中文摘要：

基于时变自回归（TVAR）方法实现了非平稳随机信号的参数化建模，提出采用最小信息准则确定模型阶数．通过多分量线性调频仿真信号的时变谱估计，表明该方法分辨率高，没有交叉项的干扰，计算速度快．在仿真分析的基础上，应用参数化时频谱和BP神经网络方法进行滚动轴承故障信号的分类和辨识，并基于能量法对时频图进行特征提取．分析结果表明，时变自回归方法的拟合精度高，能有效提取轴承故障信号特征，同时结合神经网络能对故障进行准确诊断．

英文摘要：

Parametric-modeling of nonstationary signal based on time-varying autoregression （TVAR） was realized. Akaike information criterion, which can choose the order automatically, was expatiated. Time-varying spectrum estimation of multi-ponderance linear frequency modulation signal proves that the TVAR has lots of merits, such as high resolution, without cross term and fast computing speed. The parametric time-varying spectrum and BP neural network method were used to classify and distinguish fault signal of beating.The time-varying spectrum features were extracted by energy means. Results show that the TVAR can extract the characteristic of fault signal, gain high simulation precision and identify fault types exactly.