Diagnosis of Fuel Cell Health Status with Deep Sparse Auto-Encoder Neural Network

Chenyan Fei, Dalin Zhang, Chen Melinda DangarXiv 2025 · DOI · arXiv

Abstract

Effective and accurate diagnosis of fuel cell health status is crucial for ensuring the stable operation of fuel cell stacks. Among various parameters, high-frequency impedance serves as a critical indicator for assessing fuel cell state and health conditions. However, its online testing is prohibitively complex and costly. This paper employs a deep sparse auto-encoding network for the prediction and classification of high-frequency impedance in fuel cells, achieving metric of accuracy rate above 92%. The network is further deployed on an FPGA, attaining a hardware-based recognition rate almost 90%.

Introduction

As electric vehicles (EVs) become more prevalent, maintaining the health and efficiency of their Fuel Cells (FCs) is vital. A key indicator of fuel cell health and catalyst degradation is high-frequency resistance (HFR). Traditionally, measuring HFR requires Electrochemical Impedance Spectroscopy (EIS)—a process that is prohibitively complex and costly to perform online inside a moving vehicle. While researchers have tried using predictive models to estimate HFR instead of measuring it directly, physics-based models struggle to adapt to changing operating conditions, and standard data-driven models (like LSTMs or XGBoost) tend to overfit when faced with the noisy, small datasets typical of raw electrochemical sensor readings. This paper addresses the need for a lightweight, highly accurate, and hardware-friendly model that can monitor fuel cell health in real-time.

Methodology

To overcome the overfitting and noise issues of standard neural networks, a Deep Sparse Auto-Encoder (DSAE) is proposed.

DSAE Architecture

The network takes 10 physical sensor inputs and passes them through two hidden layers (32 and 16 neurons using ReLU activation), outputting a classification into one of three health states.

The Deep Sparse Auto-Encoder (DSAE) architecture featuring input, hidden, and output layers.

A novel sparsity constraint is introduced via a penalty term in the loss function based on Kullback-Leibler (KL) divergence. This enforces sparsity on the hidden layer neurons, ensuring only the most essential neurons activate for a given input. This forces the network to learn robust, latent features rather than memorizing noisy raw data.

Dataset and Training

The model was trained and evaluated on a real-world dataset of 36,363 records collected from fuel cell OEM test benches, split into a 3:1 training-to-testing ratio. Features include 10 inputs such as output power, current density, stack voltage, and various pressures and flow rates. The target variable is HFR, categorized into three classes: Class 0 (HFR < 89 mΩ), Class 1 (89 ≤ HFR < 91 mΩ), and Class 2 (HFR ≥ 91 mΩ).

Hardware Deployment

To demonstrate real-world viability, the algorithm was translated into C++, synthesized into a hardware IP core using Vivado HLS, and deployed onto a Xilinx PYNQ-Z2 Field Programmable Gate Array (FPGA) board for high-speed, embedded inference.

Results and Discussion

The DSAE approach demonstrates that sparsity acts as a shield against noise, filtering out the inherent variability in electrochemical sensors. The feature compression inherent to the auto-encoder structure results in lower computational complexity, making it ideal for the resource limits of an FPGA.

Software and Hardware Performance

In Python simulations, the DSAE network achieved a classification accuracy of 92.13% with a Mean Squared Error (MSE) of 0.0787. Precision, recall, and F1-scores all exceeded 90%. The model converged rapidly, reaching optimal performance by the 9th training epoch.

Training and validation accuracy curves showing convergence by epoch 9.

When deployed on the Xilinx PYNQ-Z2 FPGA, the hardware-accelerated model maintained a high recognition rate, achieving an accuracy of 89.569%. The minor drop in accuracy (~2.5%) from simulation to hardware is attributed to standard quantization and hardware optimization constraints.

Conclusion

The successful deployment of the DSAE model on an FPGA proves that complex EIS hardware can be replaced by smart software-hardware co-design. This paves the way for more reliable on-board diagnostics for EVs. Future work will involve testing the FPGA system in live, on-vehicle environments to validate real-time performance under actual driving conditions.

Summary updated

Abstract

Effective and accurate diagnosis of fuel cell health status is crucial for ensuring the stable operation of fuel cell stacks. Among various parameters, high-frequency impedance serves as a critical indicator for assessing fuel cell state and health conditions. However, its online testing is prohibitively complex and costly. This paper employs a deep sparse auto-encoding network for the prediction and classification of high-frequency impedance in fuel cells, achieving metric of accuracy rate above 92\%. The network is further deployed on an FPGA, attaining a hardware-based recognition rate almost 90\%.
BibTeX
@misc{fei2025diagnosis,
  author = {Chenyan Fei and Dalin Zhang and Chen Melinda Dang},
  title = {{Diagnosis of Fuel Cell Health Status with Deep Sparse Auto-Encoder Neural Network}},
  year = {2025},
  eprint = {2510.17214},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.17214},
  howpublished = {arXiv},
}
Figure