机械故障诊断技术的进步促进了生产力的发展，提高了设备的安全性与可靠性。滚动轴承作为工业常用机械零部件，其运行状态关乎着机械设备整机安全，对滚动轴承开展故障诊断研究具有重要意义。传统滚动轴承故障诊断方法在处理复杂信号时，难以精准获取故障信息。深度学习等人工智能方法具有强大的学习能力，因此滚动轴承故障诊断趋向于基于深度学习的智能诊断。本文运用深度学习开展滚动轴承故障诊断研究，从单一故障问题入手，到复合故障，具体如下。

（1）滚动轴承振动图像样本构造方法研究。针对滚动轴承故障诊断研究，研究了4种振动信号转换图像的方法，包括：经验模态分量排列图(Intrinsic Mode Functions Arrangement Image, IMFAI)表征振动信号时域特征；EMD-PWVD时频图(EMD-PWVD Time-frequency Diagram, EPTFD)表征振动信号时频特征；对称极坐标图像(Symmetrical Polar Coordinates Image, SPCI)表征振动数据的形状特征；网格式灰度图(Grayscale Texture Map, GTM)使振动数据呈现出纹理特征。以上4种振动图像样本为基于卷积神经网络的滚动轴承智能故障诊断提供了图像样本构造方法。

（2）基于卷积神经网络(Convolutional Neural Network, CNN)的滚动轴承智能故障诊断研究。针对振动图像样本滚动轴承单一故障，采用CNN构建滚动轴承故障诊断模型。研究CNN敏感参数对诊断模型的影响，分别得到了4类振动图像样本对应的最优模型，测试准确率可达99.19%。在此基础上，采用参数迁移策略研究噪声对4类振动图像诊断模型的影响，结果表明EPTFD和SPCI振动图对噪声具有包容能力，在噪声影响下的故障诊断模型的测试准确率最高可达96.67%、模型训练耗时为170.46s。

（3）基于EMD-AADPCI振动图像的滚动轴承复合故障诊断研究。设计了滚动轴承内部零件复合故障类型，针对振动图像样本的信息容量与数据增强创新提出了一种EMD-AADPCI振动图像生成方法，该方法具有自适应特征增强优点，通过特征量化图对特征表征效果进行了验证，通过CNN实现了滚动轴承复合故障诊断，结果表明EMD-AADPCI联合CNN方法在滚动轴承复合故障诊断中准确率可达94.58%。

（4）基于多振动信息融合与CNN-SVM组合模型的复合故障诊断研究。针对多零件复合故障振动信号特点，当采用单路信号进行故障诊断时，振动图像样本特征不明显，故障诊断准确率较低。提出采用振动加速度与位移信号融合技术，使用IF-EMD-AADPCI振动图像样本，故障诊断准确率可达94.58%，较EMD-AADPCI提高了16.66%。针对诊断模型，结合支持向量机，采用数据分流的方法将3组独立的SVM作为CNN的分类模块，构造了CNN-SVM故障诊断新模型，使用该模型进行滚动轴承复合故障诊断，准确率达到96.25%，较基础CNN提高了1.67%。将IF-EMD-AADPCI振动图像样本及CNN-SVM模型推广至滚动轴承多故障-转子碰磨复合故障诊断，结果表明，诊断准确率分别达到95.00%、97.50%，较基础方法分别提高了29.42%、2.5%。

The progress of mechanical fault diagnosis technology promotes the development of productivity and improves the safety and reliability of equipment. As a common mechanical component in industry, the running state of rolling bearing is related to the safety of mechanical equipment. The research on fault diagnosis of rolling bearing is of great significance. Traditional rolling bearing fault diagnosis methods are difficult to accurately obtain fault information when dealing with complex signals. Artificial intelligence methods such as deep learning have strong learning ability, so rolling bearing fault diagnosis tends to intelligent diagnosis based on deep learning. This paper uses depth learning to carry out the research on rolling bearing fault diagnosis, from single fault problem to compound fault, as follows.

(1) Research on sample construction method of rolling bearing vibration image. Aiming at the research of rolling bearing fault diagnosis, four image conversion methods of vibration signals are studied. The time-domain characteristics of vibration signals are characterized by Empirical Mode Functions Arrangement Image (IMFAI); EMD-PWVD Time-Frequency Diagram (EPTFD) characterizes the time-frequency characteristics of vibration signals; Symmetrical Polar Coordinates Image (SPCI) represents the shape characteristics of vibration data; Grayscale Texture Map (GTM) makes vibration data present texture features. The above four vibration image samples provide an image sample construction method for intelligent fault diagnosis of rolling bearing based on Convolutional Neural Network (CNN).

(2) Research on intelligent fault diagnosis of rolling bearing based on CNN. Aiming at the single fault of rolling bearing in vibration image samples. CNN is used to build the fault diagnosis model of rolling bearing. The influence of CNN sensitive parameters on the diagnosis model is studied, and the optimal models corresponding to four kinds of vibration image samples are obtained respectively, and the test accuracy is up to 99.19%. On this basis, the parameter transfer strategy is used to study the influence of noise on four kinds of vibration image diagnosis models. The results show that EPTFD and SPCI vibration diagrams have the ability to contain noise. The test accuracy of the fault diagnosis model under the influence of noise is up to 96.67% and the training time of the model is 170.46s.

(3) Research on composite fault diagnosis of rolling bearing based on EMD-AADPCI vibration image. The composite fault types of internal parts of rolling bearing are designed. Aiming at the information capacity and data enhancement innovation of vibration image samples, the EMD-AADPCI vibration image generation method is proposed. This method has the advantages of adaptive feature enhancement. The effect of feature representation is verified by feature quantization diagram. The composite fault diagnosis of rolling bearing is realized by CNN. The results show that the accuracy of EMD-AADPCI combined with CNN method in rolling bearing composite fault diagnosis can reach 94.58%.

(4) Research on composite fault diagnosis based on multi vibration information fusion and CNN-SVM combined model. According to the characteristics of multi part composite fault vibration signal, when single signal is used for fault diagnosis, the sample characteristics of vibration image are not obvious, and the accuracy of fault diagnosis is low. Using the fusion technology of vibration acceleration and displacement signals and IF-EMD-AADPCI vibration image samples, the fault diagnosis accuracy can reach 94.58%, which is 16.66% higher than that of EMD-AADPCI. For the diagnosis model, combined with support vector machine, using the method of data diversion, three groups of independent SVM are used as the classification module of CNN, and a new CNN-SVM fault diagnosis model is constructed. Using this model for rolling bearing composite fault diagnosis, the accuracy is 96.25%, which is 1.67% higher than that of basic CNN. The IF-EMD-AADPCI vibration image sample and CNN-SVM model are extended to a composite fault diagnosis, which is multiple rolling bearing faults-rotor rubbing. The results show that the diagnosis accuracy is 95.00% and 97.50% respectively, which is 29.42% and 2.5% higher than the basic method.