Abstract
Keywords - Active Suspension, Parameter Estimation, Neuronal Network, Parity Equations, Fault Detection.
Abstract - Automobiles show an increasing number of mechatronic developments by the integration of mechanic and electronic systems. In the case of vehicle suspension systems, this comprises the application of hydraulic, pneumatic and/or electric actuators with the aim of active roll and pitch stabilization and an increase of driving safety and comfort in general. Supplementary, the integration of these actuators and the involved sensors and electronic control units enables a detailed process supervision, fault detection and identification (FDI) for the whole system. This was developed for an active vehicle suspension system on a test rig.
The inquired active suspension is a hydraulic fully loaded active suspension, whereat a hydraulic plunger is added in series with the conventional steel spring. This assembly is equivalent to the Active-Body-Control-System of DaimlerChrysler. The test rig is equipped with suitable sensors for a series passenger car. The goal is to detect faults of all involved sensors with the distinction of noise, outliers, breakdown, and also offset and gain faults.
One approach to solve FDI-problems is to use parameter estimation and parity equations. These methods are based on mathematical models of the active suspension system. The unknown parameters of these mathematical models are estimated with measured data using an estimation algorithm with low calculation effort. By this way, 27 physical parameters are estimated.
For the generation of parity equations, the concept of the local linear neuronal network LOLIMOT is applied. It represents a semi-physical modelling approach. This neuronal network is trained with measured data resulting in accurate models of the active suspension system. By comparing the online calculated outputs of the neuronal networks with the accompanying sensor signals parity equations are obtained. Moreover, these online calculated outputs can be used as analytic redundancy substituting faulty sensors.
By monitoring these estimated parameters, parity equations, and additional signal-based features, symptoms are obtained, which enable the detection of sensor faults. As each fault has a specific influence on these symptoms, each fault has a specific symptom deflection pattern. This pattern is presented in a fault-symptom classification scheme. This fault-symptom- classification shows an isolating nature, what means, that each considered fault has a specific pattern enabling an unambiguous identification of all faults. In this way, 56 different sensor faults can be distinguished.