Abstract: The amplitude mutation of multidimensional time series tension and straightening force represents the fault of fan-shaped unit.The time and space complexity of multidimensional time series data put forward higher requirements for the accuracy and efficiency of the mutation feature extraction method.Based on the temporal and spatial similarity of multidimensional time series data,a feature extraction method of multidimensional time series mutation based on wavelet correlation filtering combined with Principal Component Analysis (PCA) was proposed.Each one-dimensional time-domain noisy tension correction was transformed into the wavelet domain,and the self-similarity of tension correction was used to reduce noise.Then,the denoised tension and correction force was mapped to the orthogonal space,and the multidimensional tension and correction force were mapped to one dimension by linear transformation based on the mutual similarity between multidimensional tension and correction force.The threshold was determined according to the statistical characteristics of one-dimensional feature curve,and the mutation feature was extracted.Finally,the sliding window based on time was used to dynamically update data to realize continuous online state recognition.The analysis of field measured data showed that the accuracy and efficiency of wavelet correlation filtering-PCA model for the mutation feature extraction of multidimensional time series data were 90.39% and 97.56% respectively.

Key words: segment, multi-variate time series, principal component analysis, mutation characteristics

摘要： 多维时序拉矫力的幅值突变表征着扇形段机组的故障,多维时序数据的时间和空间复杂度对突变特征提取方法的准确率和效率提出了更高的要求。利用多维时序数据的时间和空间相似性,提出小波相关滤波结合主成分分析的多维时序突变特征提取方法。首先,将每一维时域含噪拉矫力分别变换到小波域,利用拉矫力的自相似性进行降噪;然后,将去噪后的拉矫力映射到正交空间中,利用多维拉矫力之间的互相似性,通过线性变换将多维拉矫力映射到一维,再根据一维特征曲线的统计特征确定阈值,提取突变特征;最后,应用基于时间的滑动窗口动态更新数据,实现状态的连续在线识别;通过现场实测数据分析表明,小波相关滤波主成分分析模型对多维时序数据的突变特征提取的准确率和效率分别达到90.39%和97.56%。

关键词: 扇形段, 多维时间序列, 主成分分析, 突变特征