Szczegóły publikacji
Opis bibliograficzny
Edge deployment of cross-sensor bearing fault diagnosis: accuracy–cost trade-offs in signal representation / Paweł KNAP, Jakub PRYMA, Ireneusz DOMINIK // W: MMAR 2026 [Dokument elektroniczny] : 30th international conference on Methods and Models in Automation and Robotics : 18-21 August 2026, Międzyzdroje, Poland : technical papers : on line proceedings. — Wersja do Windows. — Dane tekstowe. — [Piscataway] : IEEE, cop. 2026. — ( International Conference on Methods and Models in Automation and Robotics ; ISSN 2835-2815 ). — USB ISBN:979-8-3195-1920-7. — Print on Demand(PoD) ISBN:979-8-3195-1922-1. — e-ISBN: 979-8-3195-1921-4. — S. 185-190. — Wymagania systemowe: Adobe Reader. — Bibliogr. s. 190, Abstr. — Publikacja dostępna online od: 2026-09-02
Autorzy (3)
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 169797 |
|---|---|
| Data dodania do BaDAP | 2026-10-07 |
| Tekst źródłowy | URL |
| DOI | 10.1109/MMAR70562.2026.11667889 |
| Rok publikacji | 2026 |
| Typ publikacji | materiały konferencyjne (aut.) |
| Otwarty dostęp |  |
| Wydawca | Institute of Electrical and Electronics Engineers (IEEE) |
| Konferencja | International Conference on Methods and Models in Automation and Robotics 2026 |
| Czasopismo/seria | International Conference on Methods and Models in Automation and Robotics |
Abstract
This paper investigates the deployment of vibration-based bearing fault diagnosis on embedded edge hardware under a challenging cross-sensor setting using the Case Western Reserve University (CWRU) dataset. Models are trained on the drive-end (DE) sensor and evaluated on the fan-end (FE) sensor, introducing a realistic domain shift. Rather than proposing a new architecture, we compare three commonly used signal representations in condition monitoring pipelines: raw time-domain signals processed by a 1D CNN, frequency-domain inputs using an FFT frontend, and time-frequency inputs using an STFT frontend with a 2D CNN backend. Experiments are conducted using window sizes of 2048 and 4096 samples and overlap settings of 0.0 and 0.5. Results show that spectral and time-frequency representations generally improve cross-sensor diagnostic performance compared to raw input, with STFT achieving the highest balanced accuracy. However, deployment experiments on an NVIDIA Jetson platform reveal a practical systems constraint: while the raw CNN model exports directly to ONNX/TensorRT as a single inference engine, full FFT and STFT graphs cannot be executed end-to-end due to limited support for spectral operators. Consequently, spectral preprocessing must be performed outside the inference engine, with only the CNN backend accelerated. These findings highlight that signal representation choices for edge-based condition monitoring should be evaluated as an endto-end accuracy-efficiency trade-off rather than purely as model-level accuracy improvements.
