Scholarlink

<p>Bearing faults contribute to about 45% of all the rotating machinery failures in the industrial production. Although deep learning techniques are mathematically accurate on their benchmarks, they are physically inconsistent and decay to noisy or loaded predictions. This paper suggests a Physics-Informed Convolutional Neural Network (PI-CNN) embedding, PIDL-Edge, with physical frequency restrictions in form of a Physics Frequency Residual (PFR) loss, which can be interpreted as restricting the space of hypotheses being learned to the physically valid space ℋphysics ⊂ ℋCNN. Trained on the CWRU Bearing Dataset (0–3 HP, 12&#xa0;kHz), PIDL-Edge attained an accuracy of 98.7% (F1 = 0.987) which is 3.2 standard deviations higher than the pure-CNN baseline on clean data and 13.6 points higher on 5 dB SNR data. Generalization cross load was 87.2% during a decrease in 0 HP to 3 HP shift + 13.1%. The value of <i>r</i> = 0.92 between predictions and kinematic band energy vs. <i>r</i> = 0.62 between pure-CNN was established through Pearson correlation. Physics-aware INT8-quantization-aware training (QAT) reached 7.4 ms inference latency at 6.8&#xa0;W in NVIDIA Jetson Nano, and shows that it performs Pareto-dominant on all the dimensions measured.</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

PIDL-Edge: A Physics-Informed Convolutional Neural Network for Real-Time Bearing Fault Detection in Smart Manufacturing Using the CWRU Dataset

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • I. Arivazhagi,
  • S. Sajini,
  • S. Gopalakrishnan,
  • C. Quba Jaslin,
  • G. Siva Nageswara Rao,
  • P. Muthulakshmi

摘要

Bearing faults contribute to about 45% of all the rotating machinery failures in the industrial production. Although deep learning techniques are mathematically accurate on their benchmarks, they are physically inconsistent and decay to noisy or loaded predictions. This paper suggests a Physics-Informed Convolutional Neural Network (PI-CNN) embedding, PIDL-Edge, with physical frequency restrictions in form of a Physics Frequency Residual (PFR) loss, which can be interpreted as restricting the space of hypotheses being learned to the physically valid space ℋphysics ⊂ ℋCNN. Trained on the CWRU Bearing Dataset (0–3 HP, 12 kHz), PIDL-Edge attained an accuracy of 98.7% (F1 = 0.987) which is 3.2 standard deviations higher than the pure-CNN baseline on clean data and 13.6 points higher on 5 dB SNR data. Generalization cross load was 87.2% during a decrease in 0 HP to 3 HP shift + 13.1%. The value of r = 0.92 between predictions and kinematic band energy vs. r = 0.62 between pure-CNN was established through Pearson correlation. Physics-aware INT8-quantization-aware training (QAT) reached 7.4 ms inference latency at 6.8 W in NVIDIA Jetson Nano, and shows that it performs Pareto-dominant on all the dimensions measured.