<p>The class imbalance problem significantly impacts the performance of classification models due to differences in class frequencies. The belief rule base (BRB) is an interpretable classification model, where expert knowledge is incorporated, making it particularly effective in small-sample scenarios for alleviating the adverse effects of class imbalance, though its reasoning mechanism is often associated with high computational complexity. To leverage BRB for small-sample imbalanced multi-classification tasks, a progressive belief rule base (P-BRB) designed for parallel computing is proposed. The model performs classification progressively through a three-layer structure. First, the equilibrium layer applies multi-round, parallelizable ensemble undersampling combined with consistency constraints to reduce redundant samples while balancing class distributions and preserving key information. Second, the fusion layer employs the evidential reasoning (ER) rule to integrate the outputs of undersampled sub-models, providing more reliable coarse classification results. Finally, the binary classification layer further refines the coarse outputs by distinguishing adjacent classes. During training, optimization of undersampled sub-models can be executed in parallel; during inference and fusion, parallel reduction is adopted to significantly enhance computational efficiency. The proposed model is evaluated on five imbalanced benchmark datasets and two fault diagnosis tasks. Experimental results demonstrate that P-BRB excels in small-sample imbalanced classification, achieving superior minority class recognition and overall performance compared with baseline methods; moreover, its multi-round undersampling and fusion process exhibit strong parallelization potential, indicating feasibility for further efficiency gains in HPC environments.</p>