Research On Feature Transfer Method For Temporal Sensing Data

The rapid development and widespread application of Io T technology have driven time series analysis and mining became the research hotspot.Traditional time series analysis requires elaborate data preprocessing,feature extraction,and selection,making it difficult to be applied in large-scale scenarios.In recent years,deep learning has received significant attention for its powerful feature learning and representation ability,which was suitable for large-scale time series processing and analysis.However,effective deep learning needs to satisfy two conditions: independence identically data distribution and enough labeled data.While real-world time series often suffer from domain shift and lack of labels,which significantly affect the accuracy and generalization of classification and prediction models.Transfer learning can extract knowledge from source domain with abundant labeled data and assist in the learning task of target domain without labels.Reducing the distribution difference between the source domain and the target domain is an important paradigm of transfer learning,by cross-domain learning to promote the generalization of models in target domain.Therefore,this thesis utilizes transfer learning to explore the problem of distribution divergence and domain shift existed in time series classification and prediction tasks,respectively.The contents include:Firstly,there is an interactive effect between domain-invariant feature learning and pseudo labeling in time series classification.On the one hand,the importance of features changes with the different transfer layer pairs of the pre-trained network and the target network.On the other hand,the performance gains of the target network contributes to the pre-trained network optimization further to obtain higher-quality pseudo labels.This thesis uses consistency learning to ensure the unbiasedness during feature learning of the pre-trained network,which avoids the negative transfer caused by feature bias towards the source domain.Additionally,an interactive mechanism is designed to learn the mutual impact between domain-invariant features and pseudo labels,which can implement the iterative updates of the pre-trained and target networks.Therefore,an interactive unsupervised domain adaptation method is proposed for time series classification.Experimental results on the time series classification dataset demonstrate that the proposed method can effectively improve the classification performance.Secondly,the existing "pre-training + fine-tuning" domain adaptation model presents slow running,plenty of network parameters which results in model redundancy and difficult deployment.Therefore,this thesis further studies an end-to-end domainadaptive method for mutual learning between the pre-trained and target networks,and proposes an unsupervised domain adaptation time series classification method based on a teacher-student framework.In the proposed method,the contrast learning is used to enhance the learning of invariant features.The online distillation can strengthen the collaborative training of the two networks.And the feature matching and feature competition can achieve mutual transfer of intermediate layer features between the two networks without model pretraining.Experimental results on the time series classification dataset demonstrate the effectiveness and superiority of the proposed method.Finally,time series with non-stationarity gives rise to distribution drift,which degrades the performance of time series forecasting.This thesis proposes an unsupervised domain adaptation time series forecasting method based on invariant feature learning.The continuous data distribution of time series is quantized.Then the distribution divergence is measured and maximized to achieve sub-domain segmentation,so that the potential distribution of time series can be extracted.Additionally,a domain adaptation module is designed to minimize the difference in distribution between the two domains to extract invariant features.Experimental results on a time series forecasting dataset demonstrate that the proposed method can improve the forecasting performance.The thesis has 28 figures,23 tables,and 112 references.