Impact of standardization applied to the diagnosis of LT-PEMFC by Fuzzy C-Means clustering
Damien Chanal, Nadia Yousfi Steiner, Didier Chamagne, Marie‐Cécile Péra
Abstract
In the domain of fuel cell systems, Machine Learning diagnostic tools use signal in operation such as temperature, voltage and current or specific experiments such as Electrochemical Impedance Spectroscopy or Current Interruption. One of the most important tasks in Machine Learning is to generate high-quality features from a database. Just as the choice of features to be extracted is important, it is crucial to correctly standardize the data in order to eliminate distortions of the State of Health space that represents all the possible states of the system. Standardization permits to reduce the computation time and to improve the performance of diagnostic algorithms. In this work, a comparison of the main standardization methods is proposed for a diagnostic approach and two databases are used as study cases. A total of seven standardization approaches are compared: (i) Normalizer, (ii) Min-Max scaler, (iii) Max Absolute scaler (iv) Standard scaler, (v) Yeo-Johnson Power transformer, (vi) Uniform Quantile transformer and (vii) Normal Quantile transformer. Uniform quantile transformer provides very good performances for both datasets, making this method very attractive for potential generic use.