Study of Deep Learning Architectures for Bearing Fault Diagnosis Using STFT Spectrograms

Abstract

This thesis presents a comprehensive study on vibration-based bearing fault type and severity-level detection, this process plays a critical role in predictive maintenance for industrial systems. The proposed framework combines signal processing and image-based representations derived from short-time Fourier transform (STFT) spectrograms to classify ten fault classes encompassing various fault types and severities. Among the evaluated architectures, the pretrained ImageNet model XceptionNet-71, when fine-tuned on grayscale STFTs, achieved the best overall performance, attaining a macro F1-score of 0.9979 and a mean ROC–AUC of 0.99 across all classes. This single-channel model demonstrated superior class separability compared to both flattened 1D STFT inputs and three-channel spectrograms. Which confirms that spectrogram-based representations combined with pretrained convolutional backbones are well-suited for bearing fault diagnosis and real-time deployment. While prior studies on the CWRU dataset have improved bearing fault classification through handcrafted features, lightweight CNNs, and transformer-based models, they often suffer from dataset leakage and lack systematic benchmarking. This work addresses these gaps through a unified and reproducible framework that compares 1D and 2D CNNs, extends Delta-STFT into a cross-resolution multi-channel representation, and conducts a comprehensive evaluation to classify safe versus unsafe misclassifications, bridging the gap between high accuracy and practical deployability.