Prior-Guided Dual Stream Time-Frequency Meta Learning Network for Few-Shot Cross-Domain Fault Diagnosis

Drastic fluctuations in fault feature distributions under large-span variable speeds, compounded by target domain label scarcity, severely degrade the generalization performance of diagnostic models. To address these challenges, a Prior-Guided Time-Frequency Dual-Stream Meta-Learning network (PR-TFML) is proposed in this paper. By leveraging the deeply decoupled synergy between mechanistic knowledge and data-driven learning, PR-TFML constructs a multimodal feature extraction architecture composed of complementary time-domain and time-frequency branches. The time-domain branch employs large-receptive-field convolutions to capture nonstationary transient impact characteristics, while the time-frequency branch extracts high-dimensional evolutionary features through an embedded short-time Fourier transform layer. Building upon this, a physical prior mapping branch is introduced to construct a compact hybrid prior by combining macroscopic operating conditions (e.g., rotational speed) with intrinsic statistical and energy-based descriptors (e.g., kurtosis, envelope energy, and wavelet energy). This prior is transformed into domain-invariant semantic anchors, enabling adaptive calibration against distribution shifts. Furthermore, guided by physical priors, an explicit alignment-enhanced meta-optimization strategy couples Model-Agnostic Meta-Learning with Correlation Alignment constraints to facilitate rapid cross-domain adaptation and feature distribution alignment under extreme few-shot conditions. Extensive experiments conducted on the HUST bearing and BJTU WT-Planetary gearbox datasets demonstrate that PR-TFML maintains a high degree of feature consistency and diagnostic robustness under extreme few-shot and large-span variable operating conditions. This method effectively mitigates the feature degradation inherent in purely data-driven models under domain shifts, providing a reliable physics-guided paradigm with strong generalization capability for the cross-domain predictive maintenance of complex rotating machinery.