Dynamic Multivariate Outlier Detection Algorithm Using Ultraviolet Visible Spectroscopy for Monitoring Surface Water Contamination With Hydrological Fluctuation in Real-Time.

The contamination of surface water is of great harm. Ultraviolet-visible (UV-Vis) spectroscopy is an effective method to detect water contamination. However, surface water quality is influenced by hydrological fluctuation caused by rain, change of flow, etc., leading to changes of spectral characteristics over time. In the process of contamination detection, such changes cause confusion between hydrological fluctuation spectra and contaminated water spectra, thus increasing the false alarm rate. Besides, missing alarms of contaminated water is a common problem when the signal-to-noise ratio is low. In this paper, a dynamic multivariable outlier sampling rate detection (DM-SRD) algorithm is proposed. A dynamic updating strategy is introduced to increase adaptability to hydrological fluctuation. Additionally, multiple outlier variables are adopted as outlying degree indicators, which increases the accuracy of contamination detection. Two experiments were carried out using spectra collected from real surface water sites and hydrological fluctuation was constructed. To verify the effectiveness of the DM-SRD method, a comparison with the static SRD method and spectral match method was conducted. The results show that the accuracy of the DM-SRD method is 97.8%. Compared with the other two detection methods, DM-SRD significantly reduces false alarm rate and avoids missing alarms. Additionally, the results demonstrate that whether the database contained prior information on hydrological fluctuation or not, DM-SRD maintained high detection accuracy, which indicates great adaptability and robustness.

[Temperature effect on the noninvasive measurement of human blood glucose by NIR].

During the noninvasive measurements of body blood glucose, the result will be effected by many factors, such as the measuring conditions including temperature, contact pressure and so on, and in addition the change of body's state also will induce some error. However, among so many factors the temperature is very important and should be discussed. To find the quantitative value of the result bias caused by temperature in the wavelength range from 1 100 to 1 700 nm, the aqueous glucose with the concentration ranging from 10 mg x dL(-1) to 200 mg x dL(-1) and 10 mg x dL(-1) interval was detected at temperature of 15, 20, 25, 30, 35 and 40 degrees C. Then six different models at different temperature were founded and predicted one another. The maximum RMSEP result of models is 11. 227 9 mg x dL(-1) and the minimum is 3. 298 8 mg x dL(-1). The correlation is about 0.98. The authors have also found that 1 degrees C change of temperature will induce deltac = 2.662 (mg x dL(-1) x degrees C(-1)) change of the prediction result, so these show that the detect error will be minimal when the temperature of measuring is the same as that of modeling. Moreover, the authors put forward two approaches to decreasing or compensating the error induced by the temperature.

[Fast outlier detection for milk near-infrared spectroscopy analysis].

Near-infrared spectroscopy is a fast and efficient analytical technique based on multivariate calibration model, which correlates near-infrared spectra with the property of samples (such as concentration). The reliability of analytical results depends mostly on the accuracy of measured spectra. But outliers do not make for reliable data. The authors combined RHM (Resampling by Half-Means) with SHV (Smallest Half-Volume) method to detect the outliers of the near-infrared spectra of milk samples, and the results were satisfactory. The performance of the new method is superior to the traditional outliers detecting algorithms such as Mahalanobis distances and hat matrix leverage. And this combined method is simple and fast to use, conceptually clear, and numerically stable, so it is recommended to be used for the detection of multiple outliers in multivariate data, especially the online measurement and discriminant analysis.

Quantitative effect of temperature to the absorbance of aqueous glucose in wavelength range from 1200nm to 1700nm.

In this paper, to find the quantitative errors of aqueous glucose induced by the temperature change at every wave point ranging from 1200nm to 1700nm, the calibration curve is calculated and shown. During the measurement the temperature varies from 30 degrees to 40 degrees , at a 2 degrees interval, and aqueous glucose concentration ranges from 100mg/dL to 500mg/dL, at a interval of 100mg/dL. The absorption of aqueous glucose decreases with the increasing of temperature, also the absorbance decreases. In addition, only 1 degrees change in the temperature induces about -7x10-3 and -4x10-3 errors in the absorbance of the aqueous glucose at the wavelength of 1550nm, 1610nm respectively. So the examined result should be correct according to the data read from the calibration curve if the temperatures of modeling and measuring are not uniform. Using this method, the error caused by the temperature change can be reduced even eliminated.

[Experimental research on OGTT for non-invasive blood glucose detection through near-infrared spectroscopy ranging from 1100 nm to 1700 nm].

For non-invasive blood glucose detection through near-infrared spectroscopy, it is very important to ensure the data quantity and reliability of calibration model. In this paper, the method of sampling blood by tubing pump in OGTT (Oral Glucose Tolerance Test) is used to get reliable and adequate reference data of blood glucose concentration for calibration model, and the non-invasive blood detection system based on the AOTF (Acousto-Optic Tunable Filter) ranging from 1100 nm to 1700 nm is designed. 3 experiments were performed by the above system and method. The results showed that based on the PLS (Partial Least Square) calibration models constructed by analyzing all individual experimental data, the correlation coefficients were 0.986, 0.971 and 0.985, respectively, and the RMSEP (Root Mean Square Error of Prediction) estimated by Full Cross Validation were 0.550 mmol/l, 0.456 mmol/l and 0.520 mmol/l; respectively. The results also showed that the prediction error of the model decreased when the number of effective model data increased.