Introduction

Figure 1: Linear system to be measured

Figure 1 depicts a linear system characterized by an impulse response h(n)h(n), driven by an input signal s(n)s(n), and producing the output signal r(n)r(n). The system identification problem is to estimate h(n)h(n) given known input/output signals s(n)s(n) and r(n)r(n). This module illustrates the Golay code and swept-sine techniques for identifying finite impulse responses. The first example measures the impulse response of a highpass filter using the Golay method, and the second example measures the impulse response of a weakly nonlinear loudspeaker driver using the swept-sine method.

Summary Of Objectives

Installation

Linear System

Consider the causal, single-input single-output (SISO) system shown in Figure 2. For simplicity, we will take the system to be linear and discrete-time, so that it is characterized by its impulse response h(n)h(n) or equivalently its transfer function H(z)H(z), which is the zz transform of h(n)h(n).

We will assume that both h(n)h(n) and H(z)H(z) exist so that we can discuss measuring them interchangeably. We will further assume that h(n)h(n) has finite length so that we can measure the response to an input signal in a finite amount of time. The goal of this document is to explain how to excite the system with a signal s(n)s(n), measure the response r(n)r(n), and use s(n)s(n) and r(n)r(n) to determine h(n)h(n) (and equivalently H(z)H(z)). In particular, it is useful to pick a signal s(n)s(n) that contains a large amount of energy so that measurement noise will not significantly corrupt the measurement results.

Figure 2: Linear system

Limitations Of Sound Interfaces

Sound cards and sound interfaces are not designed for making transfer function measurements. They merely provide a cost-effective solution since almost all computers have sound cards. We demonstrate these weaknesses given measurements made on a PreSonus Firepod sound interface. The output from channel 1 was directly connected to the line input on channel 1, and the sampling rate was fS=44.1fS=44.1kHz.

Minimum Phase Systems

One further consequence of the delay is that determining the phase response of the measured system is more complicated. The delay is responsible for a linear phase term since δ(n-k)↔e-j2πfk/fSδ(n-k)↔e-j2πfk/fS. If the delay is known (or measured), then it may be removed by multiplying the measured spectrum by ej2πfk/fSej2πfk/fS. However, if the system being measured is known to be minimum phase, then the folded-cepstrum method may be applied to find the minimum phase frequency response corresponding to the measured frequency response.

The transfer function measurement toolbox assumes that the system being measured is minimum phase. This is a valid assumption in many cases. For instance, all strictly positive real transfer functions are minimum phase. Dissipative systems are strictly positive real (and therefore minimum phase) if the appropriate quantity is measured and the sensor and motor are collocated. For example, if s(n)s(n) controls a motor exerting a force on a dissipative system, and r(n)r(n) is the velocity at that same point, the corresponding transfer function will be minimum phase. This holds for other dual variable pairs such as torque and angular velocity, voltage and current, and pressure and fluid flow.

For systems that are not minimum phase, such as systems involving a transmission delay between the input and output quantities, the phase plotted by the transfer function measurement toolbox is not the system phase response, but rather the minimum phase response corresponding to the measured system phase response.

High Pass Filter Measurement

The circuit shown in Figure 6 was measured using the Golay-code method to show how the sound interface non-idealities affect a measurement. VINVIN was connected to the output of channel 1 of the PreSonus sound interface, and VOUTVOUT was connected to the line input of channel 1 of the interface.

Figure 6: High pass filter electrical circuit

The analog transfer function H(f)H(f) can be determined analytically using the voltage divider rule:

In this case, R=1R=1kΩΩ and C=0.47μC=0.47μF, so the -3dB point is about f3dB=12πRC≈340f3dB=12πRC≈340Hz. Figure 7 and Figure 8 show that the frequency response is accurately measured in the range of about 10Hz to about 9fS/209fS/20.

Figure 7: Measured magnitude response of the high pass filter

Figure 8: Measured phase response of the high pass filter

The ringing in the measured impulse response distracts from the more subtle characteristics of the ideal high pass filter impulse response. For transfer functions that pass large amounts of energy at high frequencies, it may be more instructive to inspect the frequency domain measurement results.

Figure 9: Impulse response

Sine Sweep Measurement of a Weakly Nonlinear Loudspeaker Driver

To exeraggerate the nonlinearity of a loudspeaker, we cut the cone of a mishandled driver as shown in Figure 10. We monitored the sound pressure several centimeters in front of the dustcap using an Audio Technica AT4049a microphone, which has a flat magnitude response to within 3dB from 100Hz to 5kHz. The output from channel 1 of the PreSonus sound interface was connected to the speaker via a power amplifier, and the microphone was connected to the microphone input of channel 1 on the sound interface. The following results are typical of sine sweep measurements to show how with a weakly nonlinear motor.

Figure 10: Mishandled driver with additional cuts in the cone

Inverse filtering the measured response results in Figure 11, which is a plot of nonlinear2ImpResp.wav. The linear contribution corresponds to the spike at the beginning, while the weakly nonlinear terms are clustered closer to the end of the response.

Figure 11: Full-length response

The main linear contribution is cut out and plotted in Figure 12. The measurement was not made in an anechoic chamber, so there is a reflection about 15ms after the main impact.

Figure 12: Linear impulse response term

The nonlinear terms are shown magnified in Figure 13. The lower order nonlinear terms toward the right have larger magnitude but overlap less in time (see Figure 13). Note that (Reference) implies that the overlapping could be reduced by increasing the total length TT of the sweep excitation signal.

Figure 13: Nonlinear response terms

The magnitude and phase responses corresponding to the linear impulse response term from Figure 12 are shown in Figure 14 and Figure 15 in blue. For comparison, another sine sweep measurement was made at a lower level so that the speaker behaved approximately linearly. Decreasing the level also resulted in more noise and even some systematic error, as is evidenced by the red curves in Figure 14 and Figure 15. This comparison demonstrates that making measurements at larger levels can reduce the effects of noise, while nonlinear motor effects can be overcome with the sine sweep measurement technique.

Figure 14: Proportional to magnitude response of h(n)h(n)

Figure 15: Phase response of h(n)h(n)