An Investigation Into the Behavior of Intelligent Fault Diagnostic Models Under Imbalanced Data

In solving the data imbalance problem, most of the existing studies ignored the effect of the number of samples on the diagnostic performance of intelligent fault diagnostic models. When the number of minority samples is small, the data imbalance situation becomes a compound problem of data imbalance and small sample, making it challenging to develop effective methods and identify core issues. This article aims to investigate the effects of the compound problem on intelligent diagnostic models by using the number of samples and the number of majority classes as indicators to assess the deterioration behavior of diagnostic models. Multiple datasets are systematically studied to explore three categories of influences: the influence of the number of minority samples on the model’s learning ability, the influence of the relative size of the numbers of majority samples and minority samples on the model’s convergence speed, and the influence of the relative size of the numbers of majority samples and minority samples on the model’s convergence correctness. The exacerbation of imbalance by an increase in the number of majority classes further deepens the influence of imbalance on the convergence of the model. Furthermore, this article explores the degradation mechanism of intelligent diagnostic models. Finally, a semiquantitative deterioration process is proposed to guide future studies on imbalanced learning in experimental design and method evaluation.