Abstract: To address the domain shift problem caused by variations in operating conditions between training and test data for anomalous sound detection in rotating motors—which leads to significant performance degradation under unknown working conditions—this paper proposes a motor anomaly sound detection method that integrates machine condition-aware augmentation and instance whitening. At the data level, we design a machine condition-aware augmentation strategy that recombines the magnitude and phase spectra of samples sharing the same fault type but differing in operating conditions in the frequency domain, thereby generating physically interpretable augmented samples. At the feature level, an instance whitening layer is introduced to eliminate condition-induced feature distribution discrepancies via mean-centering and decorrelation, achieving cross-condition feature alignment. Building upon this, we construct an end-to-end multi-task learning framework to jointly optimize fault classification and operating condition recognition. Experimental results on a test set comprising multiple previously unseen speed–load combinations show that the proposed method achieves 99.67% accuracy and a macro F1-score of 99.67%, outperforming the baseline model by over 4 percentage points and significantly surpassing traditional data augmentation methods such as Mixup.