Scholarlink

Purpose <p>Vibration-based fault diagnosis is increasingly deployed on resource-constrained embedded platforms using lightweight neural networks, enabling practical TinyML-based condition monitoring. While such systems often report high accuracy under matched training and deployment conditions, their robustness under realistic domain shifts remains insufficiently understood. This work aims to systematically assess how different sources of domain shift affect the deployability of TinyML vibration diagnostics.</p> Methods <p>A controlled experimental study is conducted using two real-world vibration datasets exhibiting distinct physical origins of domain variation: geometry-induced differences in a pipe system and sensor-induced differences in a gear transmission system. A fixed, compact neural architecture is used throughout as a controlled evaluation instrument to isolate the effects of domain shift, signal representation, and limited target-domain supervision. All experiments employ deterministic preprocessing, leakage-free windowing, and provenance-aware data partitioning.</p> Results <p>The results show that cross-domain generalization is strongly dependent on the physical nature of the domain shift. Geometry-induced variation preserves key temporal characteristics, enabling partial cross-domain transfer, where accuracy improves from approximately 14–19% under source-only training to 72–82% with limited target-domain supervision. In contrast, sensor-induced shifts substantially alter vibration observability, leading to consistently reduced cross-domain performance, where accuracy improves only modestly from approximately 25–27% to 36–41%, remaining well below in-domain accuracy.</p> Conclusion <p>These findings indicate that high in-domain accuracy alone is not a reliable indicator of deployable robustness in TinyML vibration diagnostics. Cross-domain performance is fundamentally constrained by sensing physics rather than model complexity, underscoring the importance of explicitly accounting for domain shift mechanisms when designing and deploying lightweight vibration-based monitoring systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Systematic Evaluation of Domain Shift Effects in TinyML Vibration Diagnostics

  • 期刊论文

  • 发表日期：

    2026年5月2日

查看全文
  • Khalid Hossen

摘要

Purpose

Vibration-based fault diagnosis is increasingly deployed on resource-constrained embedded platforms using lightweight neural networks, enabling practical TinyML-based condition monitoring. While such systems often report high accuracy under matched training and deployment conditions, their robustness under realistic domain shifts remains insufficiently understood. This work aims to systematically assess how different sources of domain shift affect the deployability of TinyML vibration diagnostics.

Methods

A controlled experimental study is conducted using two real-world vibration datasets exhibiting distinct physical origins of domain variation: geometry-induced differences in a pipe system and sensor-induced differences in a gear transmission system. A fixed, compact neural architecture is used throughout as a controlled evaluation instrument to isolate the effects of domain shift, signal representation, and limited target-domain supervision. All experiments employ deterministic preprocessing, leakage-free windowing, and provenance-aware data partitioning.

Results

The results show that cross-domain generalization is strongly dependent on the physical nature of the domain shift. Geometry-induced variation preserves key temporal characteristics, enabling partial cross-domain transfer, where accuracy improves from approximately 14–19% under source-only training to 72–82% with limited target-domain supervision. In contrast, sensor-induced shifts substantially alter vibration observability, leading to consistently reduced cross-domain performance, where accuracy improves only modestly from approximately 25–27% to 36–41%, remaining well below in-domain accuracy.

Conclusion

These findings indicate that high in-domain accuracy alone is not a reliable indicator of deployable robustness in TinyML vibration diagnostics. Cross-domain performance is fundamentally constrained by sensing physics rather than model complexity, underscoring the importance of explicitly accounting for domain shift mechanisms when designing and deploying lightweight vibration-based monitoring systems.