Data enhancement plays a vital role in bearing fault diagnosis, as diverse and comprehensive datasets contribute to more effective model training. This study presents a novel approach that enhances data by generating fused features from both vibration and current signals. Leveraging a multi-attention mechanism, the method extracts salient characteristics from each signal type and integrates them through a generative adversarial network (GAN) to synthesize new training data. These enhanced samples are then used to classify motor fault conditions. Experimental results demonstrate that the proposed technique significantly enriches the sample diversity and quality, thereby boosting diagnostic accuracy. Moreover, the approach exhibits strong adaptability and robustness when evaluated on publicly available datasets.