Study On Method Of Fetal Electrocardiogram Extraction Based On Tensor Decomposition

Fetal electrocardiogram(FECG)is an important way to monitor the health status of the fetus in the uterus.Thus,it has great significance for clinical fetal heart monitoring during perinatal period.Since its non-stationarity,low signal-to-noise ratio,and overlaps with noises and interferences in time,frequency,space and feature domains,FECG is very difficult to be extracted from an abdominal composite signal of a pregnant.To obtain better extraction results,the information from different domains should be combined and more priori information should be used.In this thesis,the tensor theory was introduced into the fetal ECG extraction,and the tensor decomposition method was improved according to the time,space and beatquency domain of ECG signal.The main contents of the research are summarized as follows:(1)The information from beatquency domain was introduced to build a tensor model for the FECG.The non-invasive abdominal ECG signal is a multi-lead time domain signal,which is usually expressed as a matrix.The majority of the FECG extraction methods are usually carried out in the matrix space,which severely limits the combination of the information from multiple domains,resulting in unsatisfactory extraction results.In this thesis,according to the multi-domain characteristics of FECG and tensor theory,the ECG signal was modeled as a third-order tensor to break the classical time-invariant linear instantaneous mixing model to realize the multi-domain joint estimation of FECG signal.(2)A new method for FECG extraction based on the CP decomposition was proposed.The statistical analyses of the existing algorithms are usually carried out in the time or space domain,resulting in low accuracy.In this thesis,the tensor was constructed due to the high-order characteristic of FECG,and the classical CP decomposition method was used for sparse representation of the tensor.The proposed method used the time-domain waveform information,spatial correlation and inherent cyclical of the ECG signal to achieve an effective estimate of fetal ECG without increasing the computational complexity.(3)A customized CP decomposition method for FECG extraction was proposed.When CP decomposition is applied directly,the efficiency can be improved by combining multi-domain information.However,it is inconsistent with the characteristics of fetal ECG due to its uniform estimation of time and space domain,resulting in that the decomposed components do not have clear explanations in physiology.In this thesis,the tensor decomposition method was improved by considering the spatial independence of ECG.Due to the different extraction environment between maternal ECG and fetal ECG,the proposed method was further modified to achieve an effective inhibition of maternal ECG and reserve the integrity and variability of fetal ECG signal.A simulation database and an open database with real records were used to evaluate the proposed method.Experimental results show that the tensor decomposition method proposed in this thesis can effectively extract the fetal ECG signal,the fetal heart rate and fetal ECG waveforms are close to the reference fetal ECG acquisited by fetal scalp electrodes,which are superior to the traditional matrix decomposition methods.