基于自适应动态无偏LSSVM的故障在线监测

Online Fault Monitoring Based on Adaptive Dynamic Non-bias LSSVM

针对无法建立精确数学模型的非线性动态系统,提出一种基于自适应动态无偏LSSVM的故障在线监测模型。该模型通过改进LSSVM的结构风险形式得到无偏的LSSVM,并能够自适应的选择滑动时间窗的长度。在此基础上根据模型动态变化过程中核函数矩阵的特点设计了基于Cholesky分解的学习算法提高了模型训练效率,实现了非线性系统的在线监测。通过系统输出预测误差的变化,利用Parzen核密度估计方法判断故障的程度。仿真结果表明该故障监测模型在系统正常工作的情况下,能够跟踪系统的动态变化趋势;在系统出现突变故障的情况下,能够快速检测系统故障;在系统出现缓变故障的前提下,能够对系统的故障进行预报。

Aiming at nonlinear dynamic systems having no accurate mathematical model, an online fault monitoring model based on adaptive dynamic non-bias LSSVM was proposed. Through improving the structure risk form of LSSVM, the non-bias LSSVM model was achieved, which could control the size of sliding time window adaptively. Then a new learning algorithm based on the Cholesky factorization was designed according to the character of kernel function matrix in the model＇s dynamic change process. The model could greatly enhance the training efficiency, so it could online monitor the nonlinear systems. The fault probability could be estimated through Parzen kernel density estimation method based on the change of prediction error. Experimental results indicate the model can track the tendency of nonlinear system when system works well. Also it can rapidly detect the system＇s fault when a fast abrupt fault occurs, and can predict the fault when system is under the slow variation fault state.