旋转机械在各类机械设备中具有重要地位，其能否安全稳定运行直接影响着企业的经济效应，甚至是人员的生命安全。因此，旋转机械的健康状态识别尤为重要，已经成为近年来的研究热点之一。传统的旋转机械健康状态识别具有振动特征提取困难、特征非线性表征差、状态等级划分不合理、健康状态识别准确率不高等问题。针对上述问题本文以旋转机械为对象，在旋转机械振动信号多尺度混合特征提取及降维选择、健康指标构建及健康状态划分和旋转机械健康状态识别模型构建三个方面进行研究，并利用公开数据集和实验室减速器平台完成了方法验证。论文主要内容如下：

（1）研究旋转机械振动信号多尺度混合特征提取及选择方法。将旋转机械振动信号进行多尺度的信号分段，并在多尺度信号分段的基础上进行经验小波分解；提取各子段振动信号和固有模态分量的时域、频域、时频域等特征，分析特征与旋转机械退化之间的关系；利用综合评价准则选择具有高判别性、退化趋势明显的退化特征，构建最优特征集。

（2）研究基于核主成分分析及指数函数滤波的旋转机械健康指标构建方法。基于核主成分分析将选择后最优的多尺度混合特征进行融合，并根据主元累积贡献率构建旋转机械健康指标；利用指数函数对健康指标上的毛刺进行去除，得到平滑的健康指标；基于健康指标并结合划分依据，完成旋转机械不同健康状态的样本划分。

（3）提出一种基于改进卷积神经网络（MSCCNN）的旋转机械健康状态识别方法。将最优的多尺度混合特征输入模型，通过多尺度卷积提升特征提取视野，丰富特征信息；引入自注意力机制，完成退化敏感特征自提取；利用自定义损失函数，保证模型参数最优化，提升模型鲁棒性。采用西安交通大学轴承退化数据集及PHM2012挑战数据集对模型有效性进行验证，模型识别准确率分别为98.48%和97.71%，并选择不同的轴承数据进行了泛化性能验证。在减速器平台退化数据集上进行验证，MSCCNN模型识别准确率为97.35%，并在不同的减速器数据下进行验证，测得的准确率达90.85%。

（4）开发旋转机械健康状态识别系统。基于前后端分离技术，开发一套集成旋转机械数据监测、多尺度混合特征提取、健康指标构建和健康状态识别等模块的旋转机械健康状态识别系统，最后通过实例完成系统测试。

综上所述，本文所提的基于多尺度混合特征的旋转机械健康状态识别方法能够对旋转机械进行准确的识别，对于旋转机械的健康状态识别提供良好的理论指导。

Rotating machinery plays an important role in all kinds of machinery and equipment, and its safe and stable operation directly affects the economic effect of enterprises, and even the life safety of personnel. Therefore, the health status recognition of rotating machinery is particularly important and has become one of the research hotspots in recent years. The traditional health status recognition of rotating machinery is difficult to extract vibration features, poor nonlinear characterization of features, unreasonable classification of state levels and low accuracy of health status recognition. In view of the above problems, this paper takes rotating machinery as the object, and researches on the multi-scale mixed feature extraction and dimensional-reduction selection of rotating machinery vibration signals, health index construction and health state division, and health state recognition model construction of rotating machinery. The method is verified by using open data sets and laboratory gear reducer platform. The main contents of the paper are as follows:

The method of multi-scale mixed feature extraction and selection of vibration signals of rotating machinery is studied. The vibration signals of rotating machinery are segmented into multi-scale signals, and empirical wavelet decomposition is carried out on the basis of multi-scale signal segmentation. The time domain, frequency domain and time frequency domain characteristics of each sub-segment vibration signal and the inherent modal components are extracted, and the relationship between the characteristics and the degradation of rotating machinery is analyzed. Using the comprehensive evaluation criteria to select the degradation characteristics with high discriminability and obvious degradation trend, the optimal collection is constructed.

The health index construction method of rotating machinery based on kernel principal component analysis and exponential function filtering is studied. Based on kernel principal component analysis, the optimal multi-scale mixing features were fused, and the health index of rotating machinery was constructed according to the cumulative contribution rate of principal components. The exponential function is used to remove the burr on the health index and get a smooth health index. Based on health index and division basis, the samples of different health states of rotating machinery were divided.

An improved convolutional neural network (MSCCNN) based health status recognition method for rotating machinery is proposed. The optimal multi-scale mixed features are input into the model, and the feature extraction field is enhanced by multi-scale convolution to enrich the feature information. The self-attention mechanism is introduced to complete the self-extraction of degenerate sensitive features. Using the custom loss function, the model parameters are optimized and the model robustness is improved. Bearing degradation data set of Xi 'an Jiaotong University and PHM2012 challenge data set were used to verify the validity of the model. The accuracy of model recognition was 98.48% and 97.71% respectively, and different bearing data were selected to verify the generalization performance. The recognition accuracy of MSCCNN model is 97.35% on the degraded data set of reducer platform, and the accuracy is 90.85% when verified under different reducer data.

A health status recognition system for rotating machinery was developed. Based on the front and back end separation technology, a health status recognition system of rotating machinery was developed, which integrated the modules of data monitoring, multi-scale mixed feature extraction, health index construction and health status recognition. Finally, the system was tested by an example.

In summary, the health status recognition method of rotating machinery based on multi-scale mixed features proposed in this paper can accurately identify rotating machinery and provide good theoretical guidance for health status recognition of rotating machinery.