MOCRN: Physics-Guided Multimodal Neural Network for Analog Circuit Fault Identification

MOCRN (Multimodal Ordinal Circuit Reliability Network) is a novel physics-guided multimodal neural network framework for analog circuit fault identification and degradation estimation. The project combines physical knowledge, multimodal data fusion, and ordinal regression methods to address the complex challenges of analog circuit fault diagnosis. Maintained by teshankj, it is open-sourced on GitHub: https://github.com/teshankj/MOCRN-Multimodal-Ordinal-Circuit-Reliability-Network, released on 2026-05-25.

Analog circuits are crucial in signal processing, power management, and other fields, but fault diagnosis faces multiple challenges:

1. Complex fault modes (parameter drift, nonlinear distortion, etc.);
2. Fuzzy fault features caused by component tolerances;
3. Multi-physics coupling affecting performance;
4. Scarce fault samples and high annotation costs. Traditional rule-based methods and single machine learning models struggle to address these problems.

Physics-Guided Fault Modeling

Integrate prior knowledge such as circuit theory and semiconductor physics into the model: feature engineering extracts physically meaningful features, model structure follows signal/fault propagation laws, loss function incorporates physical constraints, and outputs correspond to physical quantities or fault mechanisms.

Multimodal Data Fusion

Process electrical signals (voltage/current waveforms), thermal information (temperature distribution), time-series data, design parameters, etc. Extract features via dedicated encoders, then fuse them through attention mechanisms.

Ordinal Regression Network

Model the progressive degradation process of circuits (healthy → mild → moderate → severe → failure), use ordinal classifiers to estimate the probability of each level, and improve interpretability.

Encoder Design

Multi-branch structure to process different data:

• Time-series encoder (LSTM/Transformer) for time-series data;
• Image encoder (CNN) to extract thermal imaging features;
• Parameter encoder (fully connected) for design parameters;
• Frequency-domain encoder (1D convolution) to analyze spectral data.

Cross-modal Attention Fusion

Automatically learn weights for each modality via attention mechanisms, e.g., increase reliance on thermal information when noise is high.

Physical Constraint Layer

Add a differentiable physical simulation module at the backend to ensure outputs comply with circuit theory constraints (e.g., consistency between degradation level and performance indicators).

Dataset and Setup

Validated on benchmarks such as operational amplifiers, power management circuits, ADC/DAC, with data from SPICE simulations and hardware tests.

Performance Evaluation

Outperforms traditional methods in various metrics: fault identification accuracy, degradation estimation precision, early warning capability, and generalization performance are prominent. In small-sample scenarios, physics guidance compensates for data insufficiency.

Ablation Experiments:

• Removing physics guidance: Stability decreases under extreme conditions;
• Single-modal vs. multi-modal: Single-modal accuracy is significantly lower;
• Ordinal regression vs. traditional methods: Better maintains orderliness.

1. Circuit Design Verification: Evaluate reliability margins and identify weak links;
2. Manufacturing Testing: Assist in identifying manufacturing defects and improve testing efficiency;
3. In-service Monitoring: Online health monitoring to enable predictive maintenance;
4. Fault Diagnosis: Quickly locate fault causes and reduce troubleshooting time.

MOCRN is a cutting-edge achievement of AI and traditional engineering integration. Its technical contributions include:

1. Physics-guided AI design enhances reliability and interpretability;
2. Multimodal fusion strategy handles heterogeneous data;
3. Ordinal learning models ordered degradation problems;
4. Physical constraints + prior knowledge solve small-sample problems. The open-source code provides a foundation for related research and has reference value for fields such as electronic design automation and fault diagnosis.