The Research On The Method Of Filtering Dynamic Spectrum Data By Combining The Time Domain With The Frequency Domain

ObjectiveAs a kind of nondestructive real-time continuous detection method,near infrared spectroscopy analysis has played an important role in the noninvasive measurement of blood components.Dynamic spectrum theory proposed by our research group makes sense for the noninvasive detection of blood components.Sample collection and modeling are the important elements in the research of dynamic spectrum noninvasive detection,and the reliability and accuracy of acquired sample data are preconditions of modeling analysis and key points for the dynamic spectrum noninvasive detection.In the process of a large number of sample acquisition,however,because of the environment,instrument,and artificial factors,the interference and noise will appear in the samples.Thus,in order to ensure the accuracy and reliability of the collected spectral data,filtering the collect samples and further denoising the selected sample data become important parts of the in-depth research in dynamic spectrum.In this paper,after evaluating the sample quality and eliminating abnormal samples,we get reliable samples for modeling,which can significantly improve the prediction precision and reduce the prediction error.The detected near infrared spectrum range of noninvasive detection of blood component is 591.85 ~ 1119.96 nm,and its average resolution is 0.81 nm.Each sample contains spectral data with hundreds of wavelengths.In order not to lose any spectral information,we usually use the whole band data to model.But in these spectral data,some wavelengths do not carry certain blood composition information,or only contain little information,so in order to reduce model complexity,information interference and the model computation,we prioritize wavelengths in the basis of guaranteeing the model accuracy and stability.It is necessary to get main wavelength information while removing the wavelengths that contain extremely low information.Reducing the variable number of the wavelengths could lay solid foundation for the concretization and productization of noninvasive detection of dynamic spectrum theory in the future.Methods405 acquired samples are evaluated by combining valid single edge counts of dynamic spectrum pulse wave in the time domain with the quality factor Q value of dynamic spectrum pulse wave in the frequency domain.As a result,the abnormal samples are removed and 218 cases of valid samples are selected.We then use the dynamic spectrum data of the selected 218 samples as the experimental group and another 218 samples as a control group modeling analysis with the hemoglobin concentration of the corresponding samples.Each group select 200 cases of samples as a calibration set and 18 cases of sample as a prediction set.By analyzing the prediction accuracy and mean relative error of each group,we can assess the sample filtering methods of combining time domain with frequency domain.On the basis of the stable model in the experimental group,we use the method of uninformative variables eliminate(UVE)to select on the near infrared region including 586 wavelengths.With the same stability and precision as the experimental model,total number of wavelengths are reduced in the modeling process,which simplifies the model,reduces the computational cost,and improves efficiency of the model.And eventually,the cost of subsequent product development is saved.ResultsThe prediction accuracy of the experimental group is as high as 93.8%.The accuracy values for the two control group respectively evaluated by the valid single edge counts and Q value are 65.6% and 67.7%,and the corresponding values for the three unfiltered control groups are 53.7%,33.3%,and 42.6% separately.It is obvious that the prediction accuracy of the experimental group is significantly higher than that for the other control groups.The mean relative error of prediction(MSEP)of the experimental group is 0.0675,corresponding values for the other two control groups are 0.0723 and 0.0722,and values of the other three control groups are 0.0823,0.0789,and 0.0828.So compared with another control groups,the MSEP of the experimental group is the minimal.In terms of wavelength selection part,the number of variant wavelengths for each sample of experimental group are reduced from the 586 to 175,which,compared with the original variables,is decreased by 70.14% and thus greatly reduces the amount of irrelevant information variables.The prediction accuracy after the wavelength selection is 88.1%,slightly less than 93.8% which is calculated before optimization.The relative error of prediction is 0.036,and this value is less than 0.0675 calculated before prioritization.When evaluating the model prediction effect,the precision ability of the optimization model was slightly lower than before,but the error level of this optimization model is better than before.In short,the current wavelength selection method reducing the 70.14% amount of distinctive wavelengths,can possess the same prediction ability as the original model.ConclusionsBy the experimental verification,the combination of time domain and frequency domain method for spectral data sample screening is reliable and valid.This elimination method of uninformative variables is optional for wavelength selection,which can greatly decrease the number of wavelengths while guaranteeing the precision and error.The above two aspects of dynamic spectrum noninvasive detection are of great significance for the following precision and the feasibility study.