Robustness and generalisation study of a new regression-based neural network method for capacitors in power electronics

Abstract

The growing prevalence of modular multilevel converters (MMC) in various electrical engineering applications necessitates effective strategies for ensuring the robustness and reliability of key components such as capacitors. Existing data-driven capacitor lifetime estimation methods for MMC sub-module capacitors that avoid control modification, additional circuits or expert knowledge predominantly rely on machine learning, which is well applicable in theory, but practically require a significant amount of training data, which results in a lack of flexibility necessary for practical applications, where generation of large amounts of training data is difficult to attain. In order to reduce the amount of required labels to the lowest possible, a regression-based neural network approach is proposed that generates estimations for a capacitance degradation indicator on a continuous range. The model is, in contrast to similar neural network-based methods, trained using data with only two distinct labels - a factory condition and a faulty condition, greatly simplifying set-up while avoiding use of expert-knowledge or parameterisation. The novel extrapolation scheme estimates continuous capacitance values not present in the training data by combining regression with additional statistical signal processing. This generalisation performance of this approach is validated using measurement data of an MMC test bench to predict a large range of DC-link capacitance values which were not present in training data. The results show an extended generalising capability at minimal amounts of training data, which can serve as a potential basis for future developments in preventive maintenance and increase operational effectiveness in large-scale converters.

Keywords

• Fault diagnosis
• Modular Multilevel Converter (MMC)
• Predictive maintenance
• Regression
• Machine learning