Hamaidi, Lala Khadidja (2017)
Robust Distributed Multi-Source Detection and Labeling in Wireless Acoustic Sensor Networks.
Technische Universität Darmstadt
Dissertation, Erstveröffentlichung
Kurzbeschreibung (Abstract)
The growing demand in complex signal processing methods associated with low-energy large scale wireless acoustic sensor networks (WASNs) urges the shift to a new information and communication technologies (ICT) paradigm. The emerging research perception aspires for an appealing wireless network communication where multiple heterogeneous devices with different interests can cooperate in various signal processing tasks (MDMT). Contributions in this doctoral thesis focus on distributed multi-source detection and labeling applied to audio enhancement scenarios pursuing an MDMT fashioned node-specific source-of-interest signal enhancement in WASNs. In fact, an accurate detection and labeling is a pre-requisite to pursue the MDMT paradigm where nodes in the WASN communicate effectively their sources-of-interest and, therefore, multiple signal processing tasks can be enhanced via cooperation.
First, a novel framework based on a dominant source model in distributed WASNs for resolving the activity detection of multiple speech sources in a reverberant and noisy environment is introduced. A preliminary rank-one multiplicative non-negative independent component analysis (M-NICA) for unique dominant energy source extraction given associated node clusters is presented. Partitional algorithms that minimize the within-cluster mean absolute deviation (MAD) and weighted MAD objectives are proposed to determine the cluster membership of the unmixed energies, and thus establish a source specific voice activity recognition.
In a second study, improving the energy signal separation to alleviate the multiple source activity discrimination task is targeted. Sparsity inducing penalties are enforced on iterative rank-one singular value decomposition layers to extract sparse right rotations. Then, sparse non-negative blind energy separation is realized using multiplicative updates. Hence, the multiple source detection problem is converted into a sparse non-negative source energy decorrelation. Sparsity tunes the supposedly non-active energy signatures to exactly zero-valued energies so that it is easier to identify active energies and an activity detector can be constructed in a straightforward manner. In a centralized scenario, the activity decision is controlled by a fusion center that delivers the binary source activity detection for every participating energy source. This strategy gives precise detection results for small source numbers. With a growing number of interfering sources, the distributed detection approach is more promising. Conjointly, a robust distributed energy separation algorithm for multiple competing sources is proposed. A robust and regularized $t_{\nu}M$-estimation of the covariance matrix of the mixed energies is employed. This approach yields a simple activity decision using only the robustly unmixed energy signatures of the sources in the WASN. The performance of the robust activity detector is validated with a distributed adaptive node-specific signal estimation method for speech enhancement. The latter enhances the quality and intelligibility of the signal while exploiting the accurately estimated multi-source voice decision patterns. In contrast to the original M-NICA for source separation, the extracted binary activity patterns with the robust energy separation significantly improve the node-specific signal estimation.
Due to the increased computational complexity caused by the additional step of energy signal separation, a new approach to solving the detection question of multi-device multi-source networks is presented. Stability selection for iterative extraction of robust right singular vectors is considered. The sub-sampling selection technique provides transparency in properly choosing the regularization variable in the Lasso optimization problem. In this way, the strongest sparse right singular vectors using a robust $\ell_1$-norm and stability selection are the set of basis vectors that describe the input data efficiently. Active/non-active source classification is achieved based on a robust Mahalanobis classifier. For this, a robust $M$-estimator of the covariance matrix in the Mahalanobis distance is utilized. Extensive evaluation in centralized and distributed settings is performed to assess the effectiveness of the proposed approach. Thus, overcoming the computationally demanding source separation scheme is possible via exploiting robust stability selection for sparse multi-energy feature extraction.
With respect to the labeling problem of various sources in a WASN, a robust approach is introduced that exploits the direction-of-arrival of the impinging source signals. A short-time Fourier transform-based subspace method estimates the angles of locally stationary wide band signals using a uniform linear array. The median of angles estimated at every frequency bin is utilized to obtain the overall angle for each participating source. The features, in this case, exploit the similarity across devices in the particular frequency bins that produce reliable direction-of-arrival estimates for each source. Reliability is defined with respect to the median across frequencies. All source-specific frequency bands that contribute to correct estimated angles are selected. A feature vector is formed for every source at each device by storing the frequency bin indices that lie within the upper and lower interval of the median absolute deviation scale of the estimated angle. Labeling is accomplished by a distributed clustering of the extracted angle-based feature vectors using consensus averaging.
| Typ des Eintrags: | Dissertation | ||||
|---|---|---|---|---|---|
| Erschienen: | 2017 | ||||
| Autor(en): | Hamaidi, Lala Khadidja | ||||
| Art des Eintrags: | Erstveröffentlichung | ||||
| Titel: | Robust Distributed Multi-Source Detection and Labeling in Wireless Acoustic Sensor Networks | ||||
| Sprache: | Englisch | ||||
| Referenten: | Zoubir, Prof. Dr. Abdelhak ; Muma, Dr.-Ing. Michael ; Bertrand, Asst. Prof Alexander | ||||
| Publikationsjahr: | 2017 | ||||
| Ort: | Darmstadt | ||||
| Datum der mündlichen Prüfung: | 13 Dezember 2017 | ||||
| URL / URN: | http://tuprints.ulb.tu-darmstadt.de/7184 | ||||
| Kurzbeschreibung (Abstract): | The growing demand in complex signal processing methods associated with low-energy large scale wireless acoustic sensor networks (WASNs) urges the shift to a new information and communication technologies (ICT) paradigm. The emerging research perception aspires for an appealing wireless network communication where multiple heterogeneous devices with different interests can cooperate in various signal processing tasks (MDMT). Contributions in this doctoral thesis focus on distributed multi-source detection and labeling applied to audio enhancement scenarios pursuing an MDMT fashioned node-specific source-of-interest signal enhancement in WASNs. In fact, an accurate detection and labeling is a pre-requisite to pursue the MDMT paradigm where nodes in the WASN communicate effectively their sources-of-interest and, therefore, multiple signal processing tasks can be enhanced via cooperation. First, a novel framework based on a dominant source model in distributed WASNs for resolving the activity detection of multiple speech sources in a reverberant and noisy environment is introduced. A preliminary rank-one multiplicative non-negative independent component analysis (M-NICA) for unique dominant energy source extraction given associated node clusters is presented. Partitional algorithms that minimize the within-cluster mean absolute deviation (MAD) and weighted MAD objectives are proposed to determine the cluster membership of the unmixed energies, and thus establish a source specific voice activity recognition. In a second study, improving the energy signal separation to alleviate the multiple source activity discrimination task is targeted. Sparsity inducing penalties are enforced on iterative rank-one singular value decomposition layers to extract sparse right rotations. Then, sparse non-negative blind energy separation is realized using multiplicative updates. Hence, the multiple source detection problem is converted into a sparse non-negative source energy decorrelation. Sparsity tunes the supposedly non-active energy signatures to exactly zero-valued energies so that it is easier to identify active energies and an activity detector can be constructed in a straightforward manner. In a centralized scenario, the activity decision is controlled by a fusion center that delivers the binary source activity detection for every participating energy source. This strategy gives precise detection results for small source numbers. With a growing number of interfering sources, the distributed detection approach is more promising. Conjointly, a robust distributed energy separation algorithm for multiple competing sources is proposed. A robust and regularized $t_{\nu}M$-estimation of the covariance matrix of the mixed energies is employed. This approach yields a simple activity decision using only the robustly unmixed energy signatures of the sources in the WASN. The performance of the robust activity detector is validated with a distributed adaptive node-specific signal estimation method for speech enhancement. The latter enhances the quality and intelligibility of the signal while exploiting the accurately estimated multi-source voice decision patterns. In contrast to the original M-NICA for source separation, the extracted binary activity patterns with the robust energy separation significantly improve the node-specific signal estimation. Due to the increased computational complexity caused by the additional step of energy signal separation, a new approach to solving the detection question of multi-device multi-source networks is presented. Stability selection for iterative extraction of robust right singular vectors is considered. The sub-sampling selection technique provides transparency in properly choosing the regularization variable in the Lasso optimization problem. In this way, the strongest sparse right singular vectors using a robust $\ell_1$-norm and stability selection are the set of basis vectors that describe the input data efficiently. Active/non-active source classification is achieved based on a robust Mahalanobis classifier. For this, a robust $M$-estimator of the covariance matrix in the Mahalanobis distance is utilized. Extensive evaluation in centralized and distributed settings is performed to assess the effectiveness of the proposed approach. Thus, overcoming the computationally demanding source separation scheme is possible via exploiting robust stability selection for sparse multi-energy feature extraction. With respect to the labeling problem of various sources in a WASN, a robust approach is introduced that exploits the direction-of-arrival of the impinging source signals. A short-time Fourier transform-based subspace method estimates the angles of locally stationary wide band signals using a uniform linear array. The median of angles estimated at every frequency bin is utilized to obtain the overall angle for each participating source. The features, in this case, exploit the similarity across devices in the particular frequency bins that produce reliable direction-of-arrival estimates for each source. Reliability is defined with respect to the median across frequencies. All source-specific frequency bands that contribute to correct estimated angles are selected. A feature vector is formed for every source at each device by storing the frequency bin indices that lie within the upper and lower interval of the median absolute deviation scale of the estimated angle. Labeling is accomplished by a distributed clustering of the extracted angle-based feature vectors using consensus averaging. |
||||
| Alternatives oder übersetztes Abstract: |
|
||||
| URN: | urn:nbn:de:tuda-tuprints-71841 | ||||
| Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 000 Allgemeines, Informatik, Informationswissenschaft > 000 Allgemeines, Wissenschaft 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik 500 Naturwissenschaften und Mathematik > 510 Mathematik 600 Technik, Medizin, angewandte Wissenschaften > 600 Technik 600 Technik, Medizin, angewandte Wissenschaften > 620 Ingenieurwissenschaften und Maschinenbau |
||||
| Fachbereich(e)/-gebiet(e): | 18 Fachbereich Elektrotechnik und Informationstechnik 18 Fachbereich Elektrotechnik und Informationstechnik > Drahtlose Sensornetze 18 Fachbereich Elektrotechnik und Informationstechnik > Institut für Nachrichtentechnik 18 Fachbereich Elektrotechnik und Informationstechnik > Institut für Nachrichtentechnik > Signalverarbeitung |
||||
| Hinterlegungsdatum: | 04 Mär 2018 20:55 | ||||
| Letzte Änderung: | 19 Mär 2019 16:23 | ||||
| PPN: | |||||
| Referenten: | Zoubir, Prof. Dr. Abdelhak ; Muma, Dr.-Ing. Michael ; Bertrand, Asst. Prof Alexander | ||||
| Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 13 Dezember 2017 | ||||
| Export: | |||||
| Suche nach Titel in: | TUfind oder in Google | ||||
 |
Frage zum Eintrag |
Optionen (nur für Redakteure)
 |
Redaktionelle Details anzeigen |
Drucken
Impressum