Abstract:  Industrial processes often operate in multiple production modes, for the characteristics of larger variables, the center drift and the larger modal variance of the multi-modal data, a weighed k-nearest-neighbor fault detection method(FD-wkNN) is proposed. First, find the k nearest neighbor in the training data set and calculate the distance between the training sample and the k nearest neighbor, and calculate the average distance of between the k nearest neighbor and the first K local nearest neighbors, take the reciprocal of the average distance as the distance weight, take the weighted distance as statistic D, D is able to eliminate the influence of center drift and difference of modal variances. Second determine the control line using D distribution. Finally compare the calculation D statistics of online sample and control line. On-line fault detection is realized. Using multi-mode example, as well as examples of penicillin data simulation experiments, compared with the PCA, kPCA, FD-kNN method to verify the effectiveness of this method.

Key words: PCA; kPCA, kNN, multimode, fault detection

摘要： 工业过程往往运行于多个生产模态，针对多模态过程数据的空间分布特点：中心漂移和模态协方差差异明显，提出了基于权重k近邻的故障检测方法(FD-wkNN)。首先在训练数据集中寻找第k近邻并计算近邻距离；其次把此k近邻与其前K近邻集的局部近邻平均距离倒数作为权重，构建加权平均累积距离D作为统计量。加权平均累积距离可以有效降低中心漂移和协方差差异明显的影响；最后，利用核密度估计确定训练样本集统计量D的控制限，当新样本的加权平均累积距离大于控制限时，则其为故障；否则为正常。FD-wkNN具有对协方差较小模态的微弱故障的检测能力。通过模拟实例和青霉素发酵过程进行故障检测仿真实验，并与PCA，KPCA，FD-kNN等方法比较，验证了所提方法的有效性。

关键词: 主元分析, 核主元分析, k近邻, 多模态, 故障检测