A Neural Network Based Sensor Validation Scheme for Heavy-Duty Diesel Engines

This paper presents the design of a complete sensor fault detection, isolation, and accommodation (SFDIA) scheme for heavy-duty diesel engines without physical redundancy in the sensor capabilities. The analytical redundancy in the available measurements is exploited by two different banks of neural approximators that are used for the identification of the nonlinear input/output relationships of the engine system. The first set of approximators is used to evaluate the residual signals needed for fault isolation. The second set is used—following the failure detection and isolation—to provide a replacement for the signal originating from the faulty sensor. The SFDIA scheme is explained with details, and its performance is evaluated through a set of simulations in which failures are injected on measured signals. The experimental data from this study have been acquired using a test vehicle appositely instrumented to measure several engine parameters. The measurements were performed on a specific set of routes, which included a combination of highway and city driving patterns.