REPOSITORY > RESULTS

Abstract

Electric motors are produced in very high volumes and are critical parts of modern industrial and consumer systems, devices and appliances. Ensuring their reliability requires robust and automated End-of-Line (EoL) quality inspections that can detect mechanical and electromechanical defects. The EoL quality inspection is usually performed at the end of the manufacturing line and must follow the production rate of the manufacturing line. Cost constraints are also present. During quality inspection, various electrical, vibrational, acoustic, and other signals are captured. These signals are processed using predefined signal processing methods (e.g., multiple digital filtering, down-sampling and Hamming window) to extract discrete scalar values, known as features, that describe the characteristics of each inspected instance. This process is commonly known as feature engineering. However, traditional EoL inspection systems rely heavily on expert-defined features, threshold-based rules, and manual commissioning procedures. These systems are limited in their adaptability and robustness in modern, high-variability industrial environments.
This dissertation investigates applying machine learning, transfer learning, and deep learning approaches to intelligent EoL quality inspection of brushless DC motors using real industrial data from automated manufacturing lines. Three complementary approaches are developed and evaluated.
First, supervised machine learning combined with automated feature selection reduces diagnostic complexity while improving classification performance and interpretability. Second, transfer learning accelerates the commissioning of inspection systems for new motor types with limited preproduction data by reusing knowledge from mass-production models. Third, a deep learning approach based on Mel-frequency spectrograms and a hybrid convolutional neural network (CNN) and bidirectional gated recurrent unit (BiGRU) architecture is proposed to detect faults directly from raw vibration and acoustic signals, eliminating the need for manual (human expert based) feature engineering.
All methods are validated using large-scale industrial datasets obtained from fully automated EoL inspection system. The results demonstrate that automated feature selection improves efficiency and robustness; transfer learning significantly reduces data requirements and commissioning time; and deep learning improves the detection of complex fault signals. These findings confirm that AI-based inspection systems can reduce dependence on human experts and provide scalable, adaptive, and reliable quality inspection systems for modern electric motor manufacturing.