Open Access
Add to BibSonomy
Abstract
A sound source direction-of-arrival (DOA) estimation method for microphone array based on ultra-weak fiber Bragg grating (UW-FBG) distributed acoustic sensor is proposed. The principle of acoustic signal demodulation is introduced, the sound pressure sensitivity and frequency response range of a single UW-FBG microphone are analyzed, and a series linear array with three UW-FBG microphones is designed. Combined with convolutional recurrent neural networks, the DOA estimation method is developed. Log-Mel spectral features and SCOT/PHAT joint weighting generalized cross correlation features are used for DOA estimation. The corresponding system is established and experimentally verified. Results show that the measured sound pressure sensitivity of the UW-FBG microphone is in the range of 0.1032–3.306 rad/Pa within the frequency range of 1000–3000 Hz, and the peak sound pressure sensitivity is about 3.306 rad/Pa. The estimated mean error of 2D DOA estimation is about 2.85°, and the error of 3D DOA estimation is about 5.465°. This method has good application prospects in distributed sound source localization.
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Direction-of-arrival (DOA) estimation is the focus of sound source localization, speech enhancement, and position tracking fields [1–3]. Acoustic signals can compensate for the defects of blind areas in video surveillance, especially in environments where video surveillance is impossible [4], such as an environment with insufficient light, nighttime environment, an environment with many occlusions, and an environment wherein electromagnetic wave communication cannot be conducted. DOA estimation also has broad application prospect in intelligent transportation, industrial monitoring, military application, and smart city construction [5–8].
A microphone is an important pickup tool for acoustic signal. However, the traditional electronic microphones are susceptible to electromagnetic interference and have poor concealment, which cannot meet the needs of engineering applications. In recent years, optical fiber microphones based on optical fiber acoustic sensing technology have developed rapidly. They can be divided into point-type and distributed-type. Point-type optical fiber microphones generally sense the external sound pressure signal through the diaphragm, thus causing the phase change of optical signal in the optical fiber. The acoustic signals are picked up by demodulating the phase changes. For example, the diaphragm-type fiber microphone constructed by Z. J. Xia et al. requires placing the fiber into a capillary tube with an inner diameter of 126 µm, then sequentially bonding into the capillary with an inner diameter of 1.81 and 3.4 mm, and finally adhering the PET film to the end face and curing with UV. The sound pressure sensitivity of this microphone is stable at signals of 1–10 kHz, and the sound pressure sensitivity of the measured signal of 3 kHz can reach 0.63 rad/Pa [9]. Although the microphone is small and sensitive, its structure is complex and only a single probe can be made by a single fiber. In addition, the number of channels of the array constructed by this microphone is limited [10,11]. Distributed-type optical fiber microphones generally perceive the external sound pressure signal through the acoustic sensitization structure, thus causing the phase change of the optical signal in the fiber wrapped around the structure. The acoustic signals are inverted by demodulating the phase changes. For example, T. S. Jin et al. referred to the Sagnac interference principle and the time division multiplexing method to construct an all-fiber microphone array for sound source location. The array was constructed by delaying optical fiber coils and microphones, and multiple couplers were used to realize the parallel multiplexing of optical fiber microphones [12]. This kind of microphone has high sensitivity and simple structure but large volume. In addition, although the constructed array overcomes the limitation of the number of channels, the structure is complex, the cost is high, and it is difficult to expand [13,14]. Therefore, engineering application requires constructing a single-channel distributed fiber microphone array that has a simple architecture and allows for easy expansion without restriction of channel number. For example, in logging while drilling applications, the acoustic signal between the casing string and the surrounding geological layer is detected to determine the location of the lithological changes during the drilling process, or to track and locate the drill bit [15,16].
The proposed DOA estimation method has obvious advantages in angle accuracy acquisition. Traditional DOA estimation methods, including beamforming, time difference of arrival, and multiple signal classification, are susceptible to environment and noise and cannot maintain high positioning accuracy in practical application [17,18]. Therefore, a series of accuracy improvement algorithms has been proposed based on traditional methods, among which the localization method of neural network plus traditional features has become a research hotspot [19–21]. Y. Cao et al. proposed a polyphonic event detection and location method, which uses a convolutional recurrent neural network (CRNN) to train the Log-Mel and generalized cross correlation (GCC)-PATH features of acoustic signals to realize the detection and DOA estimation accuracy of various sound signals, and the estimated error is 9.85° [22]. D. Q. Zhong et al. designed a multi-channel CNN + DNN, simulated the simultaneous existence of multiple array errors and coherent signals, and improved the accuracy of signal DOA estimation to more than 95%. However, the network did not carry out relevant verification and testing of actual signals [23].
To solve the above problems, a series linear array with three ultra-weak fiber Bragg grating (UW-FBG) microphones is constructed using UW-FBG with low reflectivity and high reuse ability and an acoustic sensitization structure. On the basis of this hardware structure, signal preprocessing was added and improved features were selected to achieve high-precision DOA estimation of the sound source. It will have good development prospects in engineering application.
2. Principle
2.1 Principle of demodulation and UW-FBG microphone
In this study, a system is built based on UW-FBG distributed acoustic sensor (DAS) technology [24] to estimate the DOA of sound sources (Fig. 1). After passing through a narrow-linewidth laser (NLL), a semiconductor optical amplifier (SOA), and an erbium-doped fiber amplifier (EDFA), the optical signal is transformed into a high-power optical pulse signal with a specific pulse width. The acquisition of acoustic signal is realized by constructing a series linear array with three UW-FBG microphones. Acousto-optic demodulation mainly refers to the transmission and interference of optical pulse signal, photoelectric conversion and acquisition, and signal processing. The optical pulse signal is output to the UW-FBG microphone array by circulator 1 (CIR1), and the transmitted signal is transported to the coupler by circulator 2 (CIR2) for interference. The signal after interference is divided into three channels and output to the photodetector (PD) for conversion into electrical signals, which are then collected using analog digital converter (A/D). The entire demodulation process is hardware controlled and configured through FPGA + ARM, and data processing is also carried out within it. The principle of the system is that when the optical microphone array is subjected to the external sound pressure signal, the phase of the optical pulse signal in the optical fiber will change. After the distance compensation in the coupler, the interference can realize the phase–intensity conversion of the acoustic signal to the optical wave signal to reverse the acoustic signal.
The structure of the UW-FBG microphone mentioned above is shown in Fig. 2, which consists of UW-FBG and a sound sensitized thin-walled cylinder. The reflectivity of UW-FBG is -40 dB. The extremely low reflectivity makes it possible to achieve the same multiplexing rate, UW-FBG has much lower power and bandwidth requirements for light sources compared to FBG [25]. At the same time, it is more convenient to inscribe multiple UW-FBGs on the same fiber than to inscribe FBGs. UW-FBG's advantage of high-density multiplexing provides us with the conditions to build our series array. Considering that using UW-FBG alone cannot realize the efficient acquisition of acoustic signals, an acoustic sensitization structure [26] is used to build a high-sensitivity sensor to improve the picking ability of optical fibers for acoustic signals. The single-mode UW-FBG with high bending fatigue resistance with a grating spacing of 5 m is tightly wound onto a polycarbonate cylinder. It has a cylinder length of 100 mm, an outside diameter of 20 mm, and a wall thickness of 0.5 mm. The structure is subjected to external sound pressure in the air, which can cause elastic deformation of the cylinder and then result in elastic deformation of the optical fiber, thereby achieving the detection of acoustic signals.
The sound pressure sensitivity and frequency response range of the UW-FBG microphone are tested and analyzed, and the frequency response characteristics of the microphone and the optical fiber ring are compared (Fig. 3(a)). The UW-FBG microphone is sensitive to 1000–3000 Hz signal response, in which the sound pressure sensitivity ranges from 0.1032 rad/Pa to 3.306 rad/Pa. The sound pressure sensitivity of a signal with a frequency of 1600 Hz is 3.306 rad/Pa (Fig. 3(b)).
Fig. 3. Characteristic curve of the UW-FBG microphone ((a). frequency response characteristic curve (b). sound pressure sensitivity).
2.2 DOA estimation model and principle analysis
The DOA estimation model is constructed through a series linear array with three UW-FBG microphones (Fig. 4). The three UW-FBG microphones (M1, M2, and M3) are arranged in an isometric manner with a spacing d of 0.5 m. The coordinate axis is established with M2 as the reference center, and the distance from the sound source to M2 is 1 m. Given that the distance between the sound source and each sensor differs, the time for the acoustic signal to reach the sensor is different, resulting in different delay differences. At the same time, the signals received by each sensor have differences. Therefore, the direction of sound source can be estimated by analyzing these differences.
The azimuth is denoted as α (0°–180°), and acoustic signals are denoted as β (0°–90°). The distances (L1, L2, L3) from the sound source to the sensor are
Different delay differences (Δt12, Δt23, Δt13) can be obtained as follows:
Equation (7) and Eq. (8) indicate that Δt12 first increases and then decrease as β increases, and then decreases as α increases monotonically. When α and β are both greater than 0° but less than 90°, both Δt12’(α) and Δt12’(β) are always greater than 0 and monotonically increasing. Hence, Δt12 is monotonically increasing, and the curve gradually steepens in this interval. When β is greater than 0° but less than 90° and α is greater than 90° but less than 180°, Δt12’(α) is always greater than 0 and monotonically decreasing. Thus, Δt12 is monotonically increasing, and the curve gradually flattens out with α in this interval. Δt12’(β) is always less than 0 and monotonically increasing, so Δt12 is monotonically decreasing and the curve gradually steepens with β in this interval. In summary, when DOA is estimated, Δt12 is less obvious at both ends of α distribution, and β is larger, consequently, it will not be accurately estimated.
To intuitively demonstrate the relationship between the time delay difference of azimuth and elevation, a simulation model is established (Fig. 5). The relationship between the delay difference of M1 and M2 and different azimuth and elevation is illustrated (Fig. 5(a)). When the elevation remains constant, the value of delay difference first decreases and then increases as the azimuth increases. When the azimuth is constant, the value of delay difference decreases monotonically as the elevation increases. As shown in Fig. 5(b), the delays of curves with elevations of 0° and 10° are close to each other, the delays of curves with elevations of 70° and 80° are extremely small, and the delay of the curve with an elevation of 90° is zero. When the azimuth is 90°, elevation change will not cause delay difference change. So precise positioning cannot be achieved in these directions. Therefore, this model divides the DOA into 2D and 3D. The 2D azimuth range is from 0° to 180° with a step size of 10°, and the 3D azimuth range is from 0° to 180° and the elevation range from 10° to 60° with a step size of 10°.
Fig. 5. Simulation diagram of DOA estimation model ((a). The 3D simulation diagram (b). Simulation diagram of azimuth delay difference at different elevation)).
3. DOA estimation algorithm
To realize the DOA estimation of 2D and 3D acoustic signals, a DOA estimation algorithm is established based on CRNN (Fig. 6). This algorithm consists of two steps: one is training the model, and the other is testing it. Array UW-FBG microphones are used to obtain the acoustic signals of each training position, then the obtained acoustic signals are preprocessed. Afterward, the features are extracted and are input into the CRNN for feature learning and feature model generation. During the test, after the acoustic signals from a certain location are received through the UW-FBG microphone array, feature extraction is performed. Subsequently, the obtained feature data are input into the trained model to estimate the DOA of acoustic signals.
The CRNN algorithm is stable and has good detection and location effect in the field of sound source detection and location. Therefore, a CRNN is selected for DOA estimation in this study. Its structure is shown in Fig. 7 [22]. Input features are very important for networks, and good input features can help improve algorithm accuracy and achieve better results. However, the noise resistance of GCC-PATH feature is poor. In order to maintain signal quality, the SCOT/PHAT joint weighting GCC feature [27] is selected as one of the input features, because it is more noise resistance than the GCC-Path feature. The four-layer 2D CNN convolves the input features with a convolution kernel size of 3 × 3 and activates them by using the rectified linear unit (ReLU). The pooling layer uses 2 × 2 maximum pooling. The fourth set of pooled results is input into the bidirectional gated recurrent unit (GRU) through the hyperbolic tangent (Tanh) to obtain information in time dimension. After the DOA estimation of the fully connected layer learning, the model is finally output.
4. Experimental results and discussion
To verify the effectiveness of this line microphone array in the application of sound source location, an experimental platform is built (Fig. 8). The experimental platform consists of the array structure and a data acquisition device. The array structure is composed of three UW-FBG microphones with equal spacing (50 cm). The data acquisition device is realized using a UW-FBG DAS demodulator. The demodulator sampling rate can reach up to 33 kHz, which meets the requirements of acoustic signal acquisition, and the minimum time delay of data acquisition can reach 1/33 ms. The sound signals collected by the demodulator consist of signals from whistles, 110 sirens, and 119 sirens connected together. The total duration of the signal is 19 seconds, and the main frequency distribution range of the signal is 569.42-3351.42 Hz, 426.26-5665.42 Hz, and 605.50-2885.75 Hz, respectively.
In the experiment, the sensor position is kept unchanged, while the orientation of the sound source is changed by controlling the height and position of the tripod, thereby achieving the collection of acoustic signals from different directions and changing the playback order of the sound for repeated acquisition. The acoustic signal measured at an azimuth of 180° and an elevation angle of 10° is shown in Fig. 9(a), including a three-channel waveform and a single degree waveform for each channel. Channels 1, 2, and 3 respectively represent the signals collected by M1, M2, and M3, where the X-axis represents time, and the Y-axis represents amplitude. The red box range shown in Fig. 9(a) is enlarged locally to obtain Fig. 9(b). The signals received by the three UW-FBG microphones exhibit significant amplitude and delay differences.
Fig. 9. Schematic diagram of measured acoustic signal ((a). Waveform diagram of measured whole signal (b). Waveform diagram of partial amplification of measured signal).
The actual angles, predicted angles, and errors of the signals are shown in Table 1 and Table 2. According to the analysis of measurement data in Table 1, the average error of DOA estimation for 2D sound source is approximately 2.85°. From the analysis of measurement data in Table 2, the average error of DOA estimation for 3D sound source is about 5.465°. That is, the average error of azimuth is about 3.52°, and the average error of estimated angle is about 7.41°. In addition, to visually observe the DOA estimation results, an error distribution map is drawn between the DOA estimation angle and the actual angle (Fig. 10 and Fig. 11). The data results and error distribution map indicate that the 2D and 3D DOA estimation angles are close to the actual angles, which proves that the sound source DOA estimation method based on UW-FBG microphone array can achieve effective 2D and 3D DOA estimation.
Table 1. 2D DOA estimation of acoustic signal
Table 2. 3D DOA estimation of acoustic signal
The results of DOA estimation of 2D sound sources are visualized in Fig. 10, where the X-axis represents the actual angle and the estimated angle, and the Y-axis represents the estimation error. According to the simulation in Fig. 4, the delay difference of adjacent signals first increases and then decreases with the increase in azimuth. From Fig. 10, the signal error is large when the azimuth angle is 0°–40° and 140°–180°, whereas the signal error is small when the azimuth is 50°–130°. The actual measurement results are consistent with the simulation results.
The results of DOA estimation of 3D sound sources are visualized in Fig. 11. The XY plane shows the actual angle and estimated result of elevation and azimuth, and the Z-axis represents the estimation error. According to the simulation in Fig. 4, the delay difference of adjacent signals first increases and then decreases with the increase in azimuth and elevation. From Fig. 11, the signal error is large when the azimuth is 0°–40° and 140°–180° and the elevation is at 10° or 20°. When the azimuth is 50°–130° and the elevation is 30°–60°, the signal error is small. The actual measurement results are consistent with the simulation results.
5. Conclusion
In this study, a sound source DOA estimation method for microphone array based on UW-FBG DAS is proposed and verified. A series linear array consisting of three high sensitivity UW-FBG microphones is constructed by combining UW-FBG with an acoustic sensitization structure, and high-precision DOA estimation is performed using a CRNN. The results show that this method can effectively realize 2D and 3D DOA estimation. The average error of 2D estimation is about 2.85°, and the average error of 3D estimation is about 5.465°, indicating a high estimation accuracy. This method uses a simple linear series array structure to achieve the function of multi-channel parallel array structure. Compared with multi-channel parallel array structures, the circuit complexity and manufacturing process are significantly reduced, and it is easy to expand the number of sensors. Combined with the high-precision algorithm, precise DOA estimation of acoustic signals can be achieved. This work can provide a good idea for distributed sound source localization.
Funding
National Key Research and Development Program of China (2021YFC3001903).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. R. Wang, B. N. Khanh, D. Morikawa, and M. Daisuke, “Method of estimating three-dimensional direction-of-arrival based on monaural modulation spectrum,” Appl. Acoust. 203, 109215 (2023). [CrossRef]
2. Q. Shi, H. T. Chen, and P. Y. Zhang, “Direction-guided speech enhancement method of the spatial mixture model,” Acta. Acustica 47(01), 139–150 (2022). [CrossRef]
3. R. J. Tang, Y. X. Zou, W. Y. Liu, W. G. Zheng, L. G. Tang, and X. Fei, “Efficient energy-based orthogonal matching pursuit algorithm for multiple sound source localization with unknown source count,” Meas. Sci. Technol. 33(4), 045018 (2022). [CrossRef]
4. B. Xiao, “Research on mono source Location and Tracking Technology based on sound energy,” Guangzhou University (2022).
5. A. P. V. Benjamin, “Noise source localization in industrial facilities,” INTER-NOISE and NOISE-CON Congress and Conference Proceedings259(2), 7616–7627 (2019).
6. Y. Kwak, S. M. Park, J. Lee, and J. Park, “Rattle noise source localization through the time reversal of dispersive vibration signals on a road vehicle,” Wave Motion 93(C), 102452 (2020). [CrossRef]
7. W. D. Zhao, X. W. Xu, R. Song, and M. L. Yang, “Research Progress and Prospects of the Methods for Locating the Abnormal Noise Sources of Automobile,” Journal of Chongqing University of Technology (Natural Science) 36(10), 73–83 (2022). [CrossRef]
8. M. Zhang, G. Li, X. Zhang, and S. S. Zhang, “An infrasound source localisation algorithm for improving location accuracy of gas pipeline leakage detection system,” IJES. 15(1), 9–18 (2022). [CrossRef]
9. Z. J. Xia, Q. Liu, A. Li, Y. Y. Liu, Z. G. Jing, and W. Peng, “Sound source localization system based on diaphragm-type EFRI optical fiber microphones,” Chin. J. Laser 48(09), 0910002 (2021). [CrossRef]
10. H. S. Esat, T. T. Emre, K. Serkan, and S. O. Galip, “A simple, high sensitive fiber optic microphone based on cellulose triacetate diaphragm,” J. Lightwave Technol. 36(23), 5650–5655 (2018). [CrossRef]
11. S. J. Zhang, S. P. Wan, Y. F. Wang, B. Zhang, Z. P. Zhang, H. H. Zhong, J. L. Shi, J. Z. Sun, X. D. He, and Q. Wu, “2D sound source localization technology based on diaphragm EFPI fiber microphone array,” Opt. Commun. 519, 128435 (2022). [CrossRef]
12. T. S. Jin, G. W. Hong, and D. Zhao, “An Optical Fiber Sensing Array for Acoustic Source Localization,” Journal of Fudan University (Natural Science)58(05), 629–633 + 641 (2019).
13. M. L. Fan, “Research on Sound Source Localization with Distributed Acoustic Sensing Based on Fiber Bragg Grating,” Wuhan University of Technology (2021).
14. J. J. Liang, Z. Y. Wang, B. Lu, X. Wang, L. C. Li, Q. Ye, R. H. Qu, and H. W. Cai, “Distributed acoustic sensing for 2D and 3D acoustic source localization,” Opt. Lett. 44(7), 1690–1693 (2019). [CrossRef]
15. X. W. Cao, X. X. Li, F. T. Yue, and H. C. Sun, “Research and Application of Vibration Measurement While Drilling in the Undersea Coal Mine,” JCR. 103(sp1), 323–327 (2020). [CrossRef]
16. M. Houbiers, S. Bussat, F. Schopper, and F. Hansteen, “Seismic Drill-Bit Tracking for Improved Well Positioning,” SPE. DRILL. COMPLETION. 37(03), 206–219 (2022). [CrossRef]
17. L. M. Usman, M. H. Suliman, R. Amna, Q. Zakria, K. A. Zang, and M. M. A. Parvez, “Localization of Sound Sources: A Systematic Review,” Energies 14(13), 3910 (2021). [CrossRef]
18. D. Dhwani and M. Ninad, “A Review on Sound Source Localization Systems,” Arch. Comput. Method. E. 29(7), 4631–4642 (2022). [CrossRef]
19. K. Adam, H. Gert, and S. Ennes, “A deep learning method for grid-free localization and quantification of sound sources,” J. Acoust. Soc. Am. 146(03), EL225 (2019). [CrossRef]
20. C. Jiao, T. Zhang, and J. H. Sun, “Convolution Neural Network Based Indoor Microphon Array Sound Source Localization,” Laser. Optoelectron. 57(08), 195–200 (2020).
21. S. Y. Lee, J. H. Chang, and S. C. Lee, “Deep learning-based method for multiple sound source localization with high resolution and accuracy,” Mech. Syst. Signal. Pr. 161, 107959 (2021). [CrossRef]
22. Y. Cao, Q. Q. Kong, T. Iqbal, F. Y. An, W. W. Wang, and M. D. Plumbley, “Polyphonic Sound Event Detection and Localization using a Two-Stage Strategy,” CoRR (2019).
23. D. Q. Zhong, P. Y. He, and W. C. Yu, “Robust DOA estimation of coherent soureces based on neural network,” IEEE Sensors J. 21(6), 7498–7507 (2021). [CrossRef]
24. Z. H. Luo, Z. Yang, B. Lu, B. Xu, and J. L. Huang, “Modular DAS demodulation system based on ultra-weak fibre Bragg grating,” J. Inst. 17(10), P10037 (2022). [CrossRef]
25. Y. M. Wang, J. M. Gong, D. Y. Wang, B. Dong, W. H. Bi, and A. B. Wang, “A Quasi-Distributed Sensing Network with Time-Division-Multiplexed Fiber Bragg Gratings,” IEEE Photon. Technol. Lett. 23(2), 70–72 (2011). [CrossRef]
26. J. Xia, J. F. Jiang, S. Wang, Z. Y. Ding, G. H. Liang, P. P. Niu, Y. Z. Liu, X. Z. Zhang, and T. G. Liu, “Experimental Study on Acoustic Sensitization of Thin-walled Cylinder for Distributed Optical Fiber Acoustic Sensing,” Laser. Optoelectron. P. 1-11 (2023).
27. Z. B. Li and X. G. Zhang, “Improved microphone array sound source localization method based on generalized cross correlation,” Journal of Nanjing University (Natural Science 56(6), 917–922 (2020).
