Physics-informed domain adaptation for synthetic-to-real bearing fault diagnosis with healthy-only target data

Reliable bearing fault diagnosis is difficult when real fault data cannot be collected without risking machine damage. We study a synthetic-to-real setting in which training uses labelled synthetic faults and recordings known to represent healthy operation of the target machine, while no target-fault sample is available for training, adaptation, or model selection. We propose Healthy-Referenced Physics-informed Diagnosis (HRPD), a framework that subtracts each domain’s mean healthy envelope-order spectrum, retains a complementary log-magnitude channel, and combines a temporal convolutional encoder with a physics branch built from characteristic-order harmonics. A cosine-prototype classifier and healthy-class-restricted domain alignment use the available target information without observing target faults. Under a common-data protocol on four bearing benchmarks, HRPD ranks first on CWRU, MFPT, and Paderborn (balanced accuracy 0.643, 0.979, and 0.545) and second on XJTU-SY (0.638). The results also expose clear limitations: XJTU-SY inner-race damage is dominated by a shaft-synchronous component absent from the geometry-based source, and Paderborn operational wear remains difficult for every evaluated framework. These findings support healthy-only target data as a practical calibration resource while showing that transfer remains constrained by synthetic-source fidelity.