Novel Mahalanobis Distance and Variable Nearest Neighbors to Construct Weight Matrix based LPP: Application of Fault Diagnosis

Fault diagnosis techniques based on data-driven algorithms go mainstream gradually in industrial processes. Unfortunately, traditional data-driven algorithms cannot deal with massive high-dimensional and nonlinear strong correlation data. Based on this problem, Novel Mahalanobis Distance and Variable Nearest Neighbors to Construct Weight Matrix based LPP (MV-LPP) is proposed in this paper. MV-LPP replaces the Euclidean distance used to measure the similarity between two sample points with the Mahalanobis distance, which takes in account the correlation of the samples, so that it can exclude the interference of correlation between the variables. Besides, the MV-LPP algorithm, corresponding to the location of each sample point in the data set, optimizes the method to screen the nearest neighbor points in the locality preserving projections algorithm, to the extent that MV-LPP can obtain the appropriate number of nearest neighbor points and the weight matrix can better preserver the spatial structure shape and achieve a better mapping effect. In final, a dataset of Tennessee Eastman Process (TEP) is utilized to validate the MV-LPP method and the positive results proved its effectiveness.