Broad banded acoustic vector sensors for outdoor monitoring propeller driven aircraft

Transcription

1 Broad banded acoustic vector sensors for outdoor monitoring propeller driven aircraft Hans-Elias de Bree1,2, Jelmer Wind1, Subramaniam Sadasivan3 1 2 Microflown Technologies, The Netherlands, debree@microflown.com HAN University, Department of Vehicle Acoustics, Arnhem, The Netherlands HE.deBree@han.nl 3 Former Scientist G, ADE, Bangalore, India subramaniamsadasivan@hotmail.com Introduction Measurement case Permanent and semi-permanent sound level measurements are an increasingly common way to determine environmental noise levels. If the specific causes of sound could be identified, the measurements could be used to improve noise insulation, to tax noise polluters, and to detect possible threats. To achieve this goal, this article proposes acoustic vector sensors. These sensors measure all three components of the sound intensity and the pressure. This system is able to detect and localize noise sources in the entire frequency range. This article shows that these sensors make it possible to classify the sources and also to estimate their location. Propeller driven aircraft flying at night-time can cause nuisance in an urban environment. To be able to address a complaint it is necessary to know what aircraft (classification) was flying where and in what direction. Traditionally, only sound pressure transducers were used in acoustics. Directionality can be obtained by deploying a number of spatially positioned sound pressure transducers. The distance between the sensors determines the frequency where such a system has optimal sensitivity. For several years now, acoustic particle velocity sensors have been commercially available. Acoustic vector sensors can measure both the sound pressure and the 3D acoustic particle velocity in a single point. Such a system has the capabilities to detect noise sources across the entire audible frequency range in a 3D space. In this paper the detection, classification, localization and tracking of propeller driven aircraft is investigated. The acoustic vector sensor is able to determine the direction of arrival (DOA) of the noise that is produced by the aircraft. That information is combined with the Doppler curve, found in the time-frequency representation of the signals, in order to find the velocity, height and heading of the aircraft. Acoustic Vector Sensors An acoustic vector sensor (AVS) consists of a sound pressure transducer (a microphone) and either two or three orthogonally placed particle velocity sensors (Microflowns [1]). If the AVS is placed at a rigid surface, i.e hard ground, the three sensor version is used. This is because the particle velocity normal to the rigid surface is zero. Thus, one AVS has either three or four outputs. Acoustic vector sensors have come to play an increasingly significant role with applications focused on environmental monitoring, border control, harbour protection, gunshot localization, and situation awareness. As compared to sound pressure sensors, acoustic vector sensors have advantages acoustic bandwidth, reduced system size, low data transmission between nodes, and set up times. These benefits have led to a steep increase in practical usefulness of AVS with adoption of simple but powerful algorithms [2], [3], [4], [5], [6], [7]. The acoustic detection, classification, localization and tracking of aircraft are checked with a Mode-S receiver that decodes the radio signals that are transmitted by most aircraft. This cross check ensures the robustness of the algorithms, finds the aircrafts without Mode-S transponders and it is now possible to quantify the acoustic noise each individual plane produces on the ground: contribution of other sources (cars etc.) are excluded. An algorithm has been developed that discriminates between moving and non-moving (static) sources. This algorithm is used to filter static acoustic sources. Other sources such as cars and trucks are also acoustic sources that move. These sources are also detected, classified and discarded on the basis of their speed, sound level and altitude. Outdoor monitoring system An acoustic vector system is placed outdoors, measuring three acoustic channels with a sample frequency of 48kHz at a 16 bit resolution. The AVS is geo-referenced with a built-in GPS for localization, a 3D magnetic compass and a 3D accelerometer for orientation. The AVS module is connected by a cable to the power module; a watertight box that includes a 12V battery, a 4 channel data acquisition and an ultra miniature PC with wireless (WIFI) access. The power module measures cm and cm for the AVS module, comprising the AVS and Mode-S receiver. Figure 1: Working prototype of the outdoor AVS measurement system (Velp, NL).

2 Detection of a tonal acoustic source Detection is the first step in the procedure to find propeller driven airplanes. The use of an acoustic vector sensor is beneficial compared to a sound pressure microphone. The relevant parameters (such as heading of aircraft) are measured in a straightforward manner from the signals of the acoustic vector sensor. The cross spectrum (the real part being the sound intensity) for example can be used as the input signal for the detection. A better conditioned input signal leads to an increased detection range and less misdetection. The data rate of the measurement system is high: 23Mb per minute. The expectation is that there are only a few planes per night so the first step of the processing is the detection of a possible aircraft. It is also possible to power up the system with a trigger signal. This will be included for production versions but has not yet been integrated at this stage of R&D. A propeller driven aircraft has an acoustic signature that has a tonal character in a specific frequency band. This feature is used to detect potential aircraft. Other sources with a tonal character such as cars, motorbikes, trucks, etc. are also detected. These misdetections are discarded in the classification algorithm. Doppler algorithm: Example I A well conditioned and well described dataset is used to be able to check the Doppler algorithm and to verify a novel Doppler/DOA algorithm [4]. In the following examples more complex scenario s are described also. The Doppler frequency shift is a known phenomenon. When a sound source is approaching a listener the sound is of a higher frequency than if a source is moving away from the listener. In the time-frequency representation of the sound pressure this effect can be seen clearly. The signals of a three channel AVS caused by a low flying propeller driven aircraft are shown in the time-frequency representation in Figure 2. The upper graph shows the sound pressure signal below that the two lateral velocity signals. The harmonics of the propeller can be clearly seen as horizontal lines. The harmonics lower in frequency at 15 seconds due to the Doppler effect: the aircraft is passing. In general the sound pressure signal is cluttered with reflections, this is less observed with the velocity signals (in this example it is not seen very clearly). The sound pressure and particle velocity signals are correlated and are much better conditioned than the sound pressure signal alone. This is advantageous for the signal processing as it makes it more robust. With the equation of the Doppler shift: a) the closest point of aircraft the aircraft path; b) the true source acoustic frequency (to be able to identify it); and c) the speed of the aircraft, can all be determined. The acoustic (sound pressure) signal is not directional and it is therefore not possible to determine the heading of the aircraft. Figure 2: Measurement location (Teuge, NL). Time frequency visualization of the sound emitted by a low flying propeller driven aircraft. Upper: sound pressure. Middle and lower: lateral particle velocity, from [4].

3 Algorithms in relation with Doppler and acoustic vector sensors: Doppler/DOA It is not easy to curve-fit the Doppler shift from the sound pressure signals only. It is necessary to obtain both the highest and lowest frequencies of the Doppler shift to find a robust fitting algorithm. However, it is not possible to measure these frequencies because the sound source (airplane) is (in theory) infinitely far away, thus the signal to noise ratio at the measurement location is poor. Enhancement in relation with Doppler/DOA Detection and classification is required to be able to extract relevant information out of acoustic signals. Detection is a first step in filtering algorithms. The Doppler shift is measured three times concurrently with the AVS, (lateral particle velocities and sound pressure). With signal processing (e.g. cross spectral techniques) the signal to noise ratio increases. And of course, a clear signal increases the accuracy of any algorithm. This also increases the detection range. With an AVS the direction of arrival (DOA) from the source sound wave is directly obtained, see the upper graph of Figure 3. So the DOA of the aircraft is known regardless of the Doppler frequency shift. This is a different set of physical data (frequency for the Doppler versus signal strength ratio for the DOA algorithm) describing the speed of the aircraft combination with the aircraft-avs distance. An algorithm which uses the combination of the Doppler information and the DOA information is more robust than an algorithm based simply on Doppler information alone. The moment when the time derivative of the angle is maximal is the closest point of the aircraft to the AVS along its trajectory, see lower graph of Figure 3. This timestamp is very important for the Doppler fitting algorithm and can thus also be derived from the angle (DOA) measurement. The heading of the aircraft is directly obtained: it is perpendicular to the angle measured at the moment of the closest point of the aircraft trajectory. W N S E Figure 4: The DOA as a function of time (obtained from the data shown in Figure 3). Upper: raw data, middle: filtered response. Lower: the colour represents the angle; the brightness of the colour indicates the signal strength. The detected signal is analyzed and checked for a Doppler shift and a variation in the DOA. If these are not found in a reasonable time it can be concluded that the signal is from a static, i.e. non-moving source (e.g. a generator switched on). As can be seen in Figure 2, a continuous tone is found at approximately 115Hz coming from W-SW. A result of the static source filter is shown in Figure 4 (middle). As can be seen, the sources with a constant frequency (and angle) are removed to a large extent. Figure 3 Upper: the DOA as function of time (obtained from the data shown in Figure 3). Lower: the time derivative of the fitted angle. The maximum of this value determines the heading of the aircraft directly. In situ auto-calibration of the AVS Asymptotically the DOA of a passing aircraft converges to a 180 degree rotation and at these limits elevation is zero. These properties can be used to calibrate AVS in situ. A passing car has zero elevation and reaches a 180 degree rotation while the spectrum of the signal is broad banded. A passing car therefore can be suitable for in situ calibration.

4 Classification of a propeller driven aircraft Once a source is detected it is analyzed in the time frequency domain. From the Doppler/DOA algorithm the speed, the closest point of aircraft and the heading of the source aircraft are determined. The speed, distance and duration of the signal is used for classification. A passing car can have a speed in the order of e.g. a low flying aircraft at takeoff but the duration of such signal is very short. In other words, a plane makes a lot more noise than a car and can therefore be detected for a longer time duration. DOA Border control: Example I The AVS can be used for border control. With two lateral velocity probes and a collocated sound pressure microphone it is possible to determine the sound intensity in any direction [8]. If the source passes an AVS border control sensor, the sign of the intensity perpendicular to the border flips [4], see Figure 6. This feature provides direct tactical information. Figure 6: The intensity flips in sign if an aircraft passes a preset border [4] (obtained from the data shown in Figure 3). The upper curve represents the positive intensity and the lower curve the negative intensity. Elevation of the aircraft: Example II, Helicopter There are three independent methods of calculating the elevation of an aircraft. Firstly, one can compare the output of the particle velocity sensors to the sound pressure element. This idea is presented in [9] and is explained below. Figure 5: The Doppler/DOA algorithm is used to find the speed, heading and the closest point of aircraft. Upper: the time-frequency representation. Lower: the time-frequency-doa representation. In this example an aircraft at take-off is used, this is a data set with a good signal to noise ratio (the same data set as reported in [4]). The speed of aircraft, from the Doppler algorithm, is found to be: 116km/h and the closest point of the aircraft is 125m. If the Doppler/DOA algorithm is used a speed is found: 121km/h; and the distance is calculated to be 121m. The values of both algorithms are nearly identical indicating that the new Doppler/DOA is in agreement with the Doppler algorithm as reported in [4]. Combining the output of the Doppler and DOA algorithms however produces more (relevant) information about the aircraft: the heading and height of the aircraft. 1) The particle velocity sensors of the acoustic vector sensor system have a figure-of-eight directionality. The lateral velocity sensors have a maximal sensitivity parallel to the ground and zero sensitivity normal to the ground. The sound pressure element on the other hand is omni-directional. The sensitivity parallel to the ground is therefore the same as the sensitivity normal to the ground. The transfer function between the sound pressure element and the particle velocity sensors therefore provides a direct determination for the elevation. For lower elevation angles the normalized sensitivity of the velocity probe is almost equal to the sound pressure element. For larger elevation angles (say, more than 30 degrees) the normalized sensitivity of the velocity probe is considerably less than the sound pressure element. Therefore, the measurement of lower elevation angles is sensitive. The insitu auto calibration procedure is therefore helpful. 2) Another technique to estimate the elevation is to compare the Doppler curve and the DOA curve. The first curve provides the 3D closest point of the aircraft and the latter gives the 2D closest point of aircraft. The elevation can then be derived out of this. 3) Another idea that is reported in [4] is to use the sound level of the aircraft. The sound level should roughly fall 6dB each time as the source-to-avs distance doubles. This property, in combination with the Doppler/DOA algorithm, can be used to roughly extract the elevation of the aircraft.

5 Example III: Propeller driven aircraft measured in very unfavourable conditions The previous example was well conditioned and used to verify the algorithms. As a more challenging example an aircraft is detected during a measurement on a shooting range (shooting noise mainly south west, SW and NW). A truck with idling engine was located 40m SW from the measurement location. The measurement setup is located in a corridor through dense forest. Figure 7 Upper: time-frequency-doa representation of a helicopter. Lower: the angle derived out of the broad band signal. In the previous example the plane is taking off and the DOA has a low elevation. In this paragraph a helicopter is passing closely so the elevation of the DOA is high. In Figure 7 the time-frequency-doa representation of the passing helicopter is shown. The spectrum is tonal and broad banded. The purple lines in 0-17s, Hz (continued to 42s) are caused by a propeller plane waiting to takeoff. From the Doppler signal, a velocity of 83km/h, a closest passing distance of 92m, a heading of 6 degrees and an elevation of 89m are calculated. Using the sound level In this paper the measured sound level is not used. The property is however useful [4]. With the sound level one can determine the relative change in distance: if the sound level is increased 6dB the distance is halved. This is true under the assumption that the source is not directional and there are no reflective surfaces such as buildings. In practice it can be assumed that a plane has a directional sound radiation. Hence it is possible to extract this directionality. The sound level can be used to determine if an aircraft is approaching after detection. And once the aircraft is passed it is possible to give a real time check if the aircraft is following the (out of the Doppler/DOA) predicted track. From the Doppler algorithm one can derive the closest point of aircraft and the speed of aircraft. If the Doppler algorithm cannot be used (e.g. for a jet aircraft, a passing car), a combination of the sound level and the DOA as function of time can be used to estimate the elevation angle not to extract the closest point of aircraft. So with DOA and sound level the position of the aircraft cannot be determined. Figure 8: Time-frequency-level representation of a sound pressure signal; time-frequency-doa representation of an AVS signal (no signal enhancement processing); same but now with gunshots removed; same but now non moving sources removed.

6 Some results are shown in Figure 8. In the upper graph the sound pressure level is shown in a time-frequency representation. Below that the time-frequency-doa result of an acoustic vector sensor is shown. In the time-frequency- DOA representation, the aircraft signal shows up clearly as the blue and pink signal with the Doppler signature as it was passing S-SE to N-NE. The green horizontal line at 110Hz is the truck in idle and the vertical lines (mostly green coming from the same direction as the truck) are gunshots. In the graph below the gunshots are removed and in the lowest graph the static sources are removed from the time signals. The Doppler/DOA remains in noise with a reasonable white spectrum. This signal is well conditioned. The DOA measurement is inaccurate because of the sensor location in the trench. A Doppler analysis outcome is that the propeller driven airplane has a speed of 355km/h and was passing (CPA) at 1616m distance. Example IV: Passing car A passing car is detected and processed. As can be seen in Figure 9 (upper) most of the signal is found in the 100Hz-2kHz bandwidth and almost no Doppler shift is noticed. Only at low frequencies tonal components are found however no Doppler curve could be fitted. With an audio check it was found that the rpm of the passing car was not constant: the car was accelerating. This is an example where the Doppler algorithm can not be applied but the DOA algorithm is not affected. Conclusion In this paper a standalone single outdoor acoustic vector sensor system is presented. The system is used for the detection and classification of low flying propeller driven aircraft. A novel Doppler/DOA (direction of arrival) algorithm is presented. From the sound pressure based Doppler signal analysis alone one can derive a) the speed of aircraft, b) closest point of aircraft and c) the true source frequency. With the novel Doppler/DOA one can derive: a) speed of aircraft; b) closest point of aircraft; c) true source frequency; d) the heading of the aircraft; and e) the height/elevation of the aircraft. The airplane elevation and especially the aircraft heading are of course valuable information. References [1] H-E. de Bree et al, The Microflown; a novel device measuring acoustical flows, Sensors and Actuators: A, Physical, volume SNA054/1-3, pp , 1996 [2] T.G.H. Basten, H.E. de Bree, W.F. Druyvesteyn, Multiple incoherent sound source localization using a single vector sensor, ICSV, Krakow, 2009 [3] Jelmer Wind, HE de Bree, Emiel Tijs, Source Localization using Acoustic Vector Sensors: a MUSIC approach, submitted to NOVEM 2009, Oxford, UK. [4] Sadasivan, T. Basten, HE. de Bree, Acoustic vector sensor based intensity measurements for passive localization of small aircraft, Journal of Acoustical Society of India, 2009 [5] HE de Bree, Aero India, Bangalore, A Perspective on Acoustic Vector Sensors in Passive Surveillance - Real World Measurements, Algorithms and Applications, 2009, invited paper [6] T. Basten, HE de Bree, and E. Tijs, Localization and tracking of aircraft with ground based 3D sound probes, ERF33, Kazan, Russia, 2007 [7] Malcolm Hawkes and Arye Nehorai, Wideband Source Localization Using a Distributed Acoustic Vector-Sensor Array; IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 6, JUNE 2003 [8] H.E. de Bree, T. Basten, D. Yntema, A single broad banded 3D beamforming sound probe, DAGA, 2008 [9] H-E de Bree, C. Ostendorf, Tom Basten, An acoustic vector based approach to locate low frequency noise sources in 3D, DAGA, 2009 Figure 9 upper from left to right: time-frequency-doa representation of a passing car (colour legend same as in Figure 5). Middle: same but now in a 50Hz-100Hz bandwidth. Right: angle from the broad band signal.