When processing audio information of any kind, we have to deal with a mixture of sounds coming from a variety of sources. Signals from each source are mixed together to form a composite waveform and the frequency components of each source overlap. The challenge of working out which frequency component belongs to which sound source is highly complex.
The cocktail party scenario (top image)
At a cocktail party, the listener is able to focus on a single talker among many conversations. This processing is done by the human brain.
Separating the signals from the noise
Often, we are only interested in a single source of noise, so the challenge is to determine the target signal from all the signals in the mixture. Blind Source Separation (BSS) is a discipline within signal processing that involves separating a mixture of sounds and obtaining an estimate of each sound coming from various individual sources. The ‘blind’ refers to the fact that only a mixture of the sound is available and sometimes the number of different sources is unknown. A classic example is the ‘cocktail party scenario’, where the listener is able to focus on a single talker among many conversations. Here, the processing is done by the human brain. Using BSS, researchers have tried to capture this effect on measured microphone signals extracting any desired signal.
In the automotive industry, engineers and manufacturers must be able to evaluate the level of sound produced from many different components. For some parts of the car, the noise-level reduction is a key concern in the design process, often to comply with official legislation. Knowing the partial contributions and sources of different noises when a car is in operation is essential in prioritizing the targets for overall noise reduction.
Currently, the technique widely used in the automotive industry to identify the sources of both interior and exterior noise is referred to as Source Path Contribution (SPC) analysis. Using SPC analysis, microphones are placed at different locations to measure, for example, tire and engine noise. Typically, a so-called speaker-test is also conducted to establish noise measurements both inside and outside the car.
Although relied upon to provide good diagnostic information, SPC analysis has its drawbacks. It is labor-intensive and time-consuming to apply, and it also requires experienced personnel to correctly place microphones and speakers.
Applying Blind Source Separation
BSS is already used in various biomedical application areas and within the telecommunications industry. Historically, this technique has been driven by acoustic applications, such as hearing aids, and the subject of separating speech has been driving research in academia. What all these applications have in common is that a set of transducers is used to record a set of source mixtures and the problem is to extract each source at the transducers.
In acoustics, the mixing at the microphones is complicated by the fact that the source signals enter the microphones at different times, and reflections encountered in normal operating environments add copies of the source signals to the received microphone signals.
This means that the mixing process of sources into microphone signals is highly complicated. In addition, as the term ‘blind’ infers about the problem setup, usually only a recording of microphone outputs is available and no other information of the acoustic paths between sources and microphones is given – researchers just have to assume that the signals produced by the sources are independent.
With the complex algorithms that are involved in BSS staying very much in the world of academia, the path to applying this technique to other industries has been slow and challenging.
The mixed signals recorded at the reference microphone (set up close to the engine and tire contact patch) are blindly separated and labeled to identify their origin, that is, tire or engine. Including the receiver signal, a contribution analysis is carried out to estimate the signal portions coming from the engine and tire, respectively.
Benefits of BSS compared to SPC:
- Easier and more intuitive to perform
- No expert source separation experience required
- Superior separation results achieved
For several years, Brüel & Kjær has been researching BSS, believing it could be a more accurate and less time-consuming method for extracting different sound sources compared to SPC. Given the right algorithm, BSS could, for example, help solve the problem of separating engine and tyre noise using a set of close microphones.
Since the 1990s, a vast number of technical papers have been published on BSS. Many different algorithms have been proposed, all solving very specific separation problems. However, a very limited number of applications solving industrial problems exist and, since most algorithm development has been devoted to the separation of speech or music, it was challenging to find a useful source separation algorithm that could be applied to industrial noise challenges, including those in the automotive industry.
The right platform
An opportunity appeared when Brüel & Kjær was invited to take part in an EU-funded research project to identify the contribution of different engine modules to the exhaust noise from a turbo-shaft helicopter engine during operation. This was the platform that was needed to further develop the knowledge of BSS and help to apply it to the aerospace and automotive industries.
The plan was to install sensors into the engine and develop different methods to perform the source separation task.
Brüel & Kjær suggested the BSS approach, which was accepted by the other participants and the EU-funded project became known as TEENI (Turbo-shaft Engine Exhaust Noise Identification).
Following the kick-off meeting for TEENI, in April 2008, a systematic research approach was conducted, from a literature review to algorithm investigations and a selection of small-scale tests involving speakers inside a duct. Finally, the selected algorithm for performing the BSS task was given a full-scale test on the helicopter engine data and compared to other noise source breakdown techniques.
SPC for a full vehicle involves measurements with a sound source at around 20 microphone positions, moving the sound source for each measurement. Additionally, a vehicle operational measurement is needed to perform contribution analysis. In comparison, blind source separation only requires the vehicle operational measurement.
Verifying the algorithm
While the EU project was ongoing, Brüel & Kjær began to test the algorithm using datasets from previous projects, using the BSS processing of microphone signals to help resolve the contribution of noise from different parts of an operating car. The challenge was to validate the results. How can you, for example, accurately measure the noise from tires without the engine?
A small-scale set-up was designed using speakers around a Smart Car. The speakers played tire and engine signals and close microphones recorded the mixtures. The recorded microphone signals were used to train the BSS separation filters and perform the separation into the tire and engine contribution sounds at the microphones. The result was an improved separation, with only very little cross-talk compared to the classical separation approach based on SPC modeling.
The conclusion from this test was that BSS was superior to SPC analysis in terms of separation performance and also easier to perform.
Following this success, an SPC dataset used for contribution analysis during an indoor vehicle pass-by test was used to separate real recordings measured with microphones in an engine room and around tires during a standardized fast acceleration. Separation results, for example, listening to the tire noise component at the microphone close to the tire compared to the actual record-
ing, which was contaminated with engine and intake noise components, suggested a good separation result was obtained.
Finally, towards the end of 2013 and in early 2014, access to another, more comprehensive dataset also helped to compare the measurements from SPC processing with those from the BSS method. The dataset contained many reference measurements and assisted in further validating the results for multiple sources.
Using the same microphone positions and data for tire noise and engine noise that had been used for SPC analysis, it was possible to repeat a specific pass-by test to evaluate car noise for any source including engine, intake, exhaust, and tire, etc. This full-scale test of indoor vehicle pass-by measurements using BSS estimated the noise contribution from each tire. Comparisons to SPC modeling results were extremely positive and confirmed that the BSS method was a reliable way to separate the sources of sound and, what’s more, the separation results were superior to those obtained with SPC analysis. The Brüel & Kjær method to apply the BSS concept to determine noise source contributions of vehicles is currently patent pending.
A systematic approach to validate the estimated noise contributions is planned for spring 2015 when Brüel & Kjær will conduct a joint project with Nissan Motor Co., Ltd. in Japan. Tests will be performed to compare BSS against the traditional ‘masking’ method so that researchers can obtain direct information about the separate noise contributions.
The masking procedure involves wrapping noise sources (for example, the engine) with heavy layers or mounting silencer systems. Testing sources of sound in this way is, of course, a cumbersome contribution analysis technique that is unrealistic to use regularly, but it does offer a good way of verifying the accuracy of the BSS method.
Yoshihiro Shirahashi, Senior Engineer with the Noise and Vibration Performance Engineering Group at Nissan comments, “We have high expectations for this joint project with Brüel & Kjær and are confident that we will succeed in establishing a strong method for evaluating noise from different sources.”
Using BSS to access the partial contributions of car noises will be further validated for robustness and it is anticipated that the range of applications for BSS will expand. The BSS principle can also be applied to environmental noise from roads and airports. In this situation, the measurement microphones pick up a mixture of noise sources and BSS can be used to aid further processing and provide valuable information for traffic noise-level compliance.
BSS could also help to assess and improve the sound quality of an engine by measuring on or around the engine surface with sensors and extracting components related to hidden processes. Future plans also include pursuing cooperation with universities working on BSS methods in different contexts and attracting students to work on joint BSS projects.