Szczegóły publikacji
Opis bibliograficzny
Deployment-aware comparison of classical machine learning and 1D CNN for bearing fault diagnosis on embedded hardware / Jakub Pryma, Paweł KNAP, Krzysztof LALIK // W: 27th International Carpathian Control Conference (ICCC) [Dokument elektroniczny] : 1-3 June 2026, Szilvásvárad, Hungary : proceedings. — Wersja do Windows. — Dane tekstowe. — Piscataway : IEEE, cop. 2026. — Print on Demand(PoD) ISBN: 979-8-3195-3321-0. — e-ISBN: 979-8-3195-3320-3. — S. 384–389. — Wymagania systemowe: Adobe Reader. — Bibliogr. s. 389, Abstr. — Publikacja dostępna online od: 2026-07-07
Autorzy (3)
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 169796 |
|---|---|
| Data dodania do BaDAP | 2026-10-02 |
| Tekst źródłowy | URL |
| DOI | 10.1109/ICCC71363.2026.11593288 |
| Rok publikacji | 2026 |
| Typ publikacji | materiały konferencyjne (aut.) |
| Otwarty dostęp |  |
| Wydawca | Institute of Electrical and Electronics Engineers (IEEE) |
Abstract
Deploying vibration-based fault diagnosis models on embedded edge platforms requires balancing diagnostic accuracy with strict constraints on inference latency and memory usage. This paper investigates that trade-off through a comparative study of classical machine learning and deep learning methods for bearing fault diagnosis using the Case Western Reserve University (CWRU) dataset. The considered approaches include feature-based classifiers, namely a linear Support Vector Machine (SVM) and XGBoost, and an end-to-end one-dimensional convolutional neural network (1D CNN) operating on raw vibration signals. The models are deployed on an NVIDIA Jetson Orin Nano and evaluated using CPU execution as well as TensorRT-based GPU inference in FP32, FP16, and INT8 precision. The comparison is performed with respect to classification accuracy, inference latency, and memory consumption. The obtained results show that lightweight feature-based methods remain attractive for simple low-overhead deployment, whereas the TensorRT-optimized 1D CNN offers the best latency-accuracy trade-off under GPU acceleration. These findings highlight that model selection for edge diagnostic systems should be based on deployment-aware evaluation rather than classification accuracy alone.
