Previous Topic Back Forward Next Topic
Print Page Dr. Frank Dieterle
Ph. D. ThesisPh. D. Thesis 11. Summary and Outlook11. Summary and Outlook
About Me
Ph. D. Thesis
  Table of Contents
  1. Introduction
  2. Theory – Fundamentals of the Multivariate Data Analysis
  3. Theory – Quantification of the Refrigerants R22 and R134a: Part I
  4. Experiments, Setups and Data Sets
  5. Results – Kinetic Measurements
  6. Results – Multivariate Calibrations
  7. Results – Genetic Algorithm Framework
  8. Results – Growing Neural Network Framework
  9. Results – All Data Sets
  10. Results – Various Aspects of the Frameworks and Measurements
  11. Summary and Outlook
  12. References
  13. Acknowledgements
Research Tutorials
Site Map
Print this Page Print this Page

11.   Summary and Outlook

In this work, time-resolved measurements and time-resolved data analyses of sensor re­sponses have been introduced in the field of chemical sensing. The time-resolved measure­ments, which were systematically investigated for several sensor setups and for several analytes, can be regarded as a second major step in the field of sensor development for multianalyte determinations. The first step from selective sensors to cross-reactive sensor arrays had allowed the parallel quantifications of different analytes by the use of the signal patterns of one single array of sensors without the need of finding selective sensor materials for each analyte. This first step became very popular in the field of electronic noses during the 80s and 90s. The second step, the time-resolved evaluations of sensor signals in combi­nation with suitable sensor coatings, combines the sensory principle with the chroma­tographic principle of separating analytes in space or time. This opens the door to a completely new dimension of information in chemical sensing. This work is the first extensive and systemati­c investigation of this second step for an improved and advanced quantitative determination of analytes in the field of chemical sensing.

The time-resolved measurements of this work are all based on the microporous polymer Makrolon as sensitive layer. This polycarbonate allows a kinetic separation of the analytes during the sorption and desorption on the basis of the size of analytes. It was shown using up to quaternary mixtures of the low molecular weight alcohols as analytes that small molecules can sorb very fast into the pores whereas the larger molecules sorb only slowly into the polymer. It was demonstrated that this specific sorption into the pores is a Langmuir type sorption, which is by far more important than the unspecific Henry type sorption of the analytes into the polymer matrix. An additional effect of an expansion of the pores during long-term exposure to analyte was observed. It was demonstrated that this effect could be exploited to measure bigger analytes by expanding the pores using other carrier gases than air. It was also shown that the variation of the thickness of the sensitive layer allows the tailoring of the sensitive layer to specific analytical questions.

The time-resolved measurements have been successfully used for three different sensor setups and for many multicomponent mixtures of the low alcohols and the refrigerants R22 and R134a. It was demonstrated that the time-resolved measurement principle can be applied to single sensor setups allowing the simultaneous quantification of several analytes and consequently rendering arrays of sensors unnecessary. It was furthermore shown that the time-resolved measure­ment principle can also be applied to sensor arrays with the results of an improved calibration, a higher robustness, an increased flexibility to the number and properties of different analytes and a reduced number of sensors.

Generally speaking, the time-resolved measurement principle allows the reduction of the expenses for hardware at the cost of a more extensive and a more complicated data analysis. This leads over to the second objective of the work, the data analysis. It was shown that the best results for multianalyte quantifications are obtained when the measurements are performed with the highest possible scanning rate of the sensor responses and the highest possible number of measurements for calibration. The resulting increased number of input variables (the time points generated by the scanning of the sensor responses) and the nonlinear relationship between the sensor responses and the concentrations of the analytes put new challenges to the data analysis. It was demonstrated that most common methods for a multivariate data analysis like PLS, QPLS, INLR, CART, MARS and neural networks showed rather poor calibration results. Among these methods, the neural networks were most promising but had to struggle with the high number of correlated and redundant input variables resulting in improvable calibrations. The combination of variable selection methods and of neural networks, which is widely used in literature to solve the issue of too many redundant and correlated input variables, could not help due to the limited number of samples measured.

In order to find a calibration with the highest possible calibration and generalization ability three frameworks were innovated, implemented and optimized in this work. These frameworks use repeated runs of a combined variable selection and calibration with different subsets of the available data resulting in a very effective exploitation of the limited number of data. One framework is based on many parallel runs of genetic algorithms combined with neural networks, one framework bases on many parallel runs of growing neural networks and the third framework uses several runs of the growing neural networks in a loop. All three frameworks showed by far better calibrations than all common methods of multivariate calibration and than simple non-optimized neural networks for all data sets investigated. Additionally, the variable selection of these frameworks allowed an insight into the relationship between the time-resolved sensor responses and the concentrations of the analytes. The variable selection also suggested optimizations in terms of shorter measurements for several data sets. The variable selection quality of the parallel growing network framework could be confirmed by a brute force variable selection. The calibrations and variable selections of all three frameworks were reproducible and were not disturbed by noise in the data. All three frameworks successfully solved the problems of too many variables for too few samples and the problems caused by the nonlinearites present in the data with practically no input needed by the analyst. Thus, all three frameworks showed excellent calibration and variable selection qualities whereby each framework has its own benefits. The genetic algorithm framework is the fastest framework whereas the parallel growing neural network framework shows a slightly better calibration. The loop-based growing neural network framework shows the best calibration performance as it allows building complicated yet sparse non-uniform neural networks. All three frameworks are not limited to time-resolved sensor data, but can be used for nearly any data when a powerful variable selection and calibration are needed and when the number of samples is limited. In the area of data-mining and pattern recognition, the application of these framework has also shown excellent results for data sets from medicinal chemistry .

Together, both main focuses of this work impressively demonstrate how the combination of an advanced measurement principle and of an intelligent data analysis can improve the results of measurements at reduced hardware costs. To prevent misunderstandings, an intelligent data analysis and an advanced measurement principle cannot help if a device provides bad or senseless data. However, the amount of information provided by a device can often be dramatically increased by using advanced measurement principles (like the time-resolved measurements of this work). Yet, it was also demonstrated in this work that additionally new intelligent methods of data analysis are needed, which are able to extract and use the valuable information out of the large pool of information provided by the advanced measurement principles (such as the frameworks introduced in this work). It was also shown that the results of the data analysis give feedback how the measurement principles, the measurement parameters and the devices can be optimized and improved. This demonstrates how the interconnection of the different parts of an analysis can improve the complete analysis in a synergetic effect.

Starting with this work further research can be performed in many fields of scientific research. Beginning with the sensitive layer, the principle of different-sized pores as size-sensitive recognition elements can be further investigated. For example it was shown in [269] that there are many other polymers with pores of different sizes like the Compimide 183 with a mean pore size of 0.038 nm3, the Polyimide PI2611 with a mean pore size of 0.058nm3 and the Polyimide PI2566 AL with a mean pore size of 0.13 nm3 and many more. These polymers allow extending the range of analytes to bigger and smaller molecules. The combination of these polymers in an array should result in a powerful setup for a size-selective discrimination of a broad range of analytes. Especially the extension of the SPR-device to an array setup seems to be very promising as the SPR setup demonstrated to be the most suited device for measurements using microporous polymers as sensitive layers. Furthermore, the principle of the time-resolved measurements is not limited to optical sensor devices but can be used for practically any arbitrary (sensor)-device like electronic noses, biosensors and many more as long as the sensor responses differ in the time domain. Thereby the recognition principle is not limited to size-selective recognitions but can be of any specific type that allows time-resolved discriminations. For example in the area of biosensing, different DNA with a different number of mismatches might be quantified simultaneously by differences of the DNA-DNA binding kinetics. Also, different antibodies might be discriminated on the basis of the kinetics, if the different antibodies show different adsorption kinetics due to different sizes of the FAB fragments. This allows single sensor applications for several selective and even cross-reactive analytes [270],[271].

The combination of several sensors with different sensitive polymers for time-resolved measurements on a sensor array opens the door to second-order calibrations similar to GC-MS setups. Thereby the sensor signals represent the first order and the time represents the second order. Second-order calibrations allow the quantification of an analyte in the presence of unknown interferences, which is also known as second-order advantage. For example, the generalized rank annihilation method (GRAM) [272],[273] can already work with a single standard addition to the prediction sample. Consequently, the extensive calibrations with experimental designs can be completely avoided resulting in dramatically reduced expenses for the calibration of specific analytes. Yet, further research has to be done concerning two topics. Fist of all, more polymers are needed, which allow time-resolved measurements and which show different chemical properties, as the second order advantage requires sufficient selectivity in both orders. Additionally, the second-order methods have to be further studied in respect to dealing with nonlinear relationships, as most of the up-to-date algorithms assume linear relationships in both orders.

An interesting approach similar to time-resolved measurements is the application of temperature-resolved measurements. Kato et al. [274] demonstrated that different analytes show different dynamic sensor responses if the sensor signal is recorded during a variation of the sensor temperature of tin oxide sensors. Mielle et al. [275] used a single tin oxide sensor to discriminate 9 analytes measured at 6 different temperatures. These approaches are not limited to metal-oxide sensors but can also be used for polymer-based sensors. As long as the sorption kinetics of the various analytes depends in different ways on the temperature, the temperature-resolved measurements allow exploiting an additional information domain. A very interesting point is also the glass transition temperature of a polymer. Measurements below the glass transition temperature should show a more specific sorption behavior whereas measurements above the glass transition temperature should show a more unspecific sorption doubling the information provided by a sensor.

In summary, it may be said that this work once more demonstrates that not the lack of information is the limit for chemical sensing but the frontier of scientific research, which makes this information available and understandable for the analyst and this frontier is moving from day to day opening the doors to new possibilities in scientific research.

Page 123 © Dr. Frank Dieterle, 14.08.2006 Navigation