Enhancing Fault Diagnosis in Imbalanced Data Using Weighted GRU Architecture

Abstract

The class imbalance, characterized by an unequal distribution between normal and abnormal classes, is predominantly observed in the field of fault diagnosis. Abnormal classes typically represent a minority, leading to a biased learning process favoring the majority class. Therefore, class balancing techniques are essential when applying deep learning approaches to ensure accurate classification of minority fault classes. In this study, we investigate and propose weighted approach for the gated recurrent unit (GRU) algorithm. The proposed weighted approach adjusts all three weights-input, recurrent, and bias inside the GRU architecture. Additionally, the synthetic minority over-sampling (SMOTE) technique with vanilla GRU and long short-term memory (LSTM) as well as the combination of SMOTE and the proposed weighting technique for GRU and LSTM, are compared to the proposed weighting architecture with GRU. We evaluate the effectiveness of this technique using operational data from a real multi-stage flash desalination plant, synthesizing datasets with varying imbalance ratios (4, 9, and 14) for evaluation. Performance metrics such as accuracy is employed for evaluation. Among the models tested, the weighted GRU (WGRU), the proposed model, consistently outperforms others across all variables and imbalance ratios.