Unsupervised feature learning (UFL) has gained a growing attention in machinery fault diagnosis because of its great advancement in adaptively learning useful features over traditionally designing hand-crafted features. Since there is no available prior knowledge about the health labels, sparsity becomes the key to promote feature discrimination. In this paper, a new UFL method called sparsity optimized auto-encoder (SOAE) is proposed for machinery fault diagnosis, where SOAE is responsible for learning fault features so as to encode part-based representations of the measured signals. A sparsity optimized penalty term is integrated into the learning procedure of SOAE, which aims at optimizing two crucial sparsity properties about the hidden-layer feature distribution, i.e., lifetime sparsity and population sparsity. At the same time, a theoretical analysis of two sparsity properties is provided. Moreover, a stochastic mapping operation is employed to partly corrupt the input data of SOAE, for the purpose of improving anti-noise capability of the features. Based on SOAE, an intelligent diagnosis method is developed and it is experimentally evaluated using a sun gear and two rolling bearing datasets. The results reveal that SOAE can extract discriminative and robust features from noisy signals, and acquires significantly superior diagnosis performances than the existing state-of-the-art UFL methods.

无监督特征学习（UFL）在机械故障诊断中受到越来越多的关注，因为它在自适应学习有用特征方面比传统设计的手工特征取得了巨大进步。由于没有关于健康标签的先验知识，稀疏性成为促进特征区分的关键。本文提出了一种新的 UFL 方法，称为稀疏优化自动编码器（SOAE），用于机械故障诊断，其中 SOAE 负责学习故障特征，以便对测量信号的基于部分的表示进行编码。SOAE 的学习过程中集成了稀疏性优化惩罚项，旨在优化隐藏层特征分布的两个关键稀疏性，即寿命稀疏性和群体稀疏性。同时，提供了两个稀疏性性质的理论分析。此外，采用随机映射操作来部分破坏SOAE的输入数据，以提高特征的抗噪声能力。基于 SOAE，开发了一种智能诊断方法，并使用太阳轮和两个滚动轴承数据集进行了实验评估。结果表明，SOAE 可以从噪声信号中提取有辨别力的鲁棒特征，并获得比现有最先进的 UFL 方法显着优越的诊断性能。