SynRM Flux Linkage Identification via PINN and PCGrad

This repository contains the official PyTorch implementation for the paper:
"High-Precision Standstill Flux Linkage Identification for Synchronous Reluctance Motors Based on Gradient Projection and Implicit Inductance Modeling"

💡 Overview

Accurate flux linkage modeling is essential for the high-performance control of Synchronous Reluctance Motors (SynRM) due to their severe magnetic saturation and cross-coupling effects. However, traditional data-driven methods struggle with physical consistency, and conventional Physics-Informed Neural Networks (PINNs) easily fall into multi-task gradient conflicts when facing measurement noise.

This project proposes a novel high-precision standstill identification framework featuring:

1. Implicit Secant Inductance Architecture (SCN): Strictly satisfies the physical zero-point constraint and all-plane parity symmetry from the mathematical foundation.
2. Gradient Surgery (PCGrad): Eliminates gradient conflicts between data-fitting and Maxwell's reciprocity physical constraints without relying on noise-sensitive PDE loss.
3. Robust Evaluation: Outperforms traditional analytical formulas and baseline ANNs even under severe integral drift and rotor vibration noise.

📂 Repository Structure

SynRM_PINN_Project/
├── configs/                   # Configuration files (hyperparameters, paths)
│   └── pinn_train.yaml        # PINN training and evaluation config
├── data/                      # Dataset directory
│   ├── processed/             # Experimental data (e.g., hysteresis voltage injection)
│   └── simulation/            # Simulation data (2000/4000/8000 ideal points)
├── outputs/                   # Auto-generated results
│   └── [Dataset_Name]/
│       ├── models/            # Saved .pth model weights
│       ├── logs/              # Training logs (.csv)
│       └── evaluation_results/# Generated 3D flux surfaces, error heatmaps, loss curves
├── src/                       # Core source code
│   ├── dataset.py             # Data loading and preprocessing
│   ├── models.py              # Neural network definitions (ANN, SCN, PINN)
│   ├── baselines.py           # Traditional methods (Polynomial, Look-Up Table)
│   ├── train_engine.py        # Core training logic
│   ├── losses.py              # Loss weights and PCGrad optimizer wrapper
│   ├── utils.py               # Utilities (Seed, Logger)
│   └── visualization.py       # Academic plotting functions
├── plot_waveforms.py          # Script for plotting oscilloscope-style waveforms
├── train.py                   # Main training entry point
├── evaluate.py                # Main evaluation and comparison entry point
└── requirements.txt           # Python dependencies

⚙️ Installation

Clone this repository and install the required dependencies:

git clone [https://github.com/KrisCCeng/SynRM-PINN.git](https://github.com/KrisCCeng/SynRM-PINN.git)
cd SynRM-PINN
pip install -r requirements.txt

🚀 Quick Start

1. Training the Models

You can train three different types of models:

• ANN: Traditional Black-box Neural Network.
• SCN: Structure-Constrained Network (Implicit Inductance + Absolute Symmetry).
• PINN: Fully Constrained Network (SCN + Maxwell Reciprocity via PCGrad).

Modify the configs/pinn_train.yaml file to set your mode, data_dir, and w_phys (physical loss weight). Then run:

python train.py --config pinn.yaml

To perform ablation studies on the physical loss weight, simply change the w_phys parameter in the .yaml file.

2. Evaluation and Comparison

Evaluate all trained models and analytical baselines. This script will automatically calculate MAE, RMSE, MAPE, and generate high-resolution SVG visualizations including 3D flux surfaces and error heatmaps:

python evaluate.py --config pinn.yaml

3. Plotting Experimental Waveforms

To plot oscilloscope-style waveform figures for the standstill experimental datasets:

python plot_waveforms.py --data_dir data/processed/ExDataset1

📊 Evaluation Metrics

The framework will automatically generate an evaluation_metrics.csv report and print a formatted table in the console, comparing the training time and accuracy (d/q-axis MAE, RMSE, MAPE) across the Traditional Polynomial model, ANN, SCN, and PINN.

📝 Citation

If you find this code or our paper useful for your research, please consider citing our work:

@article{Ye2026SynRM,
  title={High-Precision Standstill Flux Linkage Identification for Synchronous Reluctance Motors Based on Gradient Projection and Implicit Inductance Modeling},
  author={Cheng Ye and Yankai Jia and Yusheng Song and Changxing Guo and Chuanwen Shen},
  journal={Submitted to Journal},
  year={2026}
}