Using Machine Learning for Predicting Interchanges

Abstract

In the bearing industry, identifying equivalent products across different manufacturers is a complex and time-consuming task, traditionally reliant on expert knowledge and manual cross-referencing. This thesis explores the feasibility of automating this process using machine learning techniques. The core challenge lies in predicting an equivalent bearing designation from a competitor’s product code, which typically lacks structured attribute data and consists of a single, information-dense string. To address this, a comprehensive machine learning pipeline was developed, encompassing feature extraction, encoding, model training, and decoding. The problem was framed as a multi-label classification task, where suffixes representing product features were predicted and then used to reconstruct a valid target designation. Several machine learning models were evaluated, including Random Forests, Support Vector Machines, k-Nearest Neighbors, and Neural Networks, using both problem transformation and method adaptation strategies. Random Forests with label powerset transformation emerged as the most effective approach, offering a strong balance between accuracy and computational efficiency. The pipeline was further optimized through domain-specific feature engineering, such as extracting product size and type indicators, which significantly improved model performance. Despite inherent limitations in the data—such as inconsistent labeling and sparse representation of certain suffixes—the results demonstrate that machine learning can reliably assist in generating bearing interchanges. This approach has the potential to significantly reduce manual effort and scale the interchange process, making it a valuable tool for industry practitioners.