Triplet attention-enhanced residual tree-inspired decision network: A hierarchical fault diagnosis model for unbalanced bearing datasets

Abstract

In fault classification tasks, deep neural networks (DNNs) have remarkable recognition performance. Nevertheless, the classification decision processes of DNNs lack hierarchical logical reasoning abilities and their diagnostic performance significantly deteriorates when dealing with imbalanced bearing fault datasets. To further address this issue, a novel model, termed the triplet attention-enhanced residual tree-inspired decision network (TARTDN) which is not a simple combination of the DNN and decision tree model, is developed to diagnose unbalanced bearing faults and provides a rational decision-making and reasoning process in this study. First, a triplet attention-enhanced residual network (TARN) is designed as the backbone network to capture key information more accurately. Second, a novel tree-inspired decision layer (TDL) is construed to infer and decide bearing data categories. Subsequently, the probability distribution values obtained by pre-training TARN are flowed into the TDL as thresholds for seed and leaf nodes. The parameters of the TARN are continuously updated with the node thresholds of the TDL, resulting in an integrated TARTDN model that combines high-quality feature extraction and inferable decision-making. In the end, the trained TARTDN progressively determines the fault types and severity levels in unbalanced bearing fault datasets. The developed model tested on two bearing fault datasets with three unbalanced ratios, has consistently achieved recognition rates exceeding 97.5%. The proposed approach has been validated through ablation experiments and comparisons with other advanced methods to exhibit higher recognition rates, superior hierarchical classification reasoning, and more stable generalization capabilities on unbalanced bearing datasets.