DE10032029A1 - Process used for analyzing products comprises generating a universal factor set from a number of materials, acquiring actual spectra, and determining and using the optimum universal factors for analyzing materials - Google Patents

Abstract

Process for evaluating spectra from a spectrometer used for investigating solid, liquid and gaseous materials comprises generating a universal factor set from a number of materials absorbing light, the factors being obtained using a known chemometric method; acquiring actual spectra; determining the optimum universal factors for differentiating the materials to be investigated; and using these optimum factors for determining the materials to be investigated, where the selection of the factors is automatically carried out.

Description

The invention relates to a method according to the preamble of claim 1.

NIR technology has established itself in many areas of analytical chemistry. She was in over the past twenty years have been intensively involved in substances such as products from the Food and feed industries, petrochemical products, Analyze pharmaceuticals or non-active environmental toxins.

However, the advantages of NIR technology are only useful if you have the overlapping broad bands that can be observed in the spectra. Therefore you need the methods of multivariate statistics. They are used to create models, with which can be used to evaluate NIR spectroscopic data. Creating such Calibration models are usually a very complex process. The optimal selection of the Factors that are relevant for a recognition task is an extensive area of specialization Science, such as in Verdú-Andres, J. / Massart, D.L. Comparison of Prediction- and Correlation-Based Methods to Select the Best Subset of Principal Component Regression and Detect Outlying Objects. Applied Spectroscopy, Vol. 52, No. 11, 1998, p. 1425.

Due to this complexity, the procedure for creating the calibration models is as follows (cf. the left part of Fig. 1): Using selected substance spectra, the so-called calibration data, trained experts create a model that forms the basis of the evaluation system, with which a sorting task is to be solved. This procedure is common to the products of all major suppliers in this area. This model formation is also of great importance in science: Most of the time it is described here how a model is created for a given application (see e.g. Espinoza, LH / Lucas, D. / Littlejohn, D. / Kyauk, S. Total Organic Carbon Content in Aqueous Samples Determined by Near-IR Spectroscopy. Applied Spectroscopy, Vol. 53, No. 1, 1999, p. 103). In some cases, there are also general suggestions for simplifying model building (see Swierenga, H. / de Weijer, AP / von Wijk, RJ / Buydens, LMC Strategy for Constructing Robust Multivariate Calibration Models. Chemometrics and Intelligent Laboratory Systems, Vol. 49, 1999, pp. 1-17), some concrete attempts at standardizing model formation, e.g. in the so-called Heisenberg model (see Höskuldsson, A. The Heisenberg Modeling Procedure and Application to Nonlinear Modeling.Chemometrics and Intelligent Laboratory Systems, Vol. 44, 1998, pp. 15-30). In addition, many scientific approaches are aimed at developing new methods to assess the properties of models such as B. to improve their robustness (see Conlin, AK / Martin, EB / Morris, AJ Data Augmentation: An Alternative Approach to the Analysis of Spectroscopic Data. Chemometrics and Intelligent Laboratory Systems, Vol. 44, 1998, pp. 161- 173).

Another focus of science is on already created models that are in practice mostly with too little data to expand (see Candolfi, A. / Massart, D.L. Model Updating for the Identification of NIR Spectra from a Pharmaceutical Excipient. Applied Spectroscopy, Vol. 54, No. 1, 2000, p. 48) or from one spectrometer to another  (see Anderson, C.E. / Kalivas, J.H. Fundamentals of Calibration Transfer through Procrustes Analysis. Applied Spectroscopy, Vol. 53, No. 10, 1999, p. 1268). Both of the latter Approaches are aimed at avoiding the time-consuming creation of a new evaluation model.

There are also so-called library search methods (see Otto, M. Chemometrie. Weinheim: Wiley-VCH, 1997). These are compared to the existing spectrum of substances Data set in a reference database. Different algorithms are used for this, which provide the most similar substances in the reference database as a result.

An automation in the creation of prediction models has so far been used Neural networks partially reached. The networks independently calculate a connection between the spectra as input data and properties of the spectra as output data. In addition to the so-called back-propagating perceptron networks, such as in Laufens, P. Computer-aided evaluation of infrared spectra for plastic identification under Use of a neural network, MARECK colloquium, Aachen, Germany, February 1997 those networks that use the FuzzyARTMAP algorithm should be mentioned (see e.g. Feldhoff, R. / Wienke, D. / Cammann, K. and Fuchs, H. On-Line Post Consumer Package Identification by NIR Spectroscopy Combined with a FuzzyARTMAP Classifier in an Industrial Environment. Applied Spectroscopy, Vol. 51, No. 3, 1997, pp. 362-368).

In addition, the LOCAL software from FOSS NirSystems is aimed in the same direction. It is based on the so-called local calibration method, as described in Aastveit, A. H. / Marum, P. Near-Infrared Reflectance Spectroscopy: Different Strategies for Local Calibrations in Analyzes of Forage Quality. Applied Spectroscopy, Vol. 47, No. 4, 1993, p. 463: For one Use case turns a database into a number of spectra that can be determined are closest to the application. These then become a prediction model calculated. In the case of the LOCAL software, this procedure is generalized so that the user does not have to intervene in the calculation of a prediction model (see www.foss nirsystems.com/SpecSheets/WinISI.htm).

A further approach for the automation of the Calibration procedure in connection with spectra of a spectrometer presented. Here are the boundary conditions of the actual evaluation process are automatically determined.

All of the methods listed above have the major disadvantage that they are either for each new use case or only for very different use cases instructed a trained user who adapts them to the application. This leaves them become time and cost intensive.

The rule in the evaluation methods used is that a model for an application is calculated for the prediction of spectra. These calculations must be done by experts are accompanied, since various boundary parameters, such as a possible data preparation or the Number of factors to be used, must be determined intuitively.

The library search procedures that do not require such a model are on it instructed that there is and is a suitable database for the present application therefore inflexible in frequently changing areas of application. In addition, you can the above-mentioned method does not independently respond to faults that are reflected in the data material, and are therefore device-dependent. Library search procedures are therefore rigid or must by pretreating the data and thus by an intervention on the part of the trained Be adjusted to disruptive influences.

Neural networks at least partially automate the process of modeling, but have the problem of complex data preparation. For the perceptron networks, for example it makes sense to subject the spectra data to a factor analysis before presenting it to the network (see, for example, Laufens, P. Computer-aided evaluation of infrared spectra for Plastic identification using a neural network, MARECK colloquium, Aachen, Germany, February 1997). However, the network only takes over part of the Modeling, namely the grouping of substances in a factor space that was previously formed must become. Alternatively, spectra data can also be presented directly to the network. For the purpose of  For a good result, a number of data pretreatments such as that are recommended Selection of wavelengths that are particularly typical for the application or simple or differentiate the spectra data twice. In addition, the network must be initialized and the Architecture of the network, the number of nodes and layers, tailored to the application become.

The fuzzy ARTMAP algorithm is also suitable for good results Data preparation instructed (see Wienke, D. / von den Broek, W. / et al. Adaptive Resonance Theory Based Neural Network for Supervised Chemical Pattern Recognition (Fuzzy ARTMAP). Chemometric & Intelligent Laboratory Systems, Vol. 32, 1996, pp. 165-176). Furthermore must Network parameters ("initialization, sizes of learning rates and vigilance parameters" see Wienke, D. / van den Broek, W. / et al. Adaptive Resonance Theory Based Neural Network for Supervised Chemical Pattern Recognition (Fuzzy ARTMAP). Chemometric & Intelligent Laboratory Systems, Vol. 32, 1996, pp. 165-176) for the application.

The very good results that can be seen especially with neural networks that use the fuzzy ARTMAP Accordingly, using an algorithm to achieve the sorting of non-active substances are also relies on complex intervention on the part of the user when building the model. For the best possible prediction with the help of local calibrations, the largest possible Spectra databases required (see Sinnaeve, G. / Dardenne, P. / Agneessens, R. Global or Local? A Choice for NIR Calibrations in Analyzes of Forage Quality. Journal of Near Infrared Spectroscopy, Vol. 2, 1994, pp. 163-175). If you want general validity with this method achieve, d. H. the spectra database must be able to sort any non-active substances Include a variety of different substances. It is therefore for a user of this method necessary to have a very large database connected to his system. Furthermore local calibrations require a lot of computation, since all spectra are in the database the use case must be compared.

The invention from the published patent application DE 198 10 917 A1 is dependent on the fact that the Spectral data, on the basis of which it carries out an automatic calibration, information for the Calibration contains, for example, the number of measuring processes, property values or Instrument data. It is therefore not possible to calibrate using spectra as provided by a device can be generated automatically. Furthermore, the primary Factors again and again calculated from the spectra of substances to be examined become.

The time and cost involved in the elaborate calibration work with the previous ones As a result of the process being created, the use of NIR spectroscopy is usually only for large, constant material flows are worthwhile.

The problem described in detail using the example of NIR spectroscopy applies to everyone spectroscopic methods that work with the absorption of electromagnetic radiation.

The object of the invention is to use spectroscopy and in particular NIR Spectroscopy for sorting substances that use the electromagnetic used Absorb radiation to simplify so much that it no longer relies on the expertise of the User is instructed. Creating models for predicting substances based on their Spectra should be automated. The spectra of the substances produced by the Evaluation system are to be sorted, presented to the system in a calibration step. This object is achieved by a method with the features from claim 1. The procedure has the name "IntellIdent".

When using this method, the costly and time-consuming step of modeling by experts is omitted (cf. the right part of FIG. 1). The evaluation system itself carries out this step, it only requires selected spectra of the substances to be determined. This makes the use of spectroscopy for the sorting of substances highly flexible, since an existing device can solve any number of sorting tasks from the area of the electromagnetic radiation absorbing substances with almost no additional effort. The purchase of a spectrometer does not have to pay off in one application, but can be applied to different applications. The measurement of spectra is influenced by the respective spectrometer. A calibration created on one device can therefore not be transferred unchanged to another device. Since the calibration of the evaluation system is fully automated in this invention, it can be carried out again with any spectrometer without time and expense. This also takes the influences of the device into account. The time-consuming transfer of a calibration from one device to another is no longer necessary.

The same applies to the expansion of a calibration. It can also be done in a short time and carried out without specialist knowledge, the effort involved is minimized by the invention. It is also advantageous to have the invention in connection with the recently on the market available mobile spectrometers. The invention enables the mobile Spectrometers can be calibrated on site in a short time and thus for the sorting of substances to be ready. This enables the immediate use of this technology on site.  

1. Description

The invention relates to a method for evaluating spectral analytical data. The invention is described in more detail below with reference to schematic representations:

Fig. 2 shows a schematic representation of a possible measuring device,

Fig. 3 Schematic representation of the inventive evaluation system,

Fig. 13 Schematic representation of the Andrassy selectivity,

Fig. 16 Schematic representation of the selection factor,

Fig. 18 Schematic representation of the determination of the four Erstfaktoren,

Fig. 20 shows a schematic representation of the determination of the optimum Faktortripels,

Fig. 22 Schematic representation of the evaluation of substances to be determined.

A possible measuring arrangement is shown in FIG . The samples are recorded by a measuring probe containing a detector. A spectrometer records the data and forwards it to a computer. The processing of the data in accordance with the invention is carried out there.

The IntellIdent evaluation system, which was developed within the scope of the invention, can be roughly divided into three sub-areas (cf. for a schematic overall representation of the system, FIG. 3):

• - The well-known factor analysis, the factorization of the reference data and the Calculation of the factor weights of calibration spectra or spectra to be determined includes.
• - The calculation of the optimal factor configuration for a given one Recognition problems.
• - The evaluation of spectra to be determined.

The mode of operation is as follows: the reference spectra, the spectra of the largest possible Number of substances, in a first step of factorization, u. a. Published in Martens, H. / Naes, T. Multivariate Calibration. Chichester: John Eiley & Sons, 1989. Ideally, the resulting factors are universal, that is, for all applications suitable. The universal factor space spanned by the factors therefore only needs to be one calculated once and is then an integral part of the system.

If substances are to be sorted subsequently, the system must first be calibrated for these substances become. To do this, he needs the so-called calibration spectra, i.e. representative spectra of the substances to be distinguished from Intellldent. These are in The universal factor space is projected: Your factor weights are related to the universal Factors determined. The system then makes a factor selection based on these factor weights: it calculates with which three of the universal factors the greatest selectivity in the Differentiation of the substances to be determined can be achieved.

After the calibration, unknown spectra to be determined are compared with those in the calibration optimal factor configuration determined evaluated.

The Intellldent evaluation system should ideally be used for all sorting tasks Measuring device used light-absorbing substances may be suitable. As an example in Following considered the organic matter using a measuring device NIR light can be examined.  

1.1. Universal factor space

The evaluation method according to the invention adjusts the similarity of the organic substances Use what is known as the largest possible number of organic substances Universal factor space forms: The organic substances are broken down on the basis of their spectra and a set of characteristic building blocks, the factors. With the help of these factors the spectra of organic substances can be put together. Because of the similarity of the organic substances should not only be able to put together the substances that are used for the Obtaining the factors were used, but ideally all organic matter. This Factors form the universal factor space, a factor space in which all organic substances are Group clearly from each other.

The similarity of organic substances will be discussed in more detail below. As the most extensive branch of chemistry, organic chemistry comprises a large part of the compounds of carbon. About 90% of the organic compounds consist of carbon (C), hydrogen (H) and oxygen (O) in varying proportions. Numerous organic compounds also contain nitrogen (N) and in rare cases also sulfur (S), phosphorus (P) and the halogens. In principle, however, every element can be incorporated into organic compounds. .. 4 Fig through 7 show, respectively, a compound of alkanes, Altrene, cycloalkanes and aromatic hydrocarbons such as inter alia occur in diesel fuels: Figures 4 to 12 show examples of the basic chemical constituents of selected organic substances... FIG. 8 and FIG. 9 show the basic structures of two plastic materials, the polyvinyl chloride and polycarbonate. In Fig. 10, the basic structure of the cellulose is illustrated. Cellulose is found as a structural substance in the plant cell wall and is part of paper. Fig. 11 and Fig. 12 show two other herbal ingredients. Firstly, the farnesol, which is a determining component of flower fragrances, e.g. B. von Rosen, Akacien und Linden (cf. Richter, Gerhard. Biochemie der Pflanzen. Stuttgart; New York: Thieme, 1996); on the other hand, chlorophyll, which is involved in photosynthesis as green leaf in every green plant.

The selection of these substances is intended to illustrate how similar organic substances are in their chemical structure. The chemical compounds that can be found in the respective substances are always the same. For example, the non-aromatic CH group is found in cellulose, farnesol, chlorophyll and in polyvinyl chloride; the aromatic CH group can be found in both benzopyrene and polycarbonate; the CH ₃ group is a component of heptamethylnonane, trimetal pentene, polycarbonate farnesol and chlorophyll. If organic substances could be broken down into their individual chemical compounds and reassembled as required, a certain repertoire of organic substances would be sufficient to be able to assemble any other organic substance completely or at least to a certain extent.

NIR spectroscopy is currently reacting to the individual chemical compounds in their respective molecular context. For example, B. the CH bond in the methyl (CH ₃ ) group is a characteristic band in an NIR spectrum.

Against this background, the central idea of this invention is to have as many as possible organic substances based on their NIR spectra in their individual compounds or in individual Disassemble groups of connections and then these individual parts as desired reassemble. This becomes possible with the help of factor analysis. The factors can be as characteristic groups of individual compounds can be interpreted. A big number of various organic substances should therefore provide factors from which any organic compound can be composed.  

1.1.1. Reference substances for the formation of the universal factor space

In Fig. 15, the materials are shown, from their NIR spectra of the universal Faktomaum of the proposed example will be calculated. The aim of the selection shown is to include as many chemical compounds as possible representative of the field of organic substances. In the example shown, there are a total of 121 different substances or mixtures of substances. In addition to a large group of plastics, it contains many biomaterials (e.g. wood, linoleum, leather) and plants (lilac, dandelion, garlic, box hedge and much more), which are particularly characterized by a variety of chemical compounds. In addition to solid organic substances, liquid ones were also taken up. Liquid substances differ from solid substances mainly in that the molecules in liquids are significantly smaller than those in solids and the bonds between the molecules are weaker than in solids. This results in a different environment for the individual chemical compounds in liquids compared to solids, which is reflected in the NIR spectra.

Water and sand are included as non-organic compounds. The Water is active even in the near infrared and very often comes into contact with organic ones Substances before, for. B. in many biomaterials or as a residue on washed plastics. The sand is not active in the near infrared and served as a carrier of organic liquids to be able to measure them even in low concentrations.

1.2. factor analysis

The factor analysis comprises two processes: first, the factorization of spectra into the Factors and on the other hand the determination of the factor weights of spectra with respect to the Factors.

The factor analysis in the case of IntellIdent is the same for all measuring devices Use electromagnetic radiation, in principle only once and is included as many different substances as possible. Since the focus of the example shown on the Examination of organic substances, as many different organic substances as possible selected and recorded as spectra. The totality of these substances, represented by Five spectra each was subjected to factor analysis. This resulted in 30 sensible ones Factors. The factors calculated beyond that contained no significant information, but only so-called noise. Theoretically, there are also a number of other factors possible. Both the calibration spectra of the substances that the system is supposed to recognize as well as unknown spectra that are to be determined after the calibration are included in the factor weights associated with universal factors.

1.3. factor selection

The analysis of substance spectra with the help of factor analysis takes place as follows: First, calibration spectra of substances to be determined are in the universal factor space projected by grouping them based on their factor weights. Any unknown The spectrum is then also projected into the universal factor space; there will be the Mahalanobis distances, u. a. described in Derde, M.P. / Massart, D.L. Analytica Chimica Acta, vol. 184, 1986, p. 33, calculated for spectra of known substances. All of this is happening automatically. With this distance measure, a statement about the affiliation of the examined spectrum for the calibration substances. The universal set up here Factor space has 30 dimensions - each spectrum is represented by 30 factor weights.

In the previously common application of factor analysis, i. H. when calculating the factors The information content of a spectrum can be determined for the substances that are to be determined afterwards  can already be represented with the help of two factors with mostly well over 80% (see e.g. Martens, H. / Naes, T. Multivariate Calibration. Chichester: John Eiley & Sons, 1989.). The at Factors used intelligently are not from the substances to be determined themselves calculated; However, due to the similarity of organic substances to each other, it is expected that for each application from the area of organic substances, some under 30 There are universal factors that are very similar to the substances to be determined themselves have calculated factors. The use of all thirty factors to describe one Spectrum was therefore not considered ideal. They would either overdetermine one Spectrum, d. H. Overvalue properties of a spectrum that are not substance-specific, but are only specific to the spectrum under consideration. That includes the so-called noise, a sequence of interferences during the measurement. Or those factors that go beyond that for that considered spectrum must also contain irrelevant information compensate each other for this superfluous information. Therefore and still because of reduced computing effort, the number of factors to be used was set to three. The thus achieved good results as well as tests carried out with two, four and five factors and have achieved significantly worse results, have made this decision as correct approved.

Against this background, it is the task of Intellldent's factor selection to determine the three of the 30 universal factors that are most suitable for a given application. The simplest procedure at this point would be to form all theoretically possible factor triples from the set of 30 universal factors and to check their suitability. The resulting computing effort is considerable. Therefore, as shown in Fig. 16, a two-step process has been developed that shortens the process.

In the first stage, the optimal four first factors are determined, i.e. H. four factors, one of which will be part of the optimal factor triple. Then in a second stage determines which two further factors together with one of the four optimal first factors that result in optimal factor triples. The system requires one for this procedure Evaluation scale with which it can evaluate the quality of a first factor or a factor triplet. Two indicators were developed for this, the Andrassy selectivity (ANSEL) and the Andrassy Stair value (ANSTAIR value), which is discussed in more detail in the following chapter.

1.3.1. Andrassy selectivity and Andrassy Stair value

Substances are determined by assigning their spectra to the spectra of Kalibrationsstoffe. This assignment happens as u. a. in Martens, H. / Naes, T. Multivariate Calibration. Chichester: John Eiley & Sons, described in 1989, in an n-dimensional Factor space. With the help of the Mahalanobis distance it is determined how far spectra are apart.

Table 1.1 shows an example of the result of a determination of five plastics ten calibration substances. The first column contains the calibration substances, i.e. H. the ten Plastics that the system was trained to recognize in this case. The ones to be recognized Fabrics are five of the ten calibration plastics. The values show the Mahalanobis distances of the to be recognized to the calibration substances. The distance to the calibration substance to the substance to be recognized is marked in each case. The shortest distance decides which calibration substance the substance to be recognized is assigned by the system becomes. The plastic 1 (PA6) to be recognized, for example, has a distance from it Calibration plastic 1 of 32.2 and a distance larger from the other calibration substances as 129. It was therefore correctly recognized as plastic 1 (PA6). The plastic 10 to be recognized (ABS) is at a distance of 6.7 from the calibration plastic 10 and from the rest Calibration plastics a distance of at least 9.3. So he would be from the system also be assigned correctly. However, the distances to calibration plastic are 6 (LF fraction) and calibration plastic 9 (PC / ABS blend) of 9.3 or 10.9 not clear  larger than the distance to plastic 10 itself. The plastic to be recognized is therefore from these three plastics are about the same distance away. A clear assignment cannot be performed.

Table 1.1 Average external and internal distances of five out of ten plastics

This example clearly shows that the respective results of the arrangement of the Material spectra in a factor space must be evaluated. For reliable evaluation the individual distances must be related to each other. For this, the Andrassy selectivity (ANSEL) developed, which is formed as follows:

It is assumed that there are n calibration substances, ie the system should subsequently recognize these n substances. Each of these n calibration substances is represented by any number of spectra. All spectra were converted into their factor weights with respect to the previously determined universal factors and were thus projected into the universal factor space. How well a calibration substance sets itself apart from the other calibration substances (counter substances) in the factor space, i.e. to what extent the spectra of a calibration substance under consideration are arranged close to one another and at the same time distant from the spectra of the other calibration substances is schematically shown in Fig. 13:

First of all, a spectrum of the calibration substance under consideration is taken into account. It will be the Mahalanobis distances to each spectrum of a counter substance are calculated and the mean value above educated. This measure is called the "foreign distance" of the substance under consideration to a counter substance. In addition, from this one spectrum to the other spectra of the considered Calibration material, the Mahalanobis distances were formed and the mean value was calculated. This measure is called the "intrinsic distance" of a substance under consideration.

For all other spectra of the calibration substance under consideration, the foreign and Distances determined. The average of all the natural distances of a substance in question gives that averaged self-distance - the average of all external distances of a substance under consideration to one The mean distance between these two substances is given by the counter substance. The averaged External distance is calculated for each of the n-1 counter substances.

A substance can be distinguished from the others if and only if its averaged self-distance is small compared to all averaged external distances. This statement is formulated mathematically in several steps and summarized in the Andrassy selectivity. It is formed according to equation 1.6 from the two key figures, K _rel (1.3) and K _abs (1.5).

K _rel is a measure of the size of the intrinsic distance (D _eigen ) of a substance relative to the external distances (D _{foreign g} ) of the counter substances. The aim of this key figure is to ascertain whether the averaged external distances differ significantly in magnitude from the internal distance. To do this, the averaged self-distance is first individually related to each averaged external distance. This is done according to equations (1.1) and (1.3). In equation (1.1), the quotient REL is formed from the averaged intrinsic and external distance. If the averaged external distance is of the same order of magnitude as the averaged self-distance, this results in a value REL of one. However, if it is of a significantly higher order of magnitude, the REL value goes to zero. If the averaged external distance is of the same or a lower order of magnitude than the averaged self-distance, it is not possible to differentiate between the two associated substances. A distinction between substances is only successful if the averaged external distance is of a higher order of magnitude than the averaged self-distance. Equation (1.2) formulates this relationship. This is a so-called S function that maps all REL values to the interval between 0 and 1. Fig. 17 illustrates the function REL All values that are close one to one or greater, are given a rating assigned krei close to unity. All values REL that go towards zero are assigned a value krei close to zero. If the spectra of a substance in question are in the n-dimensional factor space in the vicinity of another substance, the value for k _rel is accordingly close to one. If they are at a "safe" distance, the value for k _{rel is} close to zero.

The sum K _rel , which is not explicitly averaged, is formed over all values k _{rel of} a separation _task according to equation (1.3). In this way, each counter substance is included in the evaluation equally, regardless of how many counter substances are present.

This approach is based on the premise that a distinction between n substances is only considered successful if each substance can be distinguished from any other of the n-1 substances. It must be considered to have failed if at least two substances cannot be distinguished. Only if the averaged natural distance is significantly smaller than any averaged external distance will the value K _{rel be} close to zero. Since K _{rel is} inversely proportional to the Andrassy selectivity, it becomes very large with a very good distinguishability of all substances under consideration, ie with a very small K _rel . If only an averaged external distance of a similar magnitude as the averaged self-distance, the Andrassy selectivity remains comparatively small.

K _abs takes into account the absolute difference between self-distance and external distance. This is necessary because K _rel cannot differentiate between relatively large values of averaged internal and external distances and relatively small values of averaged internal and external distances. The following example serves to illustrate the problem: For an internal distance of 50 and an external distance of 100, the same relative index REL (0.5) results as for an internal distance of one and an external distance of two. A substance whose spectra are grouped with an average Mahalanobis distance of 50 from each other is more clearly different from a substance whose spectra are on average a Mahalanobis distance of 100 from the spectra of the substance under consideration than a substance with the corresponding ones Distances one and two. To correct this, the absolute index K _abs (equations 1.4 to 1.5) is introduced, which is formed from the average difference of all external distances to the intrinsic distance of a substance.

Any difference between an external distance and the self-distance is greater than 500 is the value 500, since this value is considered sufficient to make two substances clear to be distinguished from each other. Values above this do not mean a gain Selectivity and would unduly affect the rating of selectivity if using their full amount was included in the evaluation.

Table 1.2 shows the ANSEL values and ANSTAIR values for the example shown in Table 1.1. As described above, plastic 1 was recognized very well. For him there is an Andrassy Selectivity of 8.72E + 05. Plastic 2 and plastic 7, as can be seen from Table 1.1, also recognized very well. Their averaged self-distances are 3.8 and 9.8 respectively smaller than for Plastic 1, d. H. they are grouped more closely together with the expected calibration substances than Plastic 1.

Table 1.2

Andrassy selectivities and ANSTAIR values of the example from Table 1.1.

At the same time, however, the smallest averaged external distance, ie the shortest distance to one of the other calibration substances, is lower than with plastic 1: plastic 2 is 97.3 from plastic 7; Plastic 7 68.7 from plastic 6. At 129.8, plastic 1 is further away from the closest counter-material (plastic 7). The distance to the closest counter substance is particularly relevant because the detection of a substance to be determined has already failed if a counter substance is so close to the substance to be determined that it cannot be clearly differentiated from it. In this case, however, the shorter distances to the closest counter substance are more than compensated for by the much closer grouping of the substance under consideration: with ANSEL values of 4.03E + 06 for plastic 2 and 1.93E + 06 for plastic 7 their detection is therefore classified as even better than that of plastic 1. Plastic 9, with an average natural distance of 5.5, is also very close to itself as a calibration substance, but is relatively close to plastic 6 with an average external distance of 11.5. With an Andrassy selectivity of 2.46E + 04, this detection is rated significantly lower than the detection of plastics 1, 2 and 7 . As described above, plastic 10 is recognized only very poorly and, with an Andrassy selectivity of 4.04E + 02, has an even lower rating.

The Andrassy selectivity theoretically extends from 0 to ∞ and in practice is mostly in the range between 1E + 02 and 1E + 08, which makes it relatively confusing. It was also found that the selectivity does not behave linearly over this range. This makes it difficult to estimate the selectivity with the help of the Andrassy selectivity. A continuous function, the so-called Andrassy-Stair function, was therefore developed, which maps the Andrassy selectivity to an interval from 0 to 1.2 and increases approximately linearly with the selectivity of a detection. An ANSTAIR value of 1.2 indicates very good detection, a value of zero a very bad one. Equation (1.7) shows the ANSTAIR function, Fig. 14 shows it graphically.

For reasons of scaling, the count is divided into three areas. The first covers the ANSEL Values from zero to 6.00E + 04, the second ANSEL values from 6.00E + 04 to 4.00E + 05 and the third ANSEL values from 6.00E + 05 to 1.00E + 08 from. The function is characterized by a change between areas with constant ANSTAIR value and those in which the ANSTAIR value above which Andrassy selectivity increases. It consists of nine both steep and different also different levels in the range from zero to 1E + 08. That means the The selectivity of a detection remains constant over wide areas of the Andrassy selectivity only increases in nine relatively small areas.

For the example from Table 1.1, the ANSTAIR values shown in Table 1.2 result: The Plastics 1, 2 and 7 were well recognized and received ANSTAIR values of 1,100, 1,200 and 1.130. Plastic 9 was only recognized with moderate certainty and receives an ANSTAIR Value of 0.300. Plastic 10 was not recognized. The ANSTAIR value for its detection returns 0.010.

The ANSEL and ANSTAIR values have made it possible to determine the selectivity to evaluate a substance reproducibly. This allows different parameters in the Carrying out a substance determination can be compared. So, as in the following described, for example, be examined which factor space for determining a Is the most suitable. The system can perform this analysis using these key figures  carry out independently. They are therefore a necessary prerequisite for automating the entire calibration process.

1.3.2. Determination of the four first factors

As shown in FIG. 16, the factor selection takes place in two steps: first four first factors are determined and then the optimal factor triplet is calculated with these. The four first factors are determined from the universal factors (cf. schematically FIG. 18): First, a universal factor is selected. For this, the Andrassy selectivities of all calibration substances are calculated and then averaged. For example, if there are five calibration substances, substance 1 is considered first. The Andrassy selectivity is formed between it and the four other calibration substances. Then the Andrassy selectivity is determined for substance 2, etc. Thus, the Andrassy selectivity is calculated for all substances with respect to a universal factor, the mean value is formed and this mean value is stored temporarily. The same is repeated for every further universal factor. In this way, an average Andrassy selectivity is assigned to each universal factor. The four universal factors with the four largest Andrassy selectivities are used as primary factors for the calculation of the optimal factor triplet.

At this point in the investigation it would theoretically have been expected to be the only factor with the greatest Andrassy selectivity, since it is the largest possible Depicts the difference between the substances to be distinguished. However, it turned out in further course of the investigations found that the use of the first factor with the largest Andrassy selectivity as one of the factors of the optimal triple factor is not always the maximum provides the result to be achieved. This is attributed to the Andrassy selectivity is somewhat blurred because it offers the infinite possibilities of arranging the Spectra of substances in a factor space must be taken into account. This fact will Taking into account that in addition to the factor with the greatest Andrassy selectivity, the three factors with the next smaller Andrassy selectivities for determining the optimal one Factor triples can still be used.

1.3.3. Determination of the optimal triple factor

In the second stage of the factor selection, the optimal factor triple is determined, which is ultimately for the present use case is to be used. The three universal factors on the basis of which the system is best able to distinguish between the substances at hand.

Fig. 20 shows diagrammatically how this procedure: the averaged Andrassy selectivity of the present substances is formed for each theoretically possible combination of one of the four Erstfaktoren and two of the other 29 Universal factors. For each substance, the Andrassy selectivity to the other substances is calculated and the average of all Andrassy selectivities is calculated. An average Andrassy selectivity can thus be assigned to each triple factor. The four factor triples with the four largest Andrassy selectivities are considered the most suitable for the determination of the substances present and are recorded for further use. When evaluating substances to be determined, it is then determined which of the four triple factors is the optimal one.

1.4. Evaluation of substances to be determined

The evaluation of substances to be determined against the background of the calibration proceeds as follows: As shown in FIG. 22: The substance to be determined, like the calibration substances, is successively projected into the four factor spaces that are spanned by the four optimal factor triples. There, the averaged Mahalanobis distances of the substance to be determined to all calibration substances are determined. The substance to be determined is assigned to the calibration substance with the smallest averaged Mahalanobis distance (GMD). In addition, ANSEL and ANSTAIR values are formed between the smallest GMD and the remaining GMD. The result of the factor space of the four optimal factor spaces with the highest ANSEL value is recorded as the end result. The ANSTAIR value calculated for this can be used to assess the reliability with which the substance to be determined was assigned to one of the calibration substances.

Claims (12)

1. Process for evaluating spectra of a spectrometer for the investigation of solid, liquid and gaseous substances, characterized by the following process steps:

• - Generation of a universal set of factors from the largest possible number of substances that absorb the light used by the measuring device. The factors are obtained using the method of factor decomposition known per se in chemometry.
• - Acquisition of target spectra
• - Determination of the optimal universal factors for the differentiation of the substances to be examined
• - Use of the optimal universal factors for the determination of substances to be examined,
• - whereby the selection of the factors takes place automatically.

2. The method according to claim 1, characterized in that a once determined sentence universal factors is used for many applications. 3. The method according to claim 1 or 2, characterized in that for all measuring devices, that use the same electromagnetic radiation, only one set of universal factors Times is determined and subsequently for all those resulting with these measuring devices Use cases is used. 4. The method according to at least one of the preceding claims, characterized in that the Universal factors in one up to the noise limit of the measuring device used sufficient numbers can be obtained. 5. The method according to at least one of the preceding claims, characterized in that First, a number "r" of universal factors to be determined is determined, which for the Determination of the substances to be examined with only one universal factor gives the best results deliver. These are then added one after the other with a number "m" to be determined Universal factors combined in such a way that all theoretically possible combinations of the initially determined r universal factors and m others. The "n" best Factor combinations are used to determine the substances to be examined, where "n" is a number to be set. 6. The method according to claim 5, characterized in that first the four universal factors be determined for the determination of the substances to be investigated with only one Universal factor deliver the best result. Furthermore characterized by that one after the other from these four universal factors and two further universal factors theoretically possible combinations are formed and the four factor combinations the are best suited for the determination of the substances to be examined. 7. The method according to claim 5 or 6, characterized in that the retained optimal Factor combinations can be used for the determination of substances to be investigated. The Results of the optimal factor combinations are evaluated and the result the factor combination with the best result as the final result. 8. The method according to at least one of the preceding claims, characterized in that the Evaluation of a determination of substances to be examined is carried out by the fact that the distances of the factor weights of the substances determined in relation to the universal factors are related to each other. The calibration is carried out during the process Removal of the factor weights of a substance in relation to each other  Distances of the factor weights of one substance to the factor weights of all other substances. When determining the spectra of an unknown substance, the shortest distance the factor weights of the unknown substance to the factor weights of one of those in the calibration used substance related to the distances of the factor weights of the unknown Substance to the factor weights of all other substances used in the calibration. 9. The method according to claim 8, characterized in that as a distance measure per se known Mahalanobis distance is used. 10. The method according to claim 8, characterized in that the evaluation of a determination of substances to be examined is performed using the Andrassy selectivity. 11. The method according to at least one of the preceding claims, characterized in that the Selectivity of the final result is assessed. 12. The method according to claim 11, characterized in that the selectivity of the Final result is assessed using the ANSTAIR value.