Active engine mounting system based on neural network control

Abstract

In the automotive industry some components and subassemblies which were initially made of steel are now being replaced with alloys and composites which have a higher strength to weight ratio. Therefore, today's vehicles are lighter, stronger and thus have small fuel consumption. However, mounting a more powerful engine to a lighter vehicle could cause vibration induced by the dynamics of the engine and thus affecting the comfort of the passenger. One way to overcome this predicament is to modify the mounting of the engine by introducing an active engine mounting (AEM) system which consists of passive rubber mount and a linear force actuator. At the correct frequency the linear force actuator would trigger a force which has a magnitude approximately equal to the engine's disturbance force but opposite in direction. With this the force transmitted to the chassis of the vehicle would then be minimized and increases passenger's comfort. In controlling the system, especially the force actuator, numerous controllers have been introduced which include but not limited to H2 controller, hybrid of feedback and feedforward, filtered X-LMS controller, optimal controller based on Haar wavelet and other classical feedback and feedforwad controllers. Determining the controller parameters could be a major and difficult task to perform since these parameters are based on the mathematical model of the engine-chassis system which also includes the mathematical model of the engine disturbance. In this thesis an intelligent controller namely the neural network controller has been introduced to reduce controller parameters identification. The system considered in this research includes two degree and multi degree of freedom systems. The dynamics of a nonlinear actuator was also included. Two types of neural network controller that has been used in this research namely the nonlinear auto regressive moving average (NARMA-L2) and the extended minimal resource allocating network (EMRAN). The performance of the neural network based controllers was then compared with classical controller such as PID for two degree of freedom system and a Linear Quadratic Regulator (LQR) controller for the multi degree of freedom system. The ability of the EMRAN to be trained online makes it advantageous for a non-model based controller. The EMRAN neural network has the ability to add and prune hidden layer neurons and for the purpose of efficiency and additional advantage was the adoption of the 'winner-takes-all' algorithm. Results show that the EMRAN controller perform much better as compared to PID and LQR controllers for the purpose of active vibration isolation based on the reduction of the force transmitted to the chassis of the vehicle.