Generalized fault diagnostics of polymer electrolyte fuel cells using machine learning.

Area of Science:

• Electrochemistry
• Materials Science
• Energy Systems

Background:

• Polymer electrolyte fuel cells (PEFCs) offer significant potential for both mobile and stationary power generation.
• However, their widespread commercial adoption is hindered by limited operational lifetimes and susceptibility to faults, particularly water management and starvation issues.
• Effective diagnostic tools are crucial for improving PEFC reliability and longevity.

Purpose of the Study:

• To develop and evaluate a black-box diagnostic method for identifying water management and starvation faults in PEFCs.
• To assess the performance of different machine learning models (DNNs, 1D-CNNs, SVMs) in classifying PEFC operational states.
• To investigate the impact of dataset diversity on model generalization for robust fault detection.

Main Methods:

• Implementation of a black-box diagnostic approach utilizing multifrequency Walsh function perturbation signals.
• Analysis of voltage response data to detect anomalies indicative of specific faults.
• Comparative evaluation of Deep Neural Networks (DNNs), 1D Convolutional Neural Networks (1D-CNNs), and Support Vector Machines (SVMs) for fault classification.
• Testing model performance on single and multiple PEFC datasets to assess generalization capabilities.

Main Results:

• All tested models (DNNs, 1D-CNNs, SVMs) demonstrated high accuracy in classifying normal, drying, and starvation conditions within a single PEFC, with 1D-CNN and SVM achieving 100% accuracy.
• Initial model generalization to unseen PEFCs was limited when trained on data from a single cell.
• Incorporating data from multiple PEFCs significantly enhanced model performance and generalization.
• The 1D-CNN model exhibited superior generalization capabilities, even with limited training data from unseen sources.

Conclusions:

• Multifrequency Walsh function perturbation signals provide an effective, non-invasive method for diagnosing PEFC faults.
• The 1D-CNN model demonstrates the highest potential for robust and scalable PEFC diagnostics across diverse operating conditions and hardware variations.
• Dataset diversity is critical for developing reliable diagnostic models that can generalize across different PEFC units.