Method for evaluating gas analytical line spectra with a regressor, method for training the regressor and method for providing training data for training the regressor

Abstract

The invention relates to a method for providing training data for training a regressor for evaluating gas analytical line spectra measured with a gas spectrometer, comprising the following steps. - receiving data comprising a plurality of single line spectra, wherein the single line spectra are each line spectra of gases whose composition is known, in that the data also includes the known compositions; - Receiving data corresponding to one or more measuring surfaces; - Generate a variety of training datasets by generating each training dataset as follows: - Providing coefficients; - Calculating a training spectrum by linear combinations of the plurality of individual line spectra and the measurement background or backgrounds, whereby the linear combination is formed using the coefficients; - Calculating a training composition according to the coefficients and the known compositions; - Assigning the training composition to the training spectrum in the training dataset; - Providing the training data as the entirety of the training data sets.

Description

State of the art

[0001] Raman spectroscopy is well-known and is used, among other things, to detect the molecules contained in a gas. Monochromatic light is shone onto the gas to be measured, and the spectrum of the inelastically scattered light is recorded. The spectrum of the scattered light shows a characteristic spectrum for each (polarizable) gas contained in the gas. The spectrum of the gas results from the linear superposition of the individual spectra. In addition to the components of the gas, the evaluation of the spectra can also be used to determine the quantitative concentrations of the gas components.

[0002] When evaluating the spectra, it is common practice to first manually subtract a background signal from the measured signals, which is typically due to both constant and time-varying interference signals. Disclosure of the invention

[0003] The invention is based on the desire to deduce the composition of the gas in a simpler and more accurate manner based on gas analytical line spectra measured with a gas spectrometer.

[0004] For this purpose, the invention assumes the use of a regressor for the evaluation of gas analytical line spectra and the training of the regressor with training data.

[0005] According to the invention, the training data is provided by means of the following steps: - receiving data comprising a plurality of single line spectra, wherein the single line spectra are each line spectra of gases whose composition is known, in that the data also includes the known compositions; - Receiving data corresponding to one or more measuring surfaces; - Generate a plurality of training datasets by generating each training dataset as follows: - Providing coefficients; - Calculating a training spectrum by linear combinations of the plurality of individual line spectra and the measurement background or backgrounds, whereby the linear combination is formed using the coefficients; - Calculating a training composition according to the coefficients and the known compositions; - Assigning the training composition to the training spectrum in the training dataset; - Providing the training data as the entirety of the training data sets.

[0006] The regressor trained with this training data is then able to distinguish between the background in gas analytical line spectra and the parts of the gas analytical line spectra that are causally attributable to the components contained in the gas or that correlate with the concentrations of the components contained in the gas.

[0007] This eliminates the need for manual subtraction of a background signal. Likewise, this background signal no longer needs to be explicitly measured using the inventive approach. For example, so-called calibration measurements, such as those commonly used in end-of-line characterizations or for time-varying interference signals according to the state of the art, can be completely eliminated.

[0008] On the other hand, the inventive approach has the further advantage over methods that would use only single-line spectra as training data that the regressor can be trained more easily and comprehensively, even in areas with a low signal-to-noise ratio and for "unusual" gas compositions. Thus, fewer actual measurements with the gas spectrometer are required to provide a sufficient amount of training data.

[0009] In other words, the regressor can learn better with training data according to the invention to distinguish between a useful signal and an interference signal in the gas analytical line spectra and can thus subsequently better evaluate gas analytical line spectra.

[0010] Coefficients are provided to generate the linear combinations used in the calculation of the training spectra and the training compositions. For example, if the three coefficients are k1, k2, and k3, a linear combination l can be generated from two individual line spectra a and b and a measurement background m according to the following calculation rule: I=k1*a+k2*b+k3*m.

[0011] The coefficients can be "random" coefficients. For example, they can be generated in advance using a corresponding algorithm or generated during the process. They can also be arbitrary coefficients from a given interval.

[0012] The single line spectra may be signals that were previously measured with a gas spectrometer.

[0013] The measurement background(s) may be signals that were previously measured with a gas spectrometer in which no gas was present that is known to generate a measurement signal in that gas spectrometer.

[0014] The gas spectrometer can, in particular, be a Raman spectrometer, as known, for example, from the applicant's subsequently published DE 10 2022 212 942 A1. The gas analytical line spectrum is then a Raman spectrum.

[0015] Data which, in terms of their type and origin, are equivalent but not necessarily identical to the data referred to here as training data, can also be used to check the extent to which a regressor, in particular the regressor trained according to the invention, works as desired for the evaluation of gas analytical line spectra.

[0016] These data can also be referred to as test data and can be used to test a regressor for evaluating gas analytical line spectra measured with a gas spectrometer. These methods for their provision and use are expressly encompassed by the present invention.

[0017] The drawing shows: Fig. 1 a multitude of single-line spectra measured with a Raman spectrometer, Fig. 2 shows an example of a measurement background, as measured by a Raman spectrometer in which there is no gas, Fig. 3 a multitude of single line spectra measured with a Raman spectrometer, on which a measurement background is additively superimposed, Fig. 4 is a flowchart for an embodiment of the present invention.

[0018] Fig. 4 shows a flow chart of a method for evaluating a gas analytical line spectrum.

[0019] It provides that, in steps 1.1 to 1.4, training data is first provided for training a regressor. In detail, the procedure consists of the following steps: - Step 1.1: Receiving data comprising a large number of single-line spectra si, for example, 100 single-line spectra. The single-line spectra are each single-line spectra of gases measured with a known Raman spectrometer, the composition of which zi is known and can be derived from the received data. These can, for example, be single-line spectra as described in the Fig. 1 are shown as examples. - Step 1.2: Receiving data corresponding to a measurement background m or a plurality of measurement backgrounds mj. The measurement backgrounds can, for example, be signals measured with the Raman spectrometer when no gas is present in the spectrometer. For example, it can be a measurement background as in Fig. 2 is evident. - Step 1.3: Generate a large number of training data sets, for example 1000 training data sets, by generating each training data set as follows (steps 1.3.1 to 1.3.4): - Step 1.3.1: Providing coefficients ai. These can, for example, be uniformly distributed coefficients from the interval [0;1], which can be generated using a random algorithm known to those skilled in the art. - Step 1.3.2: Calculating a training spectrum T by linear combinations of the plurality of single line spectra and the measurement background or backgrounds, whereby the linear combination is formed using the coefficients, for example according to the form T = m+ ki* ai. - Step 1.3.3: Calculate a training composition t according to the coefficients and the known compositions, for example according to the form t = ki* zi. - Step 1.3.4: Assigning the training composition t to the training spectrum T in the training dataset; - Step 1.4: - Providing the training data as the entire set of training datasets.

[0020] The regressor is then trained with the training data in method step 2. The regressor can be a regressor known per se to those skilled in the art, for example, a neural network. Training can be carried out using methods known per se to those skilled in the art, for example, a gradient descent method.

[0021] The trained regressor is then ready to evaluate any gas analytical line spectra measured with the Raman spectrometer in step 3, i.e. to infer the composition of a gas from a line spectrum measured on the basis of the gas. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2022 212 942 A1

[0014]

Claims

[1] Method for providing training data for training a regressor for evaluating gas analytical line spectra measured with a gas spectrometer, comprising the following steps: - receiving data comprising a plurality of single line spectra, wherein the single line spectra are each line spectra of gases whose composition is known, in that the data also includes the known compositions; - Receiving data corresponding to one or more measuring surfaces; - Generate a plurality of training datasets by generating each training dataset as follows: - Providing coefficients; - Calculating a training spectrum by linear combinations of the plurality of individual line spectra and the measurement background or backgrounds, whereby the linear combination is formed using the coefficients; - Calculating a training composition according to the coefficients and the known compositions; - Assigning the training composition to the training spectrum in the training dataset; - Providing the training data as the entire set of training datasets. [2] Method for training a regressor for evaluating gas analytical line spectra with training data provided according to a method according to claim 1. [3] Method for evaluating gas analytical line spectra with a regressor which has been trained with a method according to claim 2. [4] A method according to any one of claims 1, 2 or 3, wherein the gas spectrometer is a Raman spectrometer and the gas analytical line spectra are Raman spectra.

Images