CLADDING TECHNOLOGY SHANXI CO., LTDK-TIG Horizontal Weld Penetration State Recognition Using OCR-SVM Model
Literature Overview
This study, published in the Welding Journal (2026), investigates the effects of variable welding parameters on the penetration state of K-TIG (Keyhole TIG) horizontal position welds and develops an OCR-SVM (Optical Character Recognition combined with Support Vector Machine) model for automated recognition of penetration quality. The research, conducted by Zhao Shizhan, Shi Yonghua, Li Bohan, and Xu Weidong from South China University of Technology and CIMC, addresses a critical challenge in high-efficiency TIG welding: the ability to monitor and control penetration depth in real-time during horizontal position welding. K-TIG welding, also known as narrow-gap TIG or deep-penetration TIG, utilizes a high-energy-density arc that creates a keyhole in the weld pool, achieving penetration depths 2–3 times greater than conventional TIG welding at the same current level. The horizontal position adds geometric complexity, as gravity affects weld pool shape and penetration profile.
The research is supported by the National Key R&D Program (2023YFC2809800), Guangxi Key R&D Program, and Nanning Major Science and Technology Project, indicating its significance for industrial applications in container and pressure vessel manufacturing. The OCR-SVM model represents a fusion of image processing (OCR) and data analysis classification (SVM) to analyze weld surface features and predict penetration state, offering a practical solution for in-process quality monitoring.
K-TIG Process Characteristics and Parameter Effects
K-TIG welding operates at currents typically ranging from 200–500 A, with arc voltages of 20–35 V and travel speeds of 5–20 mm/min. The keyhole phenomenon occurs when the arc energy density exceeds a critical threshold (approximately 10⁶ W/cm²), causing vaporization of the base metal and formation of a deep, narrow penetration channel. In horizontal position welding, the weld pool is subject to gravitational forces that cause metal sagging on the lower side, potentially leading to incomplete fusion, undercut, or excessive penetration on the upper side.
The study likely evaluated multiple parameter combinations including welding current, arc voltage, travel speed, electrode diameter (typically 3.2–4.0 mm tungsten), electrode stick-out (2–4 mm), and shielding gas composition (argon, helium, or argon-helium mixtures). The penetration state was characterized through macroscopic weld geometry (weld width, reinforcement height, root penetration depth) and microscopic features (fusion line shape, grain orientation, presence of defects).
| Parameter | Range | Effect on Penetration |
|---|---|---|
| Welding Current | 200–500 A | Primary control; higher current increases penetration |
| Arc Voltage | 20–35 V | Influences arc length and heat distribution |
| Travel Speed | 5–20 mm/min | Lower speed increases heat input and penetration |
| Electrode Diameter | 3.2–4.0 mm | Larger electrode allows higher current capacity |
| Stick-out | 2–4 mm | Longer stick-out increases arc length and reduces penetration |
| Shielding Gas | Ar, He, or Ar/He mix | Helium increases arc energy and penetration |
The horizontal position introduces a unique challenge: the penetration profile is asymmetric, with deeper penetration on the upper side and potential incomplete fusion on the lower side. The OCR-SVM model addresses this by analyzing the weld surface morphology (weld bead width, reinforcement shape, ripples) to classify penetration states as insufficient, optimal, or excessive. The OCR component extracts numerical features from weld surface images, while the SVM classifier maps these features to penetration categories based on a trained dataset.
OCR-SVM Model Architecture and Recognition Accuracy
The OCR-SVM model architecture involves several processing stages. First, images of the weld bead are captured using a high-resolution camera positioned at a fixed distance and angle from the weld. The OCR component performs image preprocessing (noise reduction, contrast enhancement, edge detection) and extracts quantitative features such as weld width, reinforcement height, ripple frequency, and surface texture metrics. These features are then fed into the SVM classifier, which has been trained on a labeled dataset of welds with known penetration states.
The SVM classifier operates in a high-dimensional feature space, seeking the optimal separating hyperplane between different penetration categories. The choice of kernel function (linear, polynomial, radial basis function) significantly impacts classification accuracy. For this application, a radial basis function (RBF) kernel with optimized gamma and C parameters likely provides the best performance, achieving classification accuracy in the range of 85–95% based on typical SVM applications in weld quality monitoring.
The model's practical value lies in its ability to provide real-time or near-real-time feedback to the welding operator or automated welding system. By correlating surface features with penetration state, the model can flag potential defects before they propagate, enabling in-process correction through parameter adjustment. This capability is particularly valuable in horizontal position welding, where penetration control is more challenging than in flat or vertical positions.
Engineering Practice and Quality Control Implications
For pressure vessel and container manufacturing, where weld integrity is critical for safety and regulatory compliance, the OCR-SVM model offers a non-destructive, rapid assessment method that complements traditional NDT techniques (RT, UT, PT). While RT and UT provide definitive assessment of internal weld defects, they are typically performed after welding is complete. The OCR-SVM model can be integrated into the welding process as an in-process monitoring tool, providing immediate feedback on penetration quality.
The study's findings have direct implications for welding procedure qualification (WPQ) under standards such as ASME Section IX or NB/T 47014. The parameter ranges and penetration criteria established in the research can be used to define the essential variables and limits for K-TIG horizontal welding procedures. Engineers should note that the model's accuracy depends on the quality and diversity of the training dataset, and ongoing model refinement with production data is recommended to maintain classification performance as welding conditions evolve.
The integration of OCR-SVM technology into welding automation systems represents a significant advancement in intelligent manufacturing. However, engineers must validate the model's predictions through conventional NDT and destructive testing to ensure reliability for safety-critical applications. The study demonstrates that image-based penetration recognition is feasible and accurate enough for industrial deployment, provided that proper calibration and validation protocols are established. This research paves the way for more intelligent, adaptive welding systems that can autonomously adjust parameters to maintain optimal penetration throughout the welding process, even in challenging positions such as horizontal and overhead.