Existing intelligent fault diagnosis methods based on two-dimensional time–frequency representation follow a single-view paradigm, which relies on adjusting convolutional or attention mechanisms for feature extraction and classification. However, this paradigm exhibits multiview perception deficiencies, including noise-masked fault impulses in the input view, corrupted timefrequency features caused by fixed padding and nonadaptive convolution in the operator view, and a lack of targeted guidance for difficult and easy samples in the training view. To address these deficiencies, we propose a multiview perception-driven network (MVPNet) for intelligent fault diagnosis. First, a noise-perception multiscale dilated gating modulator is designed to enhance noise perception in the input view. The modulator injects oriented perturbations guided by an adaptive noise-aware map and modulates them through multiscale dilated gating, forcing the network to distinguish fault impulses from noise. Second, to alleviate degraded edge perception in the operator view, a boundary-adaptive padding module (BAPM) and an edge-adaptive gradient-sensitive mask are designed to construct a boundary-edge perception module. This module preserves the topological integrity of feature boundaries through adaptive padding and enhances edge responses via gradient-sensitive masking. Finally, to overcome the absence of difficulty perception in the training view, a difficulty-perception cluster-guided inverse distillation (DCID) strategy is proposed. In this strategy, a difficulty memory mechanism is constructed, upon which intraclass multistate soft labels are generated through online clustering and inverse neighbor sampling to guide the learning of difficult and easy samples. Comprehensive experiments on the Case Western Reserve University and the self-built Anhui University of Science and Technology platforms demonstrate that MVPNet markedly surpasses existing mainstream methods.
Zhu et al. (Sat,) studied this question.