Scholarlink

Background <p>Accurate monitoring of tire wear using vibration signals is important for improving vehicle safety, ride quality, and predictive maintenance. However, many existing tire wear detection methods rely on handcrafted features that may fail to capture complex wear-related patterns and often provide limited interpretability. To address these limitations, this study proposes a decoupled deep learning framework for automatic tire wear classification using raw vibration signals.</p> Methods <p>The proposed framework consists of two parallel learning streams designed to capture complementary information from vibration signals. A 1D convolutional neuralnetwork (1D-CNN) is employed to learn transient dynamics directly from the time-domain signals, while a Short-Time Fourier Transform (STFT)-based 2D convolutional neural network (2D-CNN) extracts time–frequency resonance characteristics. Experiments were conducted using vibration data acquired at a sampling rate of 25 kHz under five tire wear conditions: 0%, 25%, 50%, 75%, and 100%. Several hyperparameter tuning experiments were carried out to improve convergence stability, computational efficiency, and generalization performance. Model evaluation was performed using 5-fold cross-validation.</p> Results <p>The proposed dual-stream framework achieved an average classification accuracy of 97.82% ± 1.46, demonstrating strong robustness and reliable generalization across all tire wear conditions under controlled laboratory settings. The analysis revealed that spectral features contributed more significantly to the identification of early-stage tire wear, whereas transient time-domain features became increasingly important for advanced wear conditions. The integration of both feature representations improved overall classification accuracy and model stability across different operating conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dual-stream Temporal–spectral Deep Learning for Vibration-based Tire Wear Classification

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • Arun Balaji M,
  • Sugumaran V

摘要

Background

Accurate monitoring of tire wear using vibration signals is important for improving vehicle safety, ride quality, and predictive maintenance. However, many existing tire wear detection methods rely on handcrafted features that may fail to capture complex wear-related patterns and often provide limited interpretability. To address these limitations, this study proposes a decoupled deep learning framework for automatic tire wear classification using raw vibration signals.

Methods

The proposed framework consists of two parallel learning streams designed to capture complementary information from vibration signals. A 1D convolutional neuralnetwork (1D-CNN) is employed to learn transient dynamics directly from the time-domain signals, while a Short-Time Fourier Transform (STFT)-based 2D convolutional neural network (2D-CNN) extracts time–frequency resonance characteristics. Experiments were conducted using vibration data acquired at a sampling rate of 25 kHz under five tire wear conditions: 0%, 25%, 50%, 75%, and 100%. Several hyperparameter tuning experiments were carried out to improve convergence stability, computational efficiency, and generalization performance. Model evaluation was performed using 5-fold cross-validation.

Results

The proposed dual-stream framework achieved an average classification accuracy of 97.82% ± 1.46, demonstrating strong robustness and reliable generalization across all tire wear conditions under controlled laboratory settings. The analysis revealed that spectral features contributed more significantly to the identification of early-stage tire wear, whereas transient time-domain features became increasingly important for advanced wear conditions. The integration of both feature representations improved overall classification accuracy and model stability across different operating conditions.