Crack Detection in Powder Compacts using Machine Learning Models

Purpose - Cracks in powder metallurgy components, being a common problem, pose significant manufacturing challenges, but are detrimental to be detected as they affect the material’s mechanical properties. To detect these cracks, Non-Destructive Testing (NDTs) methods are often used, but they come with high costs and time delays as samples need to be extracted from production in given intervals. To overcome these limitations, an indirect method of crack detection, that is, modelling it as a binary classification, is explored in this work. Design/Methodology/Approach - This study introduces a supervised machine learning approach using force signal feature extraction to detect cracks. More specifically, a supervised learning algorithm is developed and validated for the classification of samples into samples with or without crack based on the sensory data of the hydraulic press used for production. We compare different ensemble classifiers including Random Forest (RF), AdaBoost (ADA), Bagging, Gradient Boosting (GB), and Extra Trees (ET) in terms of their ability to classify workpieces using a dataset from real production. Findings - To this end, the present study deals with experimental workpieces of a specific type produced by manually adjusting the press parameters to artificially induce cracks in parts of the workpieces. The best performing model resulted in a classification accuracy as high as 99% offering a cost-effective and efficient alternative to traditional NDT methods. Originality/Value - This study provides a novel and indirect method for detecting cracks in powder metallurgy components using machine learning models trained on press sensor data, which can significantly reduce the need for costly and time-consuming NDT techniques.