Explainable Machine Learning to Improve Assembly Line Automation

Abstract

Faulty manufactured product causes huge economic loss in the manufacturing industry. A local company produces a power transfer unit (PTU) for the vehicle industry and in this production 3% of PTU are rejected due to a mismatch of shim (a small mechanical part supporting PTU). Today the dimension of a shim is predicted manually by human experts. However, there are several problems due to the manual prediction of shim dimension, automatic central control from the cloud cannot be done. Additionally, it increases rejection rates and as a consequence decreases the reliability of the systems. To solve these problems, in this study shim prediction is implemented in the manufacturing of PTU with explainable Machine Learning (ML) which automates the manual shim selection process in the assembly line and explains the ML prediction. A hybrid approach that combines support vector regression (SVR) and k nearest neighbours (kNN) for the first part of the assembly line and Partial Least Squares (PLS) and kNN for the second part of the assembly line is used for shim prediction. The hybrid approach is selected due to better performance compared to the single ML model approach. Then, the most important features of the hybrid approach were identified with SHAP (SHapley Additive exPlanations). The result indicates due to this improved automation faulty PTU rate decreased from 3% to only 1%. Additionally, it enabled control from the cloud and increased reliability. From the explanation of the hybrid approach, it is evident that one of the features values has more impact on the prediction output and controlling this feature will reduce the rejection rate.