Interpretable Solar Panel Defect Detection via Fuzzy Rule Extraction from Deep Learning Architectures

Official implementation

---

Overview

This repository contains the official code for our hybrid interpretable AI framework that automatically extracts human-readable fuzzy logic IF-THEN rules from trained deep learning models for solar panel defect detection.

Most deep learning approaches for photovoltaic (PV) defect detection are black boxes. Field technicians cannot act on a prediction they cannot understand. Our framework solves this by combining the accuracy of modern vision transformers with the transparency of fuzzy logic, giving operators both a prediction and a readable explanation.

Key Results

Model	Accuracy	Feature-Severity Correlation	Inference
**Swin-Tiny **	80.96%	0.78	22ms
ConvNeXt-Tiny	80.71%	0.82	18ms
ViT-Tiny	79.44%	—	24ms
EfficientNetB0	74.62%	0.42	11ms
ResNet50	73.60%	0.64	15ms

---

Framework Architecture

Top: Micro-architecture of the Swin Transformer block with W-MSA and SW-MSA attention.

Bottom: Full pipeline — ELPV EL image → preprocessing → deep feature extraction → Pearson correlation → fuzzy rule inference → class prediction + Grad-CAM XAI map.

How It Works

1. Deep Feature Extraction — Five architectures (ResNet50, EfficientNetB0, ConvNeXt-Tiny, ViT-Tiny, Swin-Tiny) extract penultimate-layer features from EL images
2. Fuzzy Rule Extraction — Pearson correlation identifies the single most severity-discriminative feature dimension; a threshold is optimised on the validation set
3. Hybrid Inference — Final prediction combines DL softmax output with fuzzy membership confidence: ŷ = λ·p(c|x) + (1-λ)·μ_c(g(x)) where λ=0.7

Extracted Fuzzy Rules (Best Models)

ConvNeXt-Tiny:
  IF f487 > -19.18  →  severity is MILD   (μ = 0.82)
  IF f487 < -19.18  →  severity is SEVERE (μ = 0.82)

Swin-Tiny:
  IF f101 < 1.445   →  severity is MILD   (μ = 0.78)
  IF f101 > 1.445   →  severity is SEVERE (μ = 0.78)

---

Repository Structure

Interpretable-Solar-Defect-Detection/
│
├── FINAL_NOTEBOOK.ipynb       # Complete pipeline: training, fuzzy extraction, visualisation
├── architecture.png           # Framework diagram (Fig. 1 in paper)
├── cm_swin.png                # Swin-Tiny confusion matrix
├── attn_swin.png              # Swin-Tiny attention map
├── gradcam_resnet50.png       # ResNet50 Grad-CAM map
├── tsne_resnet50.png          # t-SNE — ResNet50
├── tsne_efficientnet.png      # t-SNE — EfficientNetB0
├── tsne_convnext.png          # t-SNE — ConvNeXt-Tiny
├── tsne_swin.png              # t-SNE — Swin-Tiny
├── LICENSE
└── README.md

---

Quick Start

Clone the repo

git clone https://github.com/Latchan-Ch/Interpretable-Solar-Defect-Detection.git
cd Interpretable-Solar-Defect-Detection

Install dependencies

pip install torch torchvision timm scikit-learn matplotlib seaborn numpy pandas jupyter

Download the ELPV dataset

Download from Kaggle and place in your working directory.

Run the notebook

jupyter notebook FINAL_NOTEBOOK.ipynb

The notebook covers end-to-end:

• Data loading and preprocessing
• Training all five architectures
• Fuzzy rule extraction via Pearson correlation
• t-SNE feature space visualisation
• Grad-CAM / attention map generation
• Hybrid inference evaluation

---

Visualisations

t-SNE Feature Space Comparison

ResNet50	EfficientNetB0	ConvNeXt-Tiny	Swin-Tiny

Yellow: No Defect — Purple: Severe Defect — Cyan: Mild Defect

ConvNeXt-Tiny and Swin-Tiny produce clearly separated clusters, directly validating their higher fuzzy rule correlation scores.

Explainability Maps

ResNet50 Grad-CAM	Swin-Tiny Attention

ResNet50 fixates on narrow crack edges. Swin-Tiny attends to both defect regions and the surrounding cell grid, enabling more robust mild-defect reasoning.

---

Dataset

We use the ELPV (Electroluminescence Photovoltaic) benchmark dataset.

• 2,624 grayscale EL images of solar cells (300×300 px)
• Labels mapped to 3 severity classes: No Defect / Mild / Severe
• Split: 70% train / 15% validation / 15% test (stratified)

Class	Samples	Percentage
No Defect	1,510	57.5%
Mild Defect	299	11.4%
Severe Defect	815	31.2%

Download: Kaggle ELPV Dataset

---

Training Details

Setting	Value
Framework	PyTorch
GPU	NVIDIA T4
Input size	224 × 224
Optimizer	Adam (lr=1e-4, wd=1e-5)
Augmentation	Flip, Rotate ±15°, Brightness ±20%
Class imbalance	Weighted random sampling
Seeds	3 independent runs

---

Citations

If you use this code or find this work helpful, please cite:

@inproceedings{chhetri2026interpretable,
  title     = {Interpretable Solar Panel Defect Detection via Fuzzy Rule 
               Extraction from Deep Learning Architectures},
  author    = {Chhetri, Latchan and Das, Hrishikesh and Singh, Yogesh},
  booktitle = {2026 International Conference on Sustainable AI for Social 
               Impact and Global Development (SASIGD)},
  year      = {2026},
  publisher = {IEEE}
}

---

Contact

Latchan Chhetri Dept. of Artificial Intelligence and Data Science Sikkim Manipal Institute of Technology -> latchanchhetri19@gmail.com

---