Analysis device and method for its technical design as well as computer program product and trained machine learning system

Abstract

The present invention initially relates to a method for the technical design of an analytical device for the spectral analysis of a sample (02). The analytical device comprises a spectrometer having a plurality of measurement channels for selected frequencies according to a trained model. The analytical device is configured to apply a trained reconstruction model to reconstruct a spectrum based on individual measured values ​​(06) recorded with the spectrometer for the selected frequencies. According to the method, a first machine learning system formed jointly by the model of the spectrometer and the reconstruction model is trained (13, 14), whereby the trained model of the spectrometer and the trained reconstruction model are obtained. The invention further relates to a computer program product, a trained machine learning system, and an analytical device.

Description

[0001] The present invention initially relates to a method for the technical design of an analytical device for the spectral analysis of a sample, wherein the analytical device preferably analyzes at least one ingredient of the sample. The analytical device comprises a spectrometer for spectral measurement of the sample. The analytical device is preferably configured to apply a chemometric model to determine a concentration of at least one ingredient of the sample. The ingredient is, for example, a protein or starch. The ingredient is, for example, water, so that a moisture content can be determined by the analytical device. Furthermore, the invention relates to a computer program product and a trained machine learning system, as well as an analytical device for the spectral analysis of a sample.

[0002] DE 10 2020 116 094 A1 relates to a method for calibrating a plurality of identical spectrometers for ingredient analysis. A mathematical model of the identical spectrometers is used to generate a large number of error spectra to improve a regression model.

[0003] WO 2021 / 198247 A1 discloses a method for co-designing hardware and software for the virtual staining of a tissue sample. The method comprises iteratively acquiring multiple sets of training imaging data relating to one or more tissue samples. Each of the multiple sets of training imaging data was acquired using a different image modality from a group of image modalities. Multiple reference images are acquired, representing the tissue samples containing one or more chemical dyes. The multiple sets of training image data are processed using machine learning logic.

[0004] US 11 062 481 B2 relates to a portable device for determining the condition of one or more plants. The device comprises a digital color camera for capturing a color image of the plants within a field of view and a light source for providing broadband illumination for the plants within the field of view.

[0005] DE 10 2021 105 869 A1 discloses a spectral sensor system comprising an array of optical sensors arranged on an integrated circuit, as well as an interface between the plurality of optical sensors and a first processing device. A plurality of sets of optical filters is configured as a layer arranged on a plurality of optical sensors. Each set of optical filters contains a plurality of optical filters. Each optical filter is configured to transmit light in a different wavelength range. The spectral sensor system comprises a processing device containing an artificial neural network configured to correct a spectral response generated by the plurality of optical sensors.

[0006] US 2021 / 0 172 800 A1 relates to techniques for analyzing unknown sample compositions using a prediction model based on optical emission spectra. Initial emission spectra are received corresponding to a training sample comprising several pure elements of known concentrations. Based on these, a plurality of spectral ranges corresponding to the plurality of pure elements of known concentrations are determined. Features associated with a maximum of the spectral range are determined. A prediction model is trained to predict unknown concentrations of a plurality of components of an unknown sample based on an emission spectrum of the unknown sample.

[0007] EP 3 842 788 A1 relates to a spectral sensor for near-infrared spectroscopy, which serves to distinguish and / or detect objects and / or materials. The spectral sensor is designed to operate in a learning mode and an operating mode. In the learning mode, measurements of intensity values ​​of a first spectrum of wavelengths are performed. In the operating mode, measurements of intensity values ​​of a second spectrum of wavelengths are performed. The wavelengths of the second spectrum are selected using a machine learning method.

[0008] US 10 020 900 B2 shows a system architecture for providing spectral information for one or more devices and for providing and using the spectral information in a device.

[0009] The aforementioned state-of-the-art solutions for spectral analysis provide a spectrometer and an evaluation of the spectral measured values ​​recorded with the spectrometer. Both the spectrometer and the method for evaluating the spectral measured values ​​must be highly tailored to the given analytical task in order to produce accurate results. Such a spectral analysis is performed, for example, to determine the ingredients of agricultural products and foodstuffs. A chemometric model, developed using machine learning, for example, is used to evaluate the spectral measured values. The spectrometer must be designed to meet its spectral sensitivities; in particular, the dimensions of the mean wavelengths and bandwidths of the spectrometer's measurement channels.Selecting a spectrometer with a suboptimal configuration will impair the evaluation using the chemometric model. For example, the spectrometer may have too few or unsuitable spectral sensitivities, which limits the determination of constituents. In other cases, the spectrometer may be oversized for the given analytical task, meaning it may have spectral sensitivities that are not required for the analysis. This makes the spectrometer too expensive.

[0010] US 5,435,309 A discloses a method for determining one or more unknown values ​​of at least one known property, such as the concentration of glucose in blood. The method uses a model based on a set of samples with known values ​​of the known properties and a multivariate algorithm. Several wavelength subsets are selected from an electromagnetic spectral range suitable for determining the property for use by an algorithm. The selection of the wavelength subsets is intended to improve the suitability of the model. The selection process uses multivariate search methods that select both predictive and synergistic wavelengths within the used wavelength range. The suitability of the wavelength subsets is determined by a suitability function. A genetic algorithm is used to generate one or more counting spectra.

[0011] DE 10 2008 002 355 A1 describes a method for calibrating a spectrometer for measuring spectral components of a substance in the context of chemometrics. A spectrometer device is used to measure spectral components of the substance to be analyzed. The spectrometer device is calibrated using a calibration unit. A total of spectral components X and the corresponding substance concentrations of the substance to be analyzed are measured and stored as spectrometric measurement data in the form of a multidimensional coefficient vector. From this total of measured spectral components X, physically relevant spectral components X are determined for the respective measurement process. rel extracted using an automated iterative estimation procedure stored in the calibration unit.

[0012] The subsequently published DE 10 2022 130 044 A1 discloses a method for the technical design of an analytical device for the spectral analysis of at least one constituent of a sample. The analytical device comprises a spectrometer and is configured to apply a chemometric model formed by a trained machine learning system to determine a concentration of the constituent based on spectral measurement values ​​of the sample recorded with the spectrometer. Several steps are executed iteratively. These steps include applying a current configuration of a hardware representation of the spectrometer, thereby obtaining spectral training data. The chemometric model is trained using the spectral training data.The chemometric model created by the trained machine learning system is applied to determine a value of a suitability parameter for the current configuration of the spectrometer's hardware representation. The current configuration of the spectrometer's hardware representation is modified.

[0013] The subsequently published DE 10 2022 130 045 A1 discloses a method for the technical design of an analytical device for the spectral analysis of at least one constituent of a sample. The analytical device comprises a spectrometer and is configured to apply a chemometric model formed by a trained machine learning system to determine a concentration of the constituent based on spectral measurement values ​​of the sample recorded with the spectrometer. Several steps are executed iteratively. These steps include applying a current configuration of a model of a hardware component of the spectrometer, thereby obtaining spectral training data. A current configuration of the chemometric model formed by the machine learning system is applied to the spectral training data to determine a value of a suitability parameter for the current configurations of the models.The current configurations of the models are modified using at least the current value of at least one suitability parameter.

[0014] The object of the present invention, based on the prior art, is to be able to adapt the technical design of an analytical device for the spectral analysis of a sample more precisely to the respective given analytical task, whereby the analytical device can be designed in a less complex manner.

[0015] The stated object is achieved by a method according to the appended claim 1, by a computer program product according to the appended independent claim 14, by a trained machine learning system according to the appended independent claim 15 and by an analysis device according to the appended independent claim 16.

[0016] The method according to the invention serves for the technical design of an analytical device for the spectral analysis of a sample. The analytical device to be technically designed should be optimally suited for a given analytical task. The analytical device allows at least one spectral measurement of the sample, so that at least one spectrum of the sample is determined with the analytical device. Preferably, the analytical device is designed for the spectral analysis of at least one component of the sample.

[0017] This spectral analysis determines the concentration of at least one component of the sample based on spectral information measured on the sample. The sample comes from a material or product whose content is to be examined through analysis. The sample can, for example, be taken from the material or product, or the material or product can be fed to the analysis device as a sample. The material or product is preferably an agricultural product, a foodstuff, or a foodstuff. The agricultural product is preferably a harvested crop. The harvested crop is preferably a grain such as corn or wheat, or rapeseed, sugar beet, or soybeans. The harvested crop can also preferably be fruit plants such as peppers, tomatoes, strawberries, etc., or fruit trees such as almond trees.The sample can be formed by the harvested crop itself, as in the case of cereal grains. However, the sample can also be formed by a part of a plant that produces the harvested crop, so that, for example, the leaf of a cereal plant forms the sample. The ingredient is preferably formed by water, a protein, an oil, sugar, salt, starch or crude fiber. The ingredient is preferably formed by a single chemical element, such as nitrogen, phosphorus, potassium, calcium, magnesium, boron, molybdenum, copper, manganese, zinc, iron, chlorine or sulfur. Of particular relevance for the growth of any crop is an appropriate supply of nitrogen, as nitrogen is the most important element of chlorophyll and is therefore essential for optimal metabolism.In addition to nitrogen, other macronutrients are also relevant, such as phosphorus and potassium, which are often added to the soil together with nitrogen as so-called NPK fertilizers. Calcium and magnesium are also relevant, the latter forming the central element in the chlorophyll ring. Depending on the plant type, other nutrients are also often relevant for optimal growth and yield, but these are often required in much smaller quantities and are therefore also referred to as micronutrients. Examples of these are boron, molybdenum, copper, manganese, zinc, iron, chlorine and sulfur. Measuring the concentration of water represents a moisture measurement. The ingredient may also be an undesirable component in the product or harvest being tested, such as a pesticide or fungicide.

[0018] The analysis device comprises a spectrometer for spectral measurement of the sample. The spectrometer can be an NIR spectrometer, a VIS / NIR spectrometer, a VIS spectrometer, or a full-range spectrometer. The spectrometer can have a transmission, transflectance, or reflection design. The spectrometer is preferably designed as a compact spectrometer sensor.

[0019] The spectrometer has a configuration determined by a previously trained model of the spectrometer, for which the spectrometer was manufactured according to this model. This configuration is defined by several measurement channels for selected frequencies, whereby the measurement channels in their technical implementation also cover a preferably narrow frequency range around the respective selected frequency, which is characterized in particular by a half-width of the respective measurement channel. These frequencies are thus determined by the previously trained model of the spectrometer. These frequencies are selected from a measurement spectrum of the spectrometer. If the spectrometer is, for example, an NIR spectrometer, the frequencies are selected frequencies from the near infrared spectrum. The spectrometer allows the measurement of the measurement spectrum only for the selected frequencies, i.e.only for the technically realized narrow frequency ranges around the selected frequencies, and is therefore only spectrally thin or sparsely sampled. While the spectral resolution results from the narrowness of the measuring channels, namely from the smallest representable spectral distance between two adjacent measuring channels, the spectrometer here is characterized by spectrally sparse sampling, namely from a particularly small number of selected and individually adapted measuring channels. In this respect, it differs from other spectrometers, which can measure their respective measuring spectrum completely in spectral terms or with many measuring channels using spectrally dense sampling. By limiting the range to a small number of selected frequencies, the spectrometer can only spectrally record a portion of the measuring spectrum. This portion is preferably less than half and more preferably less than a quarter.The measurement spectrum extends from the frequency of the measurement channel available for the spectrometer with the lowest frequency to the frequency of the measurement channel available for the spectrometer with the highest frequency. In the technical implementation, this includes the narrow frequency range around the lowest frequency and the narrow frequency range around the highest frequency. The frequency of the measurement channel available for the spectrometer with the lowest frequency and the frequency of the measurement channel available for the spectrometer with the highest frequency are among the selectable frequencies, but not necessarily among the selected frequencies. The measurement channels of the spectrometer are implemented primarily by the spectrometer's hardware. Accordingly, the spectrometer's hardware was manufactured according to the previously trained model of the spectrometer.

[0020] The analyzer is also configured to apply a previously trained reconstruction model to reconstruct a spectrum based on individual sample measurements recorded with the spectrometer for the selected frequencies. The reconstruction model is used to determine the spectrum previously measured with the spectrometer across the entire measurement spectrum, using only the individual measurement values ​​recorded with the spectrometer for the selected frequencies. The spectrum thus determined, referred to as the reconstructed spectrum, should be as close as possible to the spectrum previously measured with the spectrometer—i.e., the actual spectrum of the sample—in terms of its relevant properties.

[0021] The analytical device is further preferably configured to apply a chemometric model to determine a concentration of at least one constituent of the sample based on the spectrum reconstructed using the reconstruction model. The spectrometer records spectral measurement values ​​of the sample in its measuring channels, which are determined by the constituents of the sample. By applying the chemometric model, the concentration of the at least one constituent to be analyzed is determined from the reconstructed spectrum. The result is a value for the concentration of the at least one constituent of the sample.

[0022] By means of the method according to the invention, the model of the spectrometer, the reconstruction model and preferably also the chemometric model are technically designed or developed so that they form the objects of a technical development process, as a result of which the model of the spectrometer, the reconstruction model and optionally also the chemometric model are available in a mature state.

[0023] In one step of the process, an initial configuration of the spectrometer model is selected. This configuration defines the spectrometer's measurement channels. The initial configuration represents a starting point for the technical development process, which is subsequently modified during the technical development process. The result of the technical development process is a final configuration of the spectrometer model, which serves as a technical specification for producing at least one spectrometer, which forms a component of the analytical device to be manufactured. The spectrometer model includes at least one trainable parameter.

[0024] In a further step of the process, an initial configuration of the reconstruction model is selected. This configuration defines a rule for reconstructing the actual spectrum measured with the spectrometer from the individual measured values ​​recorded with the spectrometer for the selected frequencies. The initial configuration represents a starting point for the technical development process, which is modified during the technical development process. The result of the technical development process is a final configuration of the reconstruction model, which is implemented in the analysis device to be manufactured. The reconstruction model includes at least one trainable parameter.

[0025] In a further step of the method, a first machine learning system formed jointly by the spectrometer model and the reconstruction model is trained, whereby the trained spectrometer model and the trained reconstruction model are obtained. The first machine learning system preferably comprises an artificial neural network or a linear model, which is preferably formed by a partial least squares regression model. By training the first machine learning system, the at least one trainable parameter of the spectrometer model and the at least one trainable parameter of the reconstruction model are modified.

[0026] The method according to the invention is characterized in that the spectrometer, the reconstruction model, and, if applicable, also the chemometric model are technically designed or developed together. At least the spectrometer and the reconstruction model together form the objects of a single technical development process. The hardware forming the spectrometer and the software formed by the reconstruction model and, if applicable, also by the chemometric model are developed jointly, which can be referred to as co-design of the hardware and software of the analytical device. This co-design represents a significant difference from the prior art, according to which the spectrometer is first technically designed individually, for which the available knowledge is used to the best possible extent.According to the state of the art, the next step involves developing the chemometric model based on the technically fully designed spectrometer, for example, using machine learning. In contrast, in the method according to the invention, the spectrometer, the reconstruction model, and preferably also the chemometric model are developed jointly, thus ensuring that the spectrometer, the reconstruction model, and, if applicable, the chemometric model all work together to solve the given analytical task in the best possible synergistic way.

[0027] A particular advantage of the method according to the invention is that the designed analytical device, with its components—the spectrometer, the reconstruction model, and, if applicable, the chemometric model—is very precisely adapted to the given analytical task. This ensures, on the one hand, that the designed analytical device can perform the given analytical task very precisely. On the other hand, the spectrometer, with its comparatively few measurement channels for selected frequencies, is far less complex than a reference spectrometer according to the gold standard, which spectrally records the entire measurement spectrum. The spectrometer designed as a result of the method is optimized and, in particular, has only those measurement channels with the respective bandwidths and mean wavelengths required for the given analytical task.Many analytical tasks require only a small number of measurement channels, so this number is significantly reduced compared to the number of measurement channels of the reference spectrometer according to the gold standard. This massively reduces the costs of manufacturing the analyzer's hardware. A further advantage is that, due to the reduced number of measurement channels, analyses can be performed more quickly and require less energy. Accordingly, less data needs to be transferred.

[0028] The chemometric model is preferably formed by a previously trained second machine learning system. Accordingly, the chemometric model comprises at least one trainable parameter. Training preferably occurs during the execution of the method according to the invention, but this training can also be carried out independently of the method according to the invention. By training the second machine learning system, the at least one trainable parameter of the chemometric model is modified. The second machine learning system preferably comprises an artificial neural network or a linear model, which is preferably formed by a partial least squares regression model. In these embodiments, the method preferably comprises further steps.In one of these further steps, an initial configuration of the chemometric model formed by the second machine learning system is selected. This initial configuration defines a rule for determining the concentration of at least one ingredient to be analyzed based on the reconstructed spectrum. The initial configuration, in turn, represents a starting point for the technical development process, which is modified during the technical development process. The result of the technical development process is a final configuration of the chemometric model formed by the second machine learning system, which is implemented in the analytical device to be manufactured. In a further step, the chemometric model formed by the second machine learning system is trained, whereby the chemometric model formed by the trained second machine learning system is obtained.

[0029] The first machine learning system and / or the second machine learning system are preferably trained using supervised learning. Supervised learning is preferably based on regression and / or classification. Regression is preferably used for continuously changing spectral data or values ​​of concentrations of an ingredient. An example of this is the concentration of water for determining moisture. Classification is preferably used for graded values ​​of concentrations of an ingredient. An example of this is the concentration of salt, which is classified as either sufficient or insufficient. The concentration of salt can also be classified into levels; for example, as: < 1.2 g / kg; 1.2 g / kg; 1.4 g / kg; 1.6 g / kg; 1.8 g / kg; 2.0 g / kg and > 2.0 g / kg.

[0030] Regression is preferably performed using partial least squares, which is known as partial least squares regression (PLS). This regression is robust and suitable when only a few latent variables need to be considered.

[0031] In a further preferred embodiment, the first machine learning system and / or the second machine learning system comprises an artificial neural network, which is preferably formed by a convolutional neural network (CNN) or a transformer neural network. The following parameters are preferably used: 1D convolutions, non-linear activations, 1D pooling operations, self-attention operations, fully connected operations, and / or layer normalizations. A final prediction of continuous spectral values ​​in the form of a reconstructed simulation spectrum or of values ​​of concentrations of an ingredient in the form of simulation measured values ​​of the ingredient's concentrations can be made using regression.The first machine learning system and / or the second machine learning system is preferably trained with the goal of minimizing the Euclidean norm (L2 norm) of a target prediction difference, wherein the target prediction difference is determined by comparing the simulation measured values ​​of the ingredient concentrations with reference ingredient concentration data. A final prediction of graded spectral values ​​or values ​​of concentrations of an ingredient can be made using a classification, for example, implemented as a multi-class classification, for example, by predicting a discrete class number or by predicting a multi-class probability vector. The training can, for example, be carried out in such a way that a multi-class commutation error for discrete decisions or a multi-class cross-entropy for probability vectors across multiple classes is minimized.In another implementation, the ordinal nature of the classes can be taken into account, for example, by minimizing the weighted Cohen's kappa loss. Convolutional neural networks are suitable when a large amount of training data is generated and complex relationships exist.

[0032] A first preferred embodiment comprises the following further steps. In one of these further steps, reference spectral data representing a spectrum of a reference sample is provided. This reference spectral data can, for example, be recorded using a spectrometer according to the gold standard. The steps specified below are carried out iteratively during the training of the first machine learning system. First, a spectral measurement of the reference sample is simulated using the spectrometer according to the spectrometer model currently defined by the first machine learning system. This is done by applying the spectrometer model currently defined by the first machine learning system to the reference spectral data, thereby obtaining simulation measurement data. Simulated measurement values ​​are thus obtained for the currently defined measurement channels.The simulation measurement data therefore have a low spectral resolution. The reconstruction model currently defined by the first machine learning system is then applied to the simulation measurement data, resulting in a reconstructed simulation spectrum.

[0033] In a further iterative step, the reconstructed simulation spectrum is compared with the reference spectral data to determine a reconstruction difference. The first machine learning system is trained with the goal of minimizing the reconstruction difference or minimizing or maximizing a quantity derived from the reconstruction difference. The quantity derived from the reconstruction difference is preferably formed by the Euclidean norm of a vector of the reconstruction difference (L2 norm) or by an absolute value of this vector (L1 norm). This first embodiment ensures that the reconstructed spectra in the analysis device to be designed are as close as possible to the spectra of the samples.

[0034] A second preferred embodiment is used when the analytical device is configured to apply the chemometric model to determine a concentration of at least one constituent of the sample based on a spectrum reconstructed using the reconstruction model, and comprises the following further steps. In one of these further steps, reference data is provided for at least one reference sample, wherein the reference data comprises a series of reference spectral data associated with a series of reference ingredient concentration data. The reference spectral data would ideally be acquired from a sample containing the ingredient according to the reference ingredient concentration data. This reference spectral data can, for example, be acquired using a spectrometer according to the gold standard. The steps specified below are carried out iteratively during the training of the first machine learning system.First, a spectral measurement of the reference sample is simulated with the spectrometer according to the spectrometer model currently defined by the first machine learning system. This is done by applying the spectrometer model currently defined by the first machine learning system to the reference spectral data, thereby obtaining simulation measurement data. Simulated measurement values ​​are thus obtained for the respectively defined measurement channels. The simulation measurement data accordingly have a low spectral resolution. The reconstruction model currently defined by the first machine learning system is then applied to the simulation measurement data, resulting in a reconstructed simulation spectrum. The chemometric model is applied to the reconstructed simulation spectrum, thereby obtaining simulation measurement values ​​of the ingredient's concentration.In a further iterative step, the simulation measured values ​​of the ingredient concentration are compared with the reference ingredient concentration data to determine a target prediction difference. The first machine learning system is trained with the goal of minimizing the target prediction difference or minimizing or maximizing a variable derived from the target prediction difference. The variable derived from the target prediction difference is preferably formed by a norm of the target prediction difference, particularly preferably by an Lp norm of the target prediction difference, such as preferably the L2 norm or the L1 norm, or alternatively preferably by the Huber norm of the target prediction difference. The variable derived from the target prediction difference is alternatively preferably formed by an exchange between class divisions of the target value.In this second embodiment, it is ensured that the measured values ​​of the concentration of at least one ingredient in the sample determined with the analytical device to be designed are as close as possible to the actual concentration.

[0035] In this second embodiment, the chemometric model can be formed by the second machine learning system, but the chemometric model can also have been determined in a different way. Insofar as the chemometric model is formed by the second machine learning system, in this second preferred embodiment, the training of the first machine learning system preferably takes place after the training of the chemometric model formed by the second machine learning system. However, the steps described above, which are iteratively executed during the training of the first machine learning system, can also be accompanied by training of the second machine learning system, or the second machine learning system preferably remains unchanged during these iterative steps.

[0036] In a third preferred embodiment, the chemometric model formed by the second machine learning system and the first machine learning system are trained jointly with the goal of minimizing the target prediction difference and minimizing or maximizing the quantity derived from the target prediction difference, respectively. Thus, the spectrometer model, the reconstruction model, and the chemometric model are trained jointly.

[0037] The training of the chemometric model formed by the second machine learning system and the training of the first machine learning system formed jointly by the spectrometer model and the reconstruction model can be linked in further preferred modes. In one of these preferred embodiments, in a first temporal section of the method, the first machine learning system is trained with the goal of minimizing the reconstruction difference or minimizing or maximizing the variable derived from the reconstruction difference, whereas in a second temporal section of the method, the first machine learning system is trained with the goal of minimizing the target value prediction difference or minimizing or maximizing the variable derived from the target value prediction difference. In the second temporal section, the chemometric model formed by the second machine learning system remains unchanged.The first time segment and the second time segment can be repeated multiple times. In an alternative preferred embodiment, in a first time segment of the method, the first machine learning system is trained with the goal of minimizing the reconstruction difference or minimizing or maximizing the value derived from the reconstruction difference, whereas in a second time segment, the chemometric model formed by the second machine learning system and the first machine learning system are trained jointly with the goal of minimizing the target value prediction difference or minimizing or maximizing the value derived from the target value prediction difference. Here, too, the first time segment and the second time segment can be repeated multiple times.According to a further preferred embodiment, the chemometric model formed by the second machine learning system and the first machine learning system are trained together with the aim of minimizing or maximizing the reconstruction difference or the value derived from the reconstruction difference and of minimizing or maximizing the target value prediction difference or the value derived from the target value prediction difference. In this case, the minimization of the reconstruction difference or the minimization or maximization of the value derived from the reconstruction difference and the minimization of the target value prediction difference or the minimization or maximization of the value derived from the target value prediction difference can be weighted with different priorities, wherein the weighting can be changed or adapted during training. Preferably, the minimization of the reconstruction difference orthe minimization or maximization of the variable derived from the reconstruction difference is weighted with a higher priority, whereas in a second time period, the minimization of the target value prediction difference or the minimization or maximization of the variable derived from the target value prediction difference is weighted with a higher priority. Preferably, in the first time period, the chemometric model formed by the second machine learning system and the first machine learning system are trained jointly with the aim of minimizing or maximizing the reconstruction difference or the variable derived from the reconstruction difference with a first weighting and the target value prediction difference or the variable derived from the target value prediction difference with a second weighting smaller than the first weighting.In the second time period, the chemometric model formed by the second machine learning system and the first machine learning system are trained together with the aim of minimizing or maximizing the reconstruction difference or the quantity derived from the reconstruction difference with a third weighting and the target value prediction difference or the quantity derived from the target value prediction difference with a fourth weighting greater than the third weighting.

[0038] Unless the reference data is already available from other sources, the step of providing the reference data is preferably carried out by first providing reference samples for which the concentration values ​​of the at least one ingredient are known, so that these values ​​form the reference ingredient concentration data. For this purpose, the concentration values ​​of the ingredient can be determined using a chemical analysis method. These values ​​can also be determined using a non-chemical analysis method; for example, using a high-accuracy spectral analyzer. Furthermore, the reference samples are measured using a reference spectrometer to obtain the reference spectral data.

[0039] Training the first machine learning system and / or the second machine learning system with the goal of minimizing or maximizing the reconstruction difference or the target prediction difference or the variable derived from the reconstruction or target prediction difference represents an optimization of the model formed by the respective machine learning system. For this optimization, one or more optimization techniques from the following group are preferably selected and used: Bayesian optimization, grid search, random search, gradient-free optimization such as the Nelder-Mead method, optimal experimental design, machine learning methods, greedy algorithms, evolutionary algorithms, biology-inspired optimization methods such as particle swarm optimization and FireFly optimization. Gradient-based optimization is particularly preferred.

[0040] In further preferred embodiments, it is taken into account that the process for manufacturing the spectrometer hardware is not ideal. In principle, the spectrometer hardware manufactured according to the spectrometer model will not completely resemble this model, but will differ from the model. For example, the mean wavelengths of the spectrometer's measurement channels may differ from the mean wavelengths of the measurement channels of the spectrometer model. Therefore, in one step, a manufacturing-related deviation is first determined between the spectrometer model and a spectrometer manufactured according to the spectrometer model. The determined manufacturing-related deviation is taken into account when training the first machine learning system, so that the simulation measurement data more closely correspond to the measurement data that can be acquired with the manufactured spectrometer.However, the manufacturing-related hardware deviation taken into account is preferably not changed during training.

[0041] In further preferred embodiments, the method preferably comprises further steps by which changing measurement conditions are taken into account, making the simulation measurement data as a whole more realistic. Such measurement conditions include, for example, the distance between the sample and the spectrometer, as well as environmental conditions such as humidity and temperature in the area of ​​the sample and / or the spectrometer. The changing measurement conditions are taken into account during training of the first machine learning system, which is formed jointly by the spectrometer model and the reconstruction model, by the spectrometer model also describing the changing measurement conditions. As a result, the simulation measurement data obtained thereby correspond more closely to the measurement data that can be acquired with the correspondingly manufactured spectrometer under realistic measurement conditions.The measurement conditions are varied to account for the fact that different realistic measurement conditions can occur. For example, the change in the amplitude of a measurement channel can be varied depending on the ambient temperature. This amplitude represents the sensitivity of the measurement channel.

[0042] The spectrometer model defines the multiple measurement channels of the spectrometer and includes at least one trainable parameter. The at least one trainable parameter of the spectrometer model is preferably formed by a number of measurement channels. The at least one trainable parameter of the spectrometer model is preferably formed by a wavelength difference between a wavelength parameter of one of the measurement channels and a corresponding wavelength parameter of another, in particular an adjacent, measurement channel. The number of measurement channels is preferably between 3 and 50. The individual measurement channels are preferably each also defined by at least one trainable parameter.This at least one trainable parameter is preferably formed by a peak wavelength, a centroid wavelength, a mean wavelength, a bandwidth, a half-width, a curve shape parameter and / or a peak value of the respective measurement channel. The peak value is a peak height. The measurement channels are preferably each determined by a spectral input filter. The spectral input filters each have a transmission range defined by one or more of the above-mentioned parameters. During training of the first machine learning system, one or more of the above-mentioned parameters are changed. The optimization techniques specified above are preferably used for this purpose. During training of the first machine learning system, the number of measurement channels is preferably changed. For example, the number of measurement channels can be changed from 16 to 8 or to 4, or vice versa.

[0043] The spectrometer model preferably defines at least one constraint for the at least one trainable parameter of the spectrometer model.The at least one limitation is preferably formed by a maximum number of measuring channels, by a minimum peak wavelength, by a maximum peak wavelength, by a minimum centroid wavelength, by a maximum centroid wavelength, by a minimum mean wavelength, by a maximum mean wavelength, by a minimum bandwidth, by a maximum bandwidth, by a minimum half-width, by a maximum half-width, by a minimum curve shape parameter, by a maximum curve shape parameter, by a minimum peak value, by a maximum peak value, by a minimum wavelength difference between a wavelength parameter of one of the measuring channels and a corresponding wavelength parameter of another of the measuring channels, or by a maximum wavelength difference between a wavelength parameter of one of the measuring channels and a corresponding wavelength parameter of another of the measuring channels.The minimum or maximum wavelength difference is defined in particular between wavelength parameters of two adjacent measuring channels. The wavelength parameter in question is preferably formed by the peak wavelength or by the minimum or maximum half-width. The aforementioned restrictions are preferably also defined multiple times, for example restrictions of the minimum or maximum peak wavelength in multiple intervals. The aforementioned restrictions can also relate to a group or to a proportion of the measuring channels. For example, for a proportion of one third of the measuring channels, it can be specified that their peak wavelength lies in the near infrared range. The aforementioned restrictions each restrict the possible change in the parameter in question during the training of the first machine learning system. This one restriction orHowever, these multiple constraints are preferably not changed during training of the first machine learning system.

[0044] In general, this one or more constraints can also be implemented as costs in the optimization. For example, a parameter value outside the respective constraint can be considered to have a cost of infinity. Soft constraints are also preferred. For example, for the parameter formed by the number of channels, costs can be used that lead to increasing costs with increasing number: Example 1: Number of channels 4: Cost 10 Number of channels 8: Cost 100 Number of channels 16: Cost 1,000 Number of channels 32: Cost 10,000 any other number of channels: cost of infinity.

[0045] This allows a few possible technical implementation options to be taken into account when training the first machine learning system formed jointly by the spectrometer model and the reconstruction model, which allows training to be carried out more quickly. Example 2: k channels with 1 ≤ k ≤ 128: Cost of k k > 128: cost of infinity.

[0046] The total costs during optimization then result from the preferentially weighted sum of the reconstruction difference or the target value prediction difference and the aforementioned costs resulting from the constraint.

[0047] This allows a variety of possible technical implementation options to be taken into account when training the first machine learning system formed jointly by the spectrometer model and the reconstruction model, which allows training to be carried out more quickly.

[0048] The spectrometer model preferably further defines at least one device parameter of the spectrometer, which represents signal noise, bandwidth noise, half-width noise, measurement deviation, or tolerance of the spectrometer. The measurement deviation can be temperature-dependent, for example. The aforementioned bandwidth noise is a deviation of an achieved mean bandwidth from a predetermined value for the bandwidth. The one or more tolerances are determined in particular by the manufacturing of the spectrometer hardware. The device parameter(s) are taken into account when training the first machine learning system, thereby improving the optimization of the spectrometer model and, if applicable, also of the chemometric model.This is preferably achieved through augmentation during training, for which the spectrum simulation is modified in at least one iteration, taking into account at least one device parameter. However, the device parameters are preferably not changed during training of the first machine learning system.

[0049] The computer program product according to the invention serves for the technical design or development of an analytical device for analyzing at least one component of a sample. The computer program product comprises a computer-readable storage medium having program instructions stored thereon. The program instructions are executable by one or more computers or control units and cause the one or more computers or control units to execute the method according to the invention or one of the described preferred embodiments of the method according to the invention. The program instructions include, among other things, algorithms for machine learning; namely, for training the two machine learning systems. The storage medium can be formed by an electronic medium, a magnetic medium, an optical medium, an electromagnetic medium, an infrared medium, or a semiconductor medium, such as an SSD.The program instructions can be machine-dependent or machine-independent instructions, microcode, firmware, state-defining data, or any source code or object code written, for example, in C++, Java, or similar or in conventional procedural programming languages. Electronic circuits, such as programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), can also be implemented to execute the program instructions. The resulting model of the spectrometer, the resulting reconstruction model, and, if applicable, the resulting chemometric model are preferably stored as program code to be available for later applications.

[0050] The trained machine learning system according to the invention forms a model of a spectrometer with multiple measurement channels for selected frequencies and a reconstruction model for reconstructing a spectrum based on individual measurement values ​​recorded with the spectrometer for the selected frequencies. Another trained machine learning system according to the invention preferably forms a chemometric model for determining a concentration of at least one constituent of a sample based on the reconstructed spectrum. The one or both machine learning systems according to the invention were trained by the method according to the invention or by a preferred embodiment of the method according to the invention. The machine learning system is formed by the first machine learning system. It is in the form of program code.Preferably, two of the machine learning systems are present, namely the first machine learning system and the second machine learning system.

[0051] The analytical device according to the invention serves for the spectral analysis of a sample. The analytical device comprises a spectrometer for spectral measurement of the sample. The spectrometer has, according to a model, several measurement channels for selected frequencies. The analytical device is configured for use with a reconstruction model to reconstruct a spectrum based on individual spectral measurement values ​​of the sample recorded with the spectrometer. The analytical device is also preferably configured for use with a chemometric model to determine a concentration of at least one constituent of the sample based on the reconstructed spectrum. The spectrometer, the reconstruction model, and preferably also the chemometric model are derived from the method according to the invention or one of the described preferred embodiments of the method according to the invention.The analytical device was therefore technically designed or developed using the method according to the invention. The optimization of the spectrometer, the reconstruction model, and, if applicable, also the chemometric model was carried out jointly by implementing the method according to the invention. The analytical device preferably also has further features described in connection with the method according to the invention.

[0052] In preferred embodiments, the analysis device comprises a plurality of components which are mechanically independent of one another and may have their own housings. The spectrometer is preferably designed as a handheld device, which forms one of the mechanically independent components. Another of the mechanically independent components of the analysis device is preferably formed by a computing unit, which is configured to apply the reconstruction model and / or the chemometric model. The spectrometer and the computing unit are preferably connected to one another via a wireless data connection. The computing unit can be a smartphone, for example. In a further preferred embodiment, the analysis device comprises a housing in which the remaining components of the analysis device are located, for which purpose the analysis device is preferably designed as a table-top device.The analysis device according to the invention can be realized in a smaller installation space compared to the prior art due to the smaller number of measuring channels.

[0053] Further details and developments of the invention will become apparent from the following description of preferred embodiments of the invention, with reference to the drawings. They show: Fig. 1: a flow chart illustrating the operation of a preferred embodiment of an analysis device according to the invention; Fig. 2: a flow chart of a first preferred embodiment of a method according to the invention for the technical design of the Fig. 1 illustrated analyzer; Fig. 3: a flow chart of a second preferred embodiment of the method according to the invention; and Fig. 4: a flow chart of a third preferred embodiment of the method according to the invention.

[0054] Fig. 1 shows a flow chart to illustrate the functioning of a preferred embodiment of an analysis device 01 according to the invention. The analysis device 01 is used for the spectral analysis of at least one ingredient of a sample 02. The sample 02 can be grain, for example. The sample 02 has a reflection spectrum 03. In the reflection spectrum 03, λ stands for the wavelength and R stands for the intensity of the reflection. If the reflection spectrum 03 were recorded with a high-resolution spectrometer (not shown) according to a gold standard, the recorded spectrum (not shown) would largely resemble the actual reflection spectrum 03 shown. However, the analysis device 01 comprises a spectrometer 04 with, for example, only six measuring channels for selected frequencies.The measuring channels of the spectrometer 04 were selected and manufactured according to a model which was produced by a method according to the invention, which is described with reference to the . Fig. 2 to 4. The spectrometer 04 is used to measure the reflection spectrum 03 so that measured values ​​are obtained for the six measuring channels, which are shown in a measured spectrum 06. The measured spectrum 06 comprises measured values ​​for the frequencies λ1, λ2, λ3, λ4, λ5 and λ6 of the six measuring channels. The analysis device 01 is also designed to apply a reconstruction model 07 to reconstruct a spectrum based on the individual measured values ​​06 of the sample 02 for the frequencies λ1, λ2, λ3, λ4, λ5 and λ6 of the six measuring channels recorded with the spectrometer 04. This results in a reconstructed spectrum 08 which is very close to the original reflection spectrum 03 in terms of its essential properties. The reconstruction model 07 has also been produced by the method according to the invention, which with reference to the Fig. 2 to 4. The analysis device 01 is also configured to apply a chemometric model 09 to determine a concentration of at least one ingredient of the sample 02 based on the reconstructed spectrum 08. This results in a measured value 11 for the concentration of the ingredient of the sample 02, which can be represented by way of example as x mess The chemometric model 09 was also obtained by the method according to the invention, which is Fig. 2 to 4 are explained in more detail.

[0055] Fig. 2 shows a flow chart of a first preferred embodiment of the method according to the invention for the technical design of the Fig. 1 illustrated analysis device 01. In this first preferred embodiment, a training 13 of the model of the spectrometer 04 (shown in Fig. 1) and training 14 of the reconstruction model 07 (shown in Fig. 1) by machine learning, which for the model of the spectrometer 04 (shown in Fig. 1) and the reconstruction model 07 (shown in Fig. 1) jointly. As with reference to Fig. 1, the starting point is sample 02, which represents a reference sample for the method. Sample 02 has a reflection spectrum 03, which represents a reference for the method in the form of reference spectral data and was recorded, for example, with a spectrometer (not shown) according to a gold standard. The concentration of the respective ingredient in sample 02 is determined by a reference method, for example, by a chemical analysis method, so that a reference ingredient concentration measurement value 16 is obtained for this concentration, which is exemplary represented as x referenz As stated with reference to Fig. 1, the model of the spectrometer 04 (shown in Fig. 1) is used to measure the reflection spectrum 03, so that measured values ​​are obtained for the six measurement channels, which are shown in the measured spectrum 06 and represent simulation measurement data for the process. Subsequently, the reconstruction model 07 (shown in Fig. 1) is applied for a reconstruction based on the measured spectrum 06, whereby the reconstructed spectrum 08 is obtained, which represents a reconstructed simulation spectrum for the process. The chemometric model 09 is applied to determine the concentration of the ingredient based on the reconstructed spectrum 08, whereby the measured value 11 for the concentration of the ingredient is obtained, which represents a simulation measured value for the process. In the process for the technical design of the Fig. 1, a comparison 17 is made between the measured value 11 and the reference ingredient concentration measured value 16 in order to determine parameters of the model of the spectrometer 04 (shown in Fig. 1) and the reconstruction model 07 (shown in Fig. 1) during machine learning, which allows training 13 of the model of the spectrometer 04 (shown in Fig. 1) and training 14 of the reconstruction model 07 (shown in Fig. 1) so that a target prediction difference is minimized. Training 13 of the model of the spectrometer 04 (shown in Fig. 1) and training 14 of the reconstruction model 07 (shown in Fig. 1) are carried out jointly within a common machine learning system. In this first preferred embodiment, the chemometric model 09 is used during the joint training 13, 14 of the model of the spectrometer 04 (shown in Fig. 1) and the reconstruction model 07 (shown in Fig. 1) is not changed, ie not trained. The result of the procedure is the trained model of the spectrometer 04 (shown in Fig. 1), the trained reconstruction model 07 (shown in Fig. 1) and the chemometric model 09. The spectrometer 04 (shown in Fig. 1) is to be manufactured according to his model. The reconstruction model 07 (shown in Fig. 1) and the chemometric model 09 are in the analytical device 01 to be manufactured (shown in Fig. 1) in the form of software algorithms. The resulting analyzer 01 (shown in Fig. 1) is ideally suited for the given analysis task and the spectrometer 04 (shown in Fig. 1) is far less complex than a reference spectrometer (not shown) according to the gold standard, so that the analyzer 01 (shown in Fig. 1) can be produced cost-effectively.

[0056] Fig. 3 shows a flow chart of a second preferred embodiment of the method according to the invention for the technical design of the Fig. 1 illustrated analyzer 01. The Fig. The second embodiment shown in Figure 3 is initially similar to that shown in Fig. 2. In contrast to the first embodiment shown in Fig. 2, the training 13 of the model of the spectrometer 04 (shown in Fig. 1) and training 14 of the reconstruction model 07 (shown in Fig. 1) together with a training 18 of the chemometric model 09 (shown in Fig. 1). For this purpose, the comparison 17 is carried out between the measured value 11 and the reference ingredient concentration measured value 16 in order to determine parameters of the spectrometer model 04 (shown in Fig. 1), of the reconstruction model 07 (shown in Fig. 1) and the chemometric model 09 (shown in Fig. 1) to adapt during machine learning.

[0057] Fig. 4 shows a flow chart of a third preferred embodiment of the method according to the invention for the technical design of the Fig. 1 illustrated analyzer 01. The Fig. The third embodiment shown in Figure 4 is initially similar to that shown in Fig. 2. In contrast to the first embodiment shown in Fig. 2, the training 13 of the model of the spectrometer 04 (shown in Fig. 1) and training 14 of the reconstruction model 07 (shown in Fig. 1) starting from a comparison 19 between the reconstructed spectrum 08 and the original reflection spectrum 03 of the sample 02 to determine parameters of the model of the spectrometer 04 (shown in Fig. 1) and the reconstruction model 07 (shown in Fig. 1) during machine learning so that reconstruction differences are minimized.

[0058] The Fig. 2 shows the first embodiment of the method, which in Fig. 3 shown second embodiment of the method and / or the one in Fig. 4 are carried out successively in preferred embodiments in time intervals. In further preferred embodiments, the Fig. 2 shows the first embodiment of the method, which in Fig. 3 shown second embodiment of the method and / or the one in Fig. 4, the third embodiment of the method is carried out simultaneously, wherein the results of the adjustments 17, 19 are considered together, for example by forming a weighted sum, and on this basis the training 13 of the model of the spectrometer 04 (shown in Fig. 1) and training 14 of the reconstruction model 07 (shown in Fig. 1). In further preferred embodiments, the Fig. 2 shows the first embodiment of the method, which in Fig. 3 shown second embodiment of the method and / or the one in Fig. 4, the third embodiment of the method is carried out simultaneously, wherein the results of the adjustments 17, 19 are considered together, for example by forming a product by multiplication, and on this basis the training 13 of the model of the spectrometer 04 (shown in Fig. 1) and training 14 of the reconstruction model 07 (shown in Fig.1). For example, the cost is defined by the number of channels, so that k channels result in a cost value of k, which is multiplied by the reconstruction difference. In this case, doubling the number of channels would have to at least halve the reconstruction difference to achieve a better evaluation. The two factors of the product are preferably scaled to each other using two exponents, weight1 and weight2, so that the product = cost Gewicht1 · Reconstruction difference Gewicht2 is. List of reference symbols 01 Analyzer 02 Sample 03 Reflection spectrum 04 Spectrometer 05 - 06 measured spectrum 07 Reconstruction model 08 reconstructed spectrum 09 chemometric model 10 - 11 Measured value 12 - 13 Training the spectrometer model 14 Training the reconstruction model 15 - 16 Reference ingredient concentration measurement value 17 Comparison 18 Training the chemometric model 19 Comparison

Claims

[1] Method for the technical design of an analysis device (01) for the spectral analysis of a sample (02); wherein the analysis device (01) to be technically designed must comprise a spectrometer (04) for spectral measurement of the sample (02), which has several measurement channels for selected frequencies according to a trained model of the spectrometer (04); wherein the analysis device (01) to be technically designed must be configured to apply a trained reconstruction model (07) to reconstruct a spectrum based on individual measured values ​​(06) of the sample (02) recorded for the selected frequencies with the spectrometer (04); and wherein the method comprises the following steps: - Selection of an initial configuration of a model of the spectrometer (04); - selecting an initial configuration of a reconstruction model (07); and - training (13, 14) a first machine learning system formed jointly by the model of the spectrometer (04) and the reconstruction model (07), whereby the trained model of the spectrometer (04) and the trained reconstruction model (07) are obtained. [2] Method according to claim 1, characterized by that the analysis device (01) to be technically designed must be designed for the spectral analysis of a sample which originates from an agricultural product, a foodstuff or a foodstuff. [3] Method according to claim 1 or 2, characterized bythat the model of the spectrometer comprises at least one trainable parameter which is formed by a number of the measuring channels, by a wavelength difference between a wavelength parameter of one of the measuring channels and a corresponding wavelength parameter of another of the measuring channels, by a peak wavelength of the respective measuring channel, by a center of gravity wavelength of the respective measuring channel, by a mean wavelength of the respective measuring channel, by a bandwidth of the respective measuring channel, by a half-width of the respective measuring channel, by a curve shape parameter of the respective measuring channel and / or by a peak value of the respective measuring channel, wherein the model of the spectrometer defines at least one restriction for the at least one trainable parameter. [4] Method according to one of claims 1 to 3, characterized by that it includes the following further steps: - Providing reference spectral data (03) representing a spectrum (03) of a reference sample (02); and - iteratively performing the following steps during training (13, 14) of the first machine learning system: ◯ Simulating a spectral measurement of the reference sample (02) with the spectrometer (04) according to the model of the spectrometer (04) defined by the first machine learning system by applying the model of the spectrometer (04) defined by the first machine learning system to the reference spectral data (03), thereby obtaining simulation measurement data (06); ◯ applying the reconstruction model (07) defined by the first machine learning system to the simulation measurement data (06), thereby obtaining a reconstructed simulation spectrum (08); and o Comparing the reconstructed simulation spectrum (08) with the reference spectral data (03) to determine a reconstruction difference; wherein the first machine learning system is trained (13, 14) with the aim of minimizing the reconstruction difference or minimizing or maximizing a quantity derived from the reconstruction difference. [5] Method according to one of claims 1 to 4, characterized by that the analysis device (01) to be technically designed must be designed for the spectral analysis of a constituent of the sample (02), wherein the analysis device (01) to be technically designed must be configured to apply a chemometric model (09) to determine a concentration of at least one constituent of the sample (02) based on the spectrum (08) reconstructed with the trained reconstruction model (07). [6] Method according to claim 5, characterized by that it includes the following further steps: - Providing reference data for at least one reference sample (02), wherein the reference data comprises a series of reference spectral data (03) to which a series of reference ingredient concentration data (16) is assigned; and - iteratively performing the following steps during training (13, 14) of the first machine learning system: o Simulating a spectral measurement of the at least one reference sample (02) with the spectrometer (04) according to the model of the spectrometer (04) defined by the first machine learning system, by applying the model of the spectrometer (04) defined by the first machine learning system to the reference spectral data (03), thereby obtaining simulation measurement data (06); ◯ Applying the reconstruction model (07) defined by the first machine learning system to the simulation measurement data (06), thereby obtaining a reconstructed simulation spectrum (08); ◯ applying the chemometric model (09) to the reconstructed simulation spectrum (08), whereby simulation measured values ​​(11) of the concentration of the at least one ingredient are obtained; and ◯ Comparing the simulation measured values ​​(11) of the concentration of the at least one ingredient with the reference ingredient concentration data (16) to determine a target prediction difference; wherein the first machine learning system is trained (13, 14) with the goal of minimizing the target prediction difference or minimizing or maximizing a variable derived from the target prediction difference. [7] Method according to claim 5 or 6, characterized by that the chemometric model (09) is formed by a trained second machine learning system, the method comprising the following further steps: - selecting an initial configuration of the chemometric model (09) formed by the second machine learning system; and - training (18) the chemometric model (09) formed by the second machine learning system, whereby the chemometric model (09) formed by the trained second machine learning system is obtained. [8] Method according to claim 7, which refers back to claim 6, characterized by that the chemometric model (09) formed by the second machine learning system and the first machine learning system are trained together (13, 14, 18) with the aim of minimizing the target value prediction difference or minimizing or maximizing the quantity derived from the target value prediction difference. [9] Method according to claim 8, which refers back to claim 4, characterized bythat the chemometric model (09) formed by the second machine learning system and the first machine learning system are trained together (13, 14, 18) with the aim of minimizing or maximizing the reconstruction difference or the quantity derived from the reconstruction difference and minimizing or maximizing the target value prediction difference or the quantity derived from the target value prediction difference. [10] Method according to claim 9, characterized bythat in a first time period, the chemometric model (09) formed by the second machine learning system and the first machine learning system are trained together (13, 14, 18) with the aim of minimizing or maximizing the reconstruction difference or the variable derived from the reconstruction difference with a first weighting and the target value prediction difference or the variable derived from the target value prediction difference with a second weighting smaller than the first weighting, wherein in a second time period, the chemometric model (09) formed by the second machine learning system and the first machine learning system are trained together with the aim ofto minimize or maximize the reconstruction difference or the quantity derived from the reconstruction difference with a third weighting and the target value prediction difference or the quantity derived from the target value prediction difference with a fourth weighting greater than the third weighting. [11] Method according to one of claims 7 to 9 which refer back to claims 4 and 6, characterized byin that in a first time segment the first machine learning system is trained (13, 14) with the aim of minimizing the reconstruction difference or of minimizing or maximizing the variable derived from the reconstruction difference, and in that in a second time segment the chemometric model (09) formed by the second machine learning system and the first machine learning system are trained together (13, 14, 18) with the aim of minimizing the target value prediction difference or of minimizing or maximizing the variable derived from the target value prediction difference. [12] Method according to one of claims 1 to 11, characterized by that it includes the following further steps: - determining a manufacturing-related deviation between the model of the spectrometer (04) and a spectrometer (04) manufactured according to the model of the spectrometer (04); and - Consider the manufacturing variation when training (13, 14) the first machine learning system. [13] Method according to one of claims 1 to 12, characterized by that when training (13, 14) the first machine learning system formed jointly by the model of the spectrometer (04) and the reconstruction model (07), changing measurement conditions are taken into account in that the model of the spectrometer (04) describes the changing measurement conditions, wherein the changing measurement conditions include a distance between the sample (02) and the spectrometer (04), an air humidity and / or a temperature in the area of ​​the sample (02) and / or the spectrometer (04). [14] Computer program product for the technical design of an analysis device (01) for the spectral analysis of a sample (02), wherein the computer program product comprises a computer-readable storage medium having program instructions stored thereon, wherein the program instructions are executable by one or more computers or control units and cause the one or more computers or control units to carry out the method according to one of claims 1 to 13. [15] Trained machine learning system (09) which has been trained by a method according to one of claims 1 to 13 and which is formed by the first machine learning system. [16] Analysis device (01) for the spectral analysis of a sample (02), wherein the analysis device (01) comprises a spectrometer (04) for spectral measurement of the sample (02), which has a plurality of measurement channels for selected frequencies according to a trained model of the spectrometer (04); wherein the analysis device (01) is configured to apply a reconstruction model (07) for reconstructing a spectrum based on individual spectral measurement values ​​of the sample (02) recorded with the spectrometer (04), characterized by that the spectrometer (04) and the trained reconstruction model (07) have resulted from a method according to one of claims 1 to 13.

Images