Solar Panel Fault Detection via Drone-Captured Images

Overview

This project automates the fault detection process in solar photovoltaic (PV) systems using drone-captured thermal images. By leveraging machine learning models, the goal is to identify defective modules based on temperature anomalies captured by infrared cameras.

Problem Statement

Solar PV systems are prone to module failures due to operational stresses and installation errors. Traditional inspection methods are time-consuming and labor-intensive. This project addresses these challenges by developing a machine learning pipeline to analyze drone-captured thermal images and identify defective modules efficiently.

Dataset

The dataset, sourced from Kaggle, consists of:

• Thermal (IR) images: Captured by drones.
• Annotations: Marking the corners and centers of each PV module and labeling them as "defective" or "non-defective."

Challenges

• Imbalanced Data: Few examples of faulty modules compared to non-faulty ones.
• Limited Thermal Data: Addressed using feature extraction techniques.

Features Extracted

• Temperature Statistics: Mean, standard deviation, maximum, and range.
• Distribution Metrics: Skewness and kurtosis.
• Fault Detection: Labels from annotations indicating module condition.

Tested Machine Learning Models

The following models were tested:

1. Support Vector Machine (SVM)
2. Random Forest Classifier
3. Gradient Boosting Classifier
4. K-Nearest Neighbors (KNN)
5. Logistic Regression

Evaluation Metric

Due to the imbalanced dataset, the F1-Score was chosen as the primary evaluation metric, ensuring a balance between precision and recall.

Hyperparameter Tuning

Each model was optimized using GridSearchCV with 5-fold cross-validation. The best parameters for each model were:

Logistic Regression

• C: 10
• Solver: liblinear
• Class Weight: None
• Penalty: l1

SVM (Support Vector Machine)

• C: 0.01
• Kernel: Linear
• Gamma: Scale
• Class Weight: Balanced
• Degree: 2

K-Nearest Neighbors (KNN)

• Algorithm: auto
• Leaf Size: 20
• n_neighbors: 3
• Weights: uniform

Random Forest

• Max Depth: None
• Min Samples Split: 2
• n_estimators: 200

Gradient Boosting

• Learning Rate: 0.1
• Loss: Exponential
• Max Depth: 5
• Min Samples Split: 2
• n_estimators: 50

Performance Overview

Each model was optimized using GridSearchCV with 5-fold cross-validation.

Logistic Regression

• F1-Score: [66.67, 86.96, 77.78, 78.26, 88.00]
• Recall: [53.33, 83.33, 70.00, 64.29, 84.62]
• Precision: [88.89, 90.91, 87.50, 100.00, 91.67]
• Accuracy: [97.07, 98.90, 98.53, 98.16, 98.90]

SVM (Support Vector Machine)

• F1-Score: [100.00, 81.48, 84.21, 76.92, 88.00]
• Recall: [100.00, 91.67, 80.00, 71.43, 84.62]
• Precision: [100.00, 73.33, 88.89, 83.33, 91.67]
• Accuracy: [100.00, 98.17, 98.90, 97.79, 98.90]

K-Nearest Neighbors (KNN)

• F1-Score: [80.00, 76.19, 70.59, 80.00, 91.67]
• Recall: [66.67, 66.67, 60.00, 71.43, 84.62]
• Precision: [100.00, 88.89, 85.71, 90.91, 100.00]
• Accuracy: [98.17, 98.17, 98.16, 98.16, 99.26]

Random Forest

• F1-Score: [63.64, 81.82, 84.21, 78.26, 78.26]
• Recall: [46.67, 75.00, 80.00, 64.29, 69.23]
• Precision: [100.00, 90.00, 88.89, 100.00, 90.00]
• Accuracy: [97.07, 98.53, 98.90, 98.16, 98.16]

Gradient Boosting

• F1-Score: [75.00, 81.82, 66.67, 80.00, 81.82]
• Recall: [60.00, 75.00, 60.00, 71.43, 69.23]
• Precision: [100.00, 90.00, 75.00, 90.91, 100.00]
• Accuracy: [97.80, 98.53, 97.79, 98.16, 98.53]

we clearly notice SVM outperforming the other algorithms.

Results

The SVM (Support Vector Machine) model achieved the best performance among all tested models, using the following optimal hyperparameters:

SVM (Support Vector Machine)

• C: 0.01
• Kernel: Linear
• Gamma: Scale
• Class Weight: Balanced
• Degree: 2

Upon testing, the model yielded the following average performance metrics:

• Accuracy: 98.60%
• F1-Score: 86.11%
• Precision: 86.53%
• Recall: 87.56%

Acknowledgments

• Kaggle dataset "Photovoltaic system thermography" by Marcos Gabriel. link
• Infrared thermal imaging for fault detection in solar panels dataset by Technodivesh

---

This project was made by Ouassim Milous as part of the Machine Learning for Robotics II course by professor Luca Oneto at the University of Genoa.