根据旋转机械信号循环平稳特性，以及故障出现后非平稳特征渐强的特点，提出了相对及绝对循环平稳度的概念，通过理论分析，提取出对旋转机械故障较为敏感的相对循环平稳指标。在此基础上又提出了通过对信号x（t）及其共轭乘积进行Fourier变换，代入循环频率α值获取循环自相关函数Rαxx（0）的方法，并利用本方法对不同调幅系数的正弦信号和氧压机小偏心量调幅故障信号，进行了仿真和实例验证，证明了指标的有效性。

运用小波分析的尺度函数，在进行小波分解后，进一步进行傅里叶变换，能够提取其他方法无法得到的故障特征信号，为故障诊断提供了一种新的分析手段。对Jeffcott转子的不平衡和刚度不对称等故障分析验证了该方法的可行性和实用性。最后讨论了小波变换级和噪声对尺度函数分析方法的影响，表明小波尺度函数的降噪作用是十分明显的。

在Priestley提出的基于小波分析进化谱理论基础上，对改进进化谱估计的性能进行了研究。首先，重点分析了小波分解的尺度与信号时变趋势和分辨率之间的关系，并指出了根据信号时变特征和分辨率要求确定合理的小波分解尺度。其次，讨论了权函数对进化谱估计能量泄漏现象的影响，指出了选用Hanning窗等作为权函数可以有效地抑制谱泄漏现象。仿真及在故障诊断中的应用与理论分析的结论一致。根据本文所提出的方法来估计非平稳信号的进化谱可以得到更高精度的结果。

经验模式分解方法可以将非线性非平稳信号分解为有限的固有模式函数，在故障诊断中这个固有模式函数常常就是故障信号。但当两侧端点不为极值点时，会造成三次样条拟合的极值包络线大大偏离实际值，并且随着分解的不断进行向内“污染”。提出采用时间序列建模与预测方法，对原信号两端点进行预测，有效地消除了端点效应。指出经验模式分解具有分解的自适应性特点。最后，给出了齿轮箱振动信号的应用实例。

椭球单元通过高斯分布逼近形成各模式类的决策区域，是一种非常适合于模式识别任务的前馈型人工神经网络模型。提出改进椭球单元神经网络的训练权重组，即采用多权重组，增强了椭球单元网络的抗干扰能力，提高了网络的故障诊断能力，并给出了权重选择方法。仿真和试验验证了该方法的正确性。

Neural Networks with Ellipsoidal Activation Functions close in upon a decision-making region by Gauss distribution for various patterns and are adapted for fault diagnosis well. It has the advantages, such as high precision in clssifying, few training time and well rejection for unknown pattern. For improving the antijamming performance, a multi-weight is adopted between input neuron and ellipsoidal neuron in the paper. The selection method of weights is also derived.By combining several neural networks with ellipsoidal activation functions, the multi-fault diagnosis can be diagnosed. the scale of network and training time are reduced observably and the ability for diagnosis is enhanced by the combination of networks. The simulation and experiments sustain the suggested method.

针对故障诊断对象的多故障特性，为使系统在诊断对象的各子故障域的诊断精度达到最高，给出了可以适合于在多故障情况下找到一个最优联盟结构的算法。首先，给出了各个诊断方法代理和代理联盟的诊断能力的定义，并且提出将联盟的诊断能力作为评判标准；然后，在各子故障域同时进行联盟结构的搜索，寻找出具有最高联盟诊断能力的结构形式，并改变联盟矩阵中相应单元的值；最后，根据联盟矩阵确定各子故障域的联盟结构，并在此基础上形成该诊断对象的诊断联盟。

The association deficiency of neural network affects its further application in pattern recognition. Rough set theory(RS) is excellent in disposal of small incomplete sample data. Based on this, the basic optimizing strategy of RS BP neural network fault diagnosis is presented in this paper. Using the strategy, an optimized neural network model is established. In this model, RS is employed as a preprocess course, and the results of RS preprocess decide the model structure. So the model can decrease the number of the network input nerve cells effectively, and ameliorate network inner structural. The bearing fault data and grinding condition data are analyzed for the model. The results show that the strategy has better study efficiency and diagnosis accuracy. It is estimated that the optimized strategy may be further applied in fault diagnosis.

介绍了循环平稳信号的二阶循环矩分析和处理方法、循环矩分析的抑噪特性。在给定齿轮裂纹调频信号数学模型的基础上，证明了其二阶循环平稳性。利用二阶循环矩—滞后平面图，分析了裂纹故障信号特征的循环矩提取方法，进行了仿真实验。针对齿轮裂纹故障实例，说明其对故障信号解调的有效性。

Fault diagnosis of large rotating machinery sets is very important for many industry sectors. However, it is also difficult because of its complexity. The conventional fault diagnosis methods are based on FFT spectral analysis. While its effectiveness has been approved by many applications, it also has significant limitations. In fact, it has been found that many faults have similar spectral patterns. In order to pinpoint the cause of the faults, it may be necessary to vary the speed to get additional information. Nevertheless, the speed variation makes the vibration signal non-stationary. This in turn makes the FFT spectral analysis imperative. In this paper, it is proposed to use a new technique, called evolutionary spectrum, to analyze the non-stationary signal and hence diagnose the faults. First, following a brief introduction of wavelet transform, the evolutionary spectrum is described. Next, the implementation procedure of the evolutionary spectrum is given. Then, by means of computer simulation and experiment results analysis, it is shown that the evolutionary spectrum can effective capture the fault patterns and hence, diagnose rotor unbalance and rotor stiffness asymmetric.