Physics-Guided Invariant Causal Representation Learning for Blind Domain Generalization in Bearing Fault Diagnosis

Abstract

Under conditions of limited samples, severe noise, and non-stationary operations, data-driven fault diagnosis often fails in blind domain generalization (BDG) by fitting spurious environmental correlations. To reduce the limitations of purely statistical fitting, this article proposes a Physics-guided Invariant Causal Representation Network (PICR-Net) that structurally fuses signal processing principles, bearing kinematic priors, and causal inference. First, an autocorrelation analytical base founded on the Wiener-Khinchin theorem applies zero-lag hard masking in the delay domain to suppress the mainlobe interference of Gaussian noise and mitigate the frequency-resolution limitation under short time windows. Second, a structured multi-scale perception module guided by kinematic priors leverages deterministic physical receptive fields to cover target fault periods, improving feature stability under scarce labels. Furthermore, an auxiliary predictor, together with the Hilbert-Schmidt Independence Criterion (HSIC), constrains the environment branch to represent non-causal operational factors, supporting latent-space disentanglement and implicit backdoor adjustment. Under the prior-free blind-domain protocol with 1% source labels and target-domain $\text {SNR}=-{4}$ dB, PICR-Net achieves average accuracies of 88.80% on six bidirectional OU-Bearing tasks and 83.28% on an external Mendeley Data variable-speed bearing subset. With only 8.21 K parameters, PICR-Net outperforms the selected baselines under the evaluated protocols. Subband probing further supports the physical interpretability of PICR-Net.