Physics-Informed Fourier Prior-Guided DDPM: Data Generation for Enhancing Fault Diagnosis

A physics-guided diffusion-GAN network with VMD feature transfer for imbalanced fault diagnosis

Sample imbalance across operating conditions severely constrains data-driven bearing fault detection performance. This challenge is more acute due to the complete absence of fault samples under target operating conditions. To address this limitation, this paper proposes a framework integrating variational mode decomposition (VMD), nonlinear dynamic modeling, and physics guidance. First, a two-degree-of-freedom nonlinear system is established to simulate physically realistic dynamic behaviors. Second, VMD decouples fault feature components from source-domain fault signals. These components are superimposed with target-domain normal signals and dynamic responses to form hybrid samples. Then, a diffusion-based wasserstein generative adversarial network architecture is adopted. Simultaneously, auxiliary classifiers for operating conditions and fault types are incorporated as physics-regularized losses to enforce physical consistency. The fault feature decoupling component based on VMD and the generation framework jointly mitigate the domain gap between dynamically generated response signals and measured signals. Experimental results indicate generated samples from the proposed method achieve superior scores across multiple evaluation metrics. And classification networks trained on generated samples achieve diagnostic accuracy exceeding 90% in multiple cross-condition tasks, which confirms the superiority of the developed model.

Physics-data coupled transfer learning for rolling bearing fault diagnosis under variable conditions

A Physics-Constrained Spectral Generative Adversarial Network for Rolling Bearing Fault Diagnosis

Accurate bearing fault diagnosis is often hindered by sample scarcity and class imbalance. This paper proposes the Physics-Constrained Spectral Generative Adversarial Network (PCS-GAN), which embeds prior mechanical knowledge into adversarial training to enhance physical interpretability. The PCS-GAN model employs a time-frequency dual-discriminator structure to enforce statistical and spectral realism via physics-informed loss terms. To ensure stability, it integrates WGAN-GP with a curriculum-based loss scheduling strategy. Evaluations on CWRU and MFPT datasets confirm that the PCS-GAN model achieves high spectral integrity while reducing peak GPU memory consumption by 28% compared to Transformer-based architectures. Furthermore, the model exhibits robust noise resilience and elevates macro-F1 scores to over 0.94 in extreme imbalance scenarios. These results demonstrate that the PCS-GAN model provides a computationally efficient and reliable solution for fault diagnosis in data-scarce industrial settings.

An Explainable Physics-Informed Dual-Input Deep Learning Framework for Partial Discharge Detection in Noisy Environments

Partial discharge (PD) detection is essential for assessing insulation health in high-voltage equipment but is hindered by noise and non-stationary signals. This paper proposes an explainable, physics-informed hybrid framework for PD detection under synthetic noisy conditions. The method integrates wavelet-based denoising with time-frequency feature extraction using short-time Fourier transform (STFT) and continuous wavelet transform (CWT). A dual-input deep learning architecture combining convolutional layers, bidirectional long short-term memory (BiLSTM), and multi-head attention captures spatial-temporal dependencies from raw and engineered features. Explainable artificial intelligence (XAI) techniques, including attention visualization and Integrated Gradients, are incorporated to improve interpretability by identifying signal regions that drive model decisions. On a controlled synthetic dataset, the proposed framework achieves 89.38% accuracy, demonstrating promising discrimination between PD and noise signals. These results establish a proof of concept; validation on real-world PD datasets remains an important direction for future work.

Physics informed spectrogram normalization and domain adaptation for intermittent fault diagnosis

Intermittent faults are characterized by short duration and random occurrence. They are often missed by machine learning models trained exclusively on permanent fault data. This work proposes a novel physics-informed normalization of spectrograms and unsupervised domain adaptation for efficient diagnosis of intermittent faults using permanent fault data. Domain Adversarial Neural Network (DANN), Wasserstein Distance Guided Representation Learning (WDGRL) and Conditional Domain Adaptation Network (CDAN) are used as the domain adaptation techniques. The approach is validated using a newly developed experimental setup for generating intermittent inter-turn short fault data in a synchronous generator. The results validate the effectiveness of the proposed physics-informed pre-processing and domain adaptation for diagnosing intermittent faults using permanent fault data. Further the results open up possibility of application of various types of domain adaptation techniques to solve many practical problems in the domain of intermittent fault diagnosis

A noisy small-sample fault diagnosis method based on dynamic weighted contrastive learning using multisource signals

Industrial equipment fault diagnosis faces significant challenges including severe noise interference, scarcity of annotated data, and inadequate model generalization in small-sample scenarios. Traditional transfer learning methods are limited by known fault information and categories, hindering accurate identification of all faults in target domains and compromising diagnostic performance. To address these issues, this paper proposes a Multi-source Dynamic Contrastive Learning (M-DCL) approach for noisy small-sample fault diagnosis. The framework first designs a heterogeneous temporal encoder that adapts to temporal characteristic differences between current and vibration signals using large/small convolutional kernels respectively, while incorporating channel attention to enhance critical fault features. It then constructs a time-frequency dual-domain physical fusion module that integrates spatiotemporal attention and frequency-domain gating strategies to achieve noise separation and effective feature enhancement. Finally, a Dynamic Contrastive Learning (DCL) is introduced, leveraging noise-aware dynamic weighting and multi-scale feature regularization to boost diagnostic capability against small samples and noise. Experimental results demonstrate the method's effectiveness in bearing fault diagnosis under varying operating conditions.