Underwater motors operate in complex marine environments for extended periods, making them susceptible to various external factors and leading to a high probability of failure. Therefore, achieving real-time and accurate fault diagnosis for underwater motors is crucial. Recently, Transformer has demonstrated outstanding performance in the field of fault diagnosis due to self-attention, especially in capturing the global contextual information of fault features, which has achieved remarkable results. However, the high computational complexity of the self-attention limits its deployment on resource-constrained devices, posing a challenge for real-time fault diagnosis. To address this challenge, this paper proposes an improved lightweight model EADDFormer. EADDFormer introduces an efficient additive attention based on the traditional multi-head self-attention, utilizing element-wise multiplication to simplify attention calculations, thereby reducing the original quadratic complexity to linear. This significantly enhances computational efficiency while maintaining diagnostic accuracy. Additionally, a separable multi-scale dilated convolution structure is designed. By adjusting the dilation rate of the convolution kernel, it can effectively capture features at different scales. This approach enhances the recognition of complex fault modes by expanding the receptive field. Experimental results on the underwater motor fault dataset demonstrate that the EADDFormer model not only has significant advantages in computational complexity and parameter efficiency but also exhibits strong condition transferability and generalization capabilities. These characteristics make it highly suitable for real-time fault diagnosis in resource-constrained environments.