Research

Research progress

We have been working on the following areas:

Large scale noise source mapping

A newly proposed GMPSC-DAMAS algorithm based on a deconvolution approach

Environment noise pollution in large urban areas becomes more and more serious and is known to produce significant adverse impact on health and longetivity. To effectively address the noise problem, locating environment noise sources and measuring their levels on a city- or even nation-scale is essential. The existing approaches of deploying dense microphone arrays spanning the entire region of interest, or sequential noise measurements at thousands of locations on a dense grid on this scale would be prohibitively expensive. In our previous report, we reported our newly proposed acoustic measurement scheme using a small movable array to rapidly acquire measurements at many different locations, creating a kind of non-coherent virtual array of much larger aperture. We also reported our newly proposed multiple-point sparse constrained deconvolution approach for mapping acoustic sources (MPSC-DAMAS) and multiple-point covariance matrix fitting (MP-CMF) approach.

However, MPSC-DAMAS and MP-CMF produce more errors for mapping noise sources when the number of noise sources is large, or when there are noise sources near each other, or when the ambient noise level is high. During this reporting period, we improved MPSC-DAMAS by proposing a generalized MPSC-DAMAS (GMPSC-DAMAS) approach derived based on a generic beamformer and an explicit measurement noise model. Two implementations of GMPSC-DAMAS were evaluated using the DAS beamformer and the minimum-variance-distortionless-response (MVDR) beamformer. We further proposed a new parameter-setting method based on a multiple-point MVDR (MP-MVDR) beamformer, which is not restricted to the normalized steering vectors as in the study for SC-DAMAS and has no requirement of the number of sources as in the study for SC-RDAMAS. The new GMPSC-DAMAS approach was evaluated by simulations and real data.

Performance evaluation for GMPSC-DAMAS using simulations

In the simulation setup, we considered a 30m x 20m environmental noise region, and the scanning locations were set on a 1m x 1m grid. There are a total of 600 source locations as shown in Figure 1. The x-axis was considered as the horizontal direction and the y-axis was the vertical direction. A circular array consisting of 24 microphones with a radius of 0.72m was placed at 7 locations for data acquisition. The coordinates of the array sensing locations were {(0, 10), (5, 10), …, (30, 10)} where the heights of the sources and the array were assumed the same and omitted. The acoustic sources and the additive noises were synthetic complex Gaussian zero-mean signals and the frequency of interest was 1kHz. A total of 1000 FFT segments were used at each sensing location. For all simulations, the source positions were randomly generated and the results were averaged over 500 instances. The measured absolute power error is defined as the logarithm of the absolute error.

Figure1

Figure 1. Illustration of the tested scanning region and the noise-mapping results of MP-MVDR, MPSC-DAMAS, GMPSC-DAMAS(DAS) and GMPSC-DAMAS(MVDR).

Figure 1 shows the noise-mapping results of the tested algorithms for a setting of 20 noise sources with equal source power of 85 dB and the signal to noise ratio (SNR) of 35 dB. All algorithms have accurate location estimates. MPSC-DAMAS, GMPSC-DAMAS(DAS) and GMPSC-DAMAS(MVDR) have higher accuracy for the power estimation than MP-MVDR.

 

 Figure2  Figure3
Figure 2. Absolute power error comparison of noise mapping with different number of acoustic sources.
Figure 3. Absolute power error comparison of noise mapping with different SNRs.

 

Figure 2 shows the absolute power errors obtained by the four algorithms. The number of sources was 10, 20, 30, and 40 with equal source power of 85 dB and SNR of 35 dB. It is observed that MP-MVDR has the highest absolute power errors. The other three algorithms achieve lower absolute power errors than MP-MVDR. Compared to MPSC-DAMAS, GMPSC-DAMAS(DAS) and GMPSCDAMAS(MVDR) achieve much lower absolute power errors. GMPSC-DAMAS(MVDR) has the lowest errors.

Figure 3 shows the results of 20 sources with equal power of 85 dB and SNR ranging from 5 dB to 35 dB. Similar observations are obtained that GMPSC-DAMAS(DAS) and GMPSCDAMAS(MVDR) achieve much lower absolute power errors than MPSC-DAMAS, and GMPSC-DAMAS(MVDR) has the lowest error. The above results also imply that GMPSC-DAMAS with adaptive beamformers can outperform MPSC-DAMAS with fixed beamformers, and MP-MVDR works well for the parameter settings. The averaged computational times were 0.13s for MP-MVDR, 34.76s for MPSC-DAMAS, 17.28s for GMPSC-DAMAS(DAS), and 16.69s for GMPSC-DAMAS(MVDR) on a 64-bit personal computer with a 3.0 GHz processor and 10 Gbytes of random access memory (RAM) running MATLAB.

Performance evaluation for GMPSC-DAMAS using real data

To evaluate the performance of GMPSC-DAMAS for the real situations, we built a data collection system using a 24-channel microphone array as shown in Figure 4. The system consists of an NI cDAQ-3919, 24 GRAS40 PH microphones, a portable UPS, and a monitor.

Figure4

Figure 4. A movable array setup for data collection of environmental noise sources.

The data collection was made in an open area in the front of Fusionopolis. The noise sources were played from a loudspeaker as shown in Figure 5. The experimental setting was shown in Figure 6.

 Figure5  Figure6
Figure 5. A noise source playback system using a loudspeaker and a laptop.
Figure 6. Experimental settings for data collection using a movable array.

Two loudspeakers were placed at +15m and -10m in the y direction. Several audio clips for different environmental noises including construction noise, engine noise, and jack hammer noise, etc. were played back where the spectra of the sources are illustrated in Figure 7. The microphone array was moved from 0m to 48m with a step of 8m in the x direction. At each array location, the audio clips were played and recorded individually.

 Figure7  Figure8
Figure 7. Spectral diagram of noise sources for in the data collection.
Figure 8. Illustration of noise mapping using GMPSC-DAMAS for two noise sources at 1 kHz.

Two different noise sources were mixed at +15m and -10m to create a two-source scenario. GMPSC-DAMAS was implemented on the real source data and one of the resulting maps is illustrated in Figure 8. The noise power level is the value directly resulting from the microphone output. Further calibration is required to convert the value to the sound pressure level (SPL). From the map, the noise sources are correctly located.

Acoustic noise-source mapping with beamforming tomography

We have made further improvements on our developed static tomographic imaging approach for acoustic noise-source mapping using multiple-point measurements. We consider acoustic delay-and-sum beamforming power outputs as tomographic measurements, which are computed based on the sound-pressure levels measured via a moving microphone array at multiple locations. To obtain soft-sparse solutions for the acoustic noise mapping of stationary localized noise sources, we implement a regularized version of the FOCUSS algorithm with a smoothness constraint. In September, we performed our initial outdoor experiments with the new data acquisition system we have recently purchased. We used a 24-channel microphone array being composed of two circular 12-channel subarrays as illustrated in Figure 1, and repeated measurements seven times at locations evenly separated by 8m. A preliminary experimental result from the latest experiments is also shown in Figure 9 for a stationary acoustic noise-source located at the xy-coordinates (24m, 15m). The imaging results demonstrate that the newly developed method is capable of localizing the acoustic noise-source accurately while simultaneously recovering a sound-pressure level being very close to the value measured with a SPL-meter separately.

 Figure9
Figure 9. Experimental results for sparse DAS-beamforming tomography for a stationary acoustic noise-source (dB(A): A-weighted sound pressure level(SPL)).

 Dynamic noise-source mapping of transient and moving sources

 In this reporting period, we have also been working on a dynamic imaging model for the detection of spatially-fixed time-varying (transient) acoustic noise-sources such as construction noise, and tracking of moving noise sources such as the traffic noise generated by driving vehicles using two fixed microphone arrays positioned at opposite ends. We use a Kalman-filter-based approach to recursively estimate the 2-D images of the acoustic field using MVDR beamformer outputs. Figure 10 shows the initial results of the indoor experiments performed at ADSC using two fixed 4-channel linear microphone arrays for the transient acoustic noise-source scenario simulated by hand-clapping. We also collected measurements during our initial outdoor experiments in September using two 8-channel microphone arrays for different scenarios including a moving white-noise source played from a loudspeaker, a transient moving noise-source generated by pot-banging, and the traffic noise generated by riding a motorbike. We are currently analyzing this recently collected data to further improve our model. 

Figure10

Figure 10. The initial experiments for dynamic MVDR beamforming tomography.

 
 Implementation of SRP direction finding method for I2R’s EARS robot

Besides the above proposed eigenvector clustering approach, we also implemented an alternative method for direction finding of multiple speech sources based on the microphone array setup on the EARS robot, which is an attention-directed telepresence robot developed in I2R. The configuration of the microphone array is illustrated in Figure 11. The array consists of 8 omnidirectional microphones, distributed in 3-dimensional space of size 15x9x4.5 cm. SRP is a class of methods that are based on maximizing the steered response power of a beamformer across the location space to find the source location.

In our implementation, the chosen location space is a sphere with radius of 1 meter from the center of microphone array. The SRP-PHAT outputs are projected on this sphere. The peaks of the spherical map determine the sound location. We implement our algorithm in Matlab. The input data is divided into 0.2 second frame. The time it takes to process one frame of data is 0.1s. Thus this algorithm is promising to be processed in real time. Figure 12 shows the SRP maps of real data recording two female speakers in a normal office room. The approximate directions of the speakers are 45 degree azimuth, 30 degree elevation and -45 degree azimuth, 45 degree elevation respectively. The power map illustrated a 3D hemisphere using 2D concentric circles. The map consists of 9 concentric circles corresponding to 9 elevation angles from 0 to 80 degree. The elevation angle matches the order of the circle. Each circle is plotted on the xy plane. The azimuth angle is measured from the positive x axis. The SRP map in Figure 1 shows that the algorithm is able to detect two sources with acceptable accuracy.

 Figure11  Figure12
Figure 11. The configuration of microphone array on EARS robot.
Figure 12. Direction finding of two concurrent speech sources using SRP implementation.

Research plans for the next 6 months