Robustness and Accuracy Validation of Swin Transformer-Based Predictive Maintenance for Rotating Machinery Using Vibration Spectrograms

The purpose of this study is to develop a predictive maintenance system for rotating machinery using Swin Transformer models based on vibration signals.In this research, vibration data from 11kW and 22kW motors provided by the AI Hub of South Korea were utilized.The raw vibration signals are converted into spectrograms, which serve as inputs to the Swin Transformer model.The proposed deep learning algorithm classifies normal vibration data and four types of faults: bearing failure, rotational imbalance, shaft misalignment, and loose belt conditions.The performance of the Swin Transformer model was evaluated and analyzed, demonstrating higher classification accuracy compared to traditional approaches.Furthermore, a comparative study with CNN-based models confirmed that the self-attention mechanism in the Swin Transformer effectively learns complex vibration patterns in rotating machinery.Future research will focus on enhancing predictive accuracy by integrating Fast Fourier Transform (FFT) techniques with spectrogram analysis.By combining frequency-domain features with deep learning, the reliability and accuracy of predictive maintenance systems can be further improved.The findings of this study contribute to the advancement of AI-driven predictive maintenance in smart factories, industrial equipment monitoring, and rotating machinery fault detection.This research highlights the potential of Transformer-based models for real-time fault prediction and condition monitoring in industrial applications.