摘要在现代工业生产设备不断朝着结构化、自动化和智能化方向发展的过程中，电机仍是主要的动力输出设备。若电机在运行过程中出现故障，会导致其运行效率降低，系统能耗上升等问题，严重时甚至造成电机损坏，使整体系统设备长时间停机维修，造成严重的经济损失。因此，研究电机智能故障诊断技术，对保障生产设备高效运行的稳定性、可靠性具有重要意义。随着科技的不断创新和发展，信号处理、人工智能等技术不断取得突破，故障诊断技术也更加精确化、智能化。本文结合实际生产过程中常见的电机变工况和强噪声环境下的故障诊断问题，在分析故障产生机理的基础上对电机智能故障诊断方法展开深入研究。(1)首先利用试验台采集的振动信号对电机不同故障的产生机理进行分析，探究其处于故障状态时的振动频率特性。在此基础上，研究基于信号处理和机器学习算法的故障诊断方法，分析经验模态分解(Empirical ModeDecomposition,EMD)存在的优势和不足，利用其改进算法集合经验模态分解(Ensemble Empirical Mode Decomposition,EEMD)对信号进行分解，得到反映原始信号不同频率成分的本征模态函数(ntrinsic Mode Function.IMF)分量，通过相关性与原始信号相关系数较高前4阶MF分量，最后对其进行谱分析得到多个序列作为样本信号用于特征提取。(2)对多序列样本信号中9种不同的时、频域统计特征进行计算，得到原始特征集，并在对其采用聚类算法分析的基础上，提出一种基于调整互信息和标准差的敏感特征选择方法，从原始特征集筛选特征构建敏感特征集用于电机故障诊断。并且，针对特征集中存在的特征干扰、冗余等问题，提出利用特征降维方法实现对特征的维数约简。最后分别利用支持向量机(Support Vector Machines,SVM)和极限学习机Extreme Learning Machine,ELM两种较为流行的机器学习算法实现电机故障诊断，并通过对比实验进行验证。(3)针对基于信号处理的传统智能故障诊断方法中存在的流程复杂、人为干预过多的问题，研究端到端式的卷积神经网络算法用于电机故障诊断。针对一维时序信号的特点，分析一维卷积神经网络(One-Dimensional Convolutional NeuralNetwork,1D-CNN)在以原始一维振动信号为基础进行故障诊断的优势。为了提取信号中不同尺度的互补特征，提出利用不同尺度核的卷积层提出一种多尺度融合框架，构建了基于多尺度一维卷积神经网络的电机故障诊断方法。最后通过实验验证了所提方法在变工况和噪声干扰情况下的优越性。(4)为了提升MS-1DCNN在电机变工况和强噪声干扰环境下故障诊断方法的识别效率和准确率，在残差网络结构的基础上构建多尺度特征融合框架。分别研究挤压与激励(Squeeze and Excitation,SE)模块和卷积注意力模块(ConvolutionalBlock Attention Module,CBAM)两种注意力机制算法的实现原理，设计适用于一维残差网络的注意力模块，并将其嵌入到残差模块中，构建出多尺度注意力残差网络Multi--scale attention residual network,MSA-ResNet)模型，最后利用实验台数据验证所提模型的有效性和优越性。该论文有图44幅，表18个，参考文献92篇。关键词：电机：故障诊断：多尺度特征：卷积神经网络：注意力机制AbstractIn the process of modern industrial production equipment continuously movingtowards the direction of structure,automation and intelligence,motors are still themain power output equipment.Failure of the motor during operation will causeproblems such as reduced operating efficiency and increased system energyconsumption.In severe cases,the motor will be damaged,causing the overall systemequipment to be stopped for maintenance for a long time,causing serious economiclosses.Therefore,the research of motor intelligent fault diagnosis technology is ofgreat significance for ensuring the stability and reliability of efficient operation ofproduction equipment.With the continuous innovation and development of scienceand technology,breakthroughs have been made in signal processing,artificialintelligence and other technologies,and fault diagnosis technology has become moreprecise and intelligent.The thesis combines the common fault diagnosis problems ofthe motor in the actual production process and the fault diagnosis in the strong noiseenvironment.Based on the analysis of the failure mechanism,the intelligent faultdiagnosis method of the motor is studied in depth.(1)Firstly,the intelligent fault diagnosis method based on signal processing isstudied,the mechanism of motor fault generation based on vibration signals isexplored.Based on the analysis of the motor fault characteristics under differentoperating states,the signal analysis method based on empirical mode decompositionis studied,and aimed at it Existing modal aliasing is proposed to use set empiricalmodal decomposition to analyze the motor vibration signals,select the first 4 ordersof IMF components through correlation analysis and calculate the correspondingenvelope spectrum and marginal spectrum as the feature extracted signal sequence.(2)Calculate 9 different time and frequency domain statistical features in themulti-sequence sample signal to obtain the original feature set.Based on the analysisof the clustering algorithm,a sensitive feature based on adjusting mutual informationand standard deviation is proposed.Select a method to filter features from the original