| Abstract: |
Electric motors, widely mass-produced for various industries, undergo end-of-line (EoL) quality inspections to ensure product reliability and prevent faulty units from reaching customers. This study aims to enhance the efficiency of EoL quality inspection systems by using machine learning techniques to reduce the feature space involved in motor fault detection. Traditional quality inspections rely on non-invasive measurements, such as electrical parameters, speed, torque, vibrations, and sound, collected at high frequencies and processed to identify faults. However, managing a large number of features can be inefficient due to redundancy and complexity in threshold determination, as well as high computational demands.
The core objective of this research is to streamline the feature selection process using machine learning methods, including Decision Tree, Random Forest, Bagging, and Gradient Boosting classifiers. These techniques were applied to industrial data from a leading European electric motor manufacturer, with the goal of reducing the number of features without compromising the accuracy of fault detection. By focusing on the most relevant features, the proposed approach not only reduces computational overhead but also minimizes sensor dependency, optimizing both hardware and software components of the inspection system.
The study examines the quality inspection system for brushless DC (BLDC) motors used in domestic and automotive applications. The existing EoL inspection system involves capturing motor parameters through various sensors, categorizing faults into electrical and mechanical issues. However, the current manual process of feature selection and limit value adjustment is time-consuming and heavily reliant on expert input. To address this, the study applies supervised machine learning algorithms to automate feature selection, retain 95% of useful information, and generate classification models that reduce commissioning time and improve system efficiency.
Machine learning methods were evaluated on a dataset of 37,440 motors. Classifiers were trained and tested on both full and reduced feature sets, showing that feature reduction improves inspection efficiency by decreasing the number of sensors and simplifying diagnostic steps. Bagging classifiers outperformed others in terms of misclassification costs, while Gradient Boosting required fewer features but showed higher misclassification costs. All classifiers demonstrated high accuracy and contributed to reducing the feature space, thereby optimizing database operations and simplifying the quality inspection process.
The research also highlights the practical benefits of these classification models, such as reduced data storage, lower computational load, and streamlined inspection procedures. By automating feature selection and threshold value setting, the dependency on specialized experts is minimized. Furthermore, the study proposes future research directions, including exploring the transferability of classification models across different motor types and using feature importance evaluation for condition monitoring in manufacturing processes. These advancements could significantly enhance the speed, accuracy, and cost-efficiency of quality inspection systems in industrial settings. |