Abstract
Defects such as cracks and mass reduction frequently occur during the production of powder metallurgy (PM) automotive oil pump stators, making rigorous inspection essential for reliable operation. Conventional human visual inspection is threshold-based, simple, and cost-effective, but it is limited by the presence of microscopic cracks, internal defects, and declining detection speed. In contrast, machine learning (ML) can automatically identify complex, non-linear patterns in acoustic signals, enabling more accurate and rapid defect detection. In this study, acoustic signals were recorded from 40 intact and 62 defective PM components, including 26 cracked, 16 with tooth breakage, and 20 completely fractured samples. Distinctive features were extracted from these signals and used to train multiple ML classifiers, including support vector machine, k-nearest neighbors, multilayer perceptron, and radial basis function (RBF) networks. Comparative evaluation revealed that the RBF network outperformed the other models, achieving 100% accuracy in distinguishing defective from intact components. This approach demonstrates that combining acoustic signal analysis with ML not only surpasses conventional inspection methods in accuracy and speed but also provides a scalable and reliable solution for industrial defect detection in PM components.