Bearing fault signal generation method by fusion of physical constraints and multimodal features

Abstract To address the limited generalization ability of diagnostic models resulting from the scarcity of real fault samples in bearing fault diagnosis, this paper proposes a novel physics-guided fault signal generation framework that integrates physical constraints with multimodal feature modeling. Specifically, we construct a generative adversarial network (GAN), termed time-frequency physics-guided multimodal GAN, which is guided by domain-specific physical priors. Key physical mechanisms—including the decaying impact response behavior of defective bearings and rotation-speed-induced modulation effects—are explicitly modeled and incorporated as differentiable constraints to enhance the physical plausibility of the generated signals. The generator adopts a dual-path architecture, in which the time-domain path employs a stacked Transformer encoder to capture long-range dependencies, while the frequency-domain path integrates the Morlet wavelet transform with autocorrelation analysis to extract periodic modulation features. This design enables effective decoupling and reconstruction of time-frequency representations. In contrast, the discriminator incorporates a multiscale dilated convolution module to evaluate signal authenticity across both time and frequency domains, thereby enhancing its discriminative robustness. To comprehensively assess the fidelity of the generated signals, we propose a multi-dimensional assessment framework that combines dynamic time warping (DTW), envelope spectral deviation, and statistical similarity metrics. Experimental results obtained from two publicly available bearing fault datasets—Dataset A and Dataset B—demonstrate that the generated signals exhibit high fidelity in both time and frequency domains, achieving a DTW distance of 7.4662, an envelope spectrum MSE of 0.0829 and a high Coverage of 0.70, confirming that the proposed method effectively maintains sample diversity while avoiding mode collapse. Moreover, when the generated signals are used to augment training data, downstream diagnostic models achieve significantly enhanced performance, with average F 1-scores indicating that the generated samples effectively capture the discriminative characteristics of various fault types.