DirectHF-DA integrates a lightweight network and multi-resolution STFT-RGB for edge-deployable cross-domain bearing fault diagnosis

Bearing fault diagnosis under varying operating conditions is challenged by domain distribution shifts and the high computational cost of deep models on resource-constrained edge devices. To address these coupled challenges, this study proposes DirectHF-DA, a directional high-frequency domain-adaptive network for lightweight cross-domain bearing fault diagnosis. First, a compact multi-resolution short-time Fourier transform representation (MR-STFT-RGB) is constructed by deterministically organizing short-, medium-, and long-window STFT spectrograms into a standard three-channel tensor, enabling complementary transient, modulation, and harmonic information to be jointly processed by a single lightweight backbone. Next, a lightweight feature extractor, DirectHF-Net, is developed by integrating directional depthwise separable convolution (DDSConv), directional residual refinement and transition blocks (DRRB/DRTB), and a high-frequency enhancement block (HFEB), thereby preserving fault-sensitive directional textures and high-frequency impact features with low computational complexity. Joint maximum mean discrepancy (JMMD) is further introduced to align the joint distributions of the source and target domains, enabling effective adaptation to unlabeled target operating conditions. Experiments on 12 cross-domain transfer tasks using the CWRU and HUST bearing datasets achieve average diagnostic accuracies of \(99.70\pm 0.05\%\) and \(92.31\pm 1.02\%\) , respectively. DirectHF-DA contains only 1.6 M parameters, requires 0.22 GFLOPs, and has a model size of 6.22 MB. When deployed on an NVIDIA Jetson Orin Nano Super, the proposed model achieves real-time inference throughputs of 60.54 FPS and 61.71 FPS on the CWRU and HUST tasks, respectively. These results demonstrate that DirectHF-DA achieves a favorable balance among cross-domain diagnostic accuracy, computational efficiency, and real-time edge inference capability.