Scholarlink

<p>This paper presents the Multimodal Ordinal Circuit Reliability Network (MOCRN), a unified deep learning framework for joint fault classification and progressive component degradation estimation in analog circuits. The proposed approach integrates waveform-based temporal features with statistical descriptors through a learnable late-fusion architecture. To enhance interpretability, the framework incorporates multi-dimensional fault space modeling, where degradation trajectories are generated using physically motivated degradation functions applied directly to SPICE component parameters. In addition, deterministic tolerance variations are introduced to systematically simulate manufacturing variability. Physics is embedded at both the data and model levels: parametric circuit equations guide degradation sample generation, while CORAL-based ordinal regression loss enforces the natural progression of degradation stages during training. The model contains 2.41 million parameters and is evaluated on three LTspice-simulated circuits (bridge rectifier, diode peak detector, and RC relaxation oscillator) comprising 4,800 samples across six degradation levels, including unseen tolerance conditions and ±2% measurement noise. Compared with Decision Tree, Random Forest, XGBoost, and SVM baselines, MOCRN achieves up to 100% fault classification accuracy and degradation errors as low as 0.89 MAE. The model runs at 3.96 ms on CPU and reduces to 2.5 MB after INT8 quantization without measurable accuracy loss, supporting real-time edge deployment.</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

MOCRN-A Novel Physics-Guided Fault Modeling and Multimodal Neural Network for Fault Identification and Degradation Estimation in Analog Circuits

  • 期刊论文

  • 发表日期：

    2026年4月25日

查看全文
  • J.P.Teshan Kavindu Jayasinghe,
  • Gayani Pamoda Kulathunga,
  • Rahul Thakur

摘要

This paper presents the Multimodal Ordinal Circuit Reliability Network (MOCRN), a unified deep learning framework for joint fault classification and progressive component degradation estimation in analog circuits. The proposed approach integrates waveform-based temporal features with statistical descriptors through a learnable late-fusion architecture. To enhance interpretability, the framework incorporates multi-dimensional fault space modeling, where degradation trajectories are generated using physically motivated degradation functions applied directly to SPICE component parameters. In addition, deterministic tolerance variations are introduced to systematically simulate manufacturing variability. Physics is embedded at both the data and model levels: parametric circuit equations guide degradation sample generation, while CORAL-based ordinal regression loss enforces the natural progression of degradation stages during training. The model contains 2.41 million parameters and is evaluated on three LTspice-simulated circuits (bridge rectifier, diode peak detector, and RC relaxation oscillator) comprising 4,800 samples across six degradation levels, including unseen tolerance conditions and ±2% measurement noise. Compared with Decision Tree, Random Forest, XGBoost, and SVM baselines, MOCRN achieves up to 100% fault classification accuracy and degradation errors as low as 0.89 MAE. The model runs at 3.96 ms on CPU and reduces to 2.5 MB after INT8 quantization without measurable accuracy loss, supporting real-time edge deployment.