Vibration signals from gearboxes at different rotational speeds exhibit pronounced distribution shifts, making a diagnostic model trained under one operating condition difficult to transfer directly to another. This paper proposes a DWT-BiMamba unsupervised domain adaptation method for gearbox fault diagnosis. A multilevel learnable analysis filter bank initialized with Daubechies-4 (db4) decomposes three-axis vibration signals into multiscale band components, and directional attention adaptively fuses the three responses within each band. A short-time Fourier transform (STFT) branch characterizes the temporal evolution of spectral energy and complements the multiscale discrete wavelet transform (DWT) representation. The two representations are then fed into a bidirectional Mamba (BiMamba) encoder to jointly model local impacts and long-range temporal dependencies. By combining source-domain supervision with global distribution alignment and target-domain class structure constraints, the model learns discriminative and domain-invariant shared features without using target labels. On the Huazhong University of Science and Technology gearbox dataset, the proposed method achieves an overall mean accuracy of 95.01% and a mean macro-F1 of 95.03%, demonstrating competitive overall performance while indicating that fixed-to-variable transfer remains challenging.

gearbox fault diagnosis,unsupervised domain adaptation,variable-speed conditions,discrete wavelet transform (DWT),bidirectional Mamba