Efficient fault detection in industrial equipment using PCA and SMOTE enhanced neural networks

Abstract

This research addresses the challenge of fault detection in industrial equipment using high-dimensional vibration data with limited labeled examples. The goal was to develop a neural network model capable of accurately classifying measurement vectors into normal and faulty categories. The dataset consisted of 1158 samples, each with 93,752 numerical features, representing two classes: 865 normal and 293 faulty instances. A comprehensive preprocessing pipeline was employed, including standardization, dimensionality reduction using Principal Component Analysis (PCA), and Synthetic Minority Over-sampling Technique (SMOTE) for class balancing. The developed neural network achieved a baseline accuracy of 94.40% with 100 PCA components. Further experiments demonstrated that reducing the architecture and using only 50 PCA components improved accuracy to 98.81%, highlighting the effectiveness of the proposed approach. These findings emphasize the utility of combining PCA, SMOTE, and neural networks for fault detection in industrial equipment in high-dimensional, imbalanced datasets. Future research directions include exploring advanced neural network architectures, investigating the impact of PCA component count on model performance, and studying the feasibility of training effective models on synthetic data.