Scholarlink

This paper presents a comparative study of Multilayer Perceptrons (MLPs) and Kolmogorov-Arnold Networks (KANs) for fault detection (FD) in automotive engines using audio signals. The study explores the use of two feature extraction (FE) techniques, Mel-Frequency Cepstral Coefficients (MFCCs) and Log-Mel Spectrogram (LMS), to capture the relevant characteristics of engine sound. The experimental results demonstrate that KANs consistently outperform MLPs, particularly when combined with LMS features, achieving the highest accuracy (97.67%) and perfect precision (100%). This research highlights the potential of KANs for developing low-cost, non-invasive FD systems, offering a practical alternative to traditional sensor-based methods. The findings contribute to the growing body of work on AI-driven fault detection in industrial applications, with implications for resource-constrained environments and real-time monitoring systems.

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

Performance Comparison Between Multilayer Perceptrons and Kolmogorov-Arnold Networks for Fault Detection in Automotive Engines

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Akram Houssam Eddine Hacini,
  • Farouq Zitouni,
  • Saad Harous,
  • Aridj Ferhat,
  • Rihab Lakbichi,
  • Abdelhadi Limane

摘要

This paper presents a comparative study of Multilayer Perceptrons (MLPs) and Kolmogorov-Arnold Networks (KANs) for fault detection (FD) in automotive engines using audio signals. The study explores the use of two feature extraction (FE) techniques, Mel-Frequency Cepstral Coefficients (MFCCs) and Log-Mel Spectrogram (LMS), to capture the relevant characteristics of engine sound. The experimental results demonstrate that KANs consistently outperform MLPs, particularly when combined with LMS features, achieving the highest accuracy (97.67%) and perfect precision (100%). This research highlights the potential of KANs for developing low-cost, non-invasive FD systems, offering a practical alternative to traditional sensor-based methods. The findings contribute to the growing body of work on AI-driven fault detection in industrial applications, with implications for resource-constrained environments and real-time monitoring systems.