Deep learning-based fault detection and classification in light electric vehicle motors

In recent years, Light Electric Vehicles (LEVs) have gained increasing popularity in the transportation industry due to their eco-friendly features and cost benefits. The widespread use of electric scooters, particularly for short-distance urban travel, has necessitated that LEV motors operate reliably and with high performance. Brushless Direct Current Motors are commonly used in LEVs. These motors, used in outdoor environments, fail over time due to environmental conditions as well as thermal, magnetic, and mechanical forces. This situation underscores the importance of early and reliable fault detection. In this thesis study, two different deep learning approaches were evaluated to detect motor faults with high accuracy under limited data conditions, using a multidimensional and hierarchical dataset obtained from experiments conducted on an electric scooter. The study examined healthy motor conditions, bearing failures, magnet failures, and multiple (bearing-magnet) failures. Signals for each motor condition were recorded using five sensors: flux, current, acoustic, voltage, and vibration. The hierarchical structure of the dataset was created using a two-layer folder structure, which contains fault type and sensor information, thereby enabling both the detection of fault type and the determination of the most distinctive data type for each fault. A multi-stage signal processing procedure was applied to enable the use of this collected dataset for deep learning models. One-dimensional time series signals were decomposed into Intrinsic Mode Function (IMF) components using the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise method. Subsequently, each IMF row was converted into two-dimensional spectrograms using the Short-Time Fourier Transform and replicated using data augmentation techniques. Two different deep learning approaches were applied to this generated dataset. In the first approach, the architecture of the Convolutional Neural Network (CNN) was designed from scratch using the Particle Swarm Optimization (PSO) algorithm. During this process, the number of layers, layer types, layer order, filter sizes, and training hyperparameters were optimized by PSO, and the resulting architecture was trained. In the second approach, deep network architectures such as GoogLeNet, MobileNet-v2, and ResNet-50, which were previously trained on large-scale datasets, were adapted to the target dataset using fine-tuning methods from Transfer Learning (TL). The performance of the models was evaluated using five-fold cross-validation, with metrics including accuracy, precision, recall, and F1 score. Experimental findings reveal that both the CNN architecture optimized with PSO and the pre-trained models with TL achieve high success in classifying multiple motor failure types. This study offers a unique solution for detecting LEV motor failures by comprehensively addressing signal processing, optimization, and TL methods, delivering reliable classification performance even under limited data conditions.