WO1997023776A1

WO1997023776A1 - X-ray fluorescence analysis utilizing a neural network for determining composition ratios

Info

Publication number: WO1997023776A1
Application number: PCT/IB1996/001316
Authority: WO
Inventors: Hendrik Adrianus Van Sprang
Original assignee: Philips Electronics N.V.; Philips Norden Ab
Priority date: 1995-12-21
Filing date: 1996-11-27
Publication date: 1997-07-03

Abstract

A method of determining the composition ratio of materials (notably mixtures of alloys) by radiation analysis (notably X-ray fluorescence) while utilizing a neural network operation. An X-ray fluorescence diagram is recorded for a number of mixtures of alloys having a known composition ratio, utilizing a detector (18) of limited resolution, said diagram being applied to a neural network. The neural network is adjusted by means of these known composition ratios, after which the composition ratio of mixtures of the same alloys having unknown composition ratios can be determined with this adjustment. It has been found that the method in accordance with the invention still offers suitable results when use is made of an X-ray detector (18) having an energy resolution (FWHM) of approximately 16 % of the detected energy.

Description

X-ray fluorescence analysis utilizing a neural network for determining composition ratios.

The invention relates to a method of determining the composition ratio of a composition of chemical elements to be analyzed, which method includes the recording of a radiation characteristic of the composition to be analyzed.

The invention also relates to a method of adjusting internal transform parameters of a neutral network operation in order to determine the composition ratio of a composition of chemical elements to be analyzed.

The invention also relates to an apparatus for radiation analysis of a composition of chemical elements.

In the context of the present invention a composition of chemical elements is to be understood to mean a single combination of chemical elements, a single alloy, a mixture of chemical combinations, a mixture of chemical compounds, or a mixture of alloys.

Methods of determining the composition ratio of a composition of chemical elements to be analyzed are known per se. These generally known methods utilize radiation analysis of the compositions during which a sample of the material to be analyzed is irradiated in order to obtain a radiation characteristic specific of the composition of the relevant material. Examples of such methods are:

* optical analysis, in which a sample is irradiated by infrared, visible or ultraviolet analysis radiation and in which a radiation diagram, for example a fluorescence diagram, of the relevant material is recorded in response to the analysis radiation applied; * X-ray fluorescence analysis, in which a sample is irradiated by analysis radiation in the form of X-rays and in which a radiation diagram, for example an X-ray fluorescence diagram, of the relevant material is recorded in response to the analysis radiation applied;

* diffraction analysis, in which a sample is irradiated by analysis radiation in the form of X-rays, neutron rays or electron rays, and in which a diffraction diagram, for example an X-ray diffraction diagram, of the relevant material is recorded in response to the analysis radiation applied.

A known method will be now described in detail on the basis of X-ray fluorescence. A sample of the material to be examined is irradiated by polychromatic X-rays which, generally speaking, originate from an X-ray tube. In response to this radiation, the sample emits fluorescence radiation which is wavelength-analyzed by means of a wavelength- selective detector. A relationship is thus found between the intensity of the fluorescence radiation generated in the sample and its wavelength (the fluorescence diagram), said relationship being characteristic of the nature of the composition examined. Notably the maximum values ("peaks") occurring in this diagram are characteristic of the chemical elements occurring in the sample, but the mutual influencing of the substances present in the sample also has an effect on the appearance of the diagram. This is due to the fact that fluorescent radiation generated in a given element of the sample excites another element present in the sample, thus generating fluorescent radiation which in its turn can excite other elements, etc. This mutual influencing makes interpretation of such a diagram difficult and even impossible in many circumstances.

A further complicating factor for the interpretation of said diagrams is the limited resolution in the wavelength with which these diagrams are recorded. This resolution is determined by the wavelength-analyzing element (the detection device) in the radiation analysis apparatus recording said diagrams. A few groups of wavelength analysis are distinguished, i.e. wavelength-dispersive analysis and energy-dispersive analysis, and in the latter group analysis by means of solid-state detectors and analysis by means of gas-filled detectors are distinguished again.

In a detector of the wavelength-dispersive type each photon is converted into an electric pulse whose pulse height and/or charge content are not discriminated in principle. Thus, in this detector exclusively the number of photons is determined. Such a detector is formed, for example as an assembly of an analysis crystal and an X-ray count tube. The radiation emanating from the sample is incident on the analysis crystal. In accordance with the known Bragg relation, this crystal reflects only approximately one wavelength, i.e. the wavelength associated with the angle of incidence (and a close vicinity thereof, for example 0.05°) of the selected radiation. The entire desired interval of angles of incidence is traversed, and hence also the associated range of wavelengths, by rotating the analysis crystal during the measurement. The relationship between the radiation intensity (being proportional to the count rate of the count tube) and the wavelength is thus established. Because the radiation applied to the analyzer crystal must be extremely parallel, this crystal is preceded by a collimator, for example a Soller slit. One consequence of the parallelization of the radiation emanating from the sample is that its intensity is strongly reduced, so that long measuring times, and sometimes additional steps, are required so as to enable the use of a wavelength-dispersive detector. Moreover, a comparatively expensive goniometer device is required for accurate angular adjustment of the analysis crystal and the detector movement coupled thereto. For each photon absorbed in an energy-dispersive detector the detector supplies a current pulse whose charge content equals the energy of the photon. These current pulses can be electronically selected in respect of charge content, so that the number of current pulses of a given charge content (i.e. the intensity) can be determined in dependence on the charge content (i.e. the energy of the photon) during one measuring time for all current pulses. Because the energy of an X-ray photon is inversely proportional to the wavelength of the radiation, the intensity of the X-rays incident on the detector is thus determined as a function of the wavelength. The solid state detectors of this type include the Si(Li) detector. Even though this detector has a rather attractive signal-to-noise ratio in comparison with other energy-dispersive detectors (such as the gas-filled detector), this ratio is still comparatively high for a small charge content (i.e. long X-ray wavelengths). This is due to the fact that the deviation in the charge content Q for one given photon energy is proportional to Λ Q; thus, this effect increases for low values of Q.

An energy-dispersive detector of the gas-filled type has the generally known property of a wavelength resolution of the order of magnitude of 15% of the energy to be detected which corresponds to this wavelength. This resolution is substantially lower than that of the two previously mentioned detectors. If an X-ray fluorescence diagram were recorded by means of the latter detector, the peaks situated near one another in this diagram could no longer be observed directly individually, thus making interpretation of a diagram thus formed impossible unless special steps are taken. An energy-dispersive detector of the gas-filled type is substantially less expensive than an energy-dispersive detector of the solid-state type, which in its turn is less expensive than a detector of the wavelength-dispersive type. This means that an increasing price must be paid for increasing resolution. Moreover, high-resolution analysis apparatus are often considerably more voluminous and heavier than low-resolution apparatus. If the composition ratio of a composition were to be determined by means of the known techniques and detectors, a first step of the procedure to be carried out would then be the recording of a radiation diagram. The proportions of the various chemical elements in the material to be examined should be determined on the basis of the location and the height of the peaks in the radiation diagram. It is assumed that it is known which types of compositions occur in the material to be analyzed. For example, the material to be analyzed is a mixture of wear products of a machine in its lubricant. It is known which metal alloys (i.e. which chemical ratio formules) occur in this machine, so which alloys are to be expected in the sample. However, the ratio of each of these alloys in the sample is unknown. Because the alloys (and hence the associated chemical ratio formules) are known, it could be attempted to fit the measured quantities of chemical elements in the ratio numbers stemming from the ratio formules. In the case of somewhat large numbers of alloys and elements (for example, larger than three, depending on the wavelength resolution), the execution of this determination, however, becomes unreliable and usually even impossible.

It is an object of the invention to provide a method of determining the composition ratio of a composition of chemical elements to be analyzed, which method enables the composition ratio of a comparatively large number of compositions and elements to be determined by means of comparatively inexpensive means, notably inexpensive radiation detectors.

According to a first aspect of the invention, the method of determining the composition ratio of a composition of chemical elements to be analyzed utilizes a neural network operation involving internal transform parameters adjusted for compositions of said chemical elements, the method also including the steps of:

* recording a radiation characteristic of the composition to be analyzed, which characteristic is reproduced as a set of sample points,

* subjecting the set of sample points of each characteristic recorded to a neural network operation, * obtaining, as results of the neural network operation, output values which represent the composition ratio to be determined.

The radiation characteristic of the composition to be analyzed can be recorded as a continuous diagram in a conventional radiation analysis apparatus; in such a case the diagram must be sampled so as to enable the information of the diagram to be subjected to the neural network operation. The number of sample points must be equal to the number of inputs of the neural network for carrying out the operation. However, there are also radiation analysis apparatus in which the radiation characteristic is available directly in digital form, so that sampling of the radiation characteristic is not necessary. In the context of the present invention a neural network operation is to be understood to mean an operation as it is performed by a neural network which is known per se. Generally speaking, a neural network consists of a given number of inputs whereto in this case the sample points are applied, and of a number of outputs on which in this case the ratio numbers representing the composition ratio appear. When the sample to be analyzed consists of n different components (for example, alloys), the number of outputs chosen will be equal to n-1, because the composition ratio is given by n ratio numbers, the sum of the ratio numbers being 100%.

If it is known which types of compositions occur in the material to be analyzed, the adjustment of the neural network can be based on this data. In conformity with a second aspect of the invention, a method of adjusting internal transform parameters of a neural network operation in order to determine the composition ratio of a composition of chemical elements to be analyzed includes the following steps: a) recording a radiation characteristic of each of a number of compositions having a known composition ratio, which characteristic is reproduced in the form of a set of sample points; b) subjecting the set of sample points of each characteristic recorded to a neural network operation, c) adjusting internal transform parameters of the neural network operation in dependence on the deviation between the output values resulting from the neural network operation, applied to the set of sample points of each of the characteristics, and the associated known composition ratios, d) repeating the steps b) and c) until the deviation between the output values resulting from the neural network operation, applied to the set of sample points, and the associated known composition ratios is below a predetermined threshold value.

The metal alloys appearing as wear products in the lubricant of a machine may again be taken as an example of the material to be analyzed, it being assumed that m different metal alloys could occur. On the basis of these known metal alloys, a number of samples of known composition ratio is prepared. This is preferably performed in such a manner that this ratio approximates as well as possible the composition ratios occurring in practice. Of each of these compositions of known composition ratio the radiation characteristic is recorded, sampled (if necessary) and applied to the neural network. Each set of sample points then yields a set of output values of the neural network which represent the ratio numbers of the composition. The ratio numbers obtained are compared with the known ratio numbers, after which the transform between the nodes of the neural network (the internal transform parameters) can be readjusted on the basis of the deviation between the ratio numbers obtained and the known ratio numbers. This procedure is repeated until the output values obtained suitably correspond to the known ratio numbers. It is to be noted that the recording of the radiation characteristics for adjustment of the neural network need not necessarily take place by actual irradiation of the sample. For this purpose use can also be made of Computer programs for simulating the irradiation effect which also yield the desired radiation characteristics. Therefore, "recording a radiation characteristic" is also to be understood to include the formation of radiation characteristics in this manner.

The above neural network operation can be performed by means of an electronic circuit specifically designed for use as a neural network. An embodiment of an . apparatus for carrying out the invention includes an arithmetic and processing unit which is constructed as a general purpose computer, loaded with a program for simulating a neural network.

This embodiment offers the advantage of substantially greater flexibility in comparison with a permanent electronic circuit. The number of inputs of the neural network can then be simply varied, like the number of outputs. High speed and flexibility can be achieved notably in combination with a simulation program for the recording of a radiation characteristic as described above.

These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

In the drawings:

Fig. 1 shows a relevant part of an X-ray analysis apparatus for recording an X-ray fluorescence diagram for use in accordance with the invention, including an arithmetic and processing unit for processing the data from the X-ray fluorescence diagram; Fig. 2a shows the spectral lines produced by a first single composition in response to irradiation by X-rays;

Fig. 2b shows the spectral lines produced by another single composition in response to irradiation by X-rays;

Fig. 2c shows the spectral lines produced by another single composition yet in response to irradiation by X-rays;

Fig. 3 shows the spectral lines produced by a composition in the form of a mixture in response to irradiation by X-rays, together with the detector response for various values of the energy resolution of the detector; Fig. 4a shows the deviation of the measured proportion of a first constituent in a mixture with a large number of composition ratios with respect to the actual proportion, utilizing a comparatively high energy resolution of the detector;

Fig. 4b shows the deviation of the measured proportion of a second constituent in a mixture with a large number of composition ratios with respect to the actual proportion, utilizing a comparatively high energy resolution of the detector; Fig. 5 shows the spectral lines produced by a single composition in response to irradiation by X-rays, together with the detector response for a comparatively high value and a comparatively low value of the energy resolution of the detector;

Fig. 6a shows the deviation of the measured proportion of a first constituent in a mixture with a large number of composition ratios with respect to the actual proportion, in the case of the comparatively low energy resolution of the detector as shown in Fig. 5;

Fig. 6b shows the deviation of the measured proportion of a second constituent in a mixture with a large number of composition ratios with respect to the actual proportion in the case of the comparatively low energy resolution of the detector as shown in Fig. 5.

Fig. 1 shows a relevant part of an X-ray analysis apparatus in which an X-ray fluorescence diagram can be recorded for use in accordance with the invention and in which the data from the X-ray fluorescence diagram can be further processed as will be described in detail hereinafter with reference to the subsequent Figures. An X-ray source 2 emits an X-ray beam 4 which is incident on a sample 6 to be analyzed. In the sample 6 the X-ray beam 4 excites X-rays which are to be analyzed according to wavelength. Downstream from the sample to be analyzed there is arranged a beam stopper 8 which prevents the X-ray beam 10 emanating from the sample from being scattered to the vicinity of the detector, and transmits only radiation directed onto the entrance of the X-ray detector 12. The detector 12 measures the intensity of the radiation thus selected. For the detector 12 an energy-dispersive detector is chosen, for example an energy-dispersive detector of the gas-filled type. The X- ray fluorescence diagram is determined by measuring the variation of the intensity in dependence on the wavelength. To this end, the X-ray analysis apparatus shown in Fig. 1 includes an arithmetic and processing unit 14, 16 which is connected to the detector 12. The photons incident on the detector 12 each time produce a current pulse whose charge content equals the energy of the incident photon. These current pulses can be selected according to charge content by the arithmetic and processing unit 14, 16, thus determining the number of current pulses of a given charge content (i.e. intensity) in dependence on the charge content (i.e. the energy of the photon). Because the energy of an X-ray photon is inversely proportional to the wavelength of the radiation, the intensity of the X-rays incident on the detector is thus determined as a function of the wavelength. The arithmetic and processing unit 14, 16 is also arranged to produce sample points of the radiation characteristic of the composition to be analyzed and to carry out a neural network operation on the sample points produced. The arithmetic and processing unit is also arranged to yield, as results of the neural network operation, output values which represent the composition ratio to be determined. The arithmetic and processing unit 14, 16 is constructed as a general purpose computer which is loaded inter alia with software for simulating a neural network.

The method of the invention is carried out by means of two computer programs which were applied to simulate the steps of the method in accordance with the invention. One of these computer programs is known as "FP-multi" and commercially available from Philips Nederland, identification No. 9430 977 00361. The latter computer program was first used to simulate the generating of X-rays as produced by a conventional X-ray tube. The data as used in this computer program can be found inter alia in an article by P. A. Pella et al. "An analytical algorithm for calculation of spectral distributions of X-ray tubes for quantitative X-ray fluorescence analysis", X-Rav Spectrometrv. pp. 125-135, Vol. 14, No. 3, 1985. In conjunction with the application of this computer program for the present simulation an X-ray tube was selected with a rhodium anode, operating with a voltage of 35 kV and a current of 50 mA. The radiation from the tube thus operated consists of a continuous part on which a line spectrum characteristic of the rhodium anode is superposed.

Using the X-rays thus obtained, a number of (imaginary) samples with each time a different composition ratio was irradiated during the computer simulation. The samples were formed by mixtures of three different materials (alloys) as can be found in the lubricant of a machine utilizing parts which are subject to wear and consist of these alloys. Therefore, the lubricant of such a machine contains a mixture of said materials as wear products of these parts. The alloys used by way of example are stated in the below table. In the first column ("No. "), the sequence number (Al, A2 or A3) of the alloy is stated, and in the last column ("comp. formula") the approximated composition of the relevant alloy. The second through eighth columns state the percentage by weight of the chemical element stated in the column (a metal) in the relevant alloy. All alloys stated in the table are materials customarily used in the manufacture of machines, for example jet engines for airplanes.

No. Fe Mn Cr Mo Ni V w comp. formula

Al 75.1 0.2 4.5 0.4 0 1.5 18.3 Fe₁₅WCrV

A2 90.3 0.25 4.1 4.25 0.1 1.0 0 Fe₈₂Cr₄V

A3 86.0 0.3 3.0 0 0.7 0 10.0 Fe₂₇WCr

Imaginary mixtures with various composition ratios were formed from the alloys stated in the above table. The various steps of the method in accordance with the invention were performed with these mixtures. To this end, irradiation of the mixtures of the above alloys was simulated using the radiation simulated with the described rhodium-anode X-ray tube. For each mixture this simulated irradiation produced an X-ray fluorescence diagram which was used for further processing in a neural network to be described hereinafter. Such formation of an X-ray fluorescence diagram is described in an article by D.K.G. de Boer "Calculation of X-ray fluorescense intensities from bulk and multilayer samples", X-Rav Spectrometrv. Vol. 19, pp. 145-154, 1990. By way of example, the Figs. 2a, 2b and 2c show the X-ray fluorescence diagram of the pure (i.e. non-mixed) alloys Al , A2 and A3, respectively. These diagrams show the relative intensity (in arbitrary units) as a function of the wavelength of the fluorescent radiation. An example of an X-ray fluorescence diagram of a mixture of said alloys will be described in detail hereinafter with reference to Fig. 3.

Neural networks are generally known. Such a network consists of a number of layers of network nodes (neurons) having a given transform characteristic whose value can be adjusted in dependence on the learning state of the network. Each neuron of a layer is connected to each neuron of the subsequent layer. The network used for carrying out the invention consisted of nine inputs, followed by a first layer of 13 neurons. Each input was connected to each of the neurons of the first layer. This first layer was succeeded by a second layer of 9 neurons, which itself was succeeded by a third layer of two neurons. The latter two neurons also formed the two outputs of the neural network. The transform characteristic of the first and the second layer was formed by a so-called sigmoid function. The transform characteristic of the third layer evidently was linear because this layer must be capable of yielding the output values of the network. For a description of the transform characteristics reference is made to a publication by John H. Murphy "Tutorial: Neural

Network Learning Algorithms", Proceedings of the Second Workshop on Neural Networks. WNN-AIND 91, pp. 135-142, (SPIE 1515). The cited article also describes how the value of the transform parameters is adjusted as part of the learning process as it occurs in a neural network. The learning process used for carrying out the invention is described as "Backpropagation Learning" (pp. 139-140) in the cited article. The neural network used for carrying out the invention was implemented on a general purpose computer by means of a second one of the above two computer programs used for the simulation of the steps of the method of the invention.

Fig. 3 shows the spectral lines produced by irradiation of a sample by X- rays emanating from an X-ray tube including a rhodium anode as described above. The intensity is shown in arbitrary units in this Figure. The irradiated sample is a mixture of said three alloys Al, A2 and A3, having a composition ratio of 0.3 part Al , 0.6 part A2 and 0.1 part A3, written as 0.3*A1 +0.6*A2+0.1*A3. This Figure also shows the detector response for three values of the energy resolution of the detector. Due to the limited resolution of the detector, the very narrow spectral line in the detector response is represented as a Gaussian curve having a given deviation σ, the deviation a being related in known manner to the FWHM (Full Width at Half Maximum) as FWHM = 2.35σ. The solid line in Fig. 3 corresponds to a resolution σ = 0.1 keV, the dashed line to a resolution σ = 0.15 keV, and the dash-dot line to a resolution σ = 0.25 keV. The neural network is trained by means of the above detector responses, i.e. the internal transform parameters of the neural network operation are adjusted thereby, i.e. the value of the transform characteristic. To this end, 121 radiation characteristics of each time other compositions of known composition ratio are recorded. The 121 compositions are formed by all possible combinations of the alloys Al , A2 and A3, for each of the three alloys each time steps of 0.1 being made in the concentration in the range from 0. to 1 ; evidently, values for which the requirement that the sum of the concentrations must be equal to 1 was not satisfied were excluded. During the simulation, the compositions thus formed were subjected to irradiation by X-rays from the described X-ray tube, yielding 121 diagrams with spectral fluorescence lines which are comparable to those in Fig. 3. Each of these diagrams was subsequently applied to a detector having a resolution σ = 0.25 keV ( = 0.6 keV FWHM), thus yielding the desired radiation characteristic. Thus, 121 curves comparable to the dash-dot curve of Fig. 3 were obtained. Each of the 121 curves was subsequently sampled in 9 sample points, so that a set of 9 sample points was formed for each curve, i.e. for each composition ratio. The sample points should preferably be selected in such a manner that maxima of diverse appearance are covered by the sample points. In the present embodiment the sample points 4.95 keV, 5.41 keV, 5.70 keV, 5.94 keV, 6.40 keV, 7.05 keV, 7.50 keV 8.39 keV and 9.67 keV were chosen. Each set of sample points was subsequently applied to the 9 inputs of the neural network in order to subject the set of sample points of each characteristic recorded to a neural network operation. In response to the applied set of sample points, two values then appear on the two outputs of the neural network, which two values can be interpreted as two of the three concentration values of the relevant sample. (The third concentration value is determined from the condition that the sum of the three concentrations must be equal to 1 ("must be unity")). These output values can be compared with the known concentrations and, in dependence on the deviation between the output values and the associated known composition ratios, the value of the transform parameters of the neurons of the neural network can be adjusted. The learning process used for this purpose is described as "Backpropagation Learning" in the cited article by Murphy. The above training procedure is repeated until the deviation between the output values obtained and the known composition ratios is below a predetermined threshold value.

After execution of the training procedure, a test was performed so as to determine whether the neural network was indeed capable of recognizing arbitrary composition ratios. To this end, 109 radiation characteristics were recorded of each time different compositions with a known composition ratio. The 109 compositions were formed by all feasible combinations of the alloys Al, A2 and A3; for each of the three alloys steps of each time 0.1 were made in the concentration in the range from 0.05 to 0.95; evidently, values for which the requirement that the sum of the concentrations must be equal to 1 was not satisfied were excluded. The 109 radiation characteristics were subjected to the same sampling procedure as the training sets. The sets of sample points (the test sets) thus obtained were applied to the inputs of the neural network and the values obtained in response to the test sets applied were compared with the known concentrations in the test compositions. The results of this test are given in the Figs. 4a and 4b. Fig. 4a shows the deviation of the measured concentration of the alloy Al in the mixture with respect to the actual concentration, for each of the composition ratios, denoted by the sequence numbers 1 to 109 in this Figure. Fig. 4b shows the corresponding result for the alloy A2. These Figures reveal that in practically all cases the absolute value of the deviation is less than 0.002, and is less than 0.003 in all cases.

A further embodiment of the invention consists of a method of adjusting the neural network and determining the composition ratio of the composition to be analyzed while the energy resolution of the detector is much smaller than in the case described above. According to this method, not a fixed resolution of the detector was chosen, but a resolution amounting to FWHM = 16.4% (σ = 7%) of the energy to be measured. This resolution suitably corresponds to that of a gas-filled detector; this is compatible with the object of the invention which aims to determine composition ratios by means of comparatively inexpensive radiation detectors. Fig. 5 shows the spectral lines produced by irradiation of a sample, consisting of only the alloy Al, by means of X-rays from an X-ray tube as described above. The intensity is given in arbitrary units in this Figure. The Figure also shows the detector response for two values of the energy resolution of the detector. The solid line corresponds to a resolution σ = 2% (FWHM = 4.7%) and the dashed line to a resolution σ = 1% keV (FWHM = 16.4%).

Using the 121 compositions described with reference to Fig. 3 and a detector resolution σ = 7% keV (FWHM = 16.4%), 121 radiation characteristics are recorded again and these radiation characteristics are sampled again as described with reference to Fig. 3. The training of the neural network is again entirely as described with reference to Fig. 3. Finally, after completed training of the network, 109 test sets are recorded again with a detector resolution a = 7%, after which the results of these test sets are compared again with the known concentrations in the test compositions. The results of this test are given in the Figs. 6a and 6b. Fig. 6a shows the deviation of the measured concentration of the alloy Al in the mixture with respect to the actual concentration for each of the composition ratios, denoted by the sequence numbers 1 to 109 in this Figure. Fig. 6b shows the corresponding result for the alloy A2. These Figures show that even when a detector resolution a = 7% is used, the absolute value of the deviation is less than 0.005 in substantially all cases, demonstrating that reliable determination of the composition ratio of a composition of chemical elements to be analyzed is very well possible while using a neural network operation and comparatively inexpensive radiation detectors.

Claims

CLAIMS:

1. A method of determining the composition ratio of a composition of chemical elements to be analyzed while utilizing a neural network operation with internal transform parameters adjusted for compositions of said chemical elements, which method includes the steps of: * recording a radiation characteristic of the composition to be analyzed

(Fig. 3), which characteristic is reproduced as a set of sample points;

* subjecting the set of sample points of each characteristic recorded to a neural network operation; and

* obtaining, as results of the neural network operation, output values which represent the composition ratio to be determined.

2. A method of adjusting internal transform parameters of a neural network operation in order to determine the composition ratio of a composition of chemical elements to be analyzed, including the steps of: a) recording a radiation characteristic (Fig. 3) of each of a number of compositions having a known composition ratio, which characteristic is reproduced in the form of a set of sample points; b) subjecting the set of sample points of each characteristic recorded to a neural network operation; c) adjusting internal transform parameters of the neural network operation in dependence on the deviation between the output values resulting from the neural network operation, applied to the set of sample points of each of the characteristics, and the associated known composition ratios, d) repeating the steps b) and c) until the deviation between the output values resulting from the neural network operation, applied to the set of sample points, and the associated known composition ratios is below a predetermined threshold value.

3. A method as claimed in Claim 1 or 2, in which the composition of chemical elements to be analyzed is formed by a mixture of materials.

4. A method as claimed in Claim 3, in which the materials are metal alloys.

5. A method as claimed in Claim 1 or 2, in which the radiation characteristic to be recorded for the composition to be analyzed is formed by an X-ray fluorescence spectrogram.

6. An apparatus for radiation analysis of a composition of chemical elements, including

* a sample location for mounting a sample (6) of the composition of chemical elements to be analyzed,

* a radiation source (2) for irradiating the sample location by means of analysis radiation (4) required for the recording of a radiation characteristic, * a detector (18) for detecting the radiation (16), characterizing the sample

(6) of the composition to be analyzed, in response to the analysis radiation (4) applied,

* an arithmetic and processing unit (24, 26), connected to the detector (18), for producing sample points of a radiation characteristic of the composition to be analyzed, and for executing a neural network operation on the sample points produced, said arithmetic and processing unit being arranged to yield output values as results of the neural network operation, which output values represent the composition ratio to be determined.

7. An apparatus as claimed in Claim 6, formed by an X-ray fluorescence apparatus.

8. An apparatus as claimed in Claim 6, in which the arithmetic and processing unit is constructed as a general purpose computer loaded with a program for simulating a neural network.