하디
(Nazari Mohammad Hadi)
1iD
방준호
(Junho Bang)
†iD
최철영
(Chul-Young Choi)
2iD
선로빈
(Robin Sun)
1iD
김든찬
(Deunchan Kim)
1iD
-
(Dept. of IT Applied Engineering, Jeonbuk National University, Jeonju, Rebublic of
Korea.)
-
(Dept. of Energy engineering, Jeonbuk National University, Jeonju, Rebublic of Korea.)
Copyright © The Korean Institute of Electrical Engineers
Key words
VGG16, Hybrid Model, LSTM, Dense Model, short-circuit traces
1. Introduction
Fire disasters pose a serious threat across various sectors, from residential environments
to industrial facilities [1]. Among them, electrical fires resulting from short circuits are particularly hazardous,
endangering both infrastructure and human life.
Investigating the cause of electric fires and determining their presence at fire scenes
is essential for both prevention and post-incident analysis [4]. Electrical fires are typically categorized based on ignition sources, such as electrical
overload, compression damage, tracking, insulation breakdown, and short circuits.
Each cause leaves specific physical evidence, and examining these traces allows experts
to confirm whether an electric fire occurred [5][6].
Timely and accurate detection of these events is crucial to minimizing damage and
improving emergency response effectiveness. However, conventional inspection methods
depend heavily on manual visual analysis, which can be inconsistent, time-consuming,
and prone to human error. With the increasing availability of labeled image datasets
and rapid advancements in deep learning, automated classification systems have emerged
as a promising alternative to traditional methods.
In particular, melting marks on wires or metal components caused by the heat of electric
fires are key indicators, but these are often described using various non-standardized
terms such as "melting marks" or "short-circuit marks" [7][8]. To bring clarity, this paper defines primary short-circuit marks as traces directly
linked to fire ignition and secondary short-circuit marks as those resulting indirectly
from the fire. This classification helps make the investigation process more intuitive
[9].
This paper proposes a deep learning-based approach utilizing a hybrid VGG16 framework
that combines CNN capabilities with either Dense layers or LSTM units. The VGG16 +
Dense model emphasizes simplicity and faster training, while the VGG16 + LSTM model
is designed to capture temporal and spatial features more effectively.
Using transfer learning, the pre-trained VGG16 model was fine-tuned by unfreezing
selected layers. Training data for both models consisted of labeled primary and secondary
short-circuit marks, generated under controlled laboratory conditions.
This paper presents a comparative evaluation of two convolutional neural network (CNN)-based
architectures: VGG16 combined with Dense layers and VGG16 combined with Long Short-Term
Memory (LSTM), for the classification of electric fire short-circuit trace images
[2][3]. By analyzing and comparing the performance of these hybrid architectures, the paper
aims to identify an optimal solution that balances classification accuracy with computational
efficiency, addressing the challenges posed by traditional approaches.
2. Proposed Work
2.1 Research background on model
The Visual Geometry Group at Oxford University developed the VGG architecture, which
emphasizes deeper networks with smaller (3×3) convolution filters for enhanced image
classification performance. VGG16, in particular, is known for its efficient architecture,
comprising 13 convolutional layers and 3 fully connected layers, along with pooling
and SoftMax layers [10].
In this paper, we apply the VGG16 backbone in two hybrid deep learning frameworks
for classifying electric fire short-circuit trace images: VGG16 combined with Dense
layers (VGG16-Dense) and VGG16 combined with Long Short-Term Memory layers (VGG16-LSTM).
The VGG16-Dense model adds fully connected layers after the convolutional feature
extractor to learn complex spatial patterns, offering a straightforward structure
and faster training.
In contrast, the VGG16-LSTM model incorporates an LSTM layer to capture sequential
dependencies within the spatial features extracted from the CNN, which is beneficial
for distinguishing fine grained differences in trace patterns. This architecture is
particularly effective in identifying primary and secondary short-circuit marks, crucial
for determining the root cause of electric fires.
Both models leverage transfer learning from pre-trained VGG16 weights on ImageNet.
For VGG16-LSTM, the last four convolutional layers are unfrozen for fine-tuning, allowing
better adaptation to the specific dataset, while VGG16-Dense maintains a frozen base
for simplicity and efficiency.
Overall, the paper highlights the effectiveness of hybrid CNN-based architectures
specifically those integrating Dense and LSTM layers with VGG16 in enhancing classification
accuracy for electric fire short-circuit detection. This comparative analysis demonstrates
the importance of architectural choice in safety-critical applications such as fire
investigation.
2.2. Dataset preparation
In this paper, a well-structured dataset was constructed to classify electric fire
short-circuit traces, consisting of two distinct categories: primary short-circuit
traces and secondary short-circuit traces. Each category included 867 labeled images,
summing up to a total of 1,734 samples. These images were collected from controlled
laboratory conditions and field investigations to ensure diversity and authenticity
in appearance, shape, and texture of electric fire evidence.
The experimental data were generated in the laboratory using specialized equipment.
Wires were fixed to the anode and moved with a transfer device to create contact and
induce short circuits for sample fabrication. Two copper wire types (1.5SQ and 4SQ)
were tested at 25℃ and 900℃, the latter chosen since fire temperatures typically reach
800–900℃. In the 900℃ tests, wires were either heated or exposed to flames for two
minutes before short-circuiting. The primary short-circuit traces were produced by
continuously applying current until a short circuit occurred, while the secondary
short-circuit traces were formed by applying a flame to induce the short circuit.
The thermal trace was obtained by applying a flame without current until a short circuit
appeared.
To prepare the dataset for model training and evaluation, a train-test split was performed
to ensure balanced representation across classes. Specifically, 348 images comprising
174 from each class were allocated to the test set to facilitate a fair and unbiased
assessment of model performance. The remaining 1,386 images were designated for training,
providing a substantial dataset for learning meaningful and discriminative features.
To improve the model's generalization capability and mitigate overfitting, the training
dataset underwent extensive data augmentation. This included a range of image transformations
such as rotation, scaling, horizontal and vertical flipping, zooming, shifting, and
brightness adjustment. These techniques significantly enriched the diversity of the
training samples without the need for additional data collection, thereby enhancing
the robustness and accuracy of the deep learning models.
The raw images varied widely in resolution and dimension, which posed a challenge
for input compatibility with deep learning architectures. As VGG16 requires fixed
size inputs, a preprocessing pipeline was implemented using Python. Images were first
standardized to 500×375 pixels, and subsequently resized to 224×224 pixels, which
matches the input requirement of the VGG16 backbone. This resizing ensured compatibility
with pre-trained weights and preserved essential visual features. The resizing process
was confirmed using IPython.display to verify visual integrity post-processing.
To further enrich the dataset and improve the model's generalization capacity, data
augmentation was applied using the ImageDataGenerator class from TensorFlow Keras.
This step artificially increases the size and diversity of the training set without
additional data collection. A total of 693 images per class (1,386 images) from the
training set were augmented across multiple transformations that simulate real world
distortions and variabilities. The following augmentation techniques were applied
using the ImageDataGenerator class from TensorFlow Keras.
표 1. 적용된 증강 기법
Table 1. The applied augmentation techniques
|
Technique
|
Value/Range
|
Purpose
|
|
Rescaling
|
1/255
|
Normalize pixel values to [0, 1]
|
|
Rotation
|
±30 degrees
|
Allow to learn rotational invariance
|
|
Width Shift
|
20% of width
|
Tolerate horizontal displacements
|
|
Height Shift
|
20% of height
|
Tolerate vertical displacements
|
|
Zoom
|
±30%
|
Handle variations in image size and scale
|
|
Horizontal Flip
|
True
|
Learn from mirrored images
|
|
Vertical Flip
|
True
|
Increase robustness to vertical variations
|
|
Brightness Adjustment
|
Range [0.7, 1.3]
|
Adapt to different lighting conditions
|
Data augmentation was applied throughout training to improve model generalization.
For the VGG16 + Dense model, trained over 25 epochs, this resulted in about 34,650
augmented images. For the VGG16 + LSTM model, trained over 35 epochs, approximately
48,510 augmented images were generated. This strategy helped both models learn from
more diverse data and improved their classification performance.
These preprocessing and augmentation steps formed a critical foundation for the deep
learning framework, enabling the hybrid models to achieve high accuracy in detecting
and distinguishing between primary and secondary electric fire short-circuit traces.
2.3. The two propsoed model architectures
In this paper, we utilized the VGG16 model as the backbone of two hybrid deep learning
architectures VGG16 + Dense and VGG16 + LSTM to classify primary and secondary short-circuit
trace images. Deep learning networks typically require extensive high-quality annotated
image data for effective training. However, in fire investigation scenarios, short-circuit
traces are often difficult to collect and label due to their random and irregular
appearances in real cases.
To address the challenge of limited data, we employed transfer learning by leveraging
the pre-trained weights of the VGG16 model trained on the ImageNet dataset. While
ImageNet primarily contains natural scene images unrelated to electric fire traces,
the large scale and diverse data distribution helps VGG16 extract generalized low-level
features. These features were then fine-tuned on our specialized dataset, enhancing
the learning efficiency and reducing overfitting.
The two hybrid architectures differ in their structure and objectives. The VGG16 +
Dense model employs a frozen or partially trainable VGG16 convolutional base, followed
by a Flatten layer and several Dense layers for classification. It focuses on extracting
and processing spatial features in a straightforward feedforward manner.
In contrast, the VGG16 + LSTM model takes the output of the convolutional layers,
reshapes the feature maps into sequences, and processes them through an LSTM layer.
This allows the model to capture spatial dependencies across the feature sequences,
improving its ability to recognize complex patterns in the trace images.
A notable distinction between the two models lies in the learning rate settings. The
VGG16 + Dense model uses a learning rate of 1e-4, which allows for faster convergence
due to its relatively simple architecture and fewer trainable parameters. Meanwhile,
the VGG16 + LSTM model uses a lower learning rate of 1e-5. This is because the LSTM
layer introduces more complexity and sensitivity to weight updates, requiring smaller
learning steps to ensure stable and effective training. A higher learning rate in
the LSTM-based model could result in unstable gradients or poor convergence, particularly
when learning long range spatial dependencies.
The key architectural and performance differences between the two models are summarized
below:
표 2. 제안된 두 모델의 구조 및 성능 비교
Table 2. The key architectural and performance differences between the two proposed
models
|
Aspect
|
VGG16-Dense
|
VGG16-LSTM
|
|
Feature extractor
|
Fully frozen
|
Last 4 layers fine-tuned
|
|
After CNN
|
Flatten
|
Reshape + LSTM
|
|
Type of classifier
|
Fully connected layers
|
Sequential(memory-based) layer
|
|
Learning rate
|
1e-4 (faster)
|
1e-5 (slower)
|
|
Complexity
|
Lower
|
Higher
|
|
Strength
|
Faster training, simpler
|
Can capture spatial dependencies
|
3. Result and Experimental Analysis of the Two Proposed Models
All experiments and model training were conducted on a 64-bit Windows 11 machine equipped
with an AMD Ryzen 9 5900X 12-Core Processor 3.70 GHz, 32 GB of RAM, and running on
a x64-based architecture. The models were implemented using Python, leveraging the
Keras library with TensorFlow backend. The training process was executed in an Anaconda
environment to take advantage of hardware acceleration. Although no dedicated GPU
was specified in the system details, the CPU provided sufficient performance for moderate-scale
deep learning experiments.
The dataset was categorized into two classes: primary short-circuit marks and secondary
molten marks. These were further divided into training and testing sets. To improve
model generalization and robustness, data augmentation techniques including image
rotation, shifting, zooming, flipping, and brightness adjustment were applied to the
training images.
The paper implemented two transfer learning-based hybrid models: VGG16 + Dense and
VGG16 + LSTM. These architectures utilized the VGG16 convolutional base pre-trained
on ImageNet, followed by custom classification layers suited for the specific characteristics
of electric short-circuit image data. The sigmoid activation function was used in
the final layer for binary classification, and Adam optimizer was employed for efficient
gradient-based optimization.
Model performance was evaluated using accuracy and loss metrics, with separate calculations
for training and validation datasets using standard evaluation formulas. The performance
metrics are as follows:
그림 1. 학습 및 검증 정확도 (VGG16+Dense)
Fig. 1. Training and validation accuracy (VGG16+Dense)
그림 2. 학습 및 검증 손실 (VGG16+Dense)
Fig. 2. Training and validation loss(VGG16+Dense)
그림 3. 혼동 행렬 (VGG16+Dense)
Fig. 3. Confusion matrix (VGG16+Dense)
To optimize training and avoid overfitting, the early stopping technique was applied
to both models. Although the training process for each model was initially set to
run for 50 epochs, the early stopping callback automatically halted training when
the validation loss showed no improvement over a defined patience period. As a result,
the VGG16 + Dense model stopped at epoch 25, and the VGG16 + LSTM model stopped at
epoch 35, preserving the best weights and preventing unnecessary training. This approach
not only improved model generalization but also reduced computational time and resource
usage. While both models achieved strong results, the VGG16 + LSTM consistently outperformed
the Dense variant across all metrics.
그림 4. 학습 및 검증 정확도 (VGG16+LSTM)
Fig. 4. Training and validation accuracy (VGG16+LSTM)
그림 5. 학습 및 검증 손실 (VGG16+LSTM)
Fig. 5. Training and validation loss (VGG16+LSTM)
그림 6. 혼동 행렬 (VGG16+LSTM)
Fig. 6. Confusion matrix (VGG16+LSTM)
표 3. 제안된 VGG16+LSTM 모델의 분류 성능
Table 3. Classification Performance of the Proposed VGG16+LSTM Model
|
Class
|
Precision
|
Recall
|
F1-Score
|
|
Burn marks
|
0.98
|
0.99
|
0.99
|
|
Melt Marks
|
0.99
|
0.98
|
0.99
|
To evaluate the effectiveness of the two proposed models, VGG16 + Dense and VGG16
+ LSTM, their performance on the validation set was measured using accuracy and loss
metrics. These metrics provide insight into how well each model generalizes to unseen
data. Validation accuracy indicates the proportion of correct predictions, while validation
loss reflects the model's confidence and prediction error. The results, summarized
in Table 3, demonstrate the differences in learning behavior and performance between the two
architectures.
표 4. 제안된 모델들의 검증 성능 비교
Table 4. Validation performance comparison of the proposed models
|
Metric
|
VGG16+Dense
|
VGG16+LSTM
|
|
val_accuracy
|
95.11%
|
98.85%
|
|
val_loss
|
0.1532
|
0.0365
|
4. Conclusion
In this paper, the VGG16–Dense hybrid model and the VGG16–LSTM hybrid model are proposed
for analyzing the causes of electrical fires, and the performance of these two models
is comparatively evaluated.
In conclusion, both the VGG16 + Dense and VGG16 + LSTM models proved effective for
classifying electric fire short-circuit images, leveraging the power of transfer learning
and deep neural networks. The VGG16 + Dense model offers faster training and simpler
architecture, making it suitable for deployment in resource-constrained environments.
On the other hand, the VGG16 + LSTM model, though more complex, achieved superior
performance by capturing spatial dependencies in the data more effectively.
Performance metrics validate these differences: the VGG16 + Dense model achieved a
validation accuracy of 95.11% with a validation loss of 0.1532, whereas the VGG16
+ LSTM model reached an impressive 98.85% accuracy with a significantly lower validation
loss of 0.0365. These results highlight that incorporating LSTM layers can significantly
enhance the model's ability to learn complex spatial patterns in short-circuit image
traces.
Overall, the paper confirms that combining convolutional and sequential deep learning
techniques can significantly improve the classification of electric short-circuit
marks, offering practical solutions for intelligent fire detection and forensic analysis
systems.
Acknowledgements
This work was supported by the Korea Institute of Energy Technology Evaluation and
Planning (KETEP) grant funded by the Korea government (MCEE) RS-2022-KP002707, Jeonbuk
Regional Energy Cluster Training of human resources.
References
J. Jeong, S. Bae, H. Yoon, J.-H. Kim, 2025, A novel approach for UV signal-based fire
detection and classification leveraging contrastive learning, Expert Systems with
Applications, Vol. 237, pp. 127120
S. Kharroubi, S. Boubaker, 2023, Assessment of machine and deep learning approaches
for fault diagnosis in photovoltaic systems using infrared thermography, Remote Sensing,
Vol. 15, No. 6, pp. 1686
Y. Bekele, S. Rani, D. B. Kanagachidambaresan, M. Moghaddam, S. Licea, C. Munoz, A.
Ortiz, B. S. Tadese, 2025, Hybrid CNN-LSTM detection of cyber-attacks on MPPT controllers
in PV systems, pp. 1-6
S. H. Jung, S. M. Kim, S. H. Lee, 2020, Electrical Arc Detection using Convolutional
Neural Network, J. Broadcast Eng., Vol. 25, No. 4, pp. 569-575
J. Park, J. H. Kang, E. P. Lee, Y. H. Ko, S. B. Bang, 2021, New Approach to Distinguish
Copper Molten Marks Based on Quantitative Microstructure Analysis Using Electron Backscatter
Diffraction, Fire Technol., Vol. 57, No. 4, pp. 1667-1682
L. Wang, H. W. Yao, Z. X. Wu, D. Liang, 2014, Research on evolution of electrical
fire melted marks based on digital image processing technology, pp. 793-797
V. Babrauskas, 2004, Arc beads from fires: Can ‘cause’ beads be distinguished from
‘victim’ beads by physical or chemical testing, J. Fire Prot. Eng., Vol. 14, No. 2,
pp. 125-147
A. Ayala, B. Fernandes, F. Cruz, D. Macêdo, A. L. I. Oliveira, C. Zanchettin, 2020,
KutralNet: A Portable Deep Learning Model for Fire Recognition
E. X. Li, N. Xu, N. Li, H. Ma, J. B. Shi, M. Zhao, 2018, Analysis of Primary Short-Circuited
Melted Mark Evidence Identification in Electrical Fire Case, Appl. Mech. Mater., Vol.
876, pp. 161-165
K. Chatfield, K. Simonyan, A. Vedaldi, A. Zisserman, 2014, Return of the devil in
the details: Delving deep into convolutional nets, pp. 1-11
저자소개
하디 (Nazari Mohammad Hadi)
He is currently enrolled in a Master degree program in the dept. of IT Applied System
Engineering at Jeonbuk National University. His main research interests are deep learning
and energy optimization., Artificial Intelligence, Machine learning modeling and computer
software developments.
He received B.S., M.S. and Ph.D. degrees in Department of Electrical Engineering from
Jeonbuk National University, in 1989, 1991 and 1996 respectively. He was a research
engineer with LG Semiconductor from 1997 to 1998. He is currently working as a professor
in Division of Convergence Technology Engineering and Department of Energy/Conversion
Engineering of Graduate School, Jeonbuk National University, Jeonju, Rebublic of Korea.
His main research interests include IT convergency systemdesign.
He completed his Master’s degree in the Department of Energy Engineering from March
2020 to August 2022, and he began his Ph.D. program in September 2022. His main research
interests include energy systems and offshore wind.
He completed his Master’s degree in the Department of Energy Storage Conversion Engineering
in February 2022, and he began his Ph.D. program in the Department of IT Applied System
Engineering in August 2022.
He is currently enrolled in a Ph.D. program at Jeonbuk National University with a
master's degree in the dept. of IT Applied Systems Engineering. His main research
interests are fault diagnosis and anomaly detection for power distribution facilities.