Research On The Mechanism Of Neural Network Representation-based And Topological Signature Analysis-based Machine Learning Frameworks For Few-Shot Body Sensing Time Series

The analysis and recognition of human body sensing data are an essential means of human health status identification,clinical status monitoring.Human body sensing data are acquired from different signals based on human life activities.Due to the com-plex nonlinear dynamic characteristics of the human body,and the inevitable external environmental interferences during the acquisition process,the body sensing data exhibits nonlinearities and non-stationarities.There were many pattern analysis methods based on feature engineering in the human sensing data analysis tasks in recent years.However,these static features show limitations on the ability to represent nonlinear and non-stationary data.The neural network representation learning methods show significant application values in the analysis of human sensing data due to its nonlinear structure and its excellent autonomous learning characteristics,mainly when there exists an extensive unlabeled data set.Also,the phase space reconstruction would turn the human body sensing data into data point clouds,and the topological descriptions can reveal these properties.The system characterizes the dynamics of the human body and show promising ability in the human body pattern recognition tasks.Simultaneously,the abnormal state of the human life activity process often has unusual characteristics,so the number of negative samples in the human body sensing data is relatively small.For example,the proportion of abnormal heartbeats in arrhythmia analysis is much smaller than normal heartbeats,and the frequency of freezing-of-gait abnormalities in Parkinson’s disease subjects walking experiments,which are much less than the normal gait cycles.Negative samples,such as abnormal heartbeat and frozen gait,are infrequent events.In the overall data set,they must be smaller proportion samples compared to positive samples.Furthermore,collecting multi-sample data from a limited number of subjects has the problems of high clinical risks and difficulties,uncontrollable repeatability,and high cost.The problem of human body sensing data analysis is a typical few-shot learning problem.Generally,when the samples of the data set are small,the few-shot dataset can be augmented by extracting the prior knowledge to generate a new data sample.Thus,the dataset becomes large enough for the models’ training,such as rotating,flipping,and cropping strategies in image processing.However,in the few-shot learning tasks of human sensing data,the statistical distribution is unknown,corresponding to the complexity and randomness of the human body state.It is impossible to obtain reliable data samples to achieve data augmentation through extracting prior knowledge,which made the strategies to expand few-shot dataset with similar datasets or unsupervised ones impossible in human sensing data analysis tasks.In the analysis of human sensing data,especially in the analysis of specific pathological patterns,it is often extremely costly.In recent years,the model-based approach has provided new ideas for few-shot learning.This dissertation focuses on applying model-based few-shot learning problems in the analysis of human sensing data,combined with engineering practice needs,and studies the few-shot learning based on neural network representation extraction with pretraining and re-optimization strategies.This dissertation starts with the effective feature extraction and few-shot learning in the analysis of human sensing data came from the requirements from the research projects,and the actual engineering practice needs.We studied two typical human sensing data,ECG signals,and gait information.For the research object,we analyzed the application of feature extraction based on neural network representation and the topology signature analysis for the few-shot learning tasks.We designed the corresponding computational framework and classification systems.The main work of this article in-cludes four main aspects:1.The problem of few-shot learning with a large amount of unsupervised data is studied.We proposed a few-shot learning model based on pre-training and parameter reoptimization strategy using a deep autoencoder neural network.Furthermore,we apply the model in the classification of ECG arrhythmias types.The main work includes autoencoder network pre-training method,classification system construction and parameter optimization,neural network representation extraction.Finally,the model’s arrhythmia classification and identification capabilities were verified experimentally.The results showed that the learning method using deep autoencoder neural network-based representation,which can effectively use unsupervised data to improve the modeling ability and recognition accuracy in few-shot ECG analysis problems2.We studied the application of the pre-training¶meter reoptimization few-shot learning strategy based on deep belief neural network representation in the multi-lead ECG analysis.We introduce the pre-training method,classification system construction,and parameter optimization of the deep belief network based on the restricted Boltzmann machine.We proposed a combination optimization joint decision algorithm to improve the model’s ability for analyzing complex ECG patterns.The experimental results prove the ECG characterization ability of the deep belief neural network,and the joint decision optimization algorithm formed an effective information fusion strategy3.A few-shot embedding learning mechanism based on topological signature analysis is proposed.We used the time-delay embedding based on nonlinear dynamic anal-ysis to construct point clouds for human sensing data in phase space,and then extract the topological signatures as features for the few-shot learning task.The corresponding framework of topological analysis includes phase space reconstruction of human sensing time series,topological modeling of point cloud data,extraction of topological signature,and random forest classification.We achieved promising results for the few-shot ECG learning tasks4.We proposed a topological analysis framework for human gait interval sequence analysis and its application to few-shot embedding learning in the identification of neu-rodegenerative diseases.We extracted the topological signatures of different neurode-generative diseases(Myosclerosis,Huntington’s disease,Parkinson’s disease),and then feed to random forest classifiers for the few-shot human gait classification.Experimental results show that topological signatures have significant representation ability and classification performance in the few-shot gait interval classificationIn summary,this dissertation studies the analysis method and application of human heart electrocardiogram data based on neural network representation.The proposed a deep autoencoder network structure-based classification system and a deep belief net-work decision system for ECG analysis.Both were applied to the real-world wearable heart monitoring applications,which have achieved practical values.The pattern extraction method based on topology signature analysis proposed in this dissertation has achieved excellent results in pathological analysis problems such as neurodegenerative diseases and has significant clinical value.The work in this dissertation further expands the theoretical system of health informatics.It combines the relevant research content of neural network representation analysis and topology signature analysis with health in-formatics’s practical needs.The few-shot learning mechanism introduced has essential theoretical advantages.