PhD Thesis proposal: Detection and estimation of infrasonic broadband multiple sources using convolutional recurrent neural network
Thesis subject: Detection and estimation of infrasonic broadband multiple sources using convolutional recurrent neural network
Research field: signal processing, acoustic, antenna, detection, multiple DOA estimation, source separation, learning method, neural network.
Context: CEA-DAM is a research center which exploits and analyses infrasound data (0.05-4 Hz) of the International Monitoring System (IMS) being set up to verify compliance with the Comprehensive Test-Ban Treaty (CTBT). So far, correlation-based method is applied to process continuous records from the different IMS arrays. The estimation and the angle of arrival and propagation velocity of coherent waves is achieved by measuring the propagation delays between different sensors [Cansi, 1995]. The used detection method has been essentially developed to detect and localize a single coherent source signal within a given time-frequency cell. However, localizing sources in real environments is challenging, especially when several sources are active. In the frequency band of interest, the current processing system reveals the existence of interfering signals from non-desired persistent sources of coherent noise. Such interferences can lead to erroneous detections, inaccurate estimations and fortuitous events. In order to overcome this limitation, it is needed to consider high-resolution source separation algorithms to get rid or mitigate the impact of the interfering signals.
Objectives: The main objective of this thesis is to elaborate enhanced detection algorithms. We will investigate the following items, which aim at developing, testing and implementing detection and estimation methods for the mitigation of spatially distributed wideband sources:
• Consider high-resolution methods like MUSIC (Multiple SIgnal Classification) to detect multiple narrow-band sources within the same time frequency grid.
• Consider iterative spectral methods base on Fisher statistics to extract coherent energy in incoherent noise, e.g. [Den Ouden et al., 2020].
• Consider the potential of convolutional recurrent neural network for multiple source localization without prior knowledge of the number of sources, e.g. [Chakrabarty, 2017].
• Develop statistical criteria for the estimation of the number of sources via penalized maximum likelihood approaches (such as AIC, BIC, MDL).
• Develop blind beamforming (also known as blind source separation) methods for improved detection and extraction of the signal of interest.
• Develop classification methods based on learning approaches by means of layerwise visualization technique to improve broadband DOA estimation, e.g. [Perotin et al., 2019].
For the performance assessment and validation of the different methods under investigation, we will develop and enrich databases of controlled synthetic data and records from multiple IMS infrasound arrays representative of the genuine conditions and different scenarios for infrasound source detection. In particular, these data should well represent the diversity and variability of the signals and noise/interference sources as well as different array configurations. The outcome is to improve operational infrasound monitoring procedures by elaborating a cost-effective high-resolution detection algorithm, in order to characterize sources of interest in noisy measurements in the presence of interfering signals.
Thesis organization: The PhD student will join the CEA-DAM center (http://www-dase.cea.fr/, Paris suburb, France) and will be registered at the University of Orléans (MIPTIS Doctorate School) under the supervision of Pr. Karim Abed-Meraim from PRISME laboratory. The thesis will involve the expertise of Maurice Charbit (Emeritus professor of Telecom ParisTech) and Dr. Alexis Le Pichon (CEA-DAM). All required expertise will be used to achieve an improved operational detection and characterization tool. Communications and publications will be strongly encouraged through collaborations with CEA partner institutes.
French or European citizenship is required for this research work. The scholarship (close to 2000 euros/month) is provided for the total duration of the thesis and the PhD student will be located in Paris suburb (CEA-DAM site in Arpajon) with periodical visits to the University of Orléans.
Pr. Karim Abed-Meraim (University of Orléans)
Pr. Maurice Charbit (Telecom ParisTech)
Dr. Alexis Le Pichon
CEA - Centre DAM Ile de France
Tél. : +33 (0)1 69 26 40 00
• Cansi, Y. (1995), An automatic seismic event processing for detection and location: the PMCC method, Geophys. Res. Lett., 22, https://doi.org/10.1029/95GL00468.
• Den Ouden et al. (2020), CLEAN beamforming for the enhanced detection of multiple infrasonic sources, Geophys. J. Int., 221, https://doi.org/10.1093/gji/ggaa010.
• Chakrabarty, S., and E. Habets (2017), Broadband DOA estimation using convolutional neural network trained with noise signals, IEEE workshop on applications of signal processing to audio and acoustics, 15-18.
• Perotin, L., et al. (2019). Crnn-based multiple DOA estimation using acoustic intensity features for ambisonics recordings. IEEE Journal of Selected Topics in Signal Processing, 13.