Quantitative assessments on advanced data synthesis strategies for enhancing imbalanced AHU fault diagnosis performance

Abstract

The accurate and reliable fault diagnosis of air handling units (AHUs) has profound impacts on building energy efficiency and indoor thermal comforts. Data-driven fault diagnosis methods have gained increasing popularity considering the wide availabilities of operational data and advances in data analytics. In practice, the data-driven fault diagnosis performance can be severely degraded by the imbalanced nature of building operational data, i.e., the data samples of faulty operations are much smaller than that of normal operations. This study serves as a comprehensive study to investigate the potential of different data synthesis techniques in enhancing the performance of imbalanced AHU fault diagnosis. A variety of data synthesis strategies, ranging from conventional random sampling-based to advanced variational autoencoder-based techniques, have been developed for synthetic data generation. Data experiments have been designed to quantitatively evaluate the value of data synthesis in data scenarios considering various data amounts and imbalanced ratios. The research results indicate that synthetic data can significantly enhance the performance of imbalanced AHU fault diagnosis by up to 8.94%. Optimal data synthesis strategies have been identified in different data scenarios. The research outcomes are helpful for the development of reliable data-driven tools for practical tasks in building energy management.