Defect Detection in Multiple Product Variants Using Hammering Test with Machine Learning

Various nondestructive testing (NDT) methods have been proposed to detect defects inside products. The hammering test is an NDT technique widely used for this purpose. In this test method, a worker judges whether a part is defective or not by listening to the sound after hitting the product with a hammer. Conventional research has shown that a classifier using machine learning can discriminate the hammering data with high accuracy. However, to use these machine learning methods, a lot of samples are needed for learning. In actual industrial situations, it is difficult to collect a lot of samples of defective products. Regarding the hammering test, a machine learning method that can correctly discriminate defective products without sample data has not been proposed. This study aims to construct a system that can correctly discriminate the hammering test data even when there are no defective samples. We propose a method using ‘transfer learning.’ We conducted case studies to demonstrate the effectiveness of the proposed method using two variants of a brazed product. First, we verified the effectiveness of normal machine learning in a hammering test. In this study, we succeeded in discriminating brazed products, which were not correctly discriminated by the workers. We then applied the proposed method to brazed products. We succeeded in discriminating a variant of the brazed products by transferring the knowledge learned from another variant of the brazed products.