大型转动机械在工业生产过程中，轴承故障频发，呈现出一种“大数据”的特性，并且现场获取的故障数据往往是不完备和无标签的，亟需开发具有自学习能力的诊断算法。针对该研究提出一种基于双结构深度学习的轴承故障类型与损伤程度的智能诊断算法。该方法使用不完备的数据建模，分为故障类型自学习网络和故障损伤等级识别网络两个结构。对轴承故障信号进行形态学滤波，抑制部分噪声，增强信号的脉冲特征；对消噪后信号进行S变换得到时频图，获取故障类型的共性特征；并将时频图作为卷积神经网络(CNN)的输入，利用网络的相似性度量在目标空间对同类型样本汇聚、不同类型样本分离，实现对轴承故障类型的分类与新故障类型的自学习。将实现故障类型分类的信号经归一化处理后作为深度置信网络(DBN)的输入，利用DBN对微小故障的敏感性对不同损伤程度的差异特征进行提取，之后将提取的特征作为贝叶斯分类器的输入，依据后验概率判别规则实现故障损伤等级自主识别。将该方法应用于西储大学实验平台的滚动轴承故障数据，结果表明，该方法在不完备数据建模的情况下，不仅能完成故障类型与损伤等级的准确分类，而且还能实现故障自学习和损伤等级自增长，增强了诊断过程的智能性。

An intelligent diagnosis algorithm was presented based on double structure deep learning to identify the bearing fault type and damage degree.In the method, an incomplete data model was used and divided into two structures: fault type self-learning network and fault damage degree identification network.In the first structure, the faulty signal was filtered by morphology to suppress some noises and enhance the impulse components.Then, the filtered signal was transformed by S-transform to obtain a time-frequency graph, and obtain the common features of the fault type.The time-frequency graph was used as the input of a convolution neural network (CNN) to gather the same type of samples and separate the different types of samples in the target space by virtue of the similarity of the network, and further to realize the classification of bearing fault types and the self-learning of new fault types.In the second structure, the signals of known fault types were employed as the input of a deep belief network (DBN) after normalization, in order to extract the different features of different damage degrees.Subsequently, the extracted features were utilized as the input of a Bayesian classifier to automatically recognize the fault damage degree according to the posterior probability classification rules.In the end, the proposed method was validated using the bearing fault data acquired by the Bearing Data Center supported by the Case Western Reserve.The results show the proposed method can not only accurately identify fault type and damage degree, but also realize fault type  self-learning and damage degree self-growth under the condition of incomplete data modeling, and enhance the intelligence of the diagnosis procedure.