Acoustic-based diagnosis (ABD) is a promising method for machinery fault detection due to its ability of non-contact measurement by air-couple. However, most of the ABD methods are constrained by strong and highly non-stationary background noise interference in practical industrial application. To address the shortcoming, a novel anti-noise ABD method based on recursive attention mechanism (RAM) is proposed in this paper. In proposed method, a multi-stage attention module (MSAM) is firstly designed as fundament of RAM to automatically estimate the noise interference probability within time–frequency (T-F) unit of each signal sample. Simultaneously, a recursive learning strategy is introduced to construct RAM by reusing the MSAM for multiple blocks to gradually refine the estimated probability and adaptively simulated noise interference in diagnosis model for enhancing anti-noise diagnosis ability. Then, based on RAM, a domain adaption method is established to endow the model with good cross-domain ability for further improving the anti-noise performance of the diagnosis model. The experiment result in both real-industrial noise condition and stimulated noise conditions with different SNRs indicate that the proposed method has stronger robustness and better generalization ability than other popular methods in dealing with gear fault diagnosis task under noise condition.

基于声学的诊断（ABD）是一种很有前途的机械故障检测方法，因为它能够通过空气电偶进行非接触式测量。然而，大多数ABD方法在实际工业应用中都受到强烈且高度非平稳的背景噪声干扰的限制。针对这一不足，本文提出了一种基于递归注意机制（RAM）的抗噪声ABD方法。在所提出的方法中，首先设计了多阶段注意模块（MSAM）作为 RAM 的基础，以自动估计每个信号样本的时频（TF）单元内的噪声干扰概率。同时，引入递归学习策略构建RAM，通过对多个块重用MSM，逐步细化估计概率，自适应模拟诊断模型中的噪声干扰，增强抗噪声诊断能力。然后，基于RAM，建立域适应方法，赋予模型良好的跨域能力，进一步提高诊断模型的抗噪性能。在实际工业噪声条件和不同信噪比的受激噪声条件下的实验结果表明，在处理噪声条件下的齿轮故障诊断任务时，所提出的方法比其他流行的方法具有更强的鲁棒性和更好的泛化能力。建立域自适应方法，赋予模型良好的跨域能力，进一步提高诊断模型的抗噪性能。在实际工业噪声条件和不同信噪比的受激噪声条件下的实验结果表明，在处理噪声条件下的齿轮故障诊断任务时，所提出的方法比其他流行的方法具有更强的鲁棒性和更好的泛化能力。建立域自适应方法，赋予模型良好的跨域能力，进一步提高诊断模型的抗噪性能。在实际工业噪声条件和不同信噪比的受激噪声条件下的实验结果表明，在处理噪声条件下的齿轮故障诊断任务时，所提出的方法比其他流行的方法具有更强的鲁棒性和更好的泛化能力。