针对背景噪声和不同工况下单一尺度模型特征提取能力有限，引起滚动轴承寿命状态识别率下降的问题，提出基于多尺度改进深度残差收缩网络（Multi Scale-Improved Deep Residual Shrinkage Network, MS-IDRSN）的滚动轴承寿命状态识别方法。首先，在深度残差收缩单元中引入改进的阈值函数提升网络的抗噪性能，并减小降噪过程中的寿命状态信息丢失；然后，采用不同卷积核尺寸的深度残差收缩单元构建特征提取器，避免单一尺度感受野引起在不同工况下的特征提取能力下降。最后，利用最大均值差异损失适配源域与目标域特征的特征分布，通过Softmax分类器实现在不同工况的轴承寿命状态识别。在PRONOSTIA数据集和自测轴承数据集上验证了所提方法的可行性和有效性，结果表明所提方法具有较好的抗噪性能和泛化性能，在考虑背景噪声和不同工况条件下相比对比方法的寿命状态识别率提升7.6%至46.5%。

A multi-scale improved deep residual shrinkage network (MS-IDRSN) based rolling bearing life state identification method is proposed to address the issue of reduced domain adaptation performance caused by background noise interference in spatial rolling bearing data under different operating conditions, resulting in a decrease in accuracy of life state recognition. Firstly, an improved threshold function is introduced into the contraction layer of the deep residual contraction network to enhance the network's noise resistance and reduce the loss of life state information during the denoising process; Then, using convolutional kernels of different sizes to obtain deep fusion denoising features, the generalization ability of the features in noisy backgrounds is improved. Finally, by minimizing the maximum mean difference and adapting the feature distribution of the source domain and target domain features, the recognition of bearing life status under different operating conditions is achieved. The applicability and effectiveness of MS-IDRSN were validated on the PRONOSTIA dataset and self-test bearing dataset, and the results showed that the proposed method has good noise resistance and generalization performance. Under high noise conditions, compared with the comparison method, the accuracy of life state recognition improved by 7.6% to 46.5%.