基于时序模型和自联想神经网络的齿轮故障程度评估

Fault severity assessment for gears based on AR model and auto-associative neural network

齿轮是机械传动系统中的重要零部件,在役齿轮故障程度评估是机械系统剩余寿命预测和状态维修的基础。目前广泛研究的基于概率相似度量的故障评估方法存在过早饱和等问题,不利于在线监测。提出一种基于自回归时序模型(Auto-Regressive, AR)和自联想神经网络(Auto-Associative Neural Network, AANN)的齿轮故障程度在线评估方法,其中AR模型用于齿轮振动信号特征提取,AANN用于故障程度评估。提取基准阶段(如无故障阶段)振动信号AR模型系数作为AANN的输入和输出向量,得到基准评估模型。将待评估信号AR系数构成的特征向量输入到基准AANN,得到重构AR系数。基于原始AR系数和重构AR系数组成两个自回归模型分别对待评估信号进行时序建模,分别得到各自的模型残差序列。基于残差序列之间的差异,提出了一种基于残差均方根差值的故障程度定量评估指标。离散故障程度的齿轮振动数据分析结果表明,该方法能有效区分齿轮故障的不同程度;在此基础上利用齿轮全寿命周期试验数据进一步验证该方法的有效性,结果显示提出的方法能够及时发现肉眼没有观测到的早期故障,且随着齿轮性能的不断退化,能直观反映齿轮故障程度的加深。

The gear damage severity evaluation underlines the prognostics and condition-based maintenance of mechanical systems.Being motivated by the fact that paradigms based on the probability similarity,like the hidden Markov model(HMM)and Gaussian mixed model(GMM),tend to an early saturation,a new approach for gear damage evaluation was proposed by making use of an autoregressive model(AR)and auto-associative neural network(AANN).The AR model was made to fit with gear vibration signals and the model coefficients were extracted as feature vectors which were then,fed to AANN to obtain reconstructed AR coefficients.A baseline AANN was trained by using the feature vectors from normal condition.The reconstructed AR coefficients by the baseline AANN will deviate from the original AR coefficients,if the gear condition degrade from normal condition.So,the difference between the residuals of AR models using reconstructed and original AR coefficients was exploited to formulate a gear damage indicator.Two experimental data sets involving one discrete damage-degree and one run-to-failure test were utilized to verify the proposed method.The results show the novel indicator is able to track damage progress with a consistent trend and to detect incipient damage in time.