A Novel Approach for Classifying Lev Motor Bearing Faults Using a PSO-Optimized CNN Model and Parameters Based on Transfer Learning

Abstract

Light Electric Vehicles (LEVs) have recently become a popular choice for transportation due to their environmental and economic benefits. This growing interest necessitates LEV motors to operate with higher performance and reliability. In recent years, Brushless Direct Current (BLDC) electric motors have been preferred as LEV motors to meet this need. Failures occur over time in LEV motors operating in outdoor environments. In this study, a new novel method based on CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise)-STFT (Short Time Fourier Transform) and hybrid PSO (Particle Swarm Optimization)-CNN (Convolutional Neural Networks)-TL (Transfer Learning) is proposed to diagnose LEV bearing failure. In this new method, the IMFs (Intrinsic Mode Function) of one-dimensional time series vibration signals are obtained with CEEMDAN. Each IMF matrix was converted into spectrograms using STFT. Data augmentation methods enhanced these spectrograms. Using this data set, CNN model design was performed with the PSO algorithm. Parameters were optimized with the CNN model that gave the highest accuracy. Using the fine-tuning method, which is part of the transfer learning process, the performance of the obtained hyperparameters was measured with a five-fold cross-validation on GoogleNet, ResNet-50, DarkNet-53, MobileNet-v2 and Xception deep learning architectures. These architectures were evaluated with metrics such as accuracy, precision, recall and F1 score and the DarkNet-53 model gave the highest classification accuracy of 99.53%. The results show that the proposed new method is robust for diagnosing bearing failures in LEVs with limited data.