Ali Mohajerzarrinkelk

Topic

Multi-Channel Swin Transformer Framework for Rolling Bearing Remaining Useful Life Prediction

Department of Mechanical Engineering

Date & location

• Friday, August 22, 2025
• 10:00 A.M.
• Virtual Defence

Examining Committee

Supervisory Committee

• Dr. Homayoun Najjaran, Department of Mechanical Engineering, University of Victoria (Supervisor)
• Dr. Flavio Firmani, Department of Mechanical Engineering, UVic (Member)

External Examiner

• Dr. Alex Thomo, Department of Computer Science, UVic

Chair of Oral Examination

• Prof. Merrie Klazek, School of Music, UVic

Abstract

Accurate Remaining Useful Life (RUL) prediction of rotating machinery is a central challenge in predictive maintenance, where timely interventions can significantly reduce operational downtime and prevent catastrophic failures. This thesis introduces a Multi-Channel Swin Transformer (MCSFormer) framework designed to predict the RUL of rolling bearings using dual-sensor vibration data under variable operating conditions. The proposed approach emphasizes a structured preprocessing pipeline that begins with signal denoising through a sequence of low-pass filtering, wavelet-based denoising, and Savitzky-Golay smoothing to suppress noise and preserve relevant signal structure. Vibration signals are then segmented into fixed-length windows using sliding window segmentation method and transformed into time-frequency representations using Wavelet Packet Decomposition (WPD), which enables the extraction of rich degradation features at multiple resolution levels.

To process the resulting data, a convolutional neural network is first applied separately to the WPD-based images of horizontal and vertical signals. These CNN-extracted features are then concatenated and passed to a shared Swin Transformer architecture, enabling the model to jointly capture local and global patterns associated with the progression of degradation. Additionally, a safety-aware loss function is introduced to prioritize safety by penalizing late predictions more heavily than early ones, aligning the learning process with the asymmetric risk profile of industrial failure scenarios.

The framework is evaluated on the PRONOSTIA and XJTU-SY bearing datasets under both intra-condition and cross-condition settings. Comparative experiments against multiple state-of-the-art baselines demonstrate that MCSFormer achieves superior performance in both Mean Absolute Error and a scoring metric designed to assess safety-related prediction quality. Ablation studies further highlight the importance of each component, including the denoising pipeline, segmentation strategy, and custom loss function. The results affirm the proposed framework as a robust and generalizable solution for real-world prognostic systems, offering high predictive accuracy, enhanced interpretability, and risk-aware behavior suitable for deployment in industrial environments.