts, space vehicles, and chemical processes, automotive measurement systems are going to be sensor fault tolerant [2]-[3]. Thus, hardware and/or software Instrument Fault Detection, Isolation, and Accommodation (IFDIA) schemes are finding more and more applications in automotive measurement and control systems [4]. Several approaches have been proposed to detect and isolate either sensor or actuator failures, with particular reference to engine control systems [5]-[8]. Their successful implementation demonstrates that IFDIA schemes are applicable in automotive as well as in aerospace systems, since these systems generally show significantly lower constraints on the required detection rates due to the relatively slow dynamics involved. In particular, the solution proposed in [7] was very suitable for fault sensitivity and selectivity, even if the required computational effort seems hardly compatible with an on-line on-board implementation. Authors experienced a set of different solutions to IFDIA problem, all based on the analytical redundancy relation approach [9]-[13], exploring several software implementations based on Artificial Neural Networks (ANN) and Expert Systems (ES). In [14] they present the realization and the characterization of a ruled-based IFDIA scheme applied to the engine control system [15]. Experimental tests carried out on both simulated and measured data, show good diagnostic selectivity and promptness without missed detections. However, the fault detection sensitivity proves to be poor and no fault accommodation is obtainable, due to high relationship residuals in absence of faults. This happens because the analytical redundancy relationships perform differently in steady state than in transient conditions. Consequently a more complex scheme must be thought to overcome the above mentioned modeling problems. In the paper an improved solution is presented, based on the integration between ANN layers and redundancy rule sets. In the following after some details on characteristics and analytical relationships concerning the automotive engine, the new proposed solution is accurately described. II. SOME RECALLS ON ENGINE CONTROL The proposed IFDIA technique has been developed for a FIAT 1,242 litres Spark Ignition Engine, four cylinders. The electronic control is based on a speed-density multi-point injection system by Magneti Marelli. Manifold pressure, crankshaft speed and throttle valve angle position sensors provides the basic information concerning engine states (i.e. engine load, speed). Thus, the main control actions (i.e. injection time, spark advance) are based on computations performed starting from the data measured with these sensors, while other measurements are taken to correct the basic control signals taking into