Rotating components in aviation equipment are subject to pronounced distribution shifts under different rotational speeds, loads, and installation conditions. As a result, conventional deep diagnostic models often suffer from performance degradation when deployed beyond the training operating condition. To address the practical constraint that target-domain samples are unavailable in advance and only one source operating condition can be used for training, this paper proposes a bearing fault diagnosis method that integrates frequency-domain statistical priors with class-conditional representation alignment. The one-dimensional vibration signal is transformed into both a frequency-domain amplitude spectrum and a two-dimensional time-frequency representation. The former is used to construct low-dimensional physical statistical descriptors, including center frequency, standard-deviation frequency, root-mean-square frequency, and kurtosis frequency, whereas the latter is fed into a convolutional network to extract deep discriminative representations. During training, a class-conditional prior-deep representation alignment constraint is introduced to pull the deep features of samples from the same class toward their corresponding frequency-domain prior representations while suppressing interference from priors of different classes. The two types of features are then concatenated and fed into a classifier for fault identification. Cross-condition experiments on a bearing dataset demonstrate that the proposed method improves diagnostic stability under unseen operating conditions without using target-domain training samples, thereby providing an interpretable modeling scheme for health monitoring of aviation equipment in high-cost and limited-source scenarios.