Document is current

Article History

Received: 5 December 2025

Accepted: 13 March 2026

First Online: 26 March 2026

Declarations

:

: Not applicable. This research is purely experimental and computational, involving vibration signals from mechanical systems; therefore, no ethical approval was required. No human subjects or human-related data were involved in this research.

: Not applicable. This study does not involve human participants, identifiable data, or personal information that would require consent to publish.

: The authors declare no competing interests.

: Portions of this manuscript were prepared with limited assistance from ChatGPT (OpenAI), specifically for improving language clarity, refining sentence structure, and generating preliminary schematic illustrations based on the authors’ technical descriptions. All scientific ideas, methodological formulations, experiments, analyses, interpretations, and conclusions were entirely conceived, executed, and validated by the authors. The authors thoroughly reviewed, edited, and verified all AI-assisted text and figures to ensure technical accuracy and scholarly integrity.

: The experimental results presented in this study can be replicated using the publicly available PHM Society 2009 Gearbox Fault Diagnosis Dataset, which was used as the sole source of vibration data in all experiments. The signal-processing pipeline implemented in this work, including Time Synchronous Averaging (TSA), Continuous Wavelet Transform (CWT), Spectral Kurtosis (SK), and Variational Mode Decomposition (VMD), is fully described in the manuscript with sufficient detail to enable independent reproduction.Custom scripts used for data preprocessing, feature extraction, deep learning model training (1D-CNN, CNN–LSTM, and the proposed CS-TFT), and evaluation are available from the corresponding author upon reasonable request. These scripts rely only on standard open-source scientific computing libraries. Due to institutional and licensing restrictions, code cannot be posted publicly at this time; however, all necessary algorithmic steps, hyper parameters, and training configurations have been documented to facilitate replication.All performance metrics, including Hamming loss, Micro-F1, Macro-F1, and confusion matrices, can be reproduced by applying the described models to the PHM09 dataset using the provided preprocessing and training pipeline.