Research on physics-constrained diffusion models for data augmentation of vibration signals

Abstract

Limited fault samples hinder machinery vibration diagnosis, as traditional augmentation and vanilla diffusion models produce physically inconsistent signals. This work develops a one-dimensional physics-constrained diffusion model for vibration data enhancement using Test Data Management Suite (TDMS) measurements. After overlapping segmentation and zero-mean unit-variance standardisation, an 𝑥􀬴-prediction residual one-dimensional convolutional denoiser is built with explicit noise channels and Wiener baseline correction. Multi-aspect differentiable constraints covering time-domain statistics, second-difference dynamics and discrete Fourier transform (DFT) logarithmic uniform frequency bands compose a compound physical loss. A warm-up weight schedule together with cosine annealing learning rate stabilises model training. Experiments on real industrial datasets show synthesised signals match real data in root mean square (RMS), crest factor and kurtosis, with frequency-domain Kullback-Leibler (KL) divergence of only 0.9111. Waveform, spectrum and time-frequency plots confirm faithful reproduction of original amplitude, impulse and spectral features. The proposed method generates physically consistent vibration samples and provides a potential augmentation source for subsequent fault-diagnosis tasks under limited data conditions. Downstream classifier-based validation will be conducted in future work.