Stoichiometric model generation

Abstract

The invention belongs to the field of chemometric model generation for spectrometers. The invention relates to a computer-implemented method for generating a chemometric model for a spectrometer, the method comprising: a) receiving a spectrum of an object and object data associated with a physical or chemical property of the object; b) generating at least two different candidate chemometric models, comprising: i) a pre-processing method, ii) a machine learning method, iii) a feature selection filter; c) training the candidate chemometric models; d) determining a prediction accuracy of the trained candidate chemometric models; e) selecting the trained candidate chemometric model with the highest prediction accuracy; and f) outputting the selected trained chemometric model.

Description

[0001] This invention belongs to the field of chemometric model generation for spectrometers. The invention relates to a computer-implemented method for generating a chemometric model for a spectrometer, the use of the chemometric model obtained by the method for a spectrometer, a spectrometer configured to use the chemometric model obtained by the method, a system for generating a chemometric model for a spectrometer, and a non-transient computer-readable medium comprising instructions that, when executed by one or more processors, cause the one or more processors to perform the method. Background Technology

[0002] Spectroscopy is a valuable non-destructive analytical technique suitable for various applications, such as quality control, quantitative and qualitative testing of agricultural products. Spectra contain physical or chemical information about the recorded object, but due to their nature, they cannot be directly interpreted. Therefore, a chemometric model is needed, which derives the physical or chemical properties of the object from its spectrum. To create a chemometric model, the spectrum of an object whose properties of interest are known (e.g., through laboratory analysis) is required. Generating a chemometric model from such data typically involves multiple steps, such as preprocessing and machine learning model selection. For each step, several options are available, of which there is no universally preferred option, but their effectiveness depends on the object and its properties. Traditionally, experienced data scientists manually compare and select these options. However, this is very time-consuming and often impossible due to a lack of sufficiently qualified personnel. Therefore, there is a need to facilitate the generation of chemometric models.

[0003] X. Bian et al. disclosed a method for screening infrared data preprocessing algorithms to select appropriate preprocessing methods in Chemometrics and Intelligent Laboratory Systems, Vol. 197 (2020), p. 103916. However, chemometric models require further steps, which still require the aforementioned expertise.

[0004] US 2023 / 0009725 A1 discloses a genetic algorithm for identifying processing pipelines that convert spectra into a form that can be used to generate predicted properties for corresponding samples. Feature selection may be part of the algorithm, but it is not optimized as part of finding a suitable algorithm. Therefore, this approach requires more manual expert input to be suitable for demanding applications and high-accuracy output. Summary of the Invention

[0005] The purpose of this invention is to provide a method for generating chemometric models for spectrometers that achieves high accuracy and reliability without requiring highly skilled or experienced personnel. This method is quick and easy to use, even for non-technical personnel. It is applicable to a wide variety of use cases without requiring complex adjustments.

[0006] In one aspect, the present invention relates to a computer-implemented method for generating a chemometric model for a spectrometer, the method comprising:

[0007] a) Receive the spectral data of an object and object data associated with the object's physical or chemical properties.

[0008] b) Generate at least two distinct candidate stoichiometric models, including

[0009] i) Preprocessing methods,

[0010] ii) Machine learning methods,

[0011] c) Use these spectral and object data to train these candidate chemometric models.

[0012] d) Determine the prediction accuracy of the trained candidate chemometric models.

[0013] e) Use the prediction accuracy to select the chemostometric model from these trained candidate chemostometric models, and

[0014] f) Output the selected stoichiometric model

[0015] In one aspect, the present invention relates to a computer-implemented method for generating a chemometric model for a spectrometer, the method comprising:

[0016] a) Receive the spectral data of an object and object data associated with the object's physical or chemical properties.

[0017] b) Generate at least two distinct candidate stoichiometric models, including

[0018] i) Preprocessing methods,

[0019] ii) Machine learning methods,

[0020] iii) Feature selection filter,

[0021] c) Use these spectral and object data to train these candidate chemometric models.

[0022] d) Determine the prediction accuracy of the trained candidate chemometric models.

[0023] e) Use the prediction accuracy to select the chemostometric model from these trained candidate chemostometric models, and

[0024] f) Output the selected stoichiometric model

[0025] In another aspect, the present invention relates to a computer-implemented method for generating a chemometric model for a spectrometer, the method comprising:

[0026] a) Receive the spectral data of an object and object data associated with the object's physical or chemical properties.

[0027] b) Generate at least two distinct candidate stoichiometric models, wherein the candidate stoichiometric models are configured to determine at least two physical or chemical properties based on spectra.

[0028] c) Use these spectral and object data to train these candidate chemometric models.

[0029] d) Determine the prediction accuracy of the trained candidate chemometric models.

[0030] e) Use the prediction accuracy to select the chemostometric model from these trained candidate chemostometric models, and

[0031] f) Output the selected stoichiometric model.

[0032] In another aspect, the present invention relates to the use of a chemometric model obtained by the method of any one of the preceding claims for use in a spectrometer.

[0033] In another aspect, the present invention relates to a spectrometer configured to use a stoichiometric model obtained by the method according to the invention.

[0034] In another aspect, the present invention relates to a system for generating a chemometric model for a spectrometer, the system comprising:

[0035] a) Receive the spectral data of an object and object data associated with the object's physical or chemical properties.

[0036] b) Generate at least two distinct candidate stoichiometric models, including

[0037] i) Preprocessing methods,

[0038] ii) Machine learning methods,

[0039] c) Use these spectral and object data to train these candidate chemometric models.

[0040] d) Determine the prediction accuracy of the trained candidate chemometric models.

[0041] e) Use the prediction accuracy to select the chemostometric model from these trained candidate chemostometric models, and

[0042] f) Output the selected stoichiometric model.

[0043] In another aspect, the present invention relates to a system for generating a chemometric model for a spectrometer, the system comprising:

[0044] a) Receive the spectral data of an object and object data associated with the object's physical or chemical properties.

[0045] b) Generate at least two distinct candidate stoichiometric models, including

[0046] i) Preprocessing methods,

[0047] ii) Machine learning methods,

[0048] iii) Feature selection filter

[0049] c) Use these spectral and object data to train these candidate chemometric models.

[0050] d) Determine the prediction accuracy of the trained candidate chemometric models.

[0051] e) Use the prediction accuracy to select the chemostometric model from these trained candidate chemostometric models, and

[0052] f) Output the selected stoichiometric model.

[0053] In another aspect, the present invention relates to a system for generating a chemometric model for a spectrometer, the system comprising:

[0054] a) Input terminal, used to receive the spectrum of the object and object data associated with the object's physical or chemical properties.

[0055] b) A processor configured to: generate at least two distinct candidate chemostoichiometric models, wherein the candidate chemostoichiometric models are configured to determine at least two physical or chemical properties based on spectra; train the candidate chemostoichiometric models using these spectra and object data; determine the prediction accuracy of the trained candidate chemostoichiometric models; and use the prediction accuracy to select a chemostoichiometric model from the trained candidate chemostoichiometric models.

[0056] c) Output terminal, which is used to output the selected trained stoichiometric model.

[0057] In another aspect, the present invention relates to a non-transient computer-readable medium comprising instructions that, when executed by one or more processors, cause the one or more processors to perform the following methods, including:

[0058] a) Receive the spectral data of an object and object data associated with the object's physical or chemical properties.

[0059] b) Generate at least two distinct candidate stoichiometric models, including

[0060] i) Preprocessing methods,

[0061] ii) Machine learning methods,

[0062] c) Use these spectral and object data to train these candidate chemometric models.

[0063] d) Determine the prediction accuracy of the trained candidate chemometric models.

[0064] e) Use the prediction accuracy to select the chemostometric model from these trained candidate chemostometric models, and

[0065] f) Output the selected stoichiometric model.

[0066] In another aspect, the present invention relates to a non-transient computer-readable medium comprising instructions that, when executed by one or more processors, cause the one or more processors to perform the following methods, including:

[0067] a) Receive the spectral data of an object and object data associated with the object's physical or chemical properties.

[0068] b) Generate at least two distinct candidate stoichiometric models, including

[0069] i) Preprocessing methods,

[0070] ii) Machine learning methods,

[0071] iii) Feature selection filter,

[0072] c) Use these spectral and object data to train these candidate chemometric models.

[0073] d) Determine the prediction accuracy of the trained candidate chemometric models.

[0074] e) Use the prediction accuracy to select the chemostometric model from these trained candidate chemostometric models, and

[0075] f) Output the selected stoichiometric model.

[0076] In another aspect, the present invention relates to a non-transient computer-readable medium comprising instructions that, when executed by one or more processors, cause the one or more processors to perform the following methods, including:

[0077] a) Receive the spectral data of an object and object data associated with the object's physical or chemical properties.

[0078] b) Generate at least two distinct candidate stoichiometric models, wherein the candidate stoichiometric models are configured to determine at least two physical or chemical properties based on spectra.

[0079] c) Use these spectral and object data to train these candidate chemometric models.

[0080] d) Determine the prediction accuracy of the trained candidate chemometric models.

[0081] e) Use the prediction accuracy to select the chemostometric model from these trained candidate chemostometric models, and

[0082] f) Output the selected stoichiometric model.

[0083] The generation of the chemometric model of this invention can be automated, for example, in a computer system. No expert input is required. Automation makes the method very fast and easy to use. Different users will obtain the same results without the need for experienced technicians. The obtained chemometric model is accurate and reliable. This method further enables the rapid updating of the model using new reference data, for which different settings of the chemometric model selection can produce better results. The method can be implemented in an easy-to-use front-end where users simply upload reference spectra and associated object data (e.g., analytical data) and can receive the most accurate chemometric model in a short time. This method allows for the rapid application of spectroscopic methods to new use cases.

[0084] The term "spectrometer" can refer to an instrument capable of recording the spectrum of an object. Spectrometers can be portable or stationary, such as laboratory equipment. Portable spectrometers can be handheld spectrometers or modules integrated into portable devices such as smartphones, tablets, or wearable devices (e.g., smartwatches). Portable spectrometers can be communicatively coupled to computer devices, such as cloud computers or smartphones. Such computer devices can be configured to execute chemometric models, such as those obtained by the method of the present invention. The computer device can be further configured to receive spectra from the spectrometer. The computer device can store such spectra, send them to a system according to the present invention for generating chemometric models, or use the received and / or stored spectra to execute the method of the present invention. The computer system can be further configured to use the output of the chemometric model to obtain action instructions, for example, recommending drinking water if low skin hydration is detected.

[0085] A spectrometer may include:

[0086] - An optical element configured to separate incident light radiation provided by the measuring object into a spectrum having constituent wavelength components;

[0087] - A photoelectric sensor, the photoelectric sensor including at least one photosensitive area configured to receive light radiation from the optical element, wherein the photoelectric sensor is configured to generate at least one photoelectric sensor signal based on the irradiation of the photosensitive area by the light radiation;

[0088] - A processor that processes photoelectric sensor signals into a spectrum.

[0089] The term "optical element" can refer to any element configured to influence light radiation. An optical element can be configured to at least partially disperse light radiation, at least partially filter light radiation, at least partially reflect (e.g., diffuse or direct reflection) light radiation, at least partially deflect light radiation, at least partially transmit light radiation, and at least partially absorb light radiation. An optical element can include at least one of a prism, grating, beam splitter, or interferometer (e.g., a Michelson interferometer). An optical element can be configured for use in mobile applications, such as in handheld spectrometer devices, and / or in spectrometer devices included in electronic communication devices such as smartphones or tablet computers. As another example, an optical element can include at least one optical filter element. The optical filter element can be configured to filter light radiation or, more specifically, at least one selected spectral range of light radiation. The optical filter element can be positioned in the radiation path before a photoelectric sensor. As an example, a portable spectrometer can include multiple photoelectric sensors, such as 5 to 20, for example, 8 to 12. The photoelectric sensors can be arranged in an array or matrix in the form of pixels. Portable spectrometers may include multiple optical filter elements. These optical filter elements may be positioned in the beam path prior to the photodetectors. The optical filter elements may be transmissive at different wavelengths within different wavelength regions. For example, each photodetector may be positioned relative to the beam path after an optical filter, where each optical filter is transmissive relative to the other optical filters at different wavelengths or in different wavelength regions.

[0090] A spectrometer may include one or more photoelectric sensors. A photoelectric sensor may include at least one photosensitive region. The photosensitive region may be configured to receive light radiation from optical elements. The photoelectric sensor may be configured to generate at least one photoelectric sensor signal based on the illumination of the photosensitive region by light radiation. The term "sensor" may refer to a device configured to detect at least one condition or to measure at least one measured variable. A sensor may be able to generate at least one signal (e.g., a measurement signal) that is a qualitative or quantitative indication of the measured variable and / or measured characteristic (e.g., illumination of the sensor or a portion thereof). This signal may be or include an electrical signal, such as current, specifically photocurrent. The term "photoelectric sensor" may refer to a sensor or detector configured to detect or measure light radiation, for example, by using the photoelectric effect, such as detecting illumination and / or a radiation spot generated by at least one radiation beam. A photodetector may include at least one substrate. As an example, a single photoelectric sensor may be a substrate having at least one single photosensitive region that generates a physical response, such as an electronic response, to illumination within a given wavelength range.

[0091] The term "photosensitive region" can refer to a unit of a photoelectric sensor, specifically a spatial region or volume that is part of the photoelectric sensor and is configured to be illuminated, or in other words, to receive light radiation, and to generate at least one signal, such as an electronic signal, in response to illumination. The photosensitive region can be located on the surface of the photoelectric sensor. Specifically, the photosensitive region can be a single, enclosed, and uniform photosensitive area. However, other options are also possible.

[0092] A spectrometer may include a radiating or emitting element configured to emit illumination light to illuminate an object so that the object produces detection light. The emitting element may be an incandescent lamp (e.g., a tungsten filament lamp or a halogen tungsten filament lamp), a light-emitting diode (LED), a laser diode, or a gas discharge lamp (e.g., a xenon lamp, a mercury vapor lamp, or a deuterium lamp).

[0093] The terms “light” or “radiation” can refer to electromagnetic radiation in one or more of the infrared, visible, and ultraviolet spectral ranges. The term “ultraviolet spectral range” can refer to electromagnetic radiation with wavelengths from 1 nm to 380 nm, preferably from 100 nm to 380 nm. Further, in part according to the version of ISO-21348 effective as of the date of this document, the term “visible spectral range” can refer to the spectral range from 380 nm to 760 nm. The term “spectral range” (IR) can refer to electromagnetic radiation from 760 nm to 1000 µm, wherein the range from 760 nm to 1.5 µm is generally referred to as the “near-spectral range” (NIR), the range from 1.5 µm to 15 µm is referred to as the “mid-spectral range” (MidIR), and the range from 15 µm to 1000 µm is referred to as the “far-spectral range” (FIR). Preferably, the light used for the typical purposes of this invention is light in the infrared (IR) spectral range, more preferably light in the near-infrared (NIR) and / or mid-spectral range (MidIR), especially light with wavelengths of 1 µm to 5 µm, preferably 1 µm to 3 µm, because these wavelength regions are particularly suitable for obtaining the material properties of an object.

[0094] The spectrometer may include a processor for processing photoelectric sensor signals into spectra. The processor may output the spectra, for example, to an interface for further processing or to a user interface. The processor may be further configured to apply a chemometric model and output object data obtained through the chemometric model. The processor may be configured to output both spectra and object data. The spectrometer may further include a memory. The memory may be configured to store the chemometric model. The memory may be configured to store the spectra.

[0095] The term "spectrum" can refer to a data structure in which several intensity values, or values ​​derived therefrom (such as the absorbance or reflectance of radiation), are associated with a wavelength or wavelength range of radiation. The data structure can be a vector or matrix, where each element represents an intensity, and the position in the vector or matrix represents a wavelength or wavelength range; therefore, the value at a given position represents the intensity at that wavelength or wavelength range. The data structure can also be a vector or matrix containing pairs of values, where one value represents a wavelength or wavelength range, and the other value represents the intensity at that wavelength or wavelength range. Spectra can be recorded by a spectrometer. Spectra recorded by a spectrometer can be corrected using calibration coefficients to compensate for sensor defects or drift. A spectrum can represent the absorbance or transmittance of radiation after it has penetrated an object.

[0096] The term "chemostoichiometric model" can refer to a model that uses spectra as input and outputs object data. Therefore, a chemostoichiometric model can convert the spectra of an object into corresponding object data. The object data obtained through a chemostoichiometric model can be associated with one or more physical or chemical properties (e.g., at least two or at least three physical or chemical properties). A chemostoichiometric model that outputs more than one physical or chemical property can be called a multi-label chemostoichiometric model. A multi-label chemostoichiometric model can contain several partial chemostoichiometric models, where each partial chemostoichiometric model can be configured to convert spectra into one physical or chemical property. A multi-label chemostoichiometric model can be configured to combine object data obtained through at least two partial chemostoichiometric models, for example, to output a vector including values ​​for each physical or chemical property.

[0097] A chemometric model may include preprocessing methods and machine learning models for obtaining object data. A chemometric model may include preprocessing methods, feature selection filters, and machine learning models. If a chemometric model comprises two or more partial chemometric models, each partial chemometric model may include a separate preprocessing method, feature selection filter, and machine learning model. Alternatively, these partial models may use the same preprocessing method or feature selection filter.

[0098] Chemometric models can be configured to determine physical and health information, such as skin hydration levels or blood glucose or lactose levels, based on the spectrum of a person's skin. The results can be used to provide users with recommendations, such as adjusting water intake based on skin hydration levels or adjusting training plans based on blood lactose levels.

[0099] Chemometric models can be configured to analyze agricultural products, food products, or feed based on the spectra of: agricultural products, such as fruits, vegetables, meats, and dairy products; food products, such as bread, sauces, sausages, and confectionery; and feed products, such as roughage or forage. For example, a chemometric model can determine relevant content, such as sugar or protein content. The results can be used to provide recommendations, such as providing farmers with expected harvest dates or chefs with suitable recipes.

[0100] Stoichiometric models can be configured to determine recycling-related information based on the spectrum of the object to be recycled (e.g., plastic objects like bottles or metal objects like cans). For example, it can be determined what material the object is made of, such as a bottle being made of polyethylene terephthalate (PET). The results can be used to provide recommendations, such as the most suitable method for recycling the object or the most convenient location for placing the object so that it can be recycled.

[0101] Stoichiometric models can be configured to determine product quality during a production process. Determining product quality can mean defining parameters of a specification, such as the concentration of a component. Product quality can refer to the quality of input materials (e.g., received from a supplier), the quality of intermediates (i.e., materials that have undergone certain processing steps and will undergo further processing steps), and the quality of output materials (e.g., before they are packaged and shipped to the customer).

[0102] The term "object" can refer to any object that can be measured by a spectrometer. An object can be a living or inanimate object. An object can refer to a whole object or a piece thereof, such as a sample extracted from an object. An object can have any aggregate state, such as solid, liquid, gas, or supercritical. An object can be homogeneous, i.e., it does not contain phase boundaries between regions of size equal to or larger than the irradiated radiation, such as a plastic bottle or a blood sample. An object can be non-homogeneous, containing phase boundaries between regions of size equal to or larger than the irradiated radiation, such as a soil sample containing sand.

[0103] The term "object data" can refer to data associated with the physical or chemical properties of an object. Therefore, object data can contain parameters representing the physical or chemical properties of an object. Physical properties can include thermal properties such as temperature, thermal conductivity, or specific heat capacity; macroscopic or microscopic states of matter such as crystallinity; optical properties such as refractive index, optical conductivity, or absorption coefficient; and electrical properties such as electrical conductivity or dielectric constant. The chemical properties of an object typically refer to: the object's chemical composition, including material composition such as the types and concentrations of certain chemical compounds, such as water content; the object's molecular structure, such as the presence of functional groups like carbonyl groups or hydrogen bonds; molecular weight, such as the degree of polymerization of a polymer; and the state of matter, such as crystallinity or aggregated state, such as in a dispersed phase. Object data can include one or more physical or chemical properties of an object, such as the concentrations of two chemical compounds in a sample. Physical or chemical properties can be variable, taking any value within a range, such as concentration, or they can be categorized, such as the fiber type in a fabric.

[0104] Object data can further include metadata, which can be associated with factors that may affect the physical or chemical properties of the object but are not themselves physical or chemical properties. Metadata can be associated with relevant factors of the object, spectrometer, or measurement environment. Examples of factors associated with the object could be a sample identifier, its geographical origin, or time information (such as the time the sample was acquired from a larger object). Examples of factors associated with the spectrometer could be a spectrometer identifier, the type of spectrometer, or spectrometer settings. Examples of factors associated with the measurement environment could be time information (such as the time the measurement was performed), geographical information related to the location where the measurement was performed, or information about the analytical techniques used to determine the physical or chemical properties of the object.

[0105] The method of the present invention includes receiving spectral and object data. Receiving may mean obtaining data via an interface (e.g., an interface with a storage device or a communication interface, e.g., with a cloud computing system). Spectral and object data may also be received via an interface with a user interface (e.g., a graphical user interface). For example, spectral and object data may be uploaded by a user via a user interface (e.g., using a web browser). Spectral and object data may be received directly, for example, via a communication interface (e.g., via the Internet). Alternatively, spectral and object data may be stored on a storage device after being uploaded from the location where the data was obtained via a user interface. In this way, a user can first collect all spectral and object data over a period of time, and then indicate completion, for example, by clicking a button labeled "Completed" or "Generate Chemometric Model." This indication can trigger the method of the present invention.

[0106] The received object's spectrum and object data associated with the object's physical or chemical properties are correlated; that is, the spectrum and the corresponding object data can be linked. For example, both the spectrum and the object data can include the identifier of the object from which the spectrum was taken and the identifier of the object to which the object data pertains. In this way, it can be ensured that each element of the object data can be associated with the corresponding spectrum.

[0107] The object data to be determined by a stoichiometric model can be determined based on the provided object data. If the object data contains only values ​​associated with one physical or chemical property, it can be easily determined that the stoichiometric model should determine that object data. If the object data contains values ​​associated with more than one physical or chemical property, the determination may involve predefined conditions, such as the first property appearing in the object data. Alternatively, the object data may include an identifier indicating which physical or chemical property should be determined by the stoichiometric model. This identifier may further indicate whether the physical or chemical property is associated with a variable value (e.g., a small value for concentration) or a categorical value (e.g., used to distinguish a given group of compounds, such as a type of plastic (e.g., PET) or cotton in a fabric).

[0108] The method of the present invention may include a design of experiments (DOE) for measurements designed to obtain spectral and object data received by the method of the present invention. The term "design of experiments" may refer to instructions for systematically measuring the spectra of a set of objects having specified object data. An experimental design may include:

[0109] -Receive an identifier indicating which physical or chemical property should be determined by a stoichiometric model.

[0110] -Instructions for measuring objects with different values ​​for the identified physical or chemical properties are determined by using identifiers.

[0111] - Output these instructions, and

[0112] - In response to these output instructions, receive the object's spectrum and object data associated with the object's physical or chemical properties.

[0113] Experimental design may involve DOE models using identifiers as input. These identifiers can further indicate the measurement range of physical or chemical properties. The output can be generated by the DOE model. The DOE model can be tailored to candidate chemometric models, i.e., taking into account the needs of training the chemometric model. For example, if the candidate chemometric model includes an artificial neural network, more measurements may be required compared to multiple regression.

[0114] DOE models can include full factorial designs, which involve changing all possible combinations of physical or chemical properties and their levels to systematically assess their impact on spectrometer measurements. DOE models can also include partial factorial designs, which involve selecting a subset of physical or chemical properties and their levels to systematically assess their impact on spectrometer measurements. Selecting a subset can involve randomization, statistical replication, block design, or orthogonality.

[0115] The received spectral and object data can be compared with instructions. This comparison can be designed to ensure that these instructions have been adequately followed. The comparison can generate a score indicating how much spectral and object data corresponds to the instructions. If the comparison results show insufficient compliance, such as when the score is below a preset threshold, a warning indicating insufficient received data can be output. Users can be invited to provide further spectral and object data. Users can choose to request to continue the method, for example, if external factors (such as the inability to obtain the desired object) prevent the recording of the requested data.

[0116] The method of this invention may include data wrangling of received spectral or object data. The term "data wrangling" can refer to transforming and cleaning received spectral or object data into data that can be used to train candidate chemometric models. Data wrangling is sometimes also called data cleaning. Data wrangling may include data parsing. Data parsing may involve natural language processing, such as using a large language model employing a transformer. As an example, object data may contain the term "high-density polyethylene" and its abbreviation HDPE. Data parsing can merge these two terms into a consistent term for polyethylene. Data wrangling may include data cleaning. For example, the decimal separator in numerical values ​​can be a dot or a comma, depending on the data source. Data cleaning can uniformly use a specific decimal separator. Data wrangling may include data format adjustment. For example, spectra may be received as comma-separated files, and object data as JSON files, while training candidate chemometric models requires both spectral and object data to be in vector format. Therefore, data format adjustment can convert the source format to the format required for training the chemometric model. Data wrangling may include data transformation. Data transformation can refer to changing values ​​while preserving their meaning. For example, object data may include compound concentrations in mol / L, but stoichiometric models should determine concentrations in g / L. Therefore, data conversion may involve converting a value given in one unit to a value in a different unit.

[0117] The method of this invention includes generating at least two distinct candidate chemometric models. The candidate chemometric models can be chemometric models that have not yet been trained but incorporate algorithms ready for training. The term "at least two" can mean two or more, for example, 10 to 10,000 or 50 to 1,000. The term "distinct" can mean completely or partially different, i.e., the two candidate chemometric models differ in at least one aspect. The candidate chemometric models may differ in one aspect of the preprocessing or machine learning method, or in both aspects. The two candidate chemometric models may differ on one or more parameters used in the preprocessing or machine learning method. The two candidate chemometric models should be sufficiently different, i.e., a significant difference in their predictive accuracy can be expected. The selection of candidate models can involve results generated by previous chemometric models, e.g., combinations of parameters and methods that have produced satisfactory results for similar problems in the past. The selection of candidate models can involve experimental design methods, i.e., selecting a limited number of optimized combinations of methods and parameters to systematically cover the option space.

[0118] The candidate chemometric model of this invention includes a preprocessing method. The term "preprocessing" can refer to a method that reduces or eliminates spectral interferences (such as stray light, noise, or baseline drift) to enhance subsequent machine learning. Therefore, preprocessing methods can be applied before machine learning methods. Preprocessing may include one or more of baseline correction, scattering correction, smoothing, scaling, and aggregation.

[0119] Baseline correction can be selected from linear baseline correction, polynomial baseline correction, logarithmic baseline correction, spline baseline correction, Savitzky-Golay baseline correction, Kaiser-Bessel baseline correction, Gossky line correction, Lorentz baseline correction, derivative baseline correction including first or second derivative, continuous wavelet transform (CWT), or baseline correction by local regression.

[0120] Scattering correction can be selected from multivariate scattering correction (MSC), standard normal variable (SNV), background subtraction, dark current subtraction, scattering correction using scattering plots or scattering models, and wavelet denoising.

[0121] Smoothing can be selected from Savitzky-Golay filter (SGF), Kaiser-Bessel filter (KBF), Gaussian filter (GF), median filter (MF), moving average filter (MAF), exponential smoothing (ES), or Kalman filter (KF).

[0122] Scaling can be selected from linear scaling, polynomial scaling, logarithmic scaling, mean centering (MC), autoscaling (AS), Pareto scaling (PS), or max-min scaling (MMS).

[0123] Aggregation can be selected from generating median, spatial median, or k-medoid center points on a set of spectra (e.g., all measurements of a sample measured by a device).

[0124] Preprocessing may include one preprocessing method or a combination of more than one (e.g., two, three, four, or five). For example, preprocessing may include methods for baseline correction, methods for scatter correction, methods for smoothing, methods for scaling, and methods for aggregation. When preprocessing includes more than one preprocessing method, the preprocessing methods may be applied in various orders, wherein, in the context of this invention, changing the order is considered a different preprocessing. Typically, in chemometric models, preprocessing is applied before machine learning methods.

[0125] The candidate chemometric model of this invention includes a machine learning method. The term "machine learning method" can refer to a model that converts spectra into corresponding object data. Therefore, a machine learning method can use spectra as input and derive object data from them. The machine learning method can be considered a component of the chemometric model. The machine learning method can be supervised, semi-supervised, or unsupervised. Machine learning methods can include multivariate calibration, classification, pattern recognition, clustering, ensemble methods, neural networks and deep learning, or multivariate curve resolution.

[0126] The choice of machine learning method can depend on the object data to be determined. Therefore, selecting a machine learning model for a candidate chemometric model can involve identifying the object data from which the chemometric model should output. If a variable (e.g., concentration) needs to be determined, a multivariate calibration method can be chosen. If a categorical value (e.g., the type of compound) needs to be determined, a classification model or a spatial clustering model can be chosen. Selecting a machine learning model can also involve including identifiers in the object data that indicate which object data the chemometric model should output.

[0127] Multivariate calibration methods can be classical or inverse methods. Classical multivariate calibration models can be solved to optimally describe the spectrum. Inverse multivariate calibration methods can be solved to optimally predict object data. Therefore, inverse multivariate calibration methods are preferred. Multivariate calibration can be selected from partial least squares regression (PLS), principal component regression (PCR), penalized linear regression (Lasso regression, ridge regression, elastic network), random forest regression (RFR), support vector regression (SVR), gradient boosting regressor (GBR), vector regression (VR), matrix regression (MR), artificial neural networks (ANN), and multivariate adaptive regression splines (MARS).

[0128] The classification model can be selected from Linear Discriminant Analysis (LDA), Soft Independent Analogy Modeling (SIMCA), Penalized Linear Regression (Lasso Regression, Ridge Regression, Elastic Network), Artificial Neural Network (ANN), Naive Bayes, Random Forest Classification (RFC), and Support Vector Machine (SVM).

[0129] Spatial clustering models can be selected from K-means clustering analysis (KMCA), agglomerative hierarchical clustering analysis (AHCA), principal component analysis (PCA), fuzzy C-means clustering analysis (FCMCA), vertex component analysis (VCA), and split correlation clustering analysis (DCCA).

[0130] The candidate chemometric model of this invention may include a feature selection filter. The term "feature selection filter" can refer to a method for selecting those portions of a spectrum that are relevant to the object data. Feature selection filters can facilitate machine learning methods for chemometric models, thereby avoiding overfitting and reducing the amount of training data required. A feature selection filter can use a spectrum as input, remove all unselected portions, and output a spectrum retaining only the selected portions. Therefore, the output of the feature selection filter can be a spectrum represented as a vector with a dimension lower than the input vector. The output of the feature selection filter can be used as input to a machine learning method. Therefore, the feature selection filter can be applied before the machine learning method. The input to the feature selection filter can be a received spectrum or a preprocessed spectrum, preferably a preprocessed spectrum. Therefore, the feature selection filter can be applied after the preprocessing method.

[0131] Feature selection filters can be obtained through feature selection methods (i.e., methods used to determine the relevant portions of the spectrum selected by the feature selection filter). Feature selection methods can be univariate, meaning they select those portions of the spectrum that are relevant to the object's data, typically due to the object's physical or chemical properties. Feature selection methods can be sequential, meaning they sort the portions of the spectrum in order and pair them in a forward or backward manner. A portion of the spectrum can refer to a single intensity value at a specific wavelength or multiple intensity values ​​at different wavelengths (e.g., for a certain wavelength range). Feature selection methods can be multivariate, such as multiple linear regression (MLR), principal component regression (PCR), partial least squares regression (PLSR), interactive variable selection, uninformation variable elimination (UVE), interval PLS (iPLS), significance testing of model parameters, or genetic algorithms (GA).

[0132] Chemometric models that output more than one physical or chemical property can include different preprocessing methods for different physical or chemical properties. Chemometric models that output more than one physical or chemical property can include different machine learning methods for different physical or chemical properties. Chemometric models that output more than one physical or chemical property can include different feature selection filters for different physical or chemical properties.

[0133] Candidate chemometric models can be generated by selecting a preprocessing method, a machine learning method, and a feature selection filter, and then combining them into a candidate chemometric model. This selection may be limited by the use case. For example, to determine the concentration of a compound in an object, only a regression calibration method can be selected as the machine learning method, while to distinguish certain materials (e.g., different types of plastics), only a classification method can be selected as the machine learning method. This selection may also be limited by user input or configuration files.

[0134] Two candidate chemometric models may differ in at least one aspect, including preprocessing methods, machine learning methods, or feature selection filters. A candidate chemometric model may include at least two candidate chemometric models that incorporate different preprocessing methods. A candidate chemometric model may include at least two candidate chemometric models that incorporate different machine learning methods. A candidate chemometric model may include at least two candidate chemometric models that incorporate different feature selection filters. A candidate chemometric model may include at least two candidate chemometric models that incorporate different preprocessing methods and at least two candidate chemometric models that incorporate different feature selection filters. A candidate chemometric model may include at least two candidate chemometric models that incorporate different machine learning methods and at least two candidate chemometric models that incorporate different feature selection filters. A candidate chemometric model may include at least two candidate chemometric models that incorporate different preprocessing methods, at least two candidate chemometric models that incorporate different machine learning methods, and at least two candidate chemometric models that incorporate different feature selection filters. If a method includes at least one different parameter or parameter value, the method can be considered a different method.

[0135] The method of this invention includes training a candidate chemometric model. Training may include adjusting the parameters of the chemometric model such that the output of the chemometric model best fits the provided object data associated with the spectra used as input to the chemometric model. Typically, training includes minimizing a loss function or cost function, such as the minimum mean square value of the chemometric model output relative to the provided object data. The complete set or a portion of the received spectra and associated object data may be used for training. A portion of the received spectra and associated object data may be used for training, and the remainder may be used to determine the prediction accuracy of the trained candidate chemometric model. Alternatively, cross-validation, such as K-fold cross-validation, leave-one-out cross-validation, or hierarchical cross-validation, may be applied.

[0136] Data analysis can be performed on the received spectral and associated object data. The term "data analysis" can refer to analysis performed prior to training to determine whether the received spectral and object data is suitable for training a chemometric model. This analysis can involve anomaly detection, outlier detection, or signal-to-noise ratio analysis. Spectra exceeding a preset threshold may be deemed unsuitable and therefore ignored. This analysis can also involve object data. For example, it could involve checking whether the provided object data contains the information required for the chemometric model. In this way, spectra accidentally added from other use cases can be ignored. Another example could be checking within the object data whether all spectra were recorded with the same spectrometer or a similar spectrometer to ensure that a chemometric model designed for a particular spectrometer or type of spectrometer is trained correctly. Further checks can be made to ensure that all object data is within the expected range. The range of spectral values ​​can be checked, for example, to identify whether the spectrometer has reached its limits, i.e., whether it is outside its expected measurement range.

[0137] Unsuitable spectra, i.e., those identified as inappropriate, can be ignored. All unsuitable spectra or portions thereof can be ignored. The unsuitable spectra may be displayed to the user. In response, the user can provide input indicating which unsuitable spectra should be ignored. Therefore, ignoring unsuitable spectra can involve such user input. If the suitability analysis shows that the number of suitable spectra is too small, the user can be invited to provide further data, for example, by displaying such a request to the user interface.

[0138] The method of this invention may include data augmentation for enhancing the provided spectral and associated object data. The term "data augmentation" can refer to generating additional spectral and associated object data from the provided spectral and associated object data. By using data augmentation, the chemometric model can be trained with more training data, which can reduce overfitting. Therefore, candidate chemometric models can be trained using the augmented spectral and object data (i.e., the received spectral and object data and the spectral and object data obtained through data augmentation). Moreover, the chemometric model can be more generally applicable, i.e., suitable for measurement ranges where available training data is scarce or nonexistent. Thus, by using data augmentation, robust chemometric models can be obtained, and the need for large datasets containing spectral and associated object data is reduced.

[0139] Data augmentation typically involves generating generative models that produce spectra and associated object data. The generative model can determine the correlation between the spectra and associated object data based on the provided spectra and associated object data. Therefore, the generative model can be trained using the provided spectra and associated object data.

[0140] Data augmentation can involve identifiers contained within object data that indicate which physical or chemical properties a stoichiometric model should determine. These identifiers can be used to restrict the training of the generative model to object data indicated by that identifier. This can help avoid underfitting. Alternatively, user input or profiles can be used for this purpose.

[0141] Generative models can generate new spectra and associated object data by adding noise (such as Gaussian noise) to the provided spectral and object data. For example, the model can determine the standard deviation of each wavelength in the spectrum, select noise scaled according to the standard deviation, and add Gaussian noise to the spectrum at each individual wavelength.

[0142] Generative models can generate new spectral and associated object data by blending provided spectral and object data. Blending can refer to generating a linear combination of two or more spectral and associated object data. Therefore, the new data constitutes an interpolation of the provided data.

[0143] Generative models can generate new spectra and associated object data using synthetic minority class oversampling techniques (SMOTE). SMOTE can involve selecting similar spectra, such as those that are close to each other in the feature space, and generating new spectra along trajectories between the selected spectra in the feature space.

[0144] Generative models can incorporate artificial neural networks. These networks can be trained with provided spectral and associated object data and subsequently output new spectral and associated object data. Examples of generative models incorporating artificial neural networks include auto-augment (learning augmentation policies from data), FastAutoAugment, Faster AutoAugment, Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Conditional Variational Autoencoders (CVAEs), Deterministic Regularized Autoencoders (RAEs), or generative models based on normalized flow.

[0145] The generative model can be manually selected, thus using the same generative model at all times when the method of the present invention is performed. Alternatively, several different generative models can be tested against a specific chemosiometric model, wherein, for each generative model, the change in prediction accuracy of the chemosiometric model when trained with the augmented dataset can be determined relative to when trained using only the provided dataset. The generative model can then be selected, involving determining the change in prediction accuracy caused by training the chemosiometric model with the augmented dataset. For example, the generative model that maximizes the improvement in prediction accuracy of the chemosiometric model can be selected.

[0146] In some cases, using different datasets augmented by different generative models can result in different trained candidate chemometric models having different prediction accuracies. Therefore, multiple candidate generative models can be used for different candidate chemometric models, and the prediction accuracy can be determined for each combination. The selection of the generative model and the trained candidate chemometric model can involve determining the prediction accuracy in this way. Thus, the selection of the trained candidate chemometric model can depend on the selection of the generative model, and vice versa.

[0147] The method of this invention includes determining the prediction accuracy of a trained candidate chemometric model. Prediction accuracy can be a value indicating the difference between the output of the trained candidate chemometric model and measurement data (e.g., a validation set of received object data). Generally, a higher prediction accuracy indicates a better fit between the chemometric model output and the reference object data. Values ​​indicating the difference can be slope, bias, intercept, R-squared, calibration standard error (SEC), prediction standard error (SEP), calibration root mean square error (RMSEC), mean square error (MSE), prediction root mean square error (RMSEP), mean absolute error (MAE), performance bias ratio (RPD), or range error ratio (RER), or can be derived from them. The prediction accuracy of a chemometric model including a binomial classifier can be determined based on the area under the receiver operating characteristic (ROC) curve (AUC), sensitivity and specificity based on a specific cutoff value, Youden index, etc. The prediction accuracy of a chemometric model including a multinomial classifier can be determined based on the sensitivity, specificity, precision, F1 score, and accuracy for each class, as well as the mean and weighted average of all classes.

[0148] The method of this invention includes using prediction accuracy to select a chemostometric model from a pool of trained candidate chemostometric models. This may mean selecting a model from the trained candidate chemostometric models, which will become the selected chemostometric model. The trained candidate chemostometric model with the highest prediction accuracy can be selected. However, in some cases, it may be advantageous to select a trained candidate chemostometric model with a lower prediction accuracy. For example, a trained candidate chemostometric model with the highest prediction accuracy below a preset threshold can be selected. The preset threshold can be selected to exclude abnormally high prediction accuracies. Abnormally high prediction accuracy may indicate that the trained candidate chemostometric model is overfitting. Alternatively, multiple trained candidate chemostometric models can be output to a user interface, for example, two to ten or three to five models can be output to the user interface, and the user can be prompted to select an input. The trained candidate chemostometric models and their prediction accuracies can be output. The output trained candidate chemostometric model may be the trained candidate chemostometric model with the highest prediction accuracy. The number of trained candidate chemostometric models to be output to the user interface may be a small fraction of the number of generated trained candidate chemostometric models, for example, 1% to 5%. Then, the selection of the trained candidate chemometric model can involve user input.

[0149] The method of this invention includes outputting a selected trained chemometric model. The term "output" can refer to writing the selected trained chemometric model to a non-transitory data storage medium, such as writing it to a file or database, displaying it on a user interface (e.g., a screen), or both. The selected trained chemometric model can also be output to a cloud system via an interface for storage and / or further processing. The selected trained chemometric model can further be output via an interface to a spectrometer that can perform spectroscopic measurements using the selected trained chemometric model. The selected trained chemometric model can be output along with its prediction accuracy. In this case, the user can estimate the success rate of the method and decide whether the output chemometric model should be adopted.

[0150] The present invention further relates to a system for generating a chemometric model for a spectrometer. The system includes an input for receiving the spectrum of an object and object data associated with the object's physical or chemical properties. The input may include an interface for receiving data. The input can receive data locally or remotely, for example, via an interface to a telecommunications system (such as the Internet). The input can receive data directly from the spectrometer or via a programmable logic controller or a storage medium including cloud services.

[0151] The system further includes a processor configured to perform the steps of the method of the present invention. The processor may be a local processor, including a central processing unit (CPU) and / or a graphics processing unit (GPU) and / or an application-specific integrated circuit (ASIC) and / or a tensor processing unit (TPU) and / or a field-programmable gate array (FPGA). The processor may also be an interface to a remote computer system, such as a cloud service.

[0152] The system further includes an output terminal for outputting the selected trained chemosiometric model. The output terminal may include an interface for outputting the selected trained chemosiometric model. The output terminal can send the selected trained chemosiometric model locally or remotely, for example, via an interface with a telecommunications system (such as the Internet). The output terminal can send the selected trained chemosiometric model to a programmable logic controller or a storage medium including cloud services.

[0153] The present invention further relates to a non-transient computer-readable medium comprising instructions that, when executed by one or more processors, cause the processors to perform the method according to the invention. The term "computer-readable data medium" can refer to any suitable data storage device or computer-readable storage device on which one or more sets of instructions (e.g., software) are stored, embodying any one or more of the methods or functions described herein. The instructions may also reside wholly or at least partially in main memory and / or processor during their execution by a computer, main memory, and processing device that may constitute a computer-readable storage medium. These instructions may further be transmitted or received over a network via a network interface device. Computer-readable data media include, for example, hard disk drives on servers, USB storage devices, CDs, DVDs, or Blu-ray discs. The computer program may contain all the functions and data required to perform the method according to the invention, or may provide an interface to allow portions of the method to be processed on a remote system (e.g., a cloud system). Attached Figure Description

[0154] Figure 1 It demonstrates how to obtain object data and spectra from objects.

[0155] Figure 2 An embodiment of the method of the present invention is shown.

[0156] Figure 3 An example of how to generate candidate chemometric models is shown.

[0157] Figure 4 The potential feature-selective filter for the spectrum is demonstrated.

[0158] Figure 5 An example of a stoichiometric model is shown.

[0159] Figure 6 An embodiment of the system of the present invention is shown.

[0160] Figure 7 An exemplary user interface of the system of the present invention is shown. Detailed Implementation

[0161] Figure 1This demonstrates the acquisition of object data and spectra from an object. Spectrometer 101 can illuminate object 103, such as a nutrient, drug, or body part, with a beam of light 102. The spectrometer may include a light source (e.g., an incandescent lamp or LED) and illumination optics. The beam of light 102 may contain light within the infrared wavelength range (e.g., 780 nm to 3000 nm). The beam of light may illuminate the object and be reflected back to the spectrometer. This measurement mode is commonly referred to as a reflection mode. Alternatively, the beam of light may penetrate the object and be guided back to the spectrometer, for example, by a reflector. This measurement mode is commonly referred to as a transmission mode. Spectrometer 101 may include light-collecting optics for capturing the beam of light 102 propagating from object 103 to spectrometer 101. Spectrometer 101 may include optical elements that spatially or temporally separate different wavelengths of the beam of light 102 propagating from object 103 to spectrometer 101. Spectrometer 101 may include one or more optical sensors that generate electrical signals based on the intensity of light incident on the optical sensors. For example, intensity data can be correlated with wavelength data using a processor or microcontroller in the spectrometer. Therefore, spectrometer 101 can output spectrum 104, for example, as a vector of wavelength values ​​and an associated vector of intensity values. Object 103 can be analyzed in a laboratory to obtain object data 106, such as the content of compounds in object 103 determined by gas chromatography. For example, the two can be correlated by labeling spectrum 104 and object data 106 with an object identifier.

[0162] Figure 2 An embodiment of the method of the present invention is illustrated. Spectrum 201 and object data 202 can be received, for example, via a web interface. Object data 202 is associated with spectrum 201, for example, by an identifier representing the sample from which the spectrum originates and with which the object data is associated. The received spectrum 201 and object data 202 can be analyzed at 203 to determine their suitability for training a chemometric model. Those IR spectra 201 and object data 202 determined to be unsuitable can be discarded, i.e., not used for training. The remaining IR spectra 201 and object data 202 can be augmented at 204. Data augmentation 204 can involve a generative model that is trained using spectrum 201 and object data 202 and generates additional spectral and object data, for example, spectral and object data similar to but not identical to spectrum 201 and object data 202.

[0163] 210 candidate chemometric models can be generated. This generation may include selecting a preprocessing method 211, such as a baseline correction method, a scattering correction method, a smoothing method, and / or a scaling method. The generation may further include selecting a machine learning method 212, such as multivariate calibration. The generation may further include selecting a feature selection filter 213, for example, a deletion operation that removes everything from the spectrum 201 except for a specific set of peaks. The generation may include merging the selected preprocessing method, the selected machine learning method, and the feature selection filter, for example, by concatenating the output of the preprocessing method with the input of the feature selection filter, and concatenating the output of the feature selection filter with the input of the machine learning model. Untrained candidate chemometric models can be obtained, i.e., candidate chemometric models containing parameters that have not yet been adjusted to fit the spectrum 201 and the object data 202. The generation of candidate chemometric models 210 can be repeated several times to obtain several different candidate chemometric models, for example, 200 or 400 different candidate chemometric models.

[0164] Candidate chemometric models can be trained (221), where training can use all or part of the spectra (201) and object data (202). If data augmentation (203) has been performed, the augmented dataset is used, i.e., the provided spectra (201) and object data (202) as well as newly generated spectra and object data. After the candidate chemometric models are trained, their prediction accuracy (222) can be determined, for example, by determining the root mean square error of the predicted object data relative to a validation dataset of the object data (e.g., a portion of the provided object data (202)). Alternatively, cross-validation can be applied, for example, by K-fold cross-validation. Subsequently, the trained candidate chemometric models (223) can be selected, for example, the trained candidate chemometric model with the highest prediction accuracy. The selected chemometric model (224) can be output, for example, by writing the model to a data storage device, sending the model to a spectrometer via a communication interface, or providing the model for download via a web interface.

[0165] Figure 3An example method for generating candidate chemometric models is shown. The candidate chemometric model may include a preprocessing method 310, a machine learning method 320, and a filter 330 for the relevant part of the spectrum. Preprocessing 310 may include baseline correction 311, scattering correction 312, smoothing 313, and scaling 314. As an example, baseline correction 311 can be implemented by selecting one from a first-derivative method 311a and a continuous wavelet transform method 311b, or without baseline correction. Scattering correction 312 can be implemented by selecting one from a standard normal variable method 312a and a multivariate scattering correction method 312b, or without scattering correction. Smoothing 313 can be implemented by selecting one from a moving average filter 313a and exponential smoothing 313b, or without smoothing. Scaling can be implemented by selecting one from mean centering 314a and Pareto scaling 314b, or without scaling. Machine learning 320 may include multivariate calibration 321, classification 322, and spatial clustering 323. As an example, multivariate calibration 321 can be achieved by selecting one from partial least squares regression 321a and principal component regression 321b, or by not performing multivariate calibration. Classification 322 can be achieved by selecting one from artificial neural network 322a and linear discriminant analysis 322b, or by not performing classification. Spatial clustering 323 can be achieved by selecting one from K-means clustering analysis 323a and agglomerative hierarchical clustering analysis 323b, or by not performing clustering. Feature selection filter 330 can preserve the complete spectrum 331, select a first wavelength range 332 (e.g., 1.1 to 1.4 µm), select a second wavelength range 333 (e.g., 1.6 to 1.8 µm), select a third wavelength range 334 (e.g., 1.9 to 2.3 µm), or a combination thereof.

[0166] As an illustrative example, three candidate chemometric models can be generated. The first candidate chemometric model 341 may include a first derivative model 311a as baseline correction 311, no-scattering correction 312, exponential smoothing 313b as smoothing 313, mean centering 314a as scaling 314, principal component regression 321b as multivariate calibration 321, and a complete spectrum 331 as a feature selection filter 330 (meaning that a feature selection filter is essentially not used). The second candidate chemometric model 342 may include a continuous wavelet transform method 311b as baseline correction 311, a multivariate scattering correction method 312b as scattering correction 312, no smoothing 313, Pareto scaling 314b as scaling 314, linear discriminant analysis 322b as classification 322, and selected spectral portions 332 and 333 as feature selection filters 330. The third candidate chemometric model 343 may include baseline-free correction 311, standard normal variable method 312a as scattering correction 312, moving average filter 313a as smoothing 313, no scaling 314, artificial neural network 322a as multivariate calibration 321, and selected spectral portions 332 and 334 as feature selection filters 330.

[0167] Figure 4 A potential feature-selective filter for the spectrum is illustrated. Spectrum 400 plots the relationship between radiant intensity 402 and wavelength 401. In this example, the wavelength range of 1 to 3 µm is plotted. Spectrum 400 contains three maxima or peaks, one at 1.6 µm, one at 2.2 µm, and one at 2.8 µm. Each peak can potentially be correlated with object data. Therefore, the range around the peaks can be selected as the potentially correlated portions of the spectrum, in this example, the first portion 410, the second portion 420, and the third portion 430. A feature-selective filter can be selected that removes everything from the spectrum except for one or more of the first portion 410, the second portion 420, and the third portion 430. Another feature-selective filter can be formed that leaves all or almost all of the spectrum unaffected. Another possibility is to form a feature-selective filter that removes portions of the spectrum where intensity 402 does not exceed a predefined value (e.g., 0.1). In this way, low-intensity regions that typically do not contain relevant information can be discarded. As mentioned above, more complex methods for feature-selective filters are available, but in essence, they work as follows... Figure 4 As shown.

[0168] Figure 5An example of a chemometric model is illustrated. The chemometric model 510 may include a preprocessing method 511, a feature selection filter 513, and a machine learning method 515. A spectrum 501 may be provided as input to the chemometric model 510. Spectrum 501 may be a vector of multiple intensity values ​​at different wavelengths obtained from a spectrometer. Spectrum 501 may first be preprocessed using preprocessing 511. Preprocessing may include one or more of baseline correction, scattering correction, smoothing, and scaling. The preprocessing output is a preprocessed spectrum 512. The preprocessed spectrum may be a vector of multiple intensity values ​​at different wavelengths, where these values ​​may have been modified by preprocessing compared to spectrum 501. The preprocessed spectrum 512 may then pass through a feature selection filter 513. The feature selection filter 513 may remove certain intensity values ​​that have been identified as being almost irrelevant or unrelated to the expected object data that the chemometric model should determine, in this example, I'2(λ2). Therefore, the feature selection filter 513 outputs the relevant portion 514 of the preprocessed spectrum. Spectrum 514 can be a vector with a dimension lower than spectrum 501 or 512. Spectrum 514 can be input into machine learning method 515, which then outputs object data 520. In this example, the output object data could be the concentration of the compound to be analyzed being 0.23 g / L.

[0169] Figure 6 An embodiment of the system of the present invention is illustrated. The system 620 for generating a chemometric model can be a cloud-based service. The system 620 can receive spectra 611 and object data 612 via a web interface (e.g., via an upload function). The IR spectrum 611 can be recorded using a spectrometer 601 or a similar device. The object data 612 can be obtained through laboratory analysis. The system 620 can use the spectra 611 and object data 612 to perform the method of the present invention. The system 620 can, for example, output the chemometric model 630 directly to the spectrometer 601 via a communication interface, or output it to a smartphone 605 communicatively coupled to the spectrometer 601. The smartphone 605 can use the chemometric model 630 by triggering the spectrometer 601 to measure the spectrum of object 603. The spectrometer 601 can transmit the spectrum to the smartphone 605. The processor of the smartphone 605 can execute code that uses the chemometric model 630 to convert the spectrum into object data. The object data can be displayed on the display of the smartphone 605. Alternatively or additionally, the smartphone's processor can execute an application configured to further transform object data into action recommendations. For example, if object 603 is human skin and chemometric model 630 is adjusted to detect skin moisture levels, the recommendation could be to apply a certain amount of skincare product to the skin.

[0170] Figure 7An exemplary user interface of the system of the present invention is shown. The user interface can be a computer application or a web interface. The interface can display a window of the chemometric model generator 700. The interface can have an upload area where the user can upload spectra 711 and object data 713. The interface can provide buttons 712, 714 for browsing the corresponding files. After successfully completing the upload, the user can choose to trigger the chemometric model generation method by clicking the generate model 721. The generated model can be stored in a buffer memory, and the generated model can be downloaded from the buffer memory by clicking the button marked as download model 722.

[0171] This disclosure has also been described in conjunction with various preferred embodiments and examples. However, by studying the accompanying drawings, this disclosure, and the claims, those skilled in the art, as well as those who practice the claimed invention, will understand and implement other variations.

[0172] Any steps presented in this document can be performed in any order. The methods disclosed herein are not limited to a specific order of these steps. Nor is it required that different steps be performed in a particular place or on a particular computing node in a distributed system; that is, each step can be performed on different computing nodes using different devices / data processing.

[0173] As used herein, "determine" also includes "initiating or causing determination," "generate" also includes "initiating and / or causing generation," and "provide" also includes "initiating or causing determination, generation, selection, sending, and / or receiving." "Initiating or causing an action" includes any processing signal that triggers a computing node or device to perform a corresponding action.

[0174] In the claims and specification, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude multiple. A single element or other unit can perform the function of several entities or items recited in the claims. The fact that certain measures are recited only in mutually different dependent claims does not indicate that a combination of these measures cannot be used in an advantageous implementation. In the claims and specification, the word "comprising" or similar wording does not exclude other elements or steps and should not be construed as limiting to the listed elements or steps. The indefinite article "a" or "an" does not exclude multiple. A single element or other unit can perform the function of several entities or items recited in the claims. The fact that certain measures are recited only in mutually different dependent claims does not indicate that a combination of these measures cannot be used in an advantageous implementation or can include additional elements.

[0175] Within the scope of this disclosure, provision may include any interface configured to provide data. This may include application programming interfaces, human-machine interfaces (such as displays), and / or software module interfaces. Provision may include transmitting or submitting data to an interface, particularly displaying data to a user or using data by a receiving node, entity, or interface.

[0176] Various units, circuits, entities, nodes, or other computing components may be described as being “configured to” perform one or more tasks. “Configured to” should be described as meaning “having a circuit system that performs one or more tasks during operation.” Units, circuits, entities, nodes, or other computing components may be configured to perform tasks even when the unit / circuit / component is not operating. Units, circuits, entities, nodes, or other computing components forming a structure corresponding to “configured to” may include hardware circuitry and / or memory storing executable program instructions to perform the operation. For convenience in the description, units, circuits, entities, nodes, or other computing components may be described as performing one or more tasks. This description should be interpreted as including the phrase “configured to.” Any statement of “configured to” is not expressly intended to invoke the interpretation of 35 USC § 112(f).

[0177] Generally, the methods, apparatuses, systems, computer elements, nodes, or other computing components described herein may include memory, software components, and hardware components. Memory may include volatile memory (such as static or dynamic random access memory) and / or non-volatile memory (such as optical or magnetic disk storage devices, flash memory, programmable read-only memory, etc.). Hardware components may include any combination of the following: combinational logic circuit systems, clock storage devices (such as flip-flops, registers, latches, etc.), finite state machines, memory (such as static random access memory or embedded dynamic random access memory), custom-designed circuit systems, programmable logic arrays, etc.

[0178] Any disclosures and embodiments described herein relate to the methods, systems, apparatuses, devices, chemicals, materials, and computer program elements listed above, and vice versa. Advantageously, the benefits provided by any embodiments and examples also apply to all other embodiments and examples, and vice versa. All terms and definitions used herein are to be understood broadly and have their general meanings.

Claims

1. A computer-implemented method for generating a chemometric model for a spectrometer, the method comprising: a) Receive the spectral data of an object and object data associated with the object's physical or chemical properties. b) Generate at least two distinct candidate stoichiometric models, including i) Preprocessing methods, ii) Machine learning methods, iii) Feature selection filter, c) Use these spectral and object data to train these candidate chemometric models. d) Determine the prediction accuracy of the trained candidate chemometric models. e) Use the prediction accuracy to select the chemostometric model from these trained candidate chemostometric models, and f) Output the selected stoichiometric model.

2. The method according to claim 1, wherein, The candidate chemometric models include at least two candidate chemometric models that include different preprocessing methods, at least two candidate chemometric models that include different machine learning methods, and at least two candidate chemometric models that include different feature selection filters.

3. The method according to claim 1 or 2, wherein, The method further includes data augmentation to enhance the provided spectral and associated object data, and using the enhanced spectral and object data to train the candidate chemometric model.

4. The method according to claim 3, wherein, Data augmentation involves artificial neural networks as generative models.

5. The method according to any one of claims 1 to 4, wherein, These spectra represent the absorbance or transmittance of radiation with wavelengths from 760 nm to 3 µm.

6. The method according to any one of claims 1 to 5, wherein, The object data includes data associated with the object's chemical composition.

7. The method according to any one of claims 1 to 6, wherein, Before training, the received spectral and associated object data are analyzed to determine their suitability.

8. The chemometric model obtained by the method of any one of the preceding claims is used for the purpose of a spectrometer.

9. A spectrometer configured to use a stoichiometric model obtained by the method of any one of claims 1 to 7.

10. The spectrometer according to claim 9, wherein, This spectrometer is a portable spectrometer.

11. The spectrometer according to claim 10, wherein, The spectrometer is communicatively coupled to a computer device configured to execute the chemometric model.

12. A system for generating a chemometric model for a spectrometer, the system comprising: a) Input terminal, used to receive the spectrum of the object and object data associated with the physical or chemical properties of the object. b) A processor for generating at least two distinct candidate stoichiometric models, the processor comprising: i) Preprocessing methods, ii) Machine learning methods, iii) Feature selection filter, These candidate chemometric models are trained; their prediction accuracy is determined using spectral and object data; and this prediction accuracy is used to select the chemometric model from among these trained candidate chemometric models. c) Output terminal, which is used to output the selected stoichiometric model.

13. A non-transitory computer-readable medium comprising instructions that, when executed by one or more processors, cause the one or more processors to perform the following methods: a) Receive the spectral data of an object and object data associated with the object's physical or chemical properties. b) Generate at least two distinct candidate stoichiometric models, including i) Preprocessing methods, ii) Machine learning methods, iii) Feature selection filter c) Use these spectral and object data to train these candidate chemometric models. d) Determine the prediction accuracy of the trained candidate chemometric models. e) Use the prediction accuracy to select the chemostometric model from these trained candidate chemostometric models, and f) Output the selected stoichiometric model.

14. The non-transient computer-readable medium according to claim 13, wherein, These candidate chemometric models further include feature selection filters.

15. The non-transient computer-readable medium according to claim 13 or 14, wherein, The method further includes: data augmentation for enhancing the provided spectral and associated object data.

Images