Yang Peng, Zhang Bozheng, Zhao Jianda
Computer Science and Technology, Tianjin University of Technology, Tianjin, 300384, China.
Sci Rep. 2025 Apr 3;15(1):11514. doi: 10.1038/s41598-025-96285-z.
To address the issue of low efficiency in feature extraction and model training when traditional deep learning methods handle long time-series data, this paper proposes a Time-Series Lightweight Transformer (TSL-Transformer) model. According to the data characteristics of bearing fault diagnosis tasks, the model makes lightweight improvements to the traditional Transformer model, and focuses on adjusting the encoder module (core feature extraction module), introducing multi-head attention mechanism and feedforward neural network to efficiently extract complex features of vibration signals. Considering the rich temporal features present in vibration signals, a Long Short-Term Memory (LSTM) module is introduced in parallel to the encoder module of the improved lightweight Transformer model. This enhancement further strengthens the model's ability to capture temporal features, thereby improving diagnostic accuracy. Experimental results demonstrate that the proposed TSL-Transformer model achieves a fault diagnosis accuracy of 99.2% on the CWRU dataset. Through dimensionality reduction and visualization analysis using the t-SNE method, the effectiveness of different network structures within the proposed TSL-Transformer model is elucidated.
为了解决传统深度学习方法处理长时间序列数据时特征提取和模型训练效率低下的问题，本文提出了一种时间序列轻量级Transformer（TSL-Transformer）模型。根据轴承故障诊断任务的数据特点，该模型对传统Transformer模型进行了轻量级改进，重点调整编码器模块（核心特征提取模块），引入多头注意力机制和前馈神经网络以高效提取振动信号的复杂特征。考虑到振动信号中存在丰富的时间特征，在改进的轻量级Transformer模型的编码器模块中并行引入了长短期记忆（LSTM）模块。这一增强进一步提升了模型捕捉时间特征的能力，从而提高诊断准确率。实验结果表明，所提出的TSL-Transformer模型在CWRU数据集上实现了99.2%的故障诊断准确率。通过使用t-SNE方法进行降维和可视化分析，阐明了所提出的TSL-Transformer模型内不同网络结构的有效性。
