Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Session S2C

Acoustic Imaging of Sound Sources A Junior Year Student Research Project


Gnter Bischof
Department of Automotive Engineering FH Joanneum University of Applied Sciences, Graz, Austria guenter.bischof@fh-joanneum.at Abstract In this paper the development of an acoustic camera within the scope of an undergraduate research project is presented, with a particular consideration of the projects embedding into the curriculum. At the start of the project sixteen microphones and a data acquisition device with the capability of sampling all channels simultaneously were available. In order to build an acoustic camera out of it, the students had to develop both the hardware for a microphone array and the software for the generation of an acoustic image. The physical principle that was made use of is the phase information present in the signals picked up by spatially separated microphones. For the computation of the acoustic image two different algorithms were employed. Delay-and-sum beamforming was used for the determination of the sound intensity as a function of the direction of arrival. And multiple signal classification was implemented in order to enhance the microphone arrays spatial resolution. For a flexible mounting of the microphones a portable tripod was designed, thus enabling a variety of microphone arrangements. The functionality of the acoustic camera is demonstrated by a spatially resolved noise measurement of a combustion engine in its engine compartment. Index Terms Acoustic camera, microphone array, project based learning (PBL), and undergraduate research. INTRODUCTION At the FH Joanneum University of Applied Sciences, we offer a four year automotive engineering undergraduate degree program. The faculty considers it especially important to apply modern didactical methods in the degree program as early as possible to increase the efficiency of knowledge transfer and to fortify the students' motivation to learn and to cooperate actively. For this reason a three-phase multi subject didactical method, based on the well known methodology of project based learning (PBL), was introduced [1]. It has proved to be an excellent method to demonstrate the need of mathematics and science in professional engineering. Students are confronted, complementary to their regular courses, with problems that are of a multidisciplinary nature and demand a certain degree of mathematical proficiency. A particularly suitable way of doing so turned out to be the establishment of interdisciplinary project work in the early stages of the degree program. The first phase of the project based didactical method is carried out in the second and third semester. The courses Information Systems and Programming in the second and third semester of the degree program form the basis of the projects. In the second semester the programming language Visual Basic (VB) is introduced. It enables the students to develop graphical user interfaces (GUIs) with comparatively little effort. In the third semester ANSI C, a fast machine-oriented programming language, is taught. In both semesters the students have to complete software projects as part of the requirements of both the Information Systems and Programming course, and at least one additional course within the curriculum. Generally the project is formulated within this so-called complementary course and covers a typical problem in the field of that subject. Usually a team of three works on a project. One student is designated by the team as project leader and assumes the competences and responsibilities for this position. This structure promotes the development of certain generic skills, like the ability to work in teams, to keep records and to meet deadlines. Usually two or more groups are assigned with the same task. In this way competition is generated, which in turn increases the students motivation. While in the second semester the main focus is on the acquisition of programming abilities and on soft skills, the tasks of the third semester projects focus more on the subject area of the complementary courses. The second phase of our multi subject didactical method takes place in the junior year and is not connected with any particular subject. It is, in fact, a part of the curriculum on its own. The students are encouraged to define a research and development project they would like to work on and ask faculty members to act as their advisors, or vice versa. Three hours per week are designated for that project, but experience shows that most part of the work is done in the students spare time. The development of an acoustic camera described in this paper took place within the scope of such a student R&D project. The third and last phase is inherently linked with the seventh semester of the degree program and comprises of an internship in the automotive industry. The students typically do this practical training in OEMs or supplier companies and are responsible for finding their own placements. Here they get the chance to prove their skills and knowledge in the

978-1-4244-1970-8/08/$25.00 2008 IEEE October 22 25, 2008, Saratoga Springs, NY 38th ASEE/IEEE Frontiers in Education Conference S2C-1

Session S2C
professional world and get their feet in the doors of their future employers. In this paper the development of a microphone array within the framework of the first and second stage of our three-phase multi-subject educational method is presented. First the motivation for the project is introduced, and then the mathematical background is elucidated, followed by a sketch of the projects implementation. The program is then exemplified by offline simulations and finally the paper is concluded by an example of a real measurement. MOTIVATION FOR THE PROJECT The detection of sound sources was usually done by hand with level measuring instruments or sound intensity probes. This manual scanning can nowadays be replaced by stationary arrays of microphones that allow the examination of complete structures within one measurement. The output of such an array usually results in a noise map, which can be overlaid on a digital photo of the test object. The noise measurement can be directly compared with a visual image of the object, thus significantly facilitating the interpretation of the acoustic data. Unfortunately, commercially available microphone arrays are rather costly and hardly affordable, especially for smaller universities. For this reason we started to consider building such a microphone array by ourselves. According to the two responsibilities of universities research and teaching students should be involved in the development. The first step was a student project in phase one of our project based educational scheme. A team of three students was entrusted with the development of a linear sum-anddelay beamformer consisting of four microphones. The time horizon for the project was one semester; the supervisors were the Information Systems and Programming, Engineering Mathematics and English lecturers (English, as the project working language, is not the native language of most of our students). The students had access to the departments anechoic chamber and could utilize a DEWE3010-IPC data acquisition device and four ROGA Instruments MI17 electret microphones. These microphones were arranged in a linear array facing towards the sound source. By introducing a specific delay on each microphone signal and adding the results (delay-and-sum beamforming), an acoustical receiver equivalent to an adaptive array antenna field with a main lobe along a certain angle of incident can be created. The calculation process was performed in the frequency domain for reasons of simplicity (see next chapter). By repeating the calculation for a large number of angles, a one-dimensional map of the relative sound-pressure contribution at the observation point was generated. In this first step of development the students had to get familiarized with Fourier transform and the use of complex numbers in programming. Additionally they got an idea about the resolution of this method and its possible application to the two-dimensional problem of an acoustic camera. The mathematical competencies needed for the solution of this multidisciplinary problem were slightly beyond the sophomores skills at the start of the project. The knowledge and professional skills necessary for the successful completion of the task were taught during the course of the semester. Thus, as a very welcome side-effect, an increased interest of the students in their subjects was produced. After having solved the one-dimensional problem and having gained experience with acoustic beamforming the two-dimensional problem had to be tackled. Fortunately the project leader of the linear array project (Joachim Sigl) evidenced interest in the continuation of the development of an acoustic camera. Together with his fellow student (Ren Scheucher), he launched a follow-up project within the scope of an undergraduate research project. MATHEMATICAL BACKGROUND A delay-and-sum beamformer enhances signals from a particular direction and attenuates signals from other directions. In most beamforming applications two assumptions simplify the analysis: the signal sources are located far enough away from the array that the wave fronts impinging on the array can be regarded as plane waves (farfield assumption), and the signals incident on the array are narrow-banded (narrow-band assumption). The array under consideration consists of N = 16 microphones at the positions pn (with n = 1 N) and produces a set of signals denoted by the vector
f 1 (t , p 1 ) f 2 (t , p 2 ) f (t , p) = . f (t , p ) N N

(1)

The direction of an incident plane wave approaching the array is represented by the position vector a( , ) , which is a function of polar angle and azimuthal angle (Figure 1).

FIGURE 1
POINT SOURCE AS MONITORED BY THE MICROPHONE ARRAY

978-1-4244-1970-8/08/$25.00 2008 IEEE October 22 25, 2008, Saratoga Springs, NY 38th ASEE/IEEE Frontiers in Education Conference S2C-2

Session S2C
The spherical coordinate system is centered in the origin of the microphone array. If the distance between the acoustic source and the array is large and the microphones are identical, then the gains of all microphones must be equal. The incident plane wave hits the microphones at different times depending on and , which leads to a delay with respect to a reference microphone. The individual time-delays n are chosen with the aim of achieving selective directional sensitivity in a specific direction, characterized by the unit vector a( , ) . This objective is achieved by adjusting the time delays in such a way that signals associated with a plane wave, incident from the direction a, will be aligned in time before they are summed. This can be obtained by choosing The weighting vector used in this work is simply W n = 1 N for all n (uniform shading). The beamformer output in the frequency domain, the array output power spectral density, can be written as
P( ) = Y ( ) 2 = W F( ) W F( ) = W R ( ) W

)(

(7)

n =

a pn c

(2)

where R ( ) = F ( )F ( ) represents the N N cross power spectral density matrix of the channel input signals. The asterisk (*) represents the complex conjugate transpose (the Hermitian). For the localization of the sound sources the array response function is introduced. It represents the response of a microphone array to a plane wave of frequency incident on the array at an arbitrary direction of arrival a( , ) . With the weighted array manifold vector (or steering vector)
e i k p 1 e i k p w k ( , ) = N e i k p
1

where c is the speed sound. Thus the sum of the microphone signals divided by the number of microphones, the array output y(t), represents the sound pressure of the acoustic signal 1 N (3) y (t ) = f (t n ) . N n =1 Every other set of time delays leads to an attenuated array output. The fundamental principle behind the direction of arrival estimation is the use of the phase information present in the signals picked up by the spatially separated microphones. For the processing of phase information the preceding transformation of time signals into frequency domain is most suitable.

(8)

the array output power spectral density can be defined as


PB ( , , ) = w k ( , ) R ( )w k ( , ) ,

(9)

a and c the far-field assumption taken into account the Fourier transform of the microphone signals (1) gives
F (k , ) = F ( )u k

With the introduction of the wavenumber k =

(4)

with
e i e i uk = e i
1 2

e i k p e i k p = e i k p

(5)

being the array manifold vector, which incorporates all of the spatial characteristics of the array. These microphone signals are multiplied by appropriate and in general complex weights and then summed up to get the frequency domain representation of the array output
Y ( ) = W F ( ) .

(6)

where PB ( , , ) represents the squared intensity of sound for a narrowband input signal of frequency with the direction of arrival and (see e.g. [2]). In this way the delay-and-sum beamformer is able to determine the amplitude of incident sound as a function of its frequency and direction of arrival, but unfortunately it suffers from a comparably poor resolution and from ghost images. On the other hand, subspace-based estimation techniques in array signal processing are able to locate noise sources efficiently and accurately, but do not necessarily provide information about the sound level. The combination of both methods can overcome those shortcomings and give sound intensity information at a sufficiently high resolution. The multiple signal classification (MUSIC) algorithm [3] is one of the most popular subspace based narrowband methods. In this method the cross power spectral density matrix R() is decomposed into its (ordered) eigenvalues 1 2 N and corresponding eigenvectors v1, v2,, vN. The fundamental property of the eigenvectors of a correlation matrix is that they are orthogonal to each other, since R() is self-adjoint. Assuming that there are K point sources monitored by a microphone array consisting of N microphones, then the eigenvalue decomposition of R() decomposes the N-dimensional space of the matrix into a Kdimensional signal subspace and a (N-K)-dimensional noise subspace.

978-1-4244-1970-8/08/$25.00 2008 IEEE October 22 25, 2008, Saratoga Springs, NY 38th ASEE/IEEE Frontiers in Education Conference S2C-3

Session S2C
The Euclidean distance of any arbitrary vector v from the signal subspace is the length of the projection of v into the noise subspace. The squared magnitude of this distance is given by
d 2 ( v) =

j = K +1

(v j v ) (v j v )
N

(10)

with vj being the jth eigenvector of the noise subspace of R(). A vector v that lies entirely in the signal subspace minimizes d(v). Thus the peaks of the following function will correspond to the source locations (e.g. see [4])
PM ( , , ) =

v w k ( , ) j j = K +1

(
N

)(

v w k ( , ) j

(11)

The steering vector w k ( , ) represents the spatial sampling of a plane wave with wavenumber k in the look-direction a( , ) . Thus if the look-direction of the array happens to be targeted to a sound source, w k ( , ) lies in the signal subspace and thus PM approaches a large value. The computation of PM for a grid pattern of looking directions thus produces a picture with peaks at the sound point sources. The just described procedure reveals one of the disadvantages of the MUSIC algorithm. It requires a priori knowledge of the number of sources in order to estimate the dimension of the signal subspace. The accuracy, with which the number of principal eigenvalues is determined, and the quality of the estimated subspace (e. g. a low signal-to-noise ratio) are the crucial factors for the success of subspacebased estimation techniques.

the engineering mathematics lecture the Jacobi transformation of symmetric matrices is taught, but in the case of the correlation matrix R() the complex analog, a Hermitian matrix, had to be decomposed. Generally, the students immersed themselves into the new topics. Nevertheless, in addition their learning progress was facilitated by complementary lectures. The students developed a graphical user interface (GUI) in VB that enables the initialization of the hardware with the help of pre-prepared microphone arrangements, and the visualization of the beamformer and MUSIC algorithm outputs in the form of 2-D and 3-D images. An interaction of both algorithms was implemented in order to allow the suppression of the beamformers ghost image artifacts by a conjunction with the MUSIC output. Additionally, a simulation mode was implemented that allows the creation of different microphone arrangements and virtual source fields for testing purpose. In order to facilitate the comparison of the measured sound pattern with the sound sources of the recorded object the overlay of digital images of a webcam with the contour plots of the beamformers and MUSIC algorithms outputs was also provided in the GUI. The data processing and computation part was programmed in C and developed as dynamic link libraries (DLL). The functions of the DLLs are called from the VB GUI, process the digital data and pass it back to the VB main program. As a structured programming language, C is especially suitable for numerical mathematical methods. Before the mathematical algorithms were implemented in C, simulations were performed with Mathematica [6] in order to save time and debugging effort. Moreover, the students were encouraged to make use of program libraries like the Numerical Recipes in C [7] for the computer program development.
HARDWARE DEVELOPMENT

IMPLEMENTATION
The above outlined mathematical background of an acoustic camera points out the importance of possessing good knowledge of mathematics for the realization of such a project. The curriculum of the automotive engineering degree program includes engineering mathematics courses in the first three semesters. The lectures follow typically the contents of text books like Kreyszigs Advanced Engineering Mathematics [5], with an emphasis on numerical methods in the third semester. A part of the numerical algorithms needed for the project was already covered by these lectures. Additionally, some special knowledge had to be solidified or freshly acquired in private lectures offered by the supervisors. For instance, in terms of time-frequency analysis the continuous Fourier Transform is usually introduced in the second semester, the discrete transform (DFT) as part of the numerical methods in the third. For the acoustic camera, however, a fast transform (FFT) was needed. Another example is the eigenvalue decomposition of the cross power spectral density matrix. In

s min . (12) 2 978-1-4244-1970-8/08/$25.00 2008 IEEE October 22 25, 2008, Saratoga Springs, NY 38th ASEE/IEEE Frontiers in Education Conference S2C-4

The second big part of the project was the development of an adjustable tripod and mounting for the alignment of the microphones. It should be portable, stable and dismountable for in- and outdoor applications. The characteristic dimension of a microphone array is its size in terms of operating wavelength. For high frequency (small wavelength) signals a fixed-size microphone array will appear large, thus providing high angular resolution. However, for low frequencies (large wavelengths) the same physical array appears small and the arrays spatial resolution will decrease. On the other hand, the narrowband beamformer suffers from spatial aliasing at high frequencies. Spatial aliasing comes about if a sensor spacing wider than half a wavelength is used. It is analogous to temporal aliasing in discrete-time signal processing. This leads to the condition that the spacing s between any pair of microphones in the array should not exceed half of the smallest wavelength present in the signal

Session S2C
Spatial aliasing can also be avoided when the array geometry is totally non-redundant, that is, no difference vector between any two transducer positions is repeated. For non-redundant arrays, which typically have an irregular or random geometry, the side lobes causing ghost-images are suppressed. In general, irregular arrays outperform regular array designs, but it is difficult to find out how the design should be modified to obtain the best performance. One often resorts to a tedious trial and error cycle. An irregular array design was enabled in this project by a flexible microphone mounting. A second reason for an adjustable microphone arrangement is given by the fact that the narrow-band assumption is never valid in reality and an adaptation of the microphone spacing to the respective operating frequency is required. The basic structure of the array is an aluminum tripod stand with adjustable horizontal beams for the holding devices of the microphones (Figure 2). picture are ghost image artifacts caused by the structure of the array pattern and the beamformer algorithm.

FIGURE 3
THREE SOURCES GENERATED IN THE SIMULATION MODE OF THE ACOUSTIC CAMERA AS RESOLVED BY THE DELAY-AND-SUM BEAMFORMER ALGORITHM

The MUSIC algorithms output (illustrated in Figure 4) contains no information about the sound intensity. And the ratio of the peaks amplitudes is in fact not properly represented. Nevertheless, a comparison of Figure 3 and Figure 4 shows the supremacy of MUSIC in terms of spatial resolution.

FIGURE 2
MICROPHONE ARRAY PLACED IN THE ANECHOIC CHAMBER

The implemented microphone array has an active surface of 2 m x 2 m, which gives a maximum sensor spacing of 0.67 m and a minimum of 0.02 m. These dimensions correspond to an effective frequency range from 250 Hz up to 8 kHz according to (12). The supply current of 2 - 6 mA for the 16 electret microphones is delivered by the data acquisition device.
APPLICATION

FIGURE 4

THREE SOURCES GENERATED IN THE SIMULATION MODE OF THE ACOUSTIC For a comparison of the resolution of the delay-and-sum CAMERA AS RESOLVED BY THE MUSIC ALGORITHM beamformer and the MUSIC algorithm three point sources with two different amplitudes and three (slightly) different In a real measurement taken in the departments frequencies were generated with the help of the programs anechoic chamber the combustion engine of a sports utility simulation mode. In Figure 3 the beamformers response to vehicle was used as sound source. Although the V6 Diesel the signal is illustrated, which gives both information about engine running at a medium speed was hidden in the engine the position of the sources and the amplitude of the compartment the microphone array could localize the incoming sound waves (color code on the upper right hand combustion chambers as major noise source (see Figure 5). side of the picture). The additional wiggles observable in the 978-1-4244-1970-8/08/$25.00 2008 IEEE October 22 25, 2008, Saratoga Springs, NY 38th ASEE/IEEE Frontiers in Education Conference S2C-5

Session S2C
The digital picture taken by a webcam in the center of the microphone array is displayed together with a contour plot of the beamformer algorithms output. An important factor contributing to the acceptance and success of the project was the usefulness and applicability of the outcome. Both students were highly motivated by a task that stems from a real engineering problem. The outcome of this research and development project was written up by the students in a scientific paper and submitted to the American Journal of Undergraduate Research [10]. It appears to the author that the moving spirit of the successful realization of an undergraduate research project is an excellent communication and collaboration between educators and students. The acting in concert of project team and supervisors is a key issue for success.

ACKNOWLEDGMENT
The author would like to thank his former students Joachim Sigl and Ren Scheucher for their high motivation, creativity, conscientiousness and excellent performance. In addition, he would like to express his sincere gratitude to his colleagues Dr. Emilia Bratschitsch (supervisor of the programming part of the project), Dr. Herbert Fellner (advisor for the hardware part), and Annette Casey (English language trainer) for co-supervising this undergraduate research and development project and bringing in innovative ideas.

FIGURE 5
NOISE EMISSION OF A COMBUSTION ENGINE AS MEASURED BY THE MICROPHONE ARRAY IN DELAY-AND-SUM BEAMFORMER MODE

That computational superposition of the arrays acoustic output and the digital photo best explains the use of the term acoustic imaging.

REFERENCES
[1] Bratschitsch, E., Casey, A., Bischof, G., and Rubea, D., 3-Phase Multi Subject Project Based Learning as a Didactical Method in Automotive Engineering Studies, ASEE Annual Conference Paper 1020, 2007 Brandstein, M., and Ward, D. (eds.), Microphone Arrays: Signal Processing Techniques and Applications, Springer, 1st Edition, 2001 Schmidt, R. O.., Multiple emitter location and signal parameter estimation, Proceedings of the RADC Spectral Estimation Workshop, 1979, pp. 243-258 Benesty, J., Chen, J., and Huang, Y., Microphone Array Signal Processing , Springer Topics in Signal Processing, Springer, 2008 Kreyszig, E., Advanced Engineering Mathematics, John Wiley & Sons, 9th Edition, 2005 Wolfram, S., The Mathematica Book, Wolfram Media/Cambridge University Press, 4th Edition., 1999 Press, W. H., Teukolsky, S. A., Vetterling, W. T., and Flannery, B. P., Numerical Recipes in C The Art of Scientific Computing, Cambridge University Press, 1988 Bischof, G., Bratschitsch, E., Casey, A., and Rubea, D., Facilitating Engineering Mathematics Education by Multidisciplinary Projects, , ASEE Annual Conference Paper 976, 2007 Bratschitsch, E., Casey, A., and Trzesniowski, M., Research Projects as a Part of a 3-Phase Multi Subject Project Based Learning in Vehicle Engineering Studies, ASEE Annual Conference Paper 1558, 2008

CONCLUSIONS
Starting from their freshman year, automotive engineering students at the FH Joanneum University of Applied Sciences are involved in project work within the framework of PBL. Software projects complementary to the lectures exemplify the applicability of the just learned methods and mathematical algorithms, thus increasing the students attentiveness and their appreciation for the new topics [8]. The second phase of this problem oriented curriculum framework involves student selected research and development projects that encourage investigation, team work, problem solving, and other higher order thinking skills. These R&D projects are intended for junior year students, with e.g. the development of an SAE Formula Student racing car as the departments showcase project [9]. One of the fundamental didactical aspects of these projects is the development of the students ability to tackle a complex task even though there is no predetermined way from problem to solution. The acoustic camera was developed within the scope of such an undergraduate student research and development project. The participating students were confronted, complementary to their regular courses, with a problem that was of a multidisciplinary nature and demanded advanced professional skills. Additionally, this project gave them the chance to look beyond the standard curriculum of automotive engineering education.

[2] [3]

[4] [5] [6] [7]

[8]

[9]

[10] Sigl, J., and Scheucher, R., Acoustic Imaging of Sound Sources with a student-designed Acoustic Camera, American Journal of Undergraduate Research, Vol. 6, No. 2, 2007, pp. 1-8

978-1-4244-1970-8/08/$25.00 2008 IEEE October 22 25, 2008, Saratoga Springs, NY 38th ASEE/IEEE Frontiers in Education Conference S2C-6

You might also like