Machine vision systems have served the welding industry well for decades. These systems identify edges in images by detecting sharp brightness changes and evaluate surface textures and color consistency. They require the right combination of camera, lens, lighting, and software for each specific application.

High performance and speed: Machine vision systems don't require significant computing power and consistently perform at speeds exceeding 1,000 frames per second.

Precision and accuracy: These systems are extremely precise, typically employed to measure dimensions to sub-pixel accuracy.

Explainable results: The behavior of machine vision algorithms is readily explainable, making troubleshooting and validation straightforward.

Proven applications: Laser triangulation sensors exemplify this technology's success in welding, guiding welding torches to correct positions and checking weld quality even with bright welding arcs and flying sparks present.

Limited Flexibility: Machine vision works best for simple control tasks where conditions remain consistent, parts have basic shapes and applications are based on dimensional measurements.

Environmental Sensitivity: When lighting changes due to different surface reflectivity or part shapes, systems require adjustments and re-calibration to maintain accuracy.

Complex Implementation: While simple applications can use ready-made systems, complex jobs require expert developers and specialized equipment such as telecentric lenses and custom lighting solutions.

Development Challenges: Even with expert implementation, developers need several site visits to fine-tune systems, with no guarantee of success.

Artificial Intelligence represents a set of technologies that "learn" by associating input data with outputs or states, typically through neural networks. Deep learning employs multiple layers of association for improved performance, creating systems that can adapt to varying conditions without manual re-calibration.

Adaptability: With sufficient data and variation, a single neural network model can be deployed across a wide range of shapes, scenarios (different weld joints and defects) and lighting conditions.

Simplified Programming: AI models are generally easier to program than traditional machine vision systems, reducing development time and complexity.

Robust Performance: Neural networks can handle variations in lighting, part geometry and surface conditions that would require manual adjustment in traditional machine vision systems.

Data Quality Dependency: Model performance is determined by the quality and quantity of labeled training data. Careful attention must be paid to the data annotation stage to create robust and performant models.

Explainability: The behavior of neural networks is not readily explainable, though techniques such as LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) can help understand model performance.

Training Requirements: Model training typically requires powerful GPU hardware, though deployment at the edge can occur with specialized ships that are cost, space and energy efficient.

When implementing AI solutions in welding, several challenges must be addressed:

• Image Quality and Framing: Ensuring consistent, high-quality input data
• Training Data Gaps: Avoiding insufficient or biased datasets
• Model Accuracy: Achieving reliable performance across all scenarios
• Hardware Constraints: Balancing performance with cost and space requirements
• Model Drift: Maintaining accuracy over time as conditions change

MeltTools has developed four different AI solutions, each appropriate for specific welding challenges. These models can work independently or in combination, listed from least to most complex to develop:

Purpose: Distinguish between mutually exclusive scenarios, such as determining wire presence or absence.

Applications: Flag undesirable situations, allowing operators to identify patterns when weld defects occur. This creates a foundation for quality control and process improvement.

Purpose: Simplify complex scenes by identifying and outlining specific regions of interest.

Applications: Analyze puddle shape, size and position, or asses joint geometry. This detailed analysis enables precise process control and quality assessment.

Purpose: Identify and locate specific features within welding images.

Applications: Detect silicon islands, surface droplets or subsurface heat flow issues. Real-time detection during welding can trigger alarms to stop the process, allowing operators to address faulty welds before additional defects occur.

Purpose: Identify and locate specific features within welding images.

Applications: Provide positional control that is less sensitive to image variations, ideal for torch guidance and seam tracking applications.

To choose the optimal MeltTools Edgewise model or model combination, it is important to consider the following:

• The specific welding application and environment
• The types of defects or conditions to be detected
• Required response time and accuracy
• Available computing resources
• Integration requirements with existing systems

Phase 1: Initial Experiments

Starting with a pilot project using a smaller dataset (approximately 500 images per class) to train a baseline model is a good low risk method towards a performant model. This approach allows testing of configuration settings and standardization of training data.

Phase 2: Incremental Data Addition

Gradually increase dataset size by adding more images from poorly performing components of the model. Monitor improvements in model performance to determine if additional data yields significant gains, optimizing the balance between dataset size and performance improvement.

Phase 3: Validation and Optimization

Cross-Validation: Use cross-validation and separate validation sets to ensure the model generalizes well and avoids overfitting.

Continuous Improvement: Continuously tune hyperparameters and experiment with different model architectures to optimize performance for specific welding applications.

MeltTools' MeltView® cameras are specifically designed to work with the Edgewise AI platform. However, the camera choice depends on the specific welding application, with different models optimized for various scenarios and space constraints, as well as the process requirements.

TD25 GIGE Camera: Our newest camera, designed to minimize interference of light from the welding process by synchronizing the camera and pulsed illumination with the low in the welding current. This approach significantly reduces arc visibility while maintaining clear puddle and surrounding area visualization and is particularly appropriate to GMAW, FCAW and Laser based processes. High-quality imaging across a wide weld area makes it ideal for AI integration in welding operations. While this camera is not yet available on our website, we have shared multiple videos demonstrating the camera's imaging capabilities.

HT25 SWIR GIGE Camera: Ideal for capturing thermal substrate views. Compared to other thermal cameras, the HT25 is compact and lightweight, enabling easy integration into robotic welding cells and direct mounting on robotic systems. Frame rates up to 200fps allow Edgewise software to interpret heat flow in real-time for the presence of sub surface defects.

MeltTools direct to monitor solutions: MeltTools has a wide range of very small cameras designed for various TIG welding applications such as orbital, seamer, column and boom and various cladding applications. The video feed can be brought directly into the Edgewise environment via a frame grabber.

Benefits and Implementation Considerations
Integrating AI into welding processes provides significant advantages for detecting welding defects before additional faulty welds are produced. However, successful implementation requires:

• Application-Specific Analysis: Understanding your unique welding challenges and requirements and the goals of the AI model (automated defect detection, augmented imaging, control).
• Proper Camera Selection: Choosing the right imaging system for your specific needs.
• Model Development Strategy: Following a systematic approach to data collection and model training.
• Integration Planning: Ensuring seamless integration with existing welding systems and workflows.

The integration of artificial intelligence into welding process represents a significant advancement in manufacturing quality control and process optimization. While traditional machine vision maintains advantage in specific applications requiring high speed and precision measurements, AI technologies offer unprecedented flexibility and adaptability for complex welding scenarios.

Success in implementing AI welding solutions depends on understanding the specific challenges of your application, selecting appropriate model types and following systematic development practices. The combination of advanced camera systems like MeltView® with sophisticated AI platforms like Edgewise creates powerful solutions for modern welding challenges.

This overview provides a foundation for understanding AI applications in welding. For specific implementation guidance and customized solutions for your welding operations, consider consulting with AI welding specialists who can assess your unique requirements and develop tailored solutions.

Ready to explore AI integration in your welding processes? Contact MeltTools to discuss your specific needs and learn how our Edgewise AI software can transform your welding operations.