Jose, Justin (2026) Automated Classification of Sleep-Disruptive Audio Events Under Real-World Conditions. Masterarbeit, Technische Universität Ilmenau.
|
PDF
- Nur DLR-intern zugänglich
11MB |
Kurzfassung
The process of manually labeling nocturnal sleep recordings for event by event acoustic analysis for sleep quality research can be a very resource intensive and requires an average of up to over 100 hours of an expert annotator time per each study, making it cost prohibitive for large scale longitudinal sleep studies on sleep disruptive noise exposure. This thesis investigates whether sound event detection algorithms trained on manually annotated labels can achieve sufficient accuracy of epidemiological dose response analysis on such challenging audio recordings. The challenges inherent in the sleep audio recordings include difficulty with class imbalance, varying duration of events ranging from 0.5 to 120 seconds, variability and unpredictability in signal to noise ratio and complexity of dealing with acoustically heterogeneous classes. A systematic literature review of more than 20 candidate architectures led to the choice of adapting Masked Audio Transformer for Sound Event Detection (MAT-SED) model, initially validated (phase 1) with DCASE 2023 DESED before using it to analyze any sleep audio recordings used for this study, which achieved PSDS1 = 0.648 versus published 0.587 + 10.4%. Three model configurations were adopted through a masked reconstruction pretraining followed by frozen encoder head warm up and finally end-to-end semi-supervised fine tuning on 86,315 weakly labeled clips from 31 recordings of overnight sleep collected from longitudinal sleep research studies conducted by the German Aerospace Center (DLR) - Institute of Aerospace Medicine in Cologne, which includes 27 broad classes that were narrowed down to four major distinctive classes: Airplane, Environmental, Subject Activity, and Transportation. The three models used for the performance evaluations share the same core encoder Patchout Spectrogram Transformer (PaSST), but differ in approaches to how they interpret sound patterns over time. The first model uses a Bidirectional Gated Recurrent Unit (BiGRU) as the context network that analyzes sounds both forward and backward. The second model MAT-SED replaces this with a TransformerXL based decoder designed to capture long-range temporal patterns, while the third omits the masked language model (NO-MLM) pretraining to assess its impact on final model performance. The results of all three configurations exceed the 0.85 macro-F1 threshold, BiGRU, MAT-SED, and NO-MLM achieving a macro-F1 score of 0.9056, 0.9035 and 0.9026 respectively with each of them producing Receiver Operating Characteristic – Area Under the Curve (ROC- AUC) greater than 0.97 for each class. Pairwise McNemar tests on these models confirm statistical indistinguishability, pointing to the shared PaSST encoder as the dominant performance driver. Detecting the Airplane class most consistently (F1 = 0.933) represents a key achievement, as the study focuses on how living near airports affects sleep quality. These results demonstrate that weak-label based sleep audio classification is effective in real-world conditions, laying the groundwork for automated annotation methods that support human oversight in future.
| elib-URL des Eintrags: | https://elib.dlr.de/229324/ | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Dokumentart: | Hochschulschrift (Masterarbeit) | ||||||||
| Titel: | Automated Classification of Sleep-Disruptive Audio Events Under Real-World Conditions | ||||||||
| Autoren: |
| ||||||||
| DLR-Supervisor: |
| ||||||||
| Datum: | April 2026 | ||||||||
| Open Access: | Nein | ||||||||
| Seitenanzahl: | 80 | ||||||||
| Status: | nicht veröffentlicht | ||||||||
| Stichwörter: | Audio event classification, sleep disruption, Spectrogram Transformer | ||||||||
| Institution: | Technische Universität Ilmenau | ||||||||
| Abteilung: | Dept. of Electronic Media Technology | ||||||||
| HGF - Forschungsbereich: | Luftfahrt, Raumfahrt und Verkehr | ||||||||
| HGF - Programm: | Raumfahrt | ||||||||
| HGF - Programmthema: | Technik für Raumfahrtsysteme | ||||||||
| DLR - Schwerpunkt: | Raumfahrt | ||||||||
| DLR - Forschungsgebiet: | R SY - Technik für Raumfahrtsysteme | ||||||||
| DLR - Teilgebiet (Projekt, Vorhaben): | R - Erforschung wissenschaftlicher Methoden | ||||||||
| Standort: | Jena | ||||||||
| Institute & Einrichtungen: | Institut für Datenwissenschaften > Datengewinnung und -mobilisierung Institut für Luft- und Raumfahrtmedizin > Schlaf und Humanfaktoren | ||||||||
| Hinterlegt von: | Kersten, Dr.-Ing. Jens | ||||||||
| Hinterlegt am: | 16 Nov 2026 10:51 | ||||||||
| Letzte Änderung: | 16 Nov 2026 10:51 |
Nur für Mitarbeiter des Archivs: Kontrollseite des Eintrags