工业系统设备的故障会造成严重的影响，对于故障诊断与预测技术的研究和开发成为了重要的工业研究方向，是保证设备运行安全稳定的基础。随着工业系统收集、记录系统运行数据的大量积累和近年来大数据领域知识挖掘技术如深度学习等研究的突破和广泛应用，故障诊断与预测方法的研究也进入了新的时代。本文针对工业系统中滚动轴承的故障诊断与预测面临的实际问题展开了相关研究，包括故障特征提取，非均衡数据条件下的故障诊断问题，故障诊断迁移问题以及故障预测问题。本文的主要工作包括：（1）采用深度学习中的卷积神经网络方法作为故障特征提取的工具，以传统信号分析和处理方法中的小波变换和经验模式分解作为数据信号预处理的方法，结合两种方法可以有效提取出故障特征。在提取故障特征的基础上，对一般情况下的故障诊断问题进行了研究和验证。（2）本文提出了一种基于生成对抗网络生成数据样本中少数类样本的模型来解决不均衡的问题。在分析目前非均衡条件下的机器学习分类问题研究方法相关不足的基础上，利用对少数类样本的学习生成新的模拟样本，将新生成样本和原始样本进行组合形成新的训练数据集，在此基础上通过故障诊断分类验证了本文提出方法较现有方法更高的精度和效用。（3）本文在研究迁移学习中源域和目标域样本差异的基础上，提出了一种基于改进的生成对抗网络结构用来进行故障诊断中的样本迁移。利用生成对抗网络的良好学习性能，通过建立生成器产生数据和源数据之间的差异损失函数来学习样本迁移特征，利用生成的样本进行故障诊断的预训练，提出的方法也能够良好的适应新工况下的故障诊断。（4）本文针对信号中存在不同周期性的特点，利用深度学习理论中的循环神经网络和长短期记忆网络模型实现了故障信号的预测，提出的预测方法精度较现有方法更高。精确的信号预测也保证了进一步的在数据出现故障趋势时及时提出故障告警的能力。

The failure of industrial system or equipment will cause serious impacts. The research and development of fault diagnosis and prognosis technology has become an important industrial technology and the basis for ensuring the safety and stability of equipment operation. With the accumulation of industrial system data collected, the breakthrough and extensive application of big data knowledge mining technologies such as deep learning in recent years, the research of fault diagnosis and prognosis technology has entered a new era.This paper has carried out related research on the practical problems faced by fault diagnosis and prognosis of rolling bearings in industrial systems, including fault feature extraction, fault diagnosis under unbalanced data conditions, fault diagnosis transfer learning problem and fault prognosis problem. The main work of this paper includes:(1) Convolutional neural network method in deep learning is used as the tool for fault feature extraction. The wavelet transform and empirical mode decomposition in traditional signal analysis and processing methods are used as data signal preprocessing methods. Based on the extraction of fault characteristics, the fault diagnosis problem under normal circumstances is studied and verified.(2) This paper proposes a generating model, generative adversarial network, to generate samples in minor classes and solve the problem of imbalanced data. On the basis of analyzing the shortcomings of current research methods of machine learning classification problems under imbalanced data conditions, the new simulation samples are generated by learning from the minority samples, and fault diagnosis classification is performed using the combination of the generated samples and the original samples. The proposed method shows better accuracy and performance than existing methods.(3) Based on the study of the differences between the source and target domains in transfer learning, this paper proposes an improved generative adversarial network structure for sample adaption in fault diagnosis. By using the excellent learning performance of generative adversarial network, the sample feature is learned by establishing the loss function between the generator and the source data, and the generated samples are used for the pre-training of the fault diagnosis. The proposed method can also adapt well to fault diagnosis problem under new working conditions.(4) This paper aims to predict the fault signal by using the recurrent neural network and long short term memory network model in the deep learning theory. The accuracy of the proposed prognosis method is higher than the existing methods. Accurate signal prognosis also ensures further ability for early fault detection while data is on a failure trend.