Robust rotating machinery diagnosis from Single-Source heterogeneous signals under imbalanced and noisy conditions

In practical industrial scenarios, rotating machinery fault diagnosis is often constrained by single-source measurements, where heterogeneous representations derived from the same sensor exhibit severe misalignment, strong noise interference, and highly imbalanced fault distributions. These coupled challenges significantly degrade the robustness and generalization of existing diagnostic models, which are typically designed for either balanced datasets or well-aligned multimodal inputs. To address this issue, this study proposes a robust diagnostic framework that integrates multiscale feature extraction, selective multimodal fusion, and data-level imbalance mitigation for single-source heterogeneous signals. Specifically, heterogeneous representations are constructed from raw vibration measurements in the frequency domain and time–frequency domain, enabling complementary characterization of fault signatures. A multiscale convolutional network is employed to capture non-stationary fault patterns across different temporal and spectral resolutions. To achieve effective fusion of heterogeneous features under noisy conditions, a multi-head self-attentive Mamba module is introduced, which combines attention-driven correlation modelling with selective state-space mechanisms (S6) to dynamically emphasize informative features while suppressing noise and redundancy. Furthermore, a Wasserstein generative adversarial network with gradient penalty (WGAN-GP) is utilized to alleviate severe class imbalance by generating high-fidelity synthetic fault samples. Extensive experiments conducted under varying signal-to-noise ratios and imbalanced conditions demonstrate that the proposed framework consistently outperforms state-of-the-art methods in terms of diagnostic accuracy and robustness. In particular, it maintains stable performance under low signal-to-noise environments, highlighting its suitability for real-world rotating machinery diagnosis where data heterogeneity, imbalance, and noise coexist.