Abstract— The vibration transmitted through the suspension of a car is the dominant source of structurally-generated interior noise at low frequencies. With the addition of actuators between the suspension attachment points and the car frame, we propose the design of a feedback controller for an active control of the vibrations to reduce cabin noise. A model-based controller design requires the identification of a transfer function model of the suspension, the frame, and the cabin acoustics. Here we deal with model identification of a suspension testbed. Uncertainty bounds for the flexible mode parameters of the model are estimated from experimental input-output data for a future robust mu-synthesis controller design. A preliminary controller is designed and tested by simulation for the identified nominal model. Inputs to the controllers are given by force and acceleration sensors measuring the vibration at different points on the suspension. I. INTRODUCTION any sources contribute to the ambient noise inside the passenger cabin of an automobile, such as the engine or air ducts. However, the principal source of low frequency noise is the road-induced vibration that propagates as structure-borne sound. Hermans and Van Der Auweraer [1] demonstrated it by taking structural and acoustical measurements inside an automobile, and identified the main source of the noise as road-induced vibrations causing a resonance. Similarly, Kim, Lee and Sung [2] found that car interior noise is caused in the low-frequency range by modal characteristics of the structure such as acoustic resonances, body vibration modes and structural-acoustic coupling characteristics. A noticeable reduction in the perceived level of noise can therefore be achieved by performing an active control of vibrations. Stobener and Gaul [3] implemented a modal controller by installing an array of piezoelectric sensors on the floor and centre panel of the body, with good results. This paper investigates the effectiveness of feedback- controlled actuators placed directly on the suspension, thus reducing the transmission of vibrations to the frame. A modal analysis is performed on a one-wheel car suspension to identify a model with parametric uncertainty bounds with the objective of designing a robust controller. A preliminary Manuscript received February 28, 2005. M.-P. Jolicoeur, J.-G. Roumy, S. Vanreusel, D. Dionne, B. Boulet, H. Michalska are with the McGill Centre for Intelligent Machines, Montreal, Quebec, Canada (phone: 514-398-1478; fax: 514-398-7348; e-mail: boulet@cim.mcgill.ca). H. Douville, P. Masson, and A. Berry are with the GAUS Laboratory, Université de Sherbrooke, Québec, Canada. feedback controller designed using the H ∞ control technique, which has been used in vibration control problems [4, 5], is showing some promise by providing more damping to the flexible modes of the suspension testbed. II. MODEL IDENTIFICATION The car suspension testbed can be modeled in terms of its modes of vibration. A finite-element model (FEM) would be difficult to obtain, and might yield poor results in view of the complexity of the suspension structure. Therefore, a more direct way to obtain a dynamic model in this case is to use experimentally acquired data in a system identification procedure. Figure 1 shows the one-wheel car suspension experimental testbed at the GAUS laboratory at Université de Sherbrooke. A shaker is attached to the wheel axle to simulate the road disturbance, and a force sensor records the force applied to the axle. Different acceleration and force sensors are placed on the suspension attachment points to record its behavior when excited. Figure 2 shows the location and denomination of the suspension attachment points to the car frame. Four three-axis (x, y and z) accelerometers are placed at the suspension base’s attachment points (B1, B2, B3 and B4). Three force sensors are placed at the top of the suspension (Bh) in all three axes. The suspension is excited through the shaker with white Gaussian noise for a period of six seconds during which sensor information is recorded at a sampling frequency of 1000 Hz. The type of force sensor used gives a value that is proportional to the displacement. A. FRF Computation The frequency response functions (FRF) represent, for each sensor, the ratio of the output Fourier transform X(ω) over the input Fourier Transform F(ω). ) ( ) ( ) ( ω ω ω F X H = (1) In an experimental environment, two FRFs denoted H 1 (ω) and H 2 (ω) are obtained. FRF H 1 (ω) is defined as the cross- spectrum of the input and output signals divided by the energy spectral density of the input, whereas H 2 (ω) is the energy spectral density of the output divided by the cross- spectrum of the input and output signals. These FRFs should be identical in theory [6]. However, they typically differ due to physical constraints such as noise on the input or the output sensor signal, nonlinearities on the structure or finite Reduction of Structure-Borne Noise in Automobiles by Multivariable Feedback M.-P. Jolicoeur, J.-G. Roumy, S. Vanreusel, D. Dionne, H. Douville, B. Boulet, Member, IEEE, H. Michalska, Member, IEEE, P. Masson, and A. Berry M