Feature Extraction And Analysis Based On Tensor Data

Advancement in distributed sensing,computing techniques,and the shift towards the Internet of Things(Io T)has facilitated a wide collection of complex data,leading to a spatially and temporally data-rich environment.On one hand,such spatiotemporal data that exhibit variations and dependencies across time(time-series data),space(variation over multiple locations)creates unprecedented opportunities to design more efficient systems and make more optimal decisions.On the other hand,analysis and modeling of spatiotemporal data have become a considerable challenge because of the complexity and heterogeneity of data.In this dissertation,for the purpose of feature extraction,process monitoring,and anomaly detection,new methodologies are proposed to model and analyze multi-sensor data,which can be mathematically represented as a tensor.The first part of this dissertation develops a method to characterize the correlation and variations of multi-sensor data based on multilinear algebra.The proposed method operates on the tensor data directly and does not break the tensor structure.Therefore,it can preserve more useful and compact information potentially.The method considers the interrelationship between different sensors and provides a set of eigentensors that can characterize the process variations and correlations precisely.The second part of this dissertation is to deal with the feature extraction of the multi-sensor(multi-channel)data,which is a complex spatiotemporal characteristic in this dissertation.This makes it difficult to monitor the high-dimensional multi-channel data and working process.To address this challenge,a new feature extraction method with a feature selection strategy to improve the separability of data is proposed firstly.The feature extraction-based multivariate control charts for process monitoring is then developed.The third part of this dissertation is about fault detection and classifications of multichannel data based on multilinear subspace learning.Uncorrelated multilinear discriminant analysis(UMLDA),which is exploratory research applied to face recognition and image processing,is used to analyze multi-channel data for the first time.Improved UMLDA(IUMLDA)with tensor-to-tensor projection is then proposed to reduce the effects of projection order,improve the accuracy of detection,and reduce the fluctuation of the results.The effectiveness of all proposed methods is demonstrated by using Monte Carlo simulations and real-world case studies.The proposed methodologies can be used to general industrial processes that have sensing capabilities and have broad application opportunities.