MOCRN: A Physics-Guided Multimodal Neural Network Framework for Analog Circuit Fault Detection

MOCRN (Multimodal Ordinal Circuit Reliability Network) is a physics-guided multimodal neural network framework for analog circuit fault detection. It combines physics-guided modeling and multimodal deep learning to enable dual-task learning for fault type identification and degradation level estimation.

Original Authors: J.P. Teshan Jayasinghe, Gayani Kulathunga, Rahul Thakur Source: GitHub Project MOCRN-Multimodal-Ordinal-Circuit-Reliability-Network Publication Time: April 2026 (Paper published in the Circuits, Systems, and Signal Processing journal) Original Link: https://github.com/teshankj/MOCRN-Multimodal-Ordinal-Circuit-Reliability-Network

Analog circuit reliability detection faces three key challenges:

1. Progressive Degradation: Faults exhibit continuous and progressive behavior, making it difficult for traditional binary judgment to capture early warnings;
2. Diverse Fault Modes: The same component fault may have different symptoms, while faults in different locations may have similar characteristics;
3. Limitations of Traditional Methods: Relying on expert experience, it is difficult to handle complex mixed-signal systems and SoC designs, making manual analysis increasingly impractical.

Core innovations of the MOCRN framework:

1. Physics Guidance: Embed circuit physics knowledge into data generation and feature extraction to avoid black-box fitting and ensure outputs comply with physical laws;
2. Multimodal Fusion: Late fusion of two information sources: raw waveforms (temporal information) and statistical features (frequency domain/energy distribution);
3. Ordinal Reliability Modeling: Adopt a hybrid evidence ordinal loss function to explicitly model the natural order relationship of degradation levels, reflecting the progressive process from health to severe degradation.

Dataset generation and fault injection mechanism:

• Simulation Tool: Use LTspice to generate training data, which has the advantages of repeatability, controllability, and scalability;
• Fault Injection: Structured parameter space method, defining fault types (FT) and degradation levels (0-100 to quantify severity);
• Parameter Perturbation: Replace Monte Carlo random variations with deterministic trigonometric functions (e.g., 0.1sin(run0.6283)) to ensure diversity and reproducibility;
• Test Circuits: Covers voltage rectifiers, peak detectors, and diode clamp circuits, each containing 4-6 fault types (e.g., diode leakage, resistance change, etc.);
• Simulation Settings: Transient analysis with a time step of 4-40ms, saving key node voltages and branch currents.

Model architecture and key technologies:

• Multi-Task Learning: Simultaneously output fault type classification probabilities and degradation level estimates, sharing feature representations;
• Feature Engineering: Extract time-domain (mean, variance, peak value, etc.) and frequency-domain (FFT energy spectrum) statistical features, which are input together with raw waveforms;
• Loss Function: Cross-entropy loss for fault identification, and hybrid evidence ordinal loss for degradation estimation (based on evidence theory, expressing uncertainty and respecting the order of degradation);
• Training Method: K-fold cross-validation to prevent overfitting, with configurations managed via YAML files.

Experimental validation and performance evaluation:

• Evaluation Metrics: Accuracy and F1 score for fault identification (focusing on imbalanced datasets), MAE and RMSE for degradation estimation;
• Feature Importance: Identify key circuit features through consensus analysis to improve model interpretability;
• Engineering Value: Enable early detection of soft faults, support predictive maintenance, and apply to high-reliability fields such as aerospace and medical equipment.

Technical insights and future outlook:

• AI for Science: Demonstrate a model of combining domain knowledge (physics guidance) with data-driven methods; similar physics-informed neural network ideas can be applied to multiple engineering fields;
• Hardware Testing: Simulation data generation reduces reliance on expensive hardware platforms, providing an efficient way for AI model training;
• Machine Learning: Ordinal regression methods provide references for ordered category prediction problems (e.g., disease severity, equipment aging);
• Open Source Value: The GitHub repository provides complete code (data generation, training, pre-trained models) to support reproduction and extension to other analog circuits.