Data-driven deep learning technology has gained many achievements in the field of motor fault diagnosis and prognostics. However, the application objects of those previous studies are commonly limited to the faulty data sharing the similar distribution under unvarying stable working condition. Unfortunately, this limitation is nearly invalid in the real-world scenario, where the working condition is complicated and changes invariably, resulting in the unfavourable situation that the deep representation learning methods of the previous studies always fail in extracting the effective representations for fault diagnosis in real applications. To tackle this issue, inspired by f-divergence estimation, this work takes a different route and proposes an unsupervised deep representation learning approach, named Deep Mutual Information Maximization (DMIM), using variational divergence estimation approach to maximize mutual information (MI) between the input and output of a deep neural network. Meanwhile the representation distribution is automatically tuned by matching to a prior distribution with the same philosophy of Variational Autoencoder. Opposite to previous works which learn representations basically with supervised feedback regulation or unsupervised reconstruction, the proposed unsupervised MI maximization framework aims to make representational characteristics like independence play a bigger role to capture the most unique representations. To verify the effectiveness of our proposal, faulty motor data from the motor tests under European driving cycle for simulating the real working scenario, are collected for validation. It turns out that DMIM outperforms many popular unsupervised and fully-supervised learning methods. It opens new avenues for unsupervised learning of representations for motor fault diagnosis.

数据驱动的深度学习技术在电机故障诊断和预测方面取得了许多成就。然而，这些先前研究的应用对象通常限于在稳定工作条件不变的情况下共享相似分布的故障数据。不幸的是，这种限制在实际情况下几乎是无效的，在这种情况下，工作条件复杂且不断变化，从而导致不利的情况是，先前研究的深度表示学习方法始终无法提取用于故障诊断的有效表示。实际应用。解决这个问题，受到f的启发-散度估计，这项工作采用了不同的途径，并提出了一种无监督的深度表示学习方法，称为深度互信息最大化（DMIM），使用变分散度估计方法来最大化深度神经网络的输入和输出之间的互信息（MI） 。同时，通过匹配具有可变自动编码器相同原理的先前分布，可以自动调整表示形式的分布。与以前的工作基本上是通过有监督的反馈调节或无监督的重构来学习表示的工作相反，提出的无监督的MI最大化框架旨在使诸如独立性的表示特征在捕获最独特的表示方面发挥更大的作用。为了验证我们提案的有效性，收集来自欧洲行驶周期下的汽车测试中用于模拟实际工作场景的错误电机数据进行验证。事实证明，DMIM优于许多流行的无监督和完全监督的学习方法。它为无监督学习电动机故障诊断表示法开辟了新途径。