Intelligent Identification of Grading Electrode Scale in Converter Valve Water-Cooling Systems Using X-Ray Images

I. Introduction

The reliability of converter valves in high-voltage direct current transmission systems critically depends on the thermal management efficiency of their water-cooling systems. Fouling accumulation, particularly on the surfaces of grading electrodes within these systems, presents a significant operational risk. Scale deposits on these electrodes impede heat dissipation, potentially leading to localized overheating, accelerated material degradation, voltage imbalance, and catastrophic failure of critical valve components [1–2, 11]. Timely and accurate detection of such fouling is therefore paramount for ensuring system stability and preventing costly unplanned outages.

Traditionally, monitoring fouling in the closed internal structures of converter valves, especially specific components such as voltage equalizing electrodes, has been extremely challenging. Common methods include periodic intrusive inspections that require system shutdown, indirect performance monitoring (such as monitoring temperature differences or flow resistance), or periodic fluid chemistry analysis [13, 14]. These methods have inherent limitations: intrusive inspections cause severe downtime and only provide periodic snapshots; indirect monitoring has insufficient spatial resolution and is difficult to locate the location and severity of fouling (especially for specific electrode surfaces); fluid chemistry analysis may not be directly related to localized fouling on solid surfaces [6,13]. Technologies such as ultrasonic inspection are susceptible to complex geometries around electrodes and signal interference. The subjectivity and inefficiency of these methods highlight the urgent need for non-invasive, accurate, and automated internal surface visualization inspection technologies.

X-ray imaging technology provides a feasible solution for non-destructively visualizing the internal status of opaque components such as water-cooling channels and electrodes in converter valves [12]. This technology uses the differential attenuation of X-rays by different materials (water, metal, scale) to generate detailed cross-sectional images, revealing the existence, location and shape of the scale attached to the surface [6]. Compared with other non-destructive testing methods, X-ray imaging has better penetration capabilities and spatial resolution, and is suitable for industrial metal casings. The development of high-energy industrial X-ray sources [3,7] and high-speed and high-sensitivity digital detectors [8,17] makes it possible to obtain high-contrast images in dense components, supporting in situ or near real-time detection. The basic principle is to visualize the scale layer by identifying X-ray absorption changes caused by differences in material density and composition of base metals, cooling water and scale.

However, the scale layer is usually very thin, and its density does not differ significantly from that of the electrode. This means that the engineers cannot quickly determine whether there is scale buildup. Manual interpretation of complex X-ray images to identify subtle scaling features on electrode surfaces is time-consuming and labor-intensive, requires professional experience, and is prone to human error and inconsistency, which hinders the widespread use of this technology for routine monitoring. To this end, this article solves this bottleneck by fusing the non-invasive diagnostic capabilities of X-ray imaging with the analytical capabilities of artificial intelligence (AI). Specifically, deep learning algorithms [4] that are good at identifying complex patterns in images are used to automatically detect and characterize voltage-equalizing electrode fouling in X-ray scans. At its core, computational models are trained to learn the unique visual characteristics of fouling directly from pixel data, transforming raw X-ray images into actionable insights about electrode status, reducing reliance on expert manual analysis.

The goal of this paper is to develop and verify an intelligent detection method for surface scaling of the voltage equalizing electrode in the converter valve water-cooling system, using X-ray imaging as the core detection method. In this regard, there are currently very few relevant studies. To achieve the goals, the following research was conducted in this paper: Firstly, an X-ray image dataset of electrodes was constructed and labeled. Secondly, considering that there could be multiple electrodes in a component, a target detection model is employed to identify the positions of the electrodes, and it can identify scaling based on the electrodes' surface features. Thirdly, a regressor was introduced to calculate the electrode diameter. Fourth, analyze the degree of scaling based on the diameter. This paper also conducted relevant tests and analyses to verify the feasibility of the method.

II. Methodology

a. Data acquisition and preprocessing

The dataset used in this study comprises 1,222 annotated images, including 382 samples with scaling deposits and 840 samples without scaling. The acquisition of X-ray images for scaling analysis in converter valve water-cooling systems begins with the systematic collection of high-resolution imaging data. All images were obtained during routine maintenance shutdowns of operational converter stations, ensuring alignment with real-world industrial conditions. The X-ray imaging system employed a linear electron accelerator-based high-energy X-ray source operating at 6 MeV. Key imaging parameters included a focal spot size of 2.0 mm, tube current of 3.5 mA, and exposure time optimized between 80 and 120 ms to balance resolution and radiation safety requirements.

As shown in Figure 1, images are labeled with the electrode position and diameter. Electrode position marking utilized X-AnyLabeling to create rotated bounding boxes (θ = 0–45°) around metallic components, accounting for angular variations in installation geometries. Each electrode has been classified according to scaling or not. Diameter measurements were performed within the X-ray system's software.

Figure 1:

b. Electrode detection and measurement model

The proposed deep learning model for scaling thickness identification integrates a modified YOLOv8_OBB architecture, which extends conventional object detection frameworks to address rotational bounding boxes and incorporate diameter regression [10]. Rotation recognition is necessary because the electrode rods are very close together. If the angle of the bounding box is ignored, multiple boxes will overlap and affect the analysis result. Figure 2 shows the model of this paper. Traditional object detection models often rely on axis-aligned rectangular coordinates, limiting their precision in capturing irregular or rotated scaling patterns commonly observed in X-ray images of converter valve cooling systems. To overcome this, the baseline YOLOv8_OBB algorithm utilizes DOTA-format annotations, encoding each detection box through its vertex coordinates (x₁, y₁, x₂, y₂, x₃, y₃, x₄, y₄) and a rotation angle θ. This representation aligns with the geometric complexity of scaling deposits, particularly when imaging multi-layered or angled structures within cooling channels using high-energy X-ray beamlines.

Figure 2:

The model architecture comprises three core components: a backbone network for multi-scale feature extraction, a neck module for feature fusion, and a detection head for simultaneous prediction of bounding boxes, rotation angles, and scaling thickness. The P1–P6 in Figure 2 is the backbone, which leverages a cross-stage partial Darknet design, optimized through computational efficiency analysis in industrial imaging applications. Feature pyramid networks in the neck enhance multi-scale representation by aggregating hierarchical features, critical for detecting both micron-scale scaling nucleation sites and macroscopic deposit layers [15,16]. There are three detectors to identify electrode in different feature layers.

A key innovation lies in augmenting the detection head to output thickness values alongside rotational parameters. While the original YOLOv8_OBB loss function combines classification loss L_cls, rotated box regression loss L_rot, and objectness loss L_obj, the modified framework introduces a diameter regression term L_dia. This term employs smooth L1 loss to minimize discrepancies between predicted and ground-truth diameter measurements derived from X-ray attenuation profiles. The composite loss function becomes:

(1)

$$L_{\mathit{total}}=L_{\mathit{cls}}+L_{\mathit{rot}}+L_{\mathit{obj}}+L_{\mathit{dia}}$$

Activation functions and optimization strategies are selected through ablation studies comparing ReLU, Leaky ReLU, and SiLU variants. The final implementation adopts SiLU (Sigmoid Linear Unit) activations in convolutional layers, demonstrating superior gradient propagation characteristics for imbalanced fouling thickness distributions. The model employs AdamW optimization with cosine annealing learning rate scheduling, initializing at 3 × 10⁻⁴ and decaying to 1 × 10⁻⁶ over 100 epochs. Weight decay regularization (5 × 10⁻⁵) and gradient clipping (max norm 10.0) prevent overfitting while maintaining stable convergence.

Input preprocessing pipelines incorporate histogram equalization and adaptive noise reduction, aligning with established practices for enhancing X-ray image contrast in industrial inspection scenarios. To accommodate varying imaging resolutions across different converter valve systems, all inputs are rescaled to 1,024 × 1,024 pixels with bilinear interpolation, preserving critical microstructural details above 50 μm resolution—a threshold determined through sensitivity analysis of scaling detection requirements.

The hyperparameter selection rationale draws from both computational imaging principles and empirical validation. The rotation angle prediction head utilizes circular smooth L1 loss with a periodicity-aware formulation to handle angular discontinuities at 0° and 180°, mitigating orientation estimation errors common in symmetric scaling patterns [5]. Thickness regression outputs are normalized to [0,1] using min-max scaling based on maximum observed deposit thicknesses in training data, with physical unit conversions applied during postprocessing.

This architecture addresses several limitations in existing approaches: (1) Simultaneous optimization of detection and thickness estimation avoids error propagation in cascaded models. (2) Rotation-aware detection improves measurement consistency across arbitrary imaging angles. (3) Integrated thickness regression enables direct correlation between deposit geometry and cooling performance degradation, a critical factor in predictive maintenance. The model's design principles align with emerging trends in multi-parametric industrial image analysis, particularly for systems requiring combined geometric and material property characterization [9]. Mixed-precision training and layer pruning are implemented considering potential deployment on edge devices.

c. Assessment of scaling severity

At present, there are no standards related to the assessment of electrode fouling. Paper [10] suggests classifying severity based on fouling thickness T: Level I (T < 0.1 mm), Level II (0.1 mm ≤ T < 0.3 mm), Level III (0.3 mm ≤ T < 0.8 mm), and Level IV (T ≥ 0.8 mm). Based on this calibration, this paper proposes an intelligent scaling severity assessment method as show in Figure 3.

Figure 3:

The model of this paper can regress the combined diameter of the electrode and deposit layer D. The thickness of the scale can be estimated as following:

(2)

$$\acute{T}=D-D_0$$

where T́ is the estimated value of T, D₀ is the diameter of the electrode without scaling, and let D₀ = 2 mm here.

In fact, D₀ varies depending on the different electrodes. More precisely, generally it is >2. This might be due to production errors, as it has almost no impact on the performance of the electrodes. It means that there is an error in the estimated value T′.

To overcome this limitation, this paper introduces scaling confidence α to correct the scaling severity. The α is the output of the model in classifying scaling or not. As mentioned earlier, classification is mainly achieved through features such as electrode edge shape. When α approaches 1, it is classified as scaling, and when α approaches 0 means no scaling. The correction rule is shown in Table 1.

Table 1:

Corrected scaling severity rule

T́/(mm)	α	Level
T́ < 0.1	α < 0.9	None I
	α ≥ 0.9
0.1 ≤ T́ < 0.3	α < 0.7	None II
	α ≥ 0.7
0.3 ≤ T́ < 0.8	α < 0.7	None III
	α ≥ 0.7
T́ ≥ 0.8	α < 0.7	None IV
	α ≥ 0.7

Δ₁ represents the diameter error between caliper and X-ray software, Δ₂ is error between X-ray and our CNN based model, Δ₃ is error between CNN and caliper.

The experiment proves that the reliability of α is very high. So, when α < 0.7, it is considered that there is no scaling. When α ≥ 0.7, the level is classified according to the reference paper. However, when the electrode diameter is very small, a higher threshold is adopted to prevent misjudgment.

III. Experimental Results

The proposed convolutional neural network (CNN) model was developed using the Pytorch framework. Training utilized the Adam optimizer with a learning rate of 0.001 and a batch size of 16. Model training and inference were executed on a computational server equipped with a NVIDIA RTX A6000 GPU, an AMD EPYC 7742 CPU, and 512 GB RAM, running Ubuntu 25.04 LTS.

Cross-validation, specifically a 5-fold stratified strategy, was implemented to ensure robustness and mitigate overfitting. The training dataset underwent augmentation, including random rotations (±15°), horizontal and vertical flips, and minor adjustments to brightness and contrast, effectively expanding the available training samples and enhancing model generalization. Figure 4 shows the Training process.

Figure 4.

As shown in Figure 4, after 100 epochs, precision is 98.5%, recall is 93.6%, mAP50 is 98.2%, and mAP50-95 is 81.1%. The mAP50-95 is a strict indicator with parameters used only for testing. It is relatively low but will not affect practical applications. Overall, this model has achieved a good balance between detection accuracy and positioning accuracy, indicating that the constructed dataset and model parameter configuration can effectively support the detection task of the fouling layer electrode.

As shown in Figure 5, in this test, we also selected 10 electrode samples and measured their actual diameters using a vernier caliper. Some of these electrodes had a scaling layer, while others did not. Table 2 shows the measurement results.

Figure 5:

This paper also conducted the scaling severity assessment on these 10 samples. The results proposed corrected rule is completely consistent with the actual results. The model is also deployed on an NVIDIA Jetson Orin Nano, and its inference speed reaches 42.1 FPS. It demonstrates that the proposed method is feasible for embedded AI devices.

Further validation involved comparing the proposed model against other AI architectures commonly employed for industrial image analysis. Evaluated alternatives included standard object detection models, including FasterR-CNN, MaskR-CNN and EfficientDet. The previous version YOLOv5-OBB has also been added for comparison.

As Table 3 shows, YOLOv5-OBB basically ineffective in this issue. Perhaps it cannot effectively identify electrode edge information with the default parameters. While other models delivered respectable performance, often exceeding 75% accuracy, they consistently fell short of the bespoke CNN architecture proposed herein. The MaskR-CNN, for instance, achieved a precision of 88.1% but exhibited a lower recall of 77.1% on the same test set. Its EfficientDet shows a better performance compared to other models. The proposed model's specific architectural optimizations, focusing on capturing subtle density variations characteristic of scaling in X-ray images of metallic pipes, contributed to this superior performance.

IV. Conclusion

The grading electrode is necessary for the water-cooling system of the converter valve. Scaling can significantly affect electrode performance. The existing method can only be disassembly and inspection during power outage maintenance.

The core contribution of this paper lies in developing an automated methodology for analyzing X-ray images of converter valve electrodes. The foundation for this analysis is a dataset comprising 1,222 meticulously annotated X-ray images of electrodes within operational water-cooling systems. This substantial dataset provides the empirical basis for training and validating the subsequent models. Recognizing the practical challenge of multiple electrodes potentially existing within a single component view and often exhibiting non-axis-aligned orientations, this paper employs the YOLOv8_OBB object detection model. This model is specifically chosen for its capability to detect objects with rotated bounding boxes, enabling the accurate localization and estimation of electrode inclination angles. This capability significantly enhances the reliability of identification compared to standard axis-aligned detectors. This paper integrates an electrode diameter regression function within the detection framework. This addition provides crucial dimensional information essential for subsequent fouling assessment stages. Building upon the established theoretical foundation that correlates heat transfer degradation with the thickness of deposited scale, this paper designs an automatic method for evaluating fouling severity. The primary metric remains the estimated scale thickness derived from the localized electrode region. To enhance the robustness of this assessment, a scale confidence measure is introduced, which refines the severity estimation by accounting for potential uncertainties in the image data or detection process.

The experimental results confirmed the technical feasibility of the model, and its detection accuracy and reliability were excellent. The corrected scaling severity rule helps to accurately identify scaling and reduce false positives. For example, the model achieved an average precision of 98.5% and a recall of 93.6% on the test set, showing excellent scaling anomaly identification capabilities. This performance highlights the potential of the model to be integrated into industrial condition monitoring systems, providing a more scientific, data-driven alternative to traditional practices. It will help develop online monitoring devices and achieve timely warning. In the future, the accuracy of scaling thickness can be further improved by enhancing the X-ray imaging system. This will provide support for the standardization of electrode scaling assessment.