Abstract: To address the challenges of bearing fault diagnosis for port equipment under strong noise and multi-motor interference, particularly the lack of real-world industrial acoustic datasets and insufficient algorithmic noise robustness, this paper constructs an acoustic-based bearing fault dataset(HB dataset) tailored for complex port environments. It was collected under multi-motor operating conditions and covers seven common fault types along with various industrial interferences. Based on this dataset, an efficient multi-scale attention threshold residual shrinkage context-aware masking network(EMAtrc-SCAM) is proposed. Its front-end incorporates an efficient multi-scale attention and adaptive soft threshold shrinkage(EMAtrs) module to enhance noise robustness, while its back-end employs a depthwise separable and grouped context-aware masking(SCAM) module to achieve a lightweight design. Furthermore, an adaptive boundary compensation loss function(ABC Loss) is designed, which dynamically adjusts classification boundaries through a dual-penalty mechanism to improve the discrimination of hard samples. Experimental results demonstrate that the proposed method achieves a fault identification accuracy of 99.41% and an equal error rate(EER) of 0.58% on the HB dataset, while reducing model parameters by 11.6% compared to the baseline. It maintains excellent diagnostic performance under strong noise interference.