Accelerometer intensity vector sensor network for environmental noise monitoring with source direction and location Jim Waite 1 Nanotok LLC Eastsound, WA 98245 USA

ABSTRACT AIVS (Accelerometer-based Intensity Vector Sensors, /āvs/) represent a new way to measure 3-d sound and are designed to integrate into existing noise monitoring solutions. Standard microphones measure sound pressure, which cannot alone deduce the direction of sound propagation. AIVS is based on the measurement of the velocity of a small parcel of air surrounding a triaxial accelerome- ter, from which a vector-based representation of sound intensity is calculated. AIVS combines a MEMS triaxial accelerometer with one MEMS microphone and synchronously measures particle ve- locity and pressure, resulting in a 3-d intensity vector at each AIVS node. An AIVS network is syn- chronized to GNSS time, facilitating collaboration among sensors surrounding and/or within a local measurement site. Low power AIVS nodes are location-aware and estimate azimuth and elevation angles to detected noise sources as a function of frequency. Range to source is computed when noise events are observed from multiple nodes with sufficient separation. AIVS nodes are managed by a Raspberry-Pi sensor hub in a wired CAN-bus supporting distances up to 100 m, or via the Bluetooth Low Energy (BLE) protocol. More widely separated nodes are joined through WWLAN technologies via the local RPi hub.

1. INTRODUCTION

One challenge with automated noise monitoring is ensuring that the measured data represent the site being monitored. Often, human listeners must review recorded noise data to determine if a source should be a part of the overall noise calculation or not. It is very time consuming to analyze these recordings and they can be inconclusive. A directional noise measurement system would reduce this effort by automatically determining which direction the noise is coming from, and further, to discrim- inate multiple noises from different directions by frequency. For instance, a 3-d understanding of the measured noise helps eliminate bird and aircraft noise from ground-based industrial noise.

Machine Learning (ML) methods are being used in conjunction with environmental noise meas- urements. This helps automate the process of determining what measured sounds are attributable to an industrial source, or are contaminated by other sound events near the site. Unfortunately, deep learning methods are encumbered by the need to train on datasets that have been annotated by human listeners. Because of this, large noise training datasets are rare and very expensive to generate, which ultimately limits the attainable accuracy of automated ML-based noise characterization.

AIVS (Accelerometer-based Intensity Vector Sensor) enabled noise monitoring systems comple- ments existing standardized Sound Level Meter (SLM) measurements by providing a three-dimen- sional vector that points to detected sound sources, as a function of frequency. AIVS can curate noise events that arise from different directions, and add richness to datasets from which ML features are extracted. A number of triaxial, frequency-specific outputs are available using AIVS technology,

1 jim.waite@nanotok.com

worm 2022

including narrowband (FFT) intensity, 1/3 octave band intensity, and Mel-Frequency Cepstral Coef- ficients (MFCC) based on active intensity. Adding a single AIVS to a SLM-based noise monitoring system increases the dimensionality of the measured data, supplementing time averaged, broadband energy estimators like L eq . In combination with powerful deep learning methods, direction-annotated noise values reduce data post-processing and increase overall test efficiency.

2. INTENSITY VIA PRESSURE-ACCELERATION MEASUREMENT

Acoustic intensity spectra presented in linear or 1/3 octave frequency bands are useful in under- standing the vector sound field, in addition to scalar pressure measurements commonly available using microphones. Unfortunately, due to their cost and need for precise calibration, triaxial intensity measurements have not been widely used in noise monitoring work, nor have noise monitoring stand- ards adopted acoustic vector sensor measurements.

Intensity can be expressed as:

𝐼 റ = 𝑝𝑢ሬറ ∗ /2 (1)

where p is the scalar pressure, 𝑢ሬറ is measure of the triaxial particle velocity, and u* indicates the complex conjugate of u . Measuring the 3-d intensity vector of a sound field as a function of frequency facilitates discrimination of noise sources by direction (azimuth and elevation angles), in addition to frequency and relative amplitude.

Until now, two methods have been used to estimate acoustic intensity. Both employ microphones to measure pressure, but differ in how particle velocity is estimated. One way is to observe the dif- ference in pressure between two closely spaced microphones, which is referred to as the p-p method of calculating sound intensity. A finite difference approximation, in which particle velocity is approx- imated by a differential in the pressure between microphones, is combined with the average pressure across both microphones to compute intensity. The microphone separation distance can restrict ap- plication of the p-p method to frequencies sufficiently low such that the wavelength exceeds this distance by a factor of five or more. Triaxial p-p probes are configured by spatially distributing microphones in three dimensions, ideally with a concentric reference point, with the microphones carefully phase matched by selection or calibration to avoid low frequency measurement errors.

A second particle velocity estimation method differentially measures the change in resistance of two closely spaced thin wires in a sound field. When sound propagates across the wires, it asymmet- rically alters the temperature distribution around the resistors, which after calibration is proportional to acoustic particle velocity. Three pairs of electrically stimulated wires, along with a microphone to measure acoustic pressure is generally referred to as a triaxial p-u sound intensity probe. When accu- rately calibrated, it can measure particle velocity over a wide frequency range [3]. Both the p-u and p-p probes arrive at the particle velocity estimate indirectly via estimation of an indirect physical process (heat advection and pressure differentials, respectively).

AIVS technology is a new third method that employs a triaxial accelerometer to measure acoustic particle velocity [1, 2] via integration. In contrast to the p-u or p-p approaches, the p-a method di- rectly measures the triaxial inertial velocity of a fixed volume of air. This volume is extended beyond the physical dimension of the accelerometer by placing the device within a very lightweight solid sphere which has density as close to air as possible. This solid volume encloses both the microphone and accelerometer, so this method offers improved shielding from environmental effects over both p- u and p-p systems. An accelerometer based acoustic vector sensor provides excellent noise source directional resolution by directly computing Equation 1, using 4-channel time synchronized meas-

worm 2022

urements of pressure, and (integrated) acceleration on each of three orthogonal axes. MEMS accel- erometers are DC-coupled, i.e., can sense low frequencies without phase error, and thus offer an improvement to finite difference techniques implemented using phase matched p-p intensity probes. Sensitivity to the Earth’s gravitational field provides a convenient and accurate calibration method.

There are p-a technology performance tradeoffs, of course. AIVS is sensitive to low frequency noise as the acceleration relative to velocity drops by 6 dB/octave toward lower frequencies. At higher frequencies the response is limited by the bandwidth of the triaxial accelerometer, typically around 2 kHz for accelerometers having acceptable noise floors. Between the low frequency noise limit of the sensor and the high frequency limit imposed by accelerometer bandwidth, AIVS probes experience a reduction in sensitivity proportional to the ratio of the density of the AIVS (including the solid body within which the accelerometer is mounted, and the accelerometer itself) to the density of air. AIVS are currently manufactured using closed-cell foam of density about five times that of air. When embedded within a lightweight foam sphere of 6 cm diameter, a triaxial AIVS weighs just over one gram, including mounting wires and a small circuit board. The resulting sensitivity loss is about 16 dB, per Equation 2. A S / A 0 is the acceleration sensitivity loss for the sensor, ρ S / ρ 0 is the ratio of the net density of the combined accelerometer and solid volume to that of air, k is the wavenumber, and a is the radius of the spherical volume [4].

𝐴 𝑠 𝐴 0

~ 3 1 + 2𝜌 𝑠 𝜌 0 Τ , 𝑘𝑎≪1 (2)

Presently, MEMS accelerometer technology is such that effective AIVS sensitivities are achievable at low cost. For a component cost of less than $40, a currently available triaxial MEMS accelerometer has a SPL noise floor of 22 dB at 100 Hz, and can observe signals up to 112 dB (or from 34 to 124 dB on high range). This dynamic range is not yet on par with professional grade pressure microphones, but a comparison with acoustic intensity is more illustrative given that source direction rather than overall acoustic energy is the primary objective. When the accelerometer is paired with a calibrated MEMS microphone, the p-a AIVS is comparable to p-p probe performance, given p-p finite difference and dynamic range limitations. The currently achievable AIVS frequency range (50 Hz to 2 kHz) is on par with p-p intensity probes having a fixed microphone spacing of about 50 mm. Notably, AIVS do not have a bias error toward the higher end of that frequency range.

3. NOISE MONITORING WITH AIVS

AIVS are composed of a collocated MEMS microphone and MEMS accelerometer and can measure sound pressure as well as acoustic intensity. However, due to aforementioned frequency and dynamic range limitations, AIVS cannot emulate a Sound Level Meter (SLM) and thus cannot meet international standards, such as ISO 1996 or ISO 61672. It does, however, excel as a complement to an existing noise monitoring solution by adding automated noise source direction and an estimate of range to source (for a distributed multi-sensor AIVS system). Results from a simple bearing angle test are presented in this section to highlight the core measurements provided by a single AIVS Node. First, we describe the architecture of an AIVS Node.

3.1. AIVS Nodes

The sensor portion of AIVS, consisting of a microphone and triaxial accelerometer, is integrated with processing hardware into a single small, battery-operated package known as an AIVS Node (Figure 1). AIVS Nodes are location-aware, include a low power signal processing and networking

worm 2022

framework, are synchronized to each other and satellite-based GNSS time, and can be deployed in small groups surrounding and/or within a local measurement site. AIVS nodes connect to each other, all of which are managed by a Raspberry-Pi (RPi) sensor hub in a wired CAN-bus supporting distances up to 100 m. Alternatively, the nodes communicate via the Bluetooth Low Energy (BLE) protocol, or are joined through WWLAN technologies from the local RPi hub. AIVS and its processing engine consume very little power (100 mW) even when processing full 3-axis intensity spectra and broadcasting data to the network hub.

worm 2022

An AIVS provides data interfaces to SLMs via RS-232/Bluetooth (up to one per node), or USB/Ethernet/WiFi via a Raspberry Pi interface (one per AIVS network). Thus, an environmental noise engineer can configure a noise monitoring system such that a collaborating set of AIVS nodes observes the same acoustic data as does the SLM sited nearby. The SLM and AIVS network are synchronized using GNSS timestamps, and multiple AIVS collaborate together to estimate the location of the noise source event. Even if there is only one AIVS combined with one SLM, azimuth and elevation bearing angles to the noise event are tagged to the SLM L eq measurement of the same event, including use of frequency as an event discriminator. Such scenarios can be thought of as a distributed acoustic camera for environmental noise data. Each AIVS is position and orientation aware in a geospatial frame, so noise sources are annotated in that frame as well. While a SLM makes standardized L eq measurements at a site, an AIVS network expands the metadata associated with those measurements by including azimuth and elevation angles to the detected source, possibly including range to source when observed across multiple AIVS.

Figure 1: The 24-cm high AIVS Node is a tripod-mounted integrated package including the MEMS sensor suspended above a small electronics enclosure. The hardware block diagram is anchored by an IoT processor with an ARM Cortex M4 and floating point unit (FPU).

The key advantage is that AIVS adds value to existing SLM-based environmental noise monitoring scenarios, without compromising the standardization benefits inherent from using SLMs. Interfaces on AIVS and/or the RPi support integration of a SLM into an ARES network, or conversely an ARES node (or network of nodes) into an existing SLM-based noise monitoring system, since many

manufacturers already support some form of networking between the SLM and a back-end cloud processing/analysis system.

worm 2022

3.2. Bearing Measurements

Short range calibrated angle tests were conducted outdoors, with two speakers and the sensor- source geometry as shown in Figure 2 looking down from above into the positive z-axis.

1

2

Figure 2: AIVS source direction test setup.

Source #1 on the left was swept from 220 to 2500 Hz over 30 seconds, while source #2 on the right was simultaneously swept in the reverse direction, from 2500 to 220 Hz such that the sweeps cross at about 1300 Hz. In Figure 3, the azimuth angles are correctly estimated by both sensors (SN1, SN3) except where the source frequencies overlap. Good performance is observed up to al- most 2000 Hz.

Figure 3: Azimuth angle estimation results.

In this test, the true elevation angle is zero for both sensors, to both sources, since all were set on the horizontal plane. As shown in Figure 4, the elevation angle estimates are clustered around zero (±20°), with some unexpected biases above 1500 Hz. Below 1500 Hz, deviations from zero are due to the fact that horizontal plane of the mounted foam sphere is not necessarily level. This is a static, manufacturing-time rotational calibration not yet implemented at the time of publication. Further, the results for both azimuth and elevation are not accurate below 300 Hz due to reverberation and reflections at the outdoor test site.

The elevation discrepancies above 1500 Hz are either related to calibration or accelerometer per- formance. In fact, the MEMS accelerometer is only specified to 1500 Hz in the vertical axis, since the z -axis has a somewhat lower resonant frequency than do the x - and y -axes.

Figure 4: Elevation angle estimation results.

worm 2022

3.3. Active and Reactive Intensity

For the same dataset as presented in the last section, we compute active and reactive intensity as observed at both AIVS SN01 and SN03. Except at low frequencies, and where the sine sweeps cross in the region of 1100 to 1400 Hz, the active intensity is about 10 dB above the reactive (non-propa- gating) component. This ratio can be used as a field indicator, and the factor of 10 lends confidence to the bearing measurements. Below 400 Hz, the sensors are in the near field of the sources, and the reactive component (as expected) increases relative to the active intensity.

Figure 5. Active and reactive intensity for the source direction tests.

3.4. Environmental Noise

A segment of a noise monitoring session collected from the rooftop AIVS Node pictured in Figure 1, is shown in Figure 6. The node was positioned 8 meters above street level. The system monitored a leaf-blower noise event in a park across the street in the y -direction at 22-meter range. During the 16-second time segment, the landscaper walked toward and into the street (in a negative- y direction), momentarily releasing the throttle such that the noise level dropped, and then upon reactivating the throttle, returned to the park while moving slightly in the positive x -direction.

The Figure 6 spectrograms ( x -axis not shown since nearly all the signal is in the y - z plane) represent a family of harmonically related leaf blower signatures, with a 5-second gap when the throttle is released. Bearing angle calculations are disabled when the pressure signal is below a threshold of 45 dB SPL. Triaxial intensity is computed in the frequency domain according to Equation 1, then for every pressure bin above the threshold the azimuth and elevation are computed. The values shown in the lower right of Figure 6 are for the 500 Hz harmonic of the leaf blower. Other harmonics provide similar results.

worm 2022

Figure 6: Pressure and velocity spectrograms, with detected bearing angles for leaf blower test.

3.5. Multiple AIVS Nodes

An AIVS network composed of multiple AIVS nodes deployed within or around a monitored site can observe the same event from different perspectives. This enables estimation of range to the noise source via triangulation algorithms. Figure 7 depicts the components of a multi-AIVS Node system. When the nodes are connected on a local CAN bus, accurate GNSS-derived baselines are automati- cally determined using a so-called “moving baseline” algorithm, implemented in the GNSS module. Each AIVS node computes its centimeter-accurate 3-d geospatial coordinates based on these base- lines and a reference position.

Figure 7: Typical deployment of a network of AIVS Nodes, managed by a Raspberry Pi hub.

The AIVS network hub, normally a Raspberry Pi with a CAN-bus interface, collects event data from AIVS nodes consisting of detected noise events, distinguished by their time of occurrence, fre- quency, and computed bearing angles (azimuth and elevation). Simultaneous events detected at more than one node are collected and tagged with the 3-d positions of the detecting nodes. This data is the input to a triangulation algorithm that computes the expected 3-d position of the noise source, along with a confidence interval. Simultaneous detections of multiple sources are possible, assuming some segmentation by frequency.

4. CONCLUSIONS AND FUTURE WORK

By relying on low-cost MEMS devices, triaxial acoustic intensity measurements are now practical to deploy in clusters of nodes for monitoring noise events across a network. The AIVS Node couples on-board floating-point processing to the intensity sensor, resulting in bearing angle estimates per node in addition to acoustic energy as a function of frequency. When associated with noise events, this additional metadata increases feature dimensionality, while lowering the costs of automatic data collection. Inexpensive deployments of AIVS Nodes across an environmental noise monitoring site effectively act as a distributed acoustic camera, providing a 3-d perspective of the noise.

A natural extension of the work is to perform noise source tracking for fast moving sources. Each node self-synchronizes to PPS (pulse-per-second) timing events issued by the embedded GNSS mod- ule with an accuracy of about 1 µs. Model-based Extended Kalman Filters (EKFs) or particle filters implemented on the AIVS ARM-M4 processor provide bearing angle estimation at high rates (50 Hz or more). Nanotok has prototyped EKFs for an AIVS node that can track small quadcopters over about a 50-meter range (each node), making possible continuous tracking using an AIVS network- based passive-acoustic tracking system with node spacing of about that same distance, depending on background noise levels. Tracking would thus permit an AIVS network to serve as an acoustic com- plement to a multi-modal monitoring system deployed for critical infrastructure protection.

Louder aircraft noise events from jet flyovers, takeoffs, and landings are observable from much greater distances, which require larger inter-node geometries among collaborating nodes in an AIVS network. In these cases, the nodes will still be synchronous (via GNSS), but must rely on WWLAN technologies to establish connectivity rather than the local wired CAN-bus. 5. ACKNOWLEDGEMENTS

AIVS technology development is funded in part by contracts W911NF1520082, W911QX20P0032, and W911QX21C0020, awarded to Nanotok and the Applied Physics Laboratory at the University of Washington by the U.S. Army Research Laboratory. The U.S. Government has certain rights in the technology. 6. REFERENCES

1. Dall’Osto, D., et al. Airborne vector sensor experiments within an anechoic chamber, J. Acoust.

Soc. Am . 144 , 1854 (2018). 2. Dall’Osto, D., & Dahl, P. Preliminary estimates of acoustic intensity vorticity associated with a

turbine blade rate, J. Acoust. Soc. Am. 142 , 2701 (2017). 3. Jacobsen, F., and de Bree, H. A comparison of two different sound intensity measurement prin-

ciples, J. Acoust. Soc. Am. 118 , 1510 (2005). 4. Williams, R., et al. Towards acoustic particle velocity sensors in air using entrained balloons:

Measurements and modeling, J. Acoust. Soc. Am. 143 , 780 (2018).

worm 2022