Advancing non-destructive weld assessment using autoencoder-based feature extraction

Ensuring weld quality is critical in industries, such as automotive, aerospace, and construction, where the integrity and durability of components directly affect safety and performance. Traditional non-destructive testing (NDT) methods are widely used but can be time-consuming, require specialised expertise, or be limited to surface-level evaluations. New developments in the fields of artificial intelligence and machine learning (ML) provide better, more effective, and more independent solutions. In this paper, a new approach of welding detection system that uses an autoencoder-based feature extraction method for the classification of 14 different welding conditions based on infrared (IR) image datasets is introduced. By using autoencoders, the dimensions of IR data are reduced, and important features are identified. Out of various classifiers used, Linear Discriminant Learner provided the best results and performed even better than 98% accuracy on the test data. This approach does not only help in the effective evaluation of the weld quality but also increases the flexibility of the system to accommodate different welding methods and materials. The performance of the proposed system and its simplicity also point to its usefulness as a novel technology in the management of manufacturing quality and minimising the need for time-consuming manual examination.