Weldment Classification Using ML to Replace Destructive Testing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional welding procedure qualification records require time-consuming and often destructive mechanical testing to ensure quality welds, making it inefficient to identify optimal welding parameters for producing quality welds that meet code specifications.
Innovation Solution
A machine learning system is employed to generate classification algorithms using training data, including cross-sectional images and welding parameters, to predict whether a weldment meets specific specifications such as strength, ductility, or grain structure requirements, allowing for quicker characterization and correlation of weldments without the need for extensive mechanical testing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional mechanical testing is performed to ensure quality welds, then weld quality and reliability are verified, but testing time and resource consumption increase significantly
Solution Approach 1:
The patent replaces mechanical testing systems with a machine learning-based classification system. The ML model is trained on historical weld data including images and parameters, then uses this trained model to classify new weldments as meeting or not meeting specifications, substituting physical mechanical tests with computational analysis
Solution Approach 2:
The patent creates a virtual copy of the mechanical testing process through machine learning. The ML model learns from training data that includes results from actual mechanical tests, then replicates the testing function by classifying new weldments based on their features, eliminating the need for physical copying/testing of each weldment
2Reliability
If comprehensive mechanical testing is conducted on all weldments, then specification compliance is ensured, but productivity and manufacturing efficiency decrease
Solution Approach 1:
The patent applies partial action by using the ML classification system to identify weldments that likely meet specifications without requiring full mechanical testing. The system performs classification on all weldments but only triggers actual mechanical testing on those flagged as potential non-conformities, reducing overall testing volume while maintaining compliance
Solution Approach 2:
The ML-based classification system replaces the need for comprehensive mechanical testing, substituting computational analysis for physical testing on the majority of weldments while maintaining specification compliance verification
3Measurement precision
If destructive mechanical testing is performed to verify weld quality, then accurate specification verification is achieved, but the welded part is damaged or destroyed
Solution Approach 1:
The patent uses image data and process parameter data as copies or proxies for the physical weldment properties. The ML model analyzes these data copies to predict specification compliance, eliminating the need to physically destroy the welded part for testing
Solution Approach 2:
The patent substitutes mechanical destruction-based testing with a computational classification system that analyzes image and process data to verify specification compliance without physical damage to the welded part
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments of systems and methods for characterizing weldments are disclosed. One embodiment includes a method of generating an algorithm for classifying weldments as meeting or not meeting a specification. Training data is read by a machine learning system. The training data includes cross-sectional images of training weldments, truth data indicating whether the training weldments meet the specification or not, and training weld data associated with creating the training weldments. The machine learning system trains up an algorithm using the training data such that the resultant algorithm can classify a subsequent test weldment as meeting the specification or not meeting the specification when a cross-sectional image of the test weldment and test weld data used to create the test weldment are read and processed by the classification algorithm as trained.