Fusion of Electrical and Optical Methods in the Detection of Partial Discharges in Dielectric Oils Using YOLOv8

Abstract

This study presents an innovative bimodal approach for laboratory partial discharge (PD) analysis using a YOLOv8-based convolutional neural network (CNN). The main contribution consists, first, in the transformation of a conventional DDX-type electrical detector into a smart and autonomous data source. By training the CNN, a system capable of automatically reading and interpreting the data from the detector display—discharge magnitude and applied voltage—is developed, achieving an average training accuracy of 0.91 and converting a passive instrument into a digitalized and structured data source. Second, and simultaneously, an optical visualization system captures direct images of the PDs with a high-resolution camera, allowing for their morphological characterization and spatial distribution. For electrical voltages of 10, 13, and 16 kV, PDs were detected with a confidence level of up to 0.92. The fusion of quantitative information intelligently extracted from the electrical detector with qualitative characterization from optical analysis offers a more complete and robust automated diagnosis of the origin and severity of PDs.

1. Introduction

Power transformers are fundamental components of a high-voltage electrical network, the failure of which can cause costly interruptions and long periods of downtime [1]. Dielectric oil, also known as transformer oil, is a high-purity, low-viscosity mineral oil which is essential for the operation of transformers and other electrical equipment. It serves a dual purpose, acting firstly as an electrical insulator between conductive components and secondly as a coolant, dissipating the heat generated by the core and windings. Bubbles within mineral oil are one of the most common defects in oil-paper insulation systems, weakening their structure and jeopardizing the operational safety of transformers [2]. This weakening is a precursor to the partial discharges (PDs) that can occur in them. The detection of PDs is a key aspect for preventive maintenance that helps to avoid transformer failure and for optimization of the number of scheduled transformer shutdowns.

PDs are important phenomena and one of the main indicators of the degradation of the integrity of electrical insulation, whether solid, liquid, or gaseous. They manifest as very short-duration electrical events—from tens of nanoseconds to microseconds—which the IEC 60270 [3] standard defines as localized, low-magnitude dielectric breakdowns. Although they do not cause immediate failure, their persistence over time can erode the insulation until it completely breaks down. Therefore, their early detection and accurate characterization are crucial to prevent catastrophic failures in high-voltage equipment.

Unimodal techniques (UTs) for PD detection are based on the diverse physical phenomena that accompany them. Each discharge, stochastic and aperiodic in nature, produces a current pulse. This pulse can produce acoustic waves, generate electromagnetic radiation, and release electrical charges [4]. The electrical method, standardized by IEC 60270 [3], is the reference technique for measurements in controlled laboratory environments due to its high sensitivity [5]. For monitoring under real-life operating conditions, non-conventional methods such as ultra-high frequency [6] and acoustic emission detection [7] are used. However, these techniques also present challenges such as the sensitivity of the acoustic method to the sensor location or the effects of radio frequency interference on the ultra-high frequency method [4].

Complementing electrical methods, optical techniques have emerged as a valuable diagnostic tool, focusing on the detection of the weak electroluminescence emitted during a discharge. Research in this field covers a wide spectrum of sensor technologies and analysis methodologies. To overcome this signal weakness, highly sensitive sensors such as silicon photomultipliers (SiPMs) have been developed, capable of detecting PD events with high accuracy, demonstrating their feasibility for real-time monitoring [8]. Beyond mere point detection, direct visualization of the phenomenon offers invaluable diagnostic value. Advances in imaging sensors have enabled the visual characterization of discharges, such as the corona effect, even using low-cost sensors in specific conditions such as aeronautical ones [9].

The current trend is towards combining advanced image sensors with artificial intelligence techniques to automate analysis. For example, the use of high-resolution CMOS sensors, whose rich visual information is processed by CNN, has been demonstrated to characterize and classify PDs in dielectric oils [10]. More advanced research explores the capture of information across multiple multispectral light spectra, such as ultraviolet and visible, to improve the detection, recognition, and assessment of discharges, providing a more complete optical signature of the phenomenon [11]. In turn, more sophisticated deep learning models, such as those based on the transformer architecture, are being developed to analyze composite optical sensing data and achieve even more accurate and robust PD identification [12].

Overall, the main advantage of these optical methods is their ability to provide direct and intuitive visual evidence of the physical location, morphology, and propagation of the discharge, which is vital for identifying the exact point of insulation degradation. However, optical data are often qualitative and may be less sensitive for quantifying the electrical severity of the discharge compared to the standardized electrical method.

Table 1. Comparison of representative existing methods with our proposed approach.

Approach	Sensors/Data Source	Information Obtained	Key Limitations
Traditional electrical (IEC 60270)	Coupling capacitor, Measuring impedance [3]	Quantitative (Apparent Charge)	No spatial/morphological info. Requires expert interpretation.
Acoustic & UHF Fusion	Acoustic sensors, UHF antennas [4]	Localization, Event detection	Indirect correlation with charge magnitude. Can be affected by noise/barriers.
Optical (e.g., SiPMs, multispectral cameras)	Optical [8,11]	High sensitivity, Morphological	Often qualitative. Less direct quantification of electrical severity.
Our proposed method	DDX detector screen as image source and High-Resolution CMOS camera	Quantitative (charge, voltage) and Spatial & Morphological	Discussed in Section 2

summarizes a Comparison of representative existing methods with our proposed approach, focusing on aspects like the sensors and data sources used, the information obtained, and the key limitations.

It is well known that the patterns or signatures of PDs depend on the type of PD produced [13]. However, these patterns often require expert interpretation, which introduces subjectivity, limits scalability, and makes it difficult to accurately differentiate between different types of failures. This dependence on human factors constitutes a bottleneck for large-scale, real-time monitoring.

To overcome these limitations, the scientific community has turned to the use of artificial intelligence (AI), which has demonstrated superior ability to automate feature extraction, recognize complex patterns in data, and classify discharge types with accuracy, surpassing conventional methods [14] and opening a new era in isolation status diagnosis.

Despite advances in AI using UTs, the information derived from this detection method remains one-dimensional, whether electrical, acoustic, or electromagnetic in nature. AI analysis can optimize the interpretation of such data but cannot generate more information.

Multisensory data fusion, or multimodal techniques (MTs), has emerged as a key strategy for overcoming the limitations of UT systems, representing one of the most promising frontiers in PD diagnosis. Comprehensive reviews of the state of the art, such as those presented in [15,16], provide in-depth analysis of the fundamentals, methodologies, benefits, and challenges of this discipline, demonstrating its potential for achieving more robust and reliable diagnoses.

In [17], a novel analysis method was developed that combines multimodal data and time sequences to provide rapid diagnosis in power transformers. In [12], an innovative bimodal transformer-based deep learning model was developed that uses optical data for PD classification. The system identifies sparks, corona, and surface discharges instantaneously, demonstrating near-perfect efficiency.

To improve the low accuracy of traditional PD recognition methods, in [18] a new system based on data fusion was presented. This method combines a statistical model and a CNN to analyze phase-resolved PD [19] images. The results of both are integrated using the Dempster-Shafer theory [20,21] achieving a recognition accuracy greater than 94%, which is a significant improvement over conventional approaches.

A robust diagnosis requires the synergy of two or more types of information. The present study integrates the information provided by an electrical sensor and an optical sensor. It combines the quantification of electrical severity by assessing the magnitude of the charge and the characterization of its location and physical morphology through optical imaging. Since no single method covers both dimensions, the combination of modalities is presented as an indispensable strategy for a complete understanding of the phenomenon.

Along these lines, our work proposes a bimodal approach that creates a unique synergy by combining the precise quantification of the electrical method with the spatial and morphological characterization of the optical method. Thus, while the electrical detector answers the question of the severity of the discharge magnitude, the optical detection analyzes the nature of the defect by locating its origin and shape. In both cases, AI analysis automates and enhances these two complementary procedures.

Our choice aligns with the conclusions of comprehensive state-of-the-art reviews [22], which establish that one-stage detectors like YOLO are optimized for real-time applications, unlike two-stage models like Faster R-CNN. Comparative studies such as those presented in the development of previous YOLO architectures [23] already demonstrate with benchmarks on standard datasets like COCO that they achieve a superior balance between speed and accuracy compared to models like Faster R-CNN.

In PD detection in dielectric oils, the identification of the pulses produced is of great importance. All our tests are video recorded in order to extract as much information as possible. In this way, we detect and identify these pulses in real time. To carry out this task, YOLO (You Only Look Once) [24] is used as a tool for object detection in an image. We evaluated the use of YOLO because it can be used in Python environments with OpenCV [25].

YOLO models object detection as an extension of a regression problem, dividing the image into a series of cells and predicting bounding boxes and their confidence levels for each cell. This allows for a parallel search within the entire image, making it extremely fast when using graphics processing units (GPUs) at the same time.

Object detection is a core technology in many AI applications and is the fundamental goal of computer vision. In addition to detecting the presence of objects in images, it is also desirable to determine their position within them with an acceptable level of precision, as well as a confidence score of the class to which they belong.

It is a supervised learning problem that involves providing labeled data to an algorithm for training. Subsequently, when new unlabeled data unknown to the algorithm is introduced, it will be able to recognize certain patterns in the new data.

Presented by its authors in 2015 [24], YOLO is a set of open-source algorithms for real-time object detection. Its architecture introduces a paradigm shift and marks a milestone in the study of computer vision. It is a single-pass object detector that uses a complex CNN to predict bounding boxes and class probabilities for objects of interest in input images.

Previously, the most widely used approach to analyze images was the use of CNN and the sliding window concept. This involves choosing a window of a certain size and scanning the entire image, thereby detecting any trace of an object within that window. This method is very slow because it has to scan the entire image to try to find objects whose sizes, spatial orientations, and shapes can vary greatly. In contrast, YOLO can process images in real time with acceptable average accuracy at a speed of 155 frames per second (FPS).

Early versions of YOLO used a CNN architecture with a total of 24 convolutional layers, 4 max-pooling layers, and 2 fully connected layers. To operate, YOLO resizes the image by normalizing the input to 224, 448, or 640 pixels before passing it through the CNN.

Since the launch of its first version (YOLOv1), YOLO has evolved. Version v2 was launched in 2016 [26] and v3 in 2018 [27], with the introduction of the Darknet-19 architectures in v2 and Darknet-53 in v3. Versions v4 [28] and v5 [29] were launched in 2020, v6 [30] and v7 [23] in 2022, and v8 in 2023 [31].

After evaluating the state of the art, the YOLOv8 architecture was selected. This family of models is paradigmatic in one-stage object detection, offering a balance between speed and accuracy that outperforms two-stage architectures such as Faster R-CNN for real-time applications. Although the YOLO family continues its rapid evolution (with versions such as v9–v12 already available), the choice of YOLOv8 was based on its status as a mature and stable release at the time of this project. Its robust performance, combined with broad community support and comprehensive documentation, provided the ideal framework to ensure the reliability and reproducibility of our results.

This evolution has focused on increasing the detection speed by optimizing hyperparameters through the application of genetic algorithms. A comprehensive review of YOLO architectures in computer vision from YOLOv1 to YOLOv8 can be found in [32]. At the time of writing, the most recent released version is YOLOv12. For our work, we chose YOLOv8 because it is a mature technology and fits our real-time video processing needs, facilitating the use of GPUs to increase training and inference speeds.

YOLO has played a prominent role in numerous activities including, among others, applications in agriculture [33], industry [34], and the detection of objects in flight [35].

To implement this bimodal approach, two methods are combined. This article presents a novel diagnostic system for PD analysis in dielectric oils that integrates data from a conventional DDX electrical detector with high-speed image characterization using a YOLOv8 CNN [34]. The objective is to demonstrate that this combination of methods provides a substantially more complete and robust diagnostic view than that obtained separately, laying the foundation for more reliable and intelligent monitoring.

Based on this principle, our work presents a novel bimodal diagnostic system that combines electrical quantification using a DDX detector with optical characterization using a high-quality camera (HQC). The main contribution lies in the synergistic fusion of these two electrical and optical sensors and the comprehensive automation of data interpretation using a GPU.

One of the key contributions of this study is the automation of both data sources. For the electrical sensor, a YOLOv8 model is developed that interprets and digitizes the DDX screen, transforming a conventional instrument into an intelligent data source. In parallel, for the optical sensor, a novel semi-automated methodology is presented for the generation of a high-quality dataset, which is a fundamental step for training. In both cases, intelligent inference of electrical and optical data is produced.

PD detection has advanced significantly through multi-sensor fusion strategies, which typically combine data from established sensors such as UHF, acoustic, and ground transient voltages. While these approaches are effective for detection and localization, they often lack direct correlation with quantitative electrical measurements displayed on standard test equipment, as well as a high-resolution visual context that captures the physical manifestation and morphology of the discharge.

Our work breaks away from this paradigm to address these limitations. The fundamental contribution of this study is twofold. First, we introduce a novel concept of smart sensing, through which we transform a conventional visual electrical detector—a passive display device—into an active and intelligent data source. Through advanced computer vision techniques, such as OCR and object detection, we automate the extraction of quantitative electrical values directly from the device’s screen. Second, we merge this now digitized electrical information with high-resolution optical images of the discharge phenomenon itself. This fusion creates a unique dataset that enables a morphological and spatial characterization that other sensors cannot offer, directly correlating the electrical magnitude of the event with its physical manifestation (shape, color, intensity, and precise location). Therefore, while existing studies focus on data fusion from sensors operating in distinct physical domains (acoustic, electromagnetic, etc.), our contribution lies in the creation of a new smart sensor from a visual information source and its subsequent fusion with another optical modality, offering an unprecedented level of diagnostic detail.

Unlike traditional PD diagnostic models, which typically analyze pre-processed numerical data or pattern images, the approach in this work presents two fundamental improvements in the data acquisition and analysis paradigm itself:

• Data source automation: the first and fundamental improvement lies in the automation of the data source itself. Conventional approaches typically start with already digitized data. Our system, on the other hand, uses computer vision not only to analyze but also to generate structured data in the first place. By directly reading the display of a conventional, non-smart instrument, it transforms it into a self-contained, structured, digital data source. This eliminates the dependence on manual operator interpretation or proprietary data interfaces, offering an innovative solution for the modernization and automation of existing equipment.
• Novel bimodal fusion for contextualized diagnosis: the second key improvement is the novel bimodal fusion that provides diagnostic context unattainable with unimodal models. Our approach fuses and combines this quantitative electrical information—the “how much”—automatically extracted from the detector with the qualitative and morphological characterization of direct optical images—the “where” and the “how.” Therefore, the improvement lies not only in the classification algorithm itself, but in the contextual richness of the fused data provided to it. This synergy enables a much more comprehensive diagnosis, correlating the electrical severity of an event with its precise physical manifestation in space.

The rest of this paper is structured as follows. Section 2 describes the experimental set up, Section 3 and Section 4 present the automation of the electrical detector via computer vision analyses the optical characterization of PDs via computer vision. Finally, in Section 5 the main conclusions are presented.

2. Experimental Set Up

In the present work, the PDs were measured in a dielectric oil located inside a methacrylate cell [10] containing two facing electrodes subjected to high electrical voltages (Figure 1). The experimental installation complies with the IEC60270 standard [3].

All tests lasted 45 s and were carried out at an ambient temperature of 21 °C. The PDs were monitored using a conventional DDX-9101 PD detector, Basel Switzerland, (Figure 2), and the high voltage was regulated using an OT 248 terminal, Basel Switzerland.

The DDX-9101 screen was recorded by a digital camera to store the results of each test. This camera records video in 1920 × 1080p at 30 FPS.

Preliminary tests showed that PDs were practically non-existent below 6 kV and that electric arc breakdowns occurred from 18 kV onwards.

Taking this information into account, the PD were analyzed in a first series ranging from 6 to 18 kV. The voltage was increased by 1 kV in each test, thus performing 13 independent tests. The total duration of each experimental test was 45 s. The voltage applied to the electrodes started from 0 kV and followed an ascending ramp lasting approximately 10 s until reaching the desired nominal charge. After 45 s of testing, the electrical voltage was reduced to zero. A total of four sets of tests were performed from 6 to 18 kV. Therefore, the number of tests performed in this first series is 52.

These tests showed that PDs increase significantly above 10 kV. Therefore, it was decided to conduct a second series for voltages of 10, 13, and 16 kV. Four sets were performed following the same methodology described above. Therefore, the number of tests performed in this second series amounts to 12. These additional tests were performed where the PDs are greatest, near the arc break. The total number of tests was 52 + 12 = 64.

To ensure the comparability of results across different voltage levels within a single experimental run, the same sample of dielectric oil was used. The oil was subsequently replaced before initiating a new series of tests to prevent potential degradation effects from influencing the measurements.

An HQC was used to record the PDs produced between the electrodes (Figure 3a). Detailed information about the HQC can be found in [36]. It is an affordable camera of exceptional quality with a resolution of 12.3 megapixels and a 7.9 mm diagonal sensor. This camera works especially well in low-light conditions.

In addition, two polarizers were introduced into the experimental device, rotated at a certain angle so that the light that reaches the HQC, coming from a lamp, produces the greatest possible contrast to view the PD [10] (see Figure 3b).

The proposed method is performed in a controlled laboratory environment and has proven to be an effective tool for the detection and characterization of PDs.

However, it is essential to understand the limitations inherent in their in situ application. Below, we mention some aspects to consider in these cases that are beyond the scope of this study and that may guide future lines of research.

• Sensitivity to lighting conditions and optical environment: The performance of both YOLOv8 object detection and OCR character recognition is intrinsically dependent on the quality of the captured images. The methodology is sensitive to variations in ambient lighting conditions. Factors such as reflections, shadows, or uneven lighting can introduce noise and affect the accuracy of the algorithms. A controlled environment was maintained in our laboratory, but in situ implementation would require more robust solutions, such as the use of controlled and polarized light sources or more advanced image preprocessing algorithms to normalize the captures.
• Transparency of the dielectric medium: The effectiveness of optical discharge detection relies on the assumption that the dielectric medium, in this case the oil, is optically transparent. In real-world applications, insulating oils degrade over time due to thermal and electrical stress, which can increase their turbidity, change their color, or generate suspended byproducts. This degradation would cause attenuation of the optical signal through scattering or absorption, making it difficult or even impossible to capture the discharge morphology, especially for low-intensity events.
• Direct optical access requirement: A fundamental requirement of this technique is the existence of a direct line of sight to the area where the discharges occur. Our experimental setup used a vessel with transparent walls, simulating an inspection window. However, most high-voltage electrical equipment in service is sealed metal vessels. Widespread application of this method would require the availability of equipment with inspection windows or the possibility of making significant structural modifications to install them, which is not always feasible, safe, or economically viable.
• Scalability to field equipment: The transition from a laboratory environment to on-site diagnostics on large equipment, such as power transformers, presents considerable challenges. The large internal volume of this equipment makes it complex to determine the optimal location of one or more cameras to cover all potential risk areas. Furthermore, integrating a vision system into the existing monitoring infrastructure and ensuring its durability in the harsh environmental conditions of an electrical substation are engineering hurdles that must be addressed for practical, large-scale implementation.

Despite these limitations, the proposed method sets a solid precedent for the “smart sensing” of visual sources and their fusion with optical data, opening new avenues for the in situ diagnosis of high-voltage phenomena with an unprecedented level of detail.

3. Method 1: Automating the Electrical Detector via Computer Vision

This section presents the CNN training and inference analysis in operational scenarios using images obtained from the DDX electrical detector. The dataset was manually generated for this purpose.

3.1. CNN Training

This section presents and analyzes the results obtained from the training and validation of the YOLOv8 model for PD detection and classification, as well as its quantification. Training was performed over 150 epochs using a partitioned dataset as described below.

3.1.1. Training Environment

Training and evaluation of the YOLOv8 model was carried out on a high-performance workstation running Ubuntu 22.04.5 LTS (codename: jammy). The system is equipped with a 12-core AMD Ryzen Threadripper 1920X processor (24 threads, 3.5 GHz base frequency) and 125 GB of RAM.

For the processing of deep learning tasks, an NVIDIA GeForce RTX 2070 SUPER GPU, Santa Clara, CA, USA, with 8192 MB (approximately 7.9 GB) of dedicated video memory (identified as CUDA:0) was used. This GPU operates with the NVIDIA driver version 570.133.07 and support for CUDA 12.8. The software environment was configured with Python 3.10.16, PyTorch 2.5.1 and the Ultralytics YOLOv8 framework version 8.3.111 [32]. The specific model employed presents an architecture of 92 fused layers, adding a total of 25,842,655 parameters and requiring approximately 78.7 GFLOPs for its execution in inference.

3.1.2. Manual Labeling of DDX Images

Figure 4 shows six images taken randomly during the tests at 6, 8, 10, 12, 14, and 16 kV. It can be seen how the number of pulses progressively increases with increasing voltage applied to the electrodes. A voltage limit of 16 kV was not exceeded to avoid the electric arc breakdowns that occurred in the preliminary tests from 18 kV onwards.

As mentioned above, the camera records the DDX screen at 30 FPS. Each video therefore contains 45 s × 30 FPS = 1350 frames, giving us 64 videos.

Using Image-J version 1.54g [37], the region of interest (ROI) of each video was delimited, which is necessary for the efficient analysis of the images obtained from the DDX screen. In this way, only the rectangular area of each video containing the fields that will later be labeled and analyzed was selected. The video format used for the camera was converted from MP4 to AVI, which can be imported into Image-J. Once imported, the AVI file was converted to a set of images in PNG format.

For manual image labeling, 10 images were randomly selected from each test, resulting in the labeling of a total of 640 images. The online software Roboflow [38] was used for this purpose.

The following classes were labeled in each image:

• attenuation_value
• negative_pulse
• pd_level_value
• positive_pulse
• voltaje_value

Figure 5 shows an example of an image labeled using the Roboflow program.

It is important to clarify that the approach in this work does not rely on traditional signal processing to ensure a high signal-to-noise ratio or to remove noise from the underlying electrical signal. Instead, the problem is approached as a computer vision task, where the objective is to train a YOLOv8 model to recognize the visual patterns of pulses as presented on the DDX detector screen. From this perspective, the baseline visual noise, including the sine wave and low-level fluctuations, is not filtered out but constitutes the image background.

During the training process, the CNN learns to identify the distinctive visual characteristics of the pulses, the signal, and to differentiate them from the background, the visual noise, implicitly learning to ignore it. The effectiveness of this approach is validated by the high-performance metrics obtained, as detailed later. The high accuracy scores, above 0.91, and the confusion matrix results quantitatively demonstrate that the model was able to successfully differentiate the pulses of interest from the background, validating the robustness of this image recognition-based method for the proposed task.

3.1.3. Dataset Setup and Training

An additional 200 images that had been used in pre-training the CNN were added to the initial dataset of 640 images. Hence, the final dataset comprised 840 images, managed and labeled using the Roboflow platform. For the training and evaluation process, the dataset was divided into three subsets:

• Training: 594 images (71%).
• Validation: 146 images (17%).
• Test: 100 images (12%).

The YOLOv8 model was trained for 150 epochs. Learning curves and performance metrics were monitored for both the training and validation sets.

3.1.4. Analysis of Loss Curves

Loss functions provide crucial information about how the model learns to minimize errors during training. Three main loss components were analyzed: Box Loss, Classification Loss, and Distribution Focal Loss (DFL). These three loss curves are analyzed below. In all three cases, small errors were obtained during training.

Box Loss

Figure 6 shows the evolution of Box Loss for the training and validation sets. Box Loss measures the accuracy with which the model predicts the coordinates of the object’s bounding box. A constant decrease in Training Box Loss is observed throughout the epochs, indicating that the model is learning to localize objects progressively better.

Validation Box Loss also shows a decreasing trend, although with more pronounced initial fluctuations and stabilizing at a value slightly higher than the Training Box Loss towards the end of the epochs. This behavior is typical and suggests that the model generalizes adequately, although there may be slight overfitting. Box Loss in YOLOv8 uses a CIoU (Complete Intersection over Union) metric [39] following Equation (1):

$$L_{CIoU}=1-IoU+{\displaystyle \frac{{\rho }^2\left(b,b_{gt}\right)}{c^2}}+\alpha v$$

(1)

where

$L_{CIoU}$: Value of the CIoU loss. The goal of YOLOv8 is to minimize it.

$IoU$: Intersection over Union, Equation (2). It measures the overlap between the predicted and actual boxes. Its value ranges from 0 with no overlap to 1 with perfect overlap. It is calculated as:

$$IoU={\displaystyle \frac{Area\left(b_p\cap b_{gt}\right)}{Area\left(b_p\cup b_{gt}\right)}}$$

(2)

$b$: Bounding box predicted by the model (coordinates xcenter, ycenter, width, height).

$b_{gt}$: Real bounding box (ground truth) (coordinates xcenter_gt, ycenter_gt, width, height gt).

${\rho }^2\left(b,b_{gt}\right)$: Squared Euclidean distance between the center points of the predicted box b and the actual box b_gt. $\rho$ represents the distance.

c: Length of the diagonal of the smallest bounding box that completely encloses both b and b_gt. Normalizes the distance penalty.

$\alpha$: Positive weighting parameter that adjusts the importance of the aspect ratio consistency term.

$v$: Measure of the consistency of the aspect ratio between the predicted and real box. It is calculated through Equation (3):

$$v={\left({\displaystyle \frac{4}{{\pi }^2}}arctan\left({\displaystyle \frac{w_{gt}}{h_{gt}}}\right)- arctan\left({\displaystyle \frac{w_p}{h_p}}\right)\right)}^2$$

(3)

where $w_{gt}$ and $h_{gt}$ are the width and height of the actual box, and $w_p$ and $h_p$ are the width and height of the predicted box.

Classification Loss

Figure 7 illustrates the Classification Loss. This loss quantifies the model’s error in assigning the correct class to the detected objects. Both the Training and the Validation Box Loss consistently decrease. The validation curve closely follows the training curve, also stabilizing and suggesting good generalization capability for the classification task. YOLOv8 employs a loss function such as binary cross-entropy for this task, Equation (4), [40]:

$$L_{cls}=-\sum _i\left[y_{i}log\left({\widehat{y}}_i\right)+\left(1-y_i\right)log\left(1-{\widehat{y}}_i\right)\right]$$

(4)

where the summation $\sum _i[ ]$ is performed on the classes used: attenuation_value, negative_value, pd_level_value, positive_pulse and voltage_value, and where

$L_{cls}$: Classification Loss value.

$y_i$: Ground truth label for class i. $y_i$ = 1 if the object belongs to class i, $y_i$ = 0 if it does not.

${\widehat{y}}_i$: Probability predicted by the model that the object belongs to class i. It is the result of a sigmoid function with value in [0, 1].

$log( )$: Natural logarithm.

Distribution Focal Loss

The DFL [41], shown in Figure 8, is a component that helps refine the prediction of bounding box coordinates by modeling the location of the box edges as a probability distribution. The training and validation DFL curves also show a decreasing trend and good correlation, indicating that the model is effectively learning this more detailed representation of the location.

The DFL is expressed in Equation (5):

$$L_{DFL}\left(P_{y_l},P_{y_r}\right)=-\left(\left(y_r-y\right)log\left(P_{y_l}\right)+\left(y-y_l\right)log\left(P_{y_r}\right)\right)$$

(5)

where

$L_{DFL}$: DFL value.

$y$: Continue Ground truth coordinate of a box edge.

$y_l$: Label of the discrete container immediately to the left of $y$.

$y_r$: Label of the discrete container immediately to the right of $y$.

$P_{y_l}$: Probability predicted by the model for container $y_l$.

$P_{y_r}$: Probability predicted by the model for container $y_r$.

The terms $\left(y_r-y\right)$ and $\left(y-y_l\right)$ act as weights.

3.1.5. Performance Metrics

The model’s performance was then evaluated using standard object detection metrics [42]. Specifically, the Precision, Recall, and Mean Average Precision metrics (mAP) were used. In all three cases, the growth was constant, and the stabilization of the metrics at high values indicates good detection and classification performance.

Precision and Recall in Training

Figure 9 and Figure 10 show the evolution of Precision, Equation (6), and Recall, Equation (7), respectively, the training set. Both metrics tend to increase as training progresses, stabilizing at high values of 0.91 and 0.92 for Precision and Recall, respectively, indicating that the model learns to correctly identify relevant objects while minimizing false positives and false negatives in the analyzed data.

$$Precisión={\displaystyle \frac{TP}{TP+FP}}$$

(6)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(7)

where TP are true positives, FP are false positives and FN are false negatives.

mAP in the Validation

The mAP is a key metric for evaluating the overall performance of object detectors. Figure 11 shows the mAP at an Intersection over Union (IoU) threshold of 0.5 (mAP@0.5), while Figure 12 presents the mAP averaged over multiple IoU thresholds from 0.5 to 0.95 in steps of 0.05 (mAP@0.5:0.95). Both mAP curves on the validation set show a steady increase, reaching values of 0.94 and 0.62 for (mAP@0.5) and (mAP@0.5:0.95), respectively. The more strict mAP@0.5:0.95 provides a more robust assessment of the model’s localization performance. The steady growth and stabilization at high values indicate good detection and classification performance with validation data not seen in the training set.

The Average Precision (AP) for a class, Equation (8), is calculated as the area under the Precision-Recall curve. One way to calculate it is:

$$AP=\sum _{k=1}^{N}P\left(k\right)∆r\left(k\right)$$

(8)

where

$AP$: AP for a specific class.

$k$: Index of predictions ordered by confidence (from highest to lowest).

$N$: Total number of thresholds or data points considered.

$P\left(k\right)$: Accuracy calculated at the kth Recall point or by considering the k highest confidence detections.

$∆r\left(k\right)$: Change in Recall from point (k − 1) to point k (i.e., $∆r\left(k\right)=r\left(k\right)-r\left(k-1\right))$.

3.2. Confusion Matrix

This section presents a detailed overview of the successes and errors of the classes that were considered through the study of the Confusion Matrix in the validation and in the test set.

3.2.1. Confusion Matrix on the Validation Set

The Confusion matrix [43], presented in Figure 13, provides a detailed view of the model’s classification successes for each of the five classes in the validation set. Values on the main diagonal represent correct classifications. A high number of correct predictions is observed for most classes: negative_pulse (650), positive_pulse (615), attenuation_value (145), pd_level_value (145), and voltage_value (144).

The attenuation_value, pd_level_value and voltage_value classes show very little confusion, indicating good distinction by the model. However, some confusions are identified in the background class, which is incorrectly classified as negative_pulse in 197 instances and as positive_pulse in 231 instances. Furthermore, instances of negative_pulse (98) and positive_pulse (87) are wrongly identified as background. These confusions could be due to the visual similarity of these signals with background noise or to the inherent variability of background signals, which can resemble weak pulses.

3.2.2. Confusion Matrix on the Test Set

After training the YOLOv8 model for 150 epochs, its detection and classification performance was evaluated on the test set, which consisted of 100 images not used during the training and validation phases. This set contains a total of 1025 labeled object instances belonging to the five defined classes. The evaluated model, with 92 fused layers, 25,842,655 parameters, and a complexity of 78.7 GFLOPs, was subjected to inference on this dataset.

The model’s processing speed on the test set was remarkable, with an average time of 1.7 ms for preprocessing, 11.3 ms for inference itself, and 5.3 ms for postprocessing per image. This results in an efficient overall inference time, which is crucial for applications requiring real-time responses or the processing of large volumes of data.

The overall evaluation results on the test set show robust model performance. An average Precision of 0.93 and an average Recall of 0.93 were obtained. Regarding the mAP, a value of 0.95 was achieved with an IoU threshold of 0.5 (mAP@0.5). When considering a stricter range of IoU thresholds, from 0.5 to 0.95 in steps of 0.05 (mAP@0.5:0.95), the model achieved a value of 0.62. These values suggest a good ability of the model to correctly locate and classify events in the signals, with mAP@0.5:0.95 being a stricter indicator of accuracy in locating bounding boxes.

Analyzing the performance broken down by class using the mAP@0.5:0.95 metric reveals the following values: attenuation_value 0.67, negative_pulse 0.45, pd_level_value 0.77, positive_pulse 0.48, and voltage_value 0.71.

Excellent performance is observed for the pd_level_value and voltage_value classes, followed by attenuation_value. The negative_pulse and positive_pulse classes exhibit lower mAP@0.5:0.95, indicating greater difficulty for the model in accurately localizing bounding boxes for these pulse types under strict IoU criteria, although its performance at mAP@0.5 remains high.

Figure 14 shows the confusion matrix obtained from the model’s predictions on the test set. The following correct predictions are observed on the main diagonal: attenuation_value 99, negative_pulse 327, pd_level_value 99, positive_pulse 316, and voltage_value 99.

The attenuation_value, pd_level_value, and voltage_value classes show very little or no confusion with other classes or the background, indicating excellent distinction by the model for these specific events.

The model demonstrates a high ability to correctly classify most instances. However, some significant confusions are identified, primarily related to the background class. Specifically, the true background class is incorrectly classified as negative_pulse 76 times and as positive_pulse 86 times.

On the other hand, true instances of negative_pulse 47 and positive_pulse 37 are incorrectly classified as background. That is, in these cases, the model either fails to detect them or mistakes them for the background. There is also much less confusion between negative_pulse and positive_pulse, with just one instance of negative_pulse predicted as positive_pulse. Background confusion for positive and negative pulses could be attributed to the visual similarity of low amplitude pulses to background noise or to inherent variability in the signal that makes distinction difficult.

A fundamental aspect in validating the robustness of a classification model is the analysis of the class balance in the dataset, since a severe imbalance could bias the training and evaluation metrics. To address this point, the distribution of the main classes in our test set has been examined. As evident from the confusion matrix (Figure 14), the total number of instances for the negative_pulse class is 374, while for the positive_pulse class it is 353. This distribution, with a ratio close to 1:1, confirms that there is no significant class imbalance. Therefore, it can be concluded that the high performance of the model in identifying both pulse polarities is genuine and not an artifact derived from an over-representation of one of the classes, which confers greater reliability to the presented results.

3.3. Inference in Operational Scenarios

This section analyzes the inference of images from the DDX electrical detector videos using the trained CNN. This automatically produces the five classes as the final result, and for three of them the numerical value is obtained.

3.3.1. Inference Flowchart

Figure 15 shows the flowchart that explains how inference is performed from video sequences from the DDX electrical detector for object detection and quantitative data extraction. The program was made in Python, and its architecture designed to be efficient and clear. It is divided into three phases: setup and initialization, frame-by-frame processing, and finalization.

Phase 1: Set up and initialization

This is a preliminary phase that prepares all the components necessary for the analysis. It performs three tasks sequentially:

Startup and configuration: the process begins by loading user-defined configurations, such as the input video and YOLOv8 model paths, confidence thresholds, and a list of interest classes that will trigger optical character recognition (OCR).

Engine loading: the two main inference engines, the YOLOv8 object detection model and the Python EasyOCR OCR engine, are initialized and loaded into memory. This loading is performed only once at startup to optimize system performance. The number of GPUs to be used is also determined.

Opening files: the input video stream is opened and the output files are created, including the new video with the visual annotations and the text file that will record its detailed data.

Phase 2: Frame by frame processing, main loop

This is the operating core of the system, where each frame of the video is analyzed sequentially.

YOLOv8 inference: the current frame is fed into the YOLOv8 model, which identifies and locates all classes of interest that exceed the confidence threshold, returning their bounding boxes, class labels, and confidence scores.

Detection loop: the system iterates through each of the detections found in the frame.

OCR class certification: for each detection, a decision is made based on its class label. If the class is predefined as an OCR target—pd_level_value, voltage_value, and attenuation_value—the system proceeds with OCR inference.

OCR inference: this critical step extracts the quantitative data:

ROI cropping: the exact portion of the image contained within the detection bounding box is extracted from the frame.

OCR application: the OCR engine analyzes this small ROI to recognize the textual information present.

Value interpretation: the extracted text is processed to convert it into a numerical value.

Output log: all detection data is logged. Bounding boxes and corresponding labels—confidence and OCR value, if applicable—are drawn on the output video frame. Detailed information about each detection, including the numerical value analyzed by the OCR, is added as a new line to the text file.

Phase 3: End

Once all frames have been processed, the system performs an orderly shutdown, leaving the output files ready for further analysis.

3.3.2. Results and Discussion

To test the performance and generalization capabilities of the YOLOv8 model trained and verified in the previous sections, an inference evaluation was performed on completely new data. To do this, three videos of 10 s were used, captured at voltage levels of 10 kV, 13 kV, and 16 kV, respectively. Each video, corresponding to approximately 300 images, was processed by the trained model to evaluate its effectiveness in detecting and classifying events under operating conditions not seen during training.

Figure 16a–c present representative frames of the inference for each voltage level. It is observed that the model not only successfully identifies the discharge pulsesnegative_pulse and positive_pulse, but also correctly reads and classifies the instrument numerical values—pd_level_value, voltage_value—and the attenuation value—attenuation_value. The high confidence scores, generally >0.70, for all classes demonstrate the robustness of the model in a complex task combining signal pattern detection with implicit optical character recognition.

For a deeper analysis of the relationship between PD activity and electrical magnitudes, cumulative detection images were generated for each 10 s video, as illustrated in Figure 17a–c. These images overlay all the bounding boxes of the detected pulses over the first frame of the video, providing a comprehensive view of the PD activity signature.

Analysis of these visualizations reveals a direct and physically consistent correlation between the applied voltage, the measured discharge level, and the activity detected by the model.

• At 10 kV, the model detects moderate discharge activity, with well-defined but relatively compact green clusters of negative and magenta positive pulses. This corresponds to an instrumental reading of PD Level 0.426 nC and Voltage 10.1 kV in Figure 17a.
• At 13 kV, with increasing voltage, a significant increase in the density and spatial extent of detections is observed. Both the negative and positive cumulative pulses are visibly larger and denser. This increased visual activity directly correlates with the increased discharge level measured by the instrument, which now shows PD Level 0.701 nC and Voltage 13.1 kV in Figure 17b.
• At 16 kV, the phenomenon intensifies dramatically. The cumulative image shows a much larger and more saturated area of activity, indicating a very severe PD regime. This exponential increase in visual activity is consistent with the instrumental reading, which reaches a PD Level of 3.88 nC and a Voltage of 16.6 kV, as shown in Figure 17c.

The data that our system automatically extracts (discharge magnitude and its time of occurrence) are precisely the ingredients required to construct PRPD patterns. As can be inferred from the data presented in Figure 17, these patterns could be generated and, in a subsequent research phase, be analyzed by another neural network (which could be another model from the YOLOv8 family or a different architecture) to perform a detailed classification of the PD type.

Table 2. Summary of the most relevant Pearson correlation coefficients (r).

Attribute 1	Attribute 2	Coefficient (r)
Strong positive correlations (r > 0.7)
ocr_voltage	ocr_pd_level	0.90
num_pulse_negatives	ocr_voltage	0.77
Area	Height	0.90
Area	Width	0.78
CenterX	CenterY	0.77
Significant negative correlations (r < −0.3)
ocr_voltage	Width	−0.41
ocr_pd_level	Width	−0.39
num_pulsos_negativos	Width	−0.34
Other moderate positive correlations (0.5 < r < 0.7)
num_pulse_negatives	ocr_pd_level	0.59
Confidence	Width	0.55
Area	Confidence	0.53

presents a summary of the relationships between the main attributes obtained from the inference of the trained CNN for accumulated experiments at 10 kV, 13 kV and 16 kV. It presents the most significant Pearson correlation coefficients (r) [44], with a focus on the main electrical variables, ocr_voltage and ocr_pd_level, and their relationships with other geometric characteristics such as the pulse area, pulse coordinates CenterX and CenterY, as well as the number of positive and negative pulses detected. It shows a correlation of 0.90 between the magnitudes obtained in the ocr_voltage and ocr_pd_level classes, which confirms that these variables measure strongly related aspects of the same physical phenomenon.

We also observe a strong positive relationship of 0.77 between the increase in the magnitude of ocr_voltage and the number of pulse_negatives detected, suggesting that higher voltages generate more negative discharges. On the other hand, an inverse relationship is observed between voltage and the geometric ratio. The most notable negative correlation 0.41 is between ocr_voltage and Width. This indicates that pulses tend to become narrower as voltage increases. This is a non-obvious but highly informative pattern that the machine learning model is using for classification.

A particularly interesting finding that emerges from the correlation analysis in Table 2 is the moderate negative correlation observed (−0.41) between the applied voltage, ocr_voltage, and the detected pulse width, Width. This result, although at first glance might seem counterintuitive, may have a plausible physical explanation linked to the dynamics of PDs in dielectric oils.

From a physical perspective, we hypothesize that this behavior is related to the energy and speed of the discharge process. As voltage increases, the energy injected into the dielectric medium increases. This could lead to more accelerated ionization processes and the formation of more energetic, but at the same time more ephemeral and spatially concentrated, discharge channels. A shorter duration discharge event, i.e., faster, would be directly translated on the measuring equipment screen as a visual pulse with a smaller temporal width. Therefore, the system would be capturing a morphological manifestation of the greater intensity and brevity of discharges at higher voltages.

While a comprehensive characterization of the underlying plasma dynamics to validate this hypothesis is beyond the scope of this work, this finding is significant in itself. It demonstrates the ability of our computer vision system not only to quantify parameters in isolation but also to uncover subtle correlations that link the morphology of the visual signal to the physical principles of the PD phenomenon. This is a promising result that paves the way for future analyses that can corroborate these relationships with more detailed physical models.

In conclusion, this experimental validation on data not used in the training set demonstrates the effectiveness of the trained model. Not only is it capable of generalizing and operating as a robust monitoring system, but its visual detections act as a qualitative and quantitative analogue of electrical measurements. The density, area, and frequency of bounding boxes detected by the model provide a direct visual measure of the severity of the phenomenon, validating this approach as a powerful and reliable tool for the automated diagnosis and quantification of PDs.

4. Method 2: Optical Characterization of PDs via Computer Vision

This section presents the CNN training and inference analysis in operational scenarios using images obtained using the HQ camera. The training environment in this section is the same as that used in Section 3. However, in this section a semi-automatic generation of the dataset is realized, which greatly facilitates the labeling of the training, validation, and test sets.

4.1. CNN Training

The purpose of this section is to train a CNN based on YOLOv8 architecture. The section is divided into two subsections: semi-automatic dataset generation and training results.

4.1.1. Semi-Automatic Generation of the Dataset

To train the CNN based on the YOLOv8 architecture for PD detection in videos obtained with the HQC, a Python script was developed for video processing to obtain the training, validation, and test images. This process automates the identification of candidate events, filters out known false positives, and generates a structured and labeled dataset in the format required by YOLOv8. The methodology is based on background subtraction, contour analysis, and a novel manual spatial exclusion filter that significantly improves the quality of the final dataset by reducing noise and the need for subsequent manual cleaning.

The structure of this section is as follows: first, the semi-automatic data acquisition model is configured and initialized. Next, a Python program is created for PD detection and extraction. Data filtering, validation, and collection are then performed. Finally, the dataset is generated in YOLOv8 format for CNN training. To facilitate understanding of this process, a flowchart summarizing the overall method described is included.

Configuration and Initialization

The process begins with a configuration phase where key parameters are defined. The I/O paths are defined first, and then the input video and output file paths are specified. These include debugging videos—difference, threshold and detected events—and a .dat data file with the characteristics of each PD.

Manual exclusion zones are then established. This is a crucial component of the system as it allows the user to define a priori spatial regions in the image where recurring false positives—reflections, sensor noise, etc.—are known to occur. Each zone is defined by a centroid, an exclusion radius on the x and y axes, and, optionally, an expected area with its tolerance. Any detected event whose centroid falls within one of these zones is automatically discarded.

The parameters of YOLOv8 dataset are defined below. The dataset’s root directory, the class to be detected (PD), and the ratios for dividing the data into training sets, 66.7%, validation sets, 22.0%, and test sets, 11.3%, are named.

PD Detection and Extraction

The core of the Python script processes the video frame by frame to identify events of interest. This process is broken down into four steps:

Background establishment: the first frame of the video is assumed to represent the static background of the scene. This frame is converted to grayscale and stored for reference.

Background subtraction: for each subsequent frame, the absolute difference with the background frame is calculated. The result is an image that highlights only the regions where changes have occurred (i.e., new PD).

Thresholding and morphological cleaning: the resulting image is binarized using a fixed threshold to convert subtle changes into well-defined, white-on-black regions. A morphological operation is then applied to remove noise.

Contour detection: on this last image, the OpenCV Python library algorithm [25] is applied to determine the contours of all the change regions. Each contour represents a candidate PD.

Filtering, Validation and Data Collection

This process is carried out in the following steps:

Minimum area filter: contours with an area smaller than a predefined threshold of 5 pixels are discarded to remove residual noise.

Manual exclusion filter: the contour centroid is calculated. If this centroid falls within any of the manual exclusion zones defined in the configuration, the contour is classified as a false positive and discarded.

Data collection: if a contour passes both of the above filters, it is considered a valid PD.

For each valid PD, the following is extracted and stored:

• The bounding box.
• The centroid coordinates, area and average RGB color intensity in a .dat text file for further analysis.
• A copy of the original, unprocessed frame and the list of bounding boxes for all valid events found are saved. This pair (image, labels) is the input data in YOLOv8 format.

Generating the Dataset in YOLOv8 Format

Once the entire video has been processed, the script uses the collection of frames with valid PDs to build the final dataset, and the following steps are performed:

Directory structuring: a folder structure compatible with YOLOv8 framework is created with the subdirectories train, valid, and test, each containing folders for images and labels.

Data splitting: the data collection—images and their labels—is randomly shuffled and split into training, validation, and test sets according to the ratios defined above.

File generation for each image: the original image is saved as an image_name.jpg in the corresponding images folder.

An image_label.txt file is created in the corresponding labels folder. Within this file, each line represents an event detected in that image, in the format: [class_index, x_center_norm, y_center_norm, width_norm, height_norm]. All bounding box coordinates are normalized by dividing them by the frame width and height dimensions, as required by YOLOv8.

Configuration file (data.yaml): finally, a data.yaml file is generated at the root of the dataset. This file is essential for YOLOv8 to locate the datasets and identify the number of classes and their names.

The end result is a high-quality dataset, ready to be used directly in training a YOLOv8 object detection model, minimizing manual intervention and improving labeling consistency.

Summary Flowchart of the Process

To visualize the logical flow of the script used to generate the YOLOv8 compatible dataset, a flowchart was created as shown in Figure 18.

The three main phases of the flowchart are summarized below:

• Setup and loading: in this initial phase, all resources are prepared. The script reads the file paths, uploads the video, and manually defines exclusion zones, which are key to filtering out known false positives.
• Video processing loop: this is the core of the script. It operates frame by frame, performing two main tasks in sequence:

  (a)

  PD detection and filtering: this block encapsulates all the computer vision logic, subtracts the background (see Figure 19a) to find the changes that occur, binarizes the image, finds the PD boundaries and applies filters, both the minimum area filter and the manual exclusion zones filter.

  (b)

  Temporary storage: if a frame contains at least one PD that has passed all filters, the script saves the original image of that frame along with the coordinates of the bounding boxes (see Figure 19b,c of the valid PD).

• YOLOv8 dataset generation: once the entire video has been analyzed, this final phase takes all the valid data collected and organizes it into the folder structure and file formats required by YOLOv8. This includes splitting the data into training/validation/test sets, normalizing the coordinates, and creating the .yaml configuration file.

Figure 19a shows the base or background image used as a reference, while Figure 19b,c show images with three PDs and one PD, respectively, as well as their bounding boxes.

4.1.2. Training Results

Three image series were used to train the CNN. The PD series occurring within the dielectric oil corresponds to an average voltage of 10 kV, 13 kV, and 16 kV, respectively. The total dataset consists of 4457 images, managed and labeled using the semi-automatic system explained in the flowchart represented in Figure 18. For the training and evaluation process, the dataset was divided into three subsets:

• Training set: 2967 images.
• Validation set: 982 images.
• Test set: 508 images.

The training environment is the same as for the DDX image training seen in Section 3. The YOLOv8 model was trained in four iterations: the first and second with 100 epochs, the third with 200, and the fourth, to ensure convergence, with 523. Some images from this dataset with box labeling in YOLOv8 format can be seen in Figure 19b,c. The training speed is 18.9 s per epoch. This demonstrates a very fast experimentation cycle, allowing for efficient model iteration and tuning.

The implemented object detector is based on a deep CNN architecture optimized for inference. The model consists of 92 fused computational layers, a technique that improves speed by combining operations such as convolution and batch normalization. With a total of 25,840,339 parameters, the model has a high capacity to learn and represent the complex visual characteristics of the PD of interest. Its computational load is quantified at 78.7 GFLOPs, a key metric that indicates the required processing demand and positions the model as a robust solution, suitable for running on GPU-accelerated hardware.

The evolution of performance metrics during training provides crucial information about the model’s learning process. Figure 20a,b depicts the Training vs. Validation Box Loss and Classification Loss curves, respectively. Figure 21a,b illustrates the Training vs. Validation DFL and Training Precision curves over 520 epochs, consisting of 982 images.

Consistent behavior is observed across the three loss graphs: Box Loss, Classification Loss, and DFL. During the first 375 epochs, the model demonstrates an effective learning phase. The loss curves for both training (solid blue line) and validation (dashed orange line) slope downward simultaneously. This indicates that the model is generalizing correctly, improving its ability to locate Box Losses, correctly classify PD (Classification Loss), and refine the DFL on unseen data.

However, starting at epoch 375, a clear inflection point becomes evident, signaling the onset of overfitting. While the training set loss continues its downward trend, the three validation set loss metrics reverse their trajectory and begin to increase steadily. This phenomenon is a classic indicator that the model has begun to memorize the specific characteristics and noise of the training set, losing its ability to generalize to new data. Therefore, the model with the best performance is not the one obtained at the end of training, but the one whose weights correspond to the minimum point of the validation loss, around epoch 375.

To mitigate this effect and ensure the selection of the model with the best generalization capacity, a strategy that works as an implicit early stopping mechanism was implemented. Specifically, for the final evaluation and all subsequent inferences, the model weights corresponding to the last training epoch were not used, but rather those saved from the epoch that recorded the minimum validation loss. This practice ensures that the selected model is the one that demonstrated the best performance on data not seen during training.

While this method was effective for the scope of our study, it is worth mentioning that there are additional regularization techniques that could be explored in future work to further robust the model against overfitting. Incorporating dropout layers into the architecture or applying a more aggressive data augmentation pipeline—including transformations such as random crops, stronger color variations, or mixup—could allow for longer training periods without the risk of overfitting, potentially improving the model’s generalization.

Regarding Precision, the graph in Figure 21b shows its evolution on the training set, where it stabilizes at an average value close to 0.77. This behavior suggests that, even on the training data, the model does not achieve perfect accuracy. This can be attributed to the nature of the dataset, which likely contains a subset of PDs that are intrinsically difficult to detect, such as very small-area or low-contrast PDs. The model assigns a lower confidence score to these complex detections which, when averaged over the entire set, results in an accuracy metric that does not reach higher values. The constant fluctuation in the accuracy curve reflects the model’s continuous effort to adjust its predictions to this PD variability.

In conclusion, the analysis of the training curves confirms the attainment of a functional model but also underscores the critical importance of employing an early stopping strategy or selecting the model based on the minimum validation loss to avoid deploying an overfitted and underperforming model in real-world applications.

4.2. Inference in Operational Scenarios

Once the CNN was trained, the model’s performance was evaluated on the inference task using three independent test videos, corresponding to PDs generated under voltages of 10 kV, 13 kV, and 16 kV. The model’s efficiency is remarkable, with an inference time of just 1.2 ms per frame. This translates into a theoretical processing capacity of approximately 833 FPS, confirming its suitability for real-time applications or for analyzing large volumes of video.

Figure 22 shows examples of inference on individual frames for each voltage level. The model is observed to correctly identify PDs under all conditions. The variability in the assigned confidence scores is notable. While events at 10 kV and 13 kV are detected with high confidence (0.91), 16 kV receives a more dispersed range of scores (0.92, 0.84, and even 0.39 for a weaker PD). This behavior is consistent with the analysis of the training accuracy curve and demonstrates the model’s ability to quantify the certainty of its own detections.

A more in-depth analysis is obtained by accumulating all detections over each video, lasting 45 s in this case. Figure 23 provides a visualization of the accumulated PD density over the image. In addition, Figure 24, Figure 25 and Figure 26 provide a detailed analysis of their spatial distribution, area and detection confidence.

The following conclusions can be drawn:

(a)

Correlation between voltage and discharge activity: there is a clear relationship between the voltage applied to the electrodes and the number of detected PDs. At 10 kV, 1582 PDs were accumulated (Figure 23a). As the voltage is increased to 13 kV, the activity increases significantly, recording 2050 PDs (Figure 23b). However, at 16 kV, the total number of detected PD drops slightly to 1981 (Figure 23c). A reasonable hypothesis for this small decrease is that at higher energies the PDs are larger and may merge, being detected by the model as a single PD with a larger area instead of multiple smaller PDs.

(b)

Spatial expansion of activity: the scatter plots shown in Figure 24, Figure 25 and Figure 26 visually confirm that the area of discharge activity expands with increasing voltage. The cluster of points, initially highly concentrated in the dielectric space at 10 kV, expands both vertically and horizontally at 13 kV and, more pronouncedly, at 16 kV. This suggests that at higher voltage levels in the dielectric, PDs are not only more frequent but also occupy a larger volume.

(c)

Increasing the detection area and correlation with PD confidence: the most revealing analysis comes from the direct comparison between the area and confidence of the PD in Figure 24, Figure 25 and Figure 26:

Area distribution (Figure 24a, Figure 25a and Figure 26a): at 10 kV, the vast majority of PDs are small in area (blue and green dots). At 13 kV, a slight increase in the average area is observed. The change is important at 16 kV, where a significant presence of large-area PDs appears, represented by yellow and orange colors.

Confidence distribution (Figure 24b, Figure 25b and Figure 26b): complementarily, the analysis of detection confidence provides a new layer of information. A strong positive correlation is observed between the area of a PD and the confidence with which it is detected. Larger PDs with warm colors in Figure 24a, Figure 25a and Figure 26a consistently correspond to high-confidence detections, with warm colors close to 1.0 in Figure 24b, Figure 25b and Figure 26b. This is physically consistent. Larger and more energetic PDs are visually clearer and therefore more confidently identified by the model. Conversely, low-confidence points—cool colors in Figure 24b, Figure 25b and Figure 26b—tend to correspond to smaller PDs, which are harder to distinguish from background noise.

The results of this study should be interpreted considering the substantial improvements our model offers compared to traditional approaches. The innovation of this work lies not simply in the application of an advanced classification algorithm, but in a redesign of the PD diagnostic workflow.

First, we have demonstrated that it is possible to convert a conventional measuring instrument into an intelligent, autonomous sensor. Unlike models that require already digitized data, our system automates the extraction of information directly from a screen, eliminating the need for manual intervention or costly hardware upgrades. This intelligence applied to legacy equipment represents a practical and scalable contribution.

Second, the fusion of quantified electrical information with the spatial and morphological characterization of optical images offers an enriched diagnosis. While a traditional unimodal model can identify the severity of a discharge (the “quantum”), our bimodal approach adds crucial context by also answering the “where” and “how” questions. This synergy between quantitative and qualitative data is the system’s main strength, allowing for a much deeper and more complete understanding of the PD phenomenon.

5. Conclusions

This work presents an innovative bimodal approach for laboratory PD analysis through training of a CNN based on YOLOv8.

Firstly, a conventional DDX-type PD electrical detector is enhanced by endowing it with smart capabilities. A system is developed capable of automatically reading and interpreting data displayed on the electrical detector screen, such as discharge magnitude, pulse count, and applied voltage. In this way, we transform a passive conventional instrument into a smart and autonomous source of digitized and structured data. The mean precision in the training was 0.91.

Concurrently, an optical visualization system using a high-quality camera is employed to capture direct images of PDs occurring in the dielectric oil. In addition, the training dataset for the camera is generated semi-automatically using a Python program. These images provide complementary qualitative and quantitative information, enabling the classification of discharge types based on their visual characteristics. This offers a new and complementary dimension providing the spatial location and morphology of PDs. Image analysis makes it possible to identify exactly where the PDs originate and how they propagate between the electrodes, vital information for diagnosing the exact point of failure or insulation degradation. For electrical voltages of 10 kV, 13 kV and 16 kV, PDs were detected with confidence scores of up to 0.92.

This synergy offers a more complete, accurate, and automated diagnosis of PD behavior in dielectric oils, improving the understanding of degradation mechanisms and the operational reliability of electrical assets. In this way, both systems, operating in parallel, enhance each other. The DDX electrical detector quantifies the charge, providing a measure of the magnitude of the problem, while the optical detector finds the location of the source of the problem. The fusion of this bimodal information, the electrical magnitude and the spatiotemporal distribution, allows for a much more complete and robust diagnosis of the dielectric insulation oil condition than could be achieved with either system alone. This approach represents a significant advance toward smarter and more accurate monitoring systems, capable of not only detecting the presence of PDs but also identifying their root cause and predicting failures more effectively.