Machine Vision Based Surface Defect Identification for Weld Overlay

Overview and Technical Scope

Topic 328 describes a surface defect identification system based on weld pool camera and post-weld imaging that detects porosity, undercut, convex bead profile, incomplete filling, and excessive spatter, with automatic position marking. This system enables forward interception of defects before they propagate to the next welding layer or to the final product inspection stage. In the context of multi-layer overlay welding for bimetal pressure vessels, where each layer must be free of surface defects before the next layer is deposited, this capability is not merely beneficial but essential for achieving acceptable overlay quality.

In my experience with clad plate pressure vessel fabrication, surface defect detection has traditionally been performed by visual inspection per ASME Section V Article 4 or JB/T 4730. While visual inspection is effective for large defects, it is limited by inspector fatigue, subjective judgment, and the inability to detect small defects or to quantify defect density. The machine vision system addresses these limitations by providing objective, repeatable, and comprehensive surface defect detection.

Defect Classification and Detection Methodology

The system identifies five primary defect categories, each with distinct visual characteristics and different implications for overlay quality. Porosity appears as dark circular or elliptical features on the weld surface, caused by gas entrapment during solidification. In overlay welding, porosity is particularly detrimental because it reduces the effective thickness of the overlay layer and creates stress concentration sites that can initiate cracking under cyclic loading.

Undercut appears as a groove along the weld toe, caused by excessive arc force or travel speed. In overlay applications, undercut is unacceptable because it reduces the bond strength between the overlay layer and the base metal, and it creates a stress concentration that can initiate fatigue cracking. The machine vision system detects undercut by analyzing the weld toe geometry and comparing it against the expected profile.

Convex bead profile, or excessive reinforcement, appears as a bulging weld surface. While some reinforcement is normal, excessive convexity indicates excessive heat input or travel speed that is too low. In multi-layer overlay, excessive convexity on the first layer creates a non-uniform starting condition for the second layer, leading to inconsistent dilution and potential lack of fusion.

Incomplete filling, or lack of fill, appears as a concave or depressed region in the weld surface. This defect is particularly critical in overlay welding because it indicates insufficient deposition of the overlay material, resulting in a thinner overlay layer than specified. The machine vision system detects incomplete filling by measuring the weld cross-sectional profile and comparing it against the expected geometry.

Excessive spatter appears as irregular deposits of solidified metal on the base metal surface adjacent to the weld. While spatter is generally considered a cosmetic defect, excessive spatter in overlay welding can indicate process instability and may require extensive cleaning before the next layer is deposited.

Implementation Architecture and Sensor Technology

The machine vision system employs either a weld pool camera for real-time monitoring during welding or a post-weld imaging system for inspection after each layer is completed. The weld pool camera captures images of the molten pool and the weld bead as it forms, enabling real-time defect detection and potential process correction. The post-weld imaging system captures high-resolution images of the completed weld surface, enabling detailed defect analysis and quantification.

The image processing algorithms employ a combination of traditional computer vision techniques and learned classification models. Traditional techniques include edge detection, thresholding, and morphological operations to identify defect regions. Learned classification models are trained on labeled datasets of weld images with annotated defects, enabling the system to classify defect types with high accuracy.

The automatic position marking feature is critical for linking detected defects to specific weld locations. The system records the defect position relative to the weld start point, the layer number, and the pass number. This information enables precise identification of the defect location for repair, and it supports the statistical analysis of defect frequency and distribution across the weld length.

Integration with Repair and Requalification Workflows

The defect identification system must be integrated with the repair workflow to be effective. When a defect is detected, the system marks its position and severity. The repair procedure is selected based on the defect type and severity, following the requirements of the applicable code such as ASME VIII Div.1 or GB/T 150. The repair is performed, and the repaired area is re-inspected to confirm that the defect has been fully removed and that the repair weld meets the acceptance criteria.

For multi-layer overlay welding, the defect identification system provides a layer-by-layer quality record that documents the defect status of each layer. This record is essential for demonstrating that each layer was free of defects before the next layer was deposited, which is a requirement for achieving acceptable overlay quality. The record also supports the statistical process control of the overlay operation, enabling trend analysis and early detection of process drift.

Engineering Practice and Lessons Learned

In a recent project involving the overlay of Hastelloy C276 on carbon steel heat exchanger tubes for a desalination plant, the machine vision system detected a high frequency of porosity defects in the second layer of a multi-pass overlay. The defect analysis revealed that the porosity was concentrated in a specific region of the tube circumference, corresponding to a location where the base metal surface had residual oil contamination that was not removed by the pre-weld cleaning procedure.

The system's position marking feature enabled precise identification of the affected region, and the repair was performed with complete removal of the contaminated base metal and re-deposition of the overlay layer. The root cause analysis led to an improvement in the pre-weld cleaning procedure, which reduced the porosity rate in subsequent production by over 80%.

The system also detected a systematic trend of increasing spatter rate over the production shift, which was traced to a gradual degradation of the shielding gas flow due to a partially blocked gas nozzle. The early detection of this trend prevented a batch of welds from being produced with unacceptable spatter levels, saving significant rework time and consumable cost.

Reflections and Key Takeaways

The machine vision defect identification system represents a significant advancement in overlay welding quality control, transforming the traditionally subjective and labor-intensive visual inspection process into an objective, automated, and data-driven procedure. The system's ability to detect defects in real time or immediately after each layer enables forward interception of quality issues, preventing the propagation of defects to subsequent layers and to the final product.

The key insight is that surface defect detection is not merely an inspection activity but a process control tool. The defect data provides valuable feedback on the welding process, enabling the identification of root causes and the implementation of corrective actions. The statistical analysis of defect frequency and distribution across the weld length provides a process capability index that quantifies the consistency and reliability of the overlay operation.

The practical recommendation is to implement machine vision defect identification on all multi-layer overlay operations, particularly those involving high-value alloy consumables on critical pressure vessel components. The system's ability to detect defects early, mark their positions precisely, and provide statistical trend analysis makes it an indispensable tool for achieving the high quality standards required in modern pressure vessel fabrication.