Thermal Image-Based Fault Detection and

Localization Using Deep Learning

Created By:
Raveesha Anuruddha Ranamukage

1
1. Abstract
This report outlines a deep-learning approach for detecting and locating faults in electrical
systems through thermal imaging. While traditional thermography is a non-intrusive method, it
often suffers from the limitations of manual analysis and environmental factors that can impact
accuracy. This project incorporates YOLOv8, a cutting-edge object detection model, to
automate the fault detection process, improving diagnostic accuracy, operational safety, and
preventive maintenance. The proposed solution was tested on a dataset of annotated thermal
images, showing strong performance in identifying common electrical faults. Future
improvements will focus on expanding applications to high-voltage systems and addressing the
limitations of the dataset.

2. Introduction
Electrical systems can experience faults like loose connections, overloaded circuits, and failing
components, which often show up as increased heat. Thermography, which uses infrared
imaging to detect temperature differences, is a non-invasive technique for spotting these issues.
However, traditional thermography methods can be time-consuming, depend heavily on manual
interpretation, and may lead to inaccuracies.
A significant drawback of conventional thermography is that it is not performed frequently
enough. Inspections are usually done once or twice a year, which can result in missing faults
that develop outside of these scheduled checks. By automating and increasing the frequency of
inspections with the help of artificial intelligence, this gap can be filled, allowing for real-time
detection and preventive maintenance. Identifying faults early enables prompt corrective
actions, which helps minimize damage and cut operational costs.
Integrating AI greatly improves thermography by facilitating automation, real-time analysis,
and standardization. Deep learning models, especially, can enhance the accuracy of fault
detection and lessen the reliance on human expertise. This project utilizes YOLOv8, a real-time
object detection model, to identify and locate electrical faults, thereby boosting safety measures,
enhancing efficiency, and overcoming the limitations of manual inspections.

3. Methodology
3.1 Generalization Across Sites
To ensure adaptability, the deep learning model was trained using data gathered from various
sites, which included a range of environmental and operational conditions. By incorporating
images from different panels and locations, the model was able to learn how to handle variations
like emissivity, reflectivity, and temperature discrepancies. This training method allows the
model to generalize fault detection patterns across new and previously unseen panels, making
it suitable for a wide range of applications.

2
3.2 Thermography Process Overview
Thermography is based on the detection of infrared radiation emitted by objects that are above
absolute zero. FLIR T335 infrared cameras were utilized to capture thermal images, which
convert temperature differences into visual representations. During data collection, factors such
as emissivity, reflectivity, humidity, and proximity to heat sources were considered to ensure
accuracy.
3.3 Data Collection and Preparation
• Source: Over 1,000 infrared images of electrical panels were collected from Amithi
Power Consultants (Pvt.) Ltd. The dataset focused on medium- and low-voltage
systems, but the approach can be extended to other components with appropriate
datasets.
• Annotation: Faults were categorized and labeled using Roboflow, with bounding
boxes drawn around fault areas.
• Augmentation: Techniques such as flipping and color adjustments expanded the
dataset to 1,676 images.
• Dataset Split: The dataset was divided into 70% training, 20% validation, and 10%
testing.
Link to the Dataset

Figure 1:FLIR Camera

Therma Image Capturing

3.4 Standards and Assumptions

The project followed specific standards to ensure its practical relevance. During the Manual
Annotation of the Dataset, we relied on these standards and assumptions to identify relevant
fault classes
Standards:
• IEC 61439-1: This standard specifies the allowable temperature rises for low-voltage
panel boards, including busbars, conductors, and insulated enclosures. Thermography

3
 surveys based on this standard help confirm that component temperatures stay within
safe limits.
• IEC 60502: This outlines the temperature limits for power cables under both normal
and fault conditions.
• IEC 62271: This provides guidelines for conducting thermography surveys of high-
voltage switchgear, ensuring compliance and safety.

Assumptions:
• The reflection temperature was consistent across all images.
• The temperature range was restricted to 30°C–60°C. In this case, the Temperature
Variation Bar on the right side of the image was disregarded to train the model on heat
patterns.
• Factors such as emissivity, reflectivity, relative humidity, and proximity to heat-emitting
sources were considered to reduce errors in thermal image readings.

3.5 Fault Categories

The dataset included images of various fault categories:
• Overload Circuits: Indicated by excessive heat patterns on wires and components.

• Connectivity Issue: Loose connections or corrosion visible through irregular heat

patterns.

4
• Deteriorated Insulation: Characterized by irregular heat patterns near connection
points.

• Phase Imbalance: Shown by unequal heat distribution across phases.

• Faulty Circuit Breakers: Identified by unusual heat concentrations.

5
 • Normal Operation: Represented by expected and uniform heat patterns.

3.6 Deep Learning Model Selection

This project is a computer vision problem so we consider computer vision task models for this
project.
YOLOv8 was selected for its real-time object detection capabilities, balancing speed and
accuracy. A comparison of potential models is shown below:

Model Advantages Limitations Suitability

High classification Not specialized for Not ideal for fault
ResNet
accuracy localization localization
High object Slower inference Impractical for real-
Faster R-CNN
detection accuracy speed time use
Fast and accurate Requires a large Chosen for real-time
YOLOv8
detection dataset fault detection

4. Implementation

4.1 Training Environment

• Platform: Google Colab
• Framework: PyTorch
The reason to select Pytorch is the easiness of using the Yolo model using Pytorch.

4.2 Training Data and Model Configuration

The YOLOv8 model was trained on a carefully prepared dataset comprising 668 thermal images
with 6 fault classes: Connectivity Issues (Loose Connection/ Corroded), Deteriorated
Insulation, Faulty Circuit Breakers, Normal Operation Conditions, Overloaded Circuits, and
Phase Imbalance. The dataset was divided into training (504 images), validation (102 images),
and testing (62 images), ensuring a balanced evaluation across all stages.
6
Preprocessing involved resizing images to 640×640 pixels. This dimension is a well-balanced
size, being computational-efficient and at the same time allowing the resolution needed to detect
complicated fault features in thermal images. The square format adheres well to the structure
of YOLOv8, where the network functionality is optimized with constant input sizes.
• Augmentation Techniques: Horizontal and vertical flipping, cropping, and little
blurring were applied to enhance the generalization of the model and help in the
detection of faults under differing conditions.
• Epochs and Training Time: The model was trained through 50 epochs, a time length
decided on the basis of seeing convergence loss and accuracy measures through
preliminary runs. This process was conducted on Google Colab with GPU acceleration
which took up to about 29 minutes. This is a fair balance between the computational
cost and performance of the model.

Link to the code

4.3 Performance Metrics
The model’s performance was evaluated using the following metrics:
• Precision: 72.8% (correct fault detections).
• Recall: 79.3% (proportion of actual faults detected).
• mAP50: 81.1% (mean average precision at 50% Intersection Over Union).

4.4 Performance Curve on Validation

The subsequent Precision-Recall Curve delineates the detection performance across various
fault classes, emphasizing the trade-offs between precision and recall for each category. The

7
[email protected] score of 81.1% signifies the model's robust overall accuracy:

• Fault classes such as "Phase Imbalance" and "Overloaded Circuits" attain high precision
and recall, which indicates the model's reliability in identifying these issues.

• the lower performance observed for "Connectivity Issue - Loose Connection" and
"Deteriorated Insulation" suggests challenges that may arise because of subtle
temperature variations or dataset imbalance.

• the graph showcases the model's ability to generalize across diverse fault types, it
simultaneously maintains consistent detection accuracy. This dual capability is crucial
for practical applications in real-world scenarios.

5. Results and Discussion

8
Fault Class Predicted by Model
Connectivity Issue

Deteriorated Insulation

Overload Circuits

Faulty Circuit Breakers

Phase Imbalance

9
Normal Operation Condition

10
5.1 Key Findings
• The model illustrated strong capabilities in fault detection for typical electrical
problems.
• Real-time processing features underscore its practical applicability in preventive
maintenance.
• Rare faults showed decreased detection accuracy because of dataset imbalance.
• Automating fault detection not only reduces reliance on trained personnel but also
expedites inspections, thereby enhancing scalability for large facilities.
• AI integration enhances thermography by identifying subtle patterns or anomalies that
could be overlooked during manual inspections, although this can be challenging at
times.
5.2 Limitations
• Dataset limitations: Predominantly focused on medium- and low-voltage systems,
with limited data for rare faults.
• Environmental factors: inconsistent temperature evaluations, which significantly
impacted certain readings.

6. Practical Applications
The combination of AI and thermography is proving to be a game-changer for spotting faults,
with several key applications:
• Preventive Maintenance: Using thermal cameras for real-time fault detection helps cut
down on downtime and maintenance expenses.
• Scalability and Efficiency: Automating inspections allows for large facilities to be
monitored without needing a lot of manpower.
• Enhanced Safety Systems: Automated alerts notify operators about possible hazards,
which helps lower the risks tied to electrical failures.
• Standardized Inspections: This approach ensures consistency and adherence to
industry standards, reducing the inconsistencies often found in manual assessments.

7. Future Work
• Broaden the dataset to cover high-voltage systems and uncommon fault types.
• Boost the accuracy of fault categorization and detection using advanced augmentation
techniques.
• Look into integrating with wider electrical diagnostic systems for more thorough
monitoring.

11
8. Conclusion

This project has effectively showcased how deep learning can be combined with
thermography for detecting electrical faults. The YOLOv8 model achieved precise and real-
time fault localization, tackling significant challenges faced by traditional methods. Moving
forward, we aim to strengthen the model’s reliability and expand its range of applications,
making it a crucial asset for diagnosing and maintaining electrical systems.

12