旋转机械作为工业领域最常用的机械设备之一，其一旦发生故障，轻则造成经济损失，重则影响人员生命安全。轴承、齿轮作为旋转机械关键组成部件，长期处于高转速或重负载的复杂环境下，易产生疲劳损伤导致退化失效。因此，识别旋转机械的健康状态具有重要的理论价值和现实意义。振动信号因其包含丰富的状态信息而被广泛用于研究机械设备健康状态。本文基于旋转机械振动信号，从多域特征提取、健康指标构建及健康状态识别三方面展开本文研究。论文主要工作如下：

       针对振动信号非线性特点及单一特征信息难以全面描述旋转机械不同阶段退化状态问题，在时域、频域特征基础上，利用残差占比确定变分模态分解（Variational Mode Decomposition，VMD）层数，提取时频域、熵等多域特征。构建了基于单调性、相关性和鲁棒性的综合评价准则，筛选准确表征旋转机械退化状态的最优特征。

       针对旋转机械健康指标构建过程中忽略特征时序性、特征间关联性以及主观因素参与较多的问题，提出一种基于长短期记忆网络与卷积降噪自编码器（Long Short Term Memory-Convolution Denoising Autoencoder, LSTM-CDAE）的旋转机械健康指标构建方法。将筛选的多域特征首先输入长短期记忆网络（Long Short Term Memory, LSTM）挖掘特征时序性，之后利用卷积降噪自编码器（Convolution Denoising Autoencoder, CDAE）提取特征间关联关系、降低噪声干扰，以无监督方式构建旋转机械健康指标。提出方法在西安交通大学轴承数据集上构建的健康指标相较于单一的均方根指标平均早14.7分钟发现早期劣化点、较其他同类方法构建的健康指标平均早3分钟到达早期退化点。

       针对深度神经网络层数加深模型出现退化导致识别准确率不高的问题，提出一种基于深度残差收缩网络（Deep Residual Shrinkage Network, DRSN）的旋转机械健康状态识别方法。利用格拉姆角场（Gramian Angular Field，GAF）构造健康状态样本，增强特征表达；改进网络结构、优化模型参数，引入高斯误差线性单元（Gaussian Error Linear Unit，GELU）激活函数及正则化技术，避免梯度消失、抑制模型过拟合，提高模型识别准确率。在西安交通大学轴承数据集上分别与深度置信网络、卷积神经网络及残差神经网络进行对比，结果表明提出方法的识别准确率为99.52%，平均高于其他三者3.05%。

       最后，设计并开发了一套旋转机械健康状态识别系统，并在实验室搭建的减速器实验平台数据上完成上述方法的实验验证及系统测试。实验结果表明LSTM-CDAE方法构建的减速器健康指标综合评分为0.77，DRSN模型对减速器不同健康状态的识别准确率达98.87%。

Rotating machinery is one of the most commonly used mechanical equipment in the industrial field. Once it breaks down, it will cause economic loss and affect the life safety of personnel. Bearing and gear, as the key components of rotating machinery, are easy to cause fatigue damage and degradation failure under the complex environment of high speed or heavy load for a long time. Therefore, it is of great theoretical value and practical significance to identify the health state of rotating machinery. Vibration signal is widely used to study the health state of mechanical equipment because of their rich state information. Based on vibration signals of rotating machinery, this paper carries out the research from three aspects: multi-domain feature extraction, health index construction, and health status recognition. The main work of this paper is as follows:

Because of the nonlinear characteristics of vibration signal and the fact that it is difficult for single feature information to fully describe the degradation state of rotating machinery at different stages, based on the characteristics of the time domain and frequency domain, the residual proportion is used to determine the layer number of Variational Mode Decomposition (VMD), and extract the multi-domain features such as time-frequency domain and entropy. A comprehensive evaluation criterion based on monotone, correlation, and robustness was established to screen the best features that accurately represent the degenerate state of rotating machinery.

Aiming at the problems of ignoring feature timing, inter-feature correlation, and subjective factors in the construction of health indicators of rotating machinery, A method for constructing a health index of rotating machinery based on Long Short-Term Memory-Convolution Denoising Autoencoder (LSTM-CDAE) was proposed. The selected multi-domain features were first input into the Long Short-Term Memory (LSTM) network to mine the feature timing, and then Convolution Denoising Autoencoder (CDAE) was used to extract the correlation between features to reduce noise interference, and the health index of rotating machinery was constructed in an unsupervised way. In this paper, the health indicators constructed on the XJTU-SY bearing data set found the early deterioration point 14.7 minutes earlier on average than the single Root Mean Square index, and reached the early degradation point 3 minutes earlier than the health indicators constructed by other similar methods.

Aiming at the problem of low recognition accuracy due to the degradation of the layer-deepening model of the Deep neural Network, a method for health status identification of rotating machinery based on Deep Residual Shrinkage Network (DRSN) is proposed. The Gramian Angular Field (GAF) was used to construct health samples and enhance feature expression. The network structure was improved, the model parameters were optimized, and the Gaussian Error Linear Unit (GELU) activation function and regularization technique were introduced to avoid gradient disappearance and inhibit model overfitting, to improve the model recognition accuracy. Compared with the deep belief network, convolutional neural network, and residual neural network on the Xi'an Jiaotong University bearing data set, the results show that the recognition accuracy of the proposed method is 99.52%, 3.05% higher than the other three methods on average.

Finally, a set of rotating machinery health recognition systems is designed and developed, and the experimental verification and system test of the above method is completed on the data of the reducer experimental platform built in the laboratory. The experimental results show that the comprehensive score of the health index of the reducer constructed by the LSTM-CDAE method is 0.77, and the accuracy of the DRSN model for the different health states of the reducer is 98.87%.