Fault detection and classification for predictive maintenance using machine learning and deep learning methods

Predictive maintenance efficiently manages machinery and equipment maintenance schedules by predicting potential faults using smart sensors, data analysis, and machine learning algorithms. Fault classification categorizes these faults as either failures or normal variations, thereby enhancing maintenance planning and operational efficiency while ensuring timely interventions and smoother operations. This thesis study comprises two main parts. The first section involves classification investigations conducted using a one-dimensional (1D) predictive maintenance dataset. The subsequent section focuses on transforming the original data into a two-dimensional (2D) image data, followed by further classification analyses. The first part of thesis introduces a study of different methodologies for classifying the failures in original 1D dataset. We present the performance evaluation of fault classification performed by traditional machine learning methods such as decision tree, support vector machine, k-nearest neighbors, and 1D deep learning techniques like 1D-LeNet, 1D-AlexNet, and 1D-VGG16. From the results, using 1D-LeNet classifiers and data normalization achieved the highest accuracy and F1-score performance. In the second part of thesis, a robust fault classification system was employed using eight different convolutional neural network (CNN) models (AlexNet, VGG16, MobileNetV2, VGG19, DarkNet19, DarkNet53, ResNet50 and ResNet18) and 2D image data generated from original 1D dataset comprising two different classes (machine failure and normal). Ant colony algorithm (ACO), particle swarm optimization (PSO) and grey wolf optimization (GWO) were applied within the feature selection methods to combined features, which were attained from features extracted from CNN models, and their performances were subsequently compared. The results were acquired using four different classifier such as support vector machine, k-nearest neighbors, decision tree and naive bayes. Considering the experimental results obtained, it has been seen that using the model based on support vector machines and ACO provides the highest classification performance for 2D image data.