Multi-Scale Feature-Based Fault Classification in Proton Exchange Membrane Fuel Cell Systems using WPT and Random Forest

Uyar M., Akgün M. U.

III. International Congress on Engineering and Sciences, Ankara, Türkiye, 22 - 24 Ekim 2026, ss.377-387, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: Ankara
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.377-387
  • Bursa Uludağ Üniversitesi Adresli: Evet

Özet

Proton exchange membrane fuel cells (PEMFCs) are widely used in transportation, stationary power generation, and portable energy applications due to their high energy conversion efficiency, low 

emissions, and flexible operating characteristics. However, their performance and reliability can be significantly affected by abnormal operating conditions such as membrane drying, cathode flooding, 

hydrogen starvation, thermal imbalance, and sensor-related faults. Therefore, accurate and early fault diagnosis is essential for improving system reliability and preventing unexpected failures. This study 

proposes an explainable and data-driven fault diagnosis framework for PEMFC systems by integrating the wavelet packet transform (WPT), analysis of variance (ANOVA), synthetic minority over-sampling 

technique (SMOTE), kernel Fisher discriminant analysis (KFDA), and random forest (RF). The analysis is performed using real PEMFC measurement data comprising 206,360 samples, with system operation 

categorized into three states: Normal, Unknown/Transitional, and Faulty. For feature extraction, a three-level WPT decomposition using the Daubechies-4 (db4) mother wavelet is applied to the raw sensor 

signals. Ten statistical measures are extracted from each of the eight terminal nodes, resulting in 80 discriminative features. ANOVA-based significance analysis is then used to identify the 20 most 

informative variables, while SMOTE is employed to address class imbalance. The resulting feature space is reduced to two dimensions using KFDA and subsequently classified using the RF algorithm. 

The proposed WPT–ANOVA–KFDA–RF framework achieves an overall classification accuracy of 95.08% and an F1-score of 93.28%. The results further indicate that uncertain transitional states and 

early performance degradation can be detected before they develop into severe faults, demonstrating the potential of the proposed framework for reliable PEMFC fault diagnosis and predictive maintenance 

applications.