Estimation of elemental content in molten material at the opening of a metallurgical vessel

Abstract

A system (1000) including a computer (200) estimates the content of a specific chemical element in molten material available at an opening (110) of a metallurgical vessel (100). In the system (1000), a pyrometer (140-P) monitors the opening (110) and provides radiation data (150-P) representing thermal radiation from the molten material (300) at the opening (110) for at least two wavelengths (λ1, λ2). The computer (200) processes the radiation data (250) to estimate the content using a pre-trained estimation module (250). Because phenomena (160) at the opening (110), such as deposits, fumes, or reflections, can interfere with the estimation, the system (1000) further includes a camera (140-C) monitoring the opening (110 and an image classification module (260) for identifying the phenomena (160). The computer (200) obtains a phenomenon-based reliability classification (RC) and improves the overall estimation accuracy by filtering reliable emission data from being processed or ignoring content based on unreliable emission data.

Description

[Technical Field]

[0001] Generally, the present disclosure relates to metallurgical vessels or furnaces, and more particularly, the present disclosure relates to computer systems, methods, and computer program products for processing measurement data taken during the loading and casting process. [Background technology]

[0002] The metallurgical industry ("iron and steel industry") is known for using all kinds of large containers to hold liquid materials during at least many stages of production. Briefly, a metallurgical vessel is a device for producing and transporting metals. Typically, the vessel holds the material at a relatively high temperature for several hours to melt the material, cause chemical reactions (e.g., converting ore to metal), and the like. Openings in the vessel allow the molten material to flow out of (or into) the vessel.

[0003] A typical example of a metallurgical vessel comprises a furnace, which has several openings and sections. Again, simply stated, the furnace has an inlet opening at the top and a corresponding section for supplying material (e.g., a "feed inlet," usually raw material that is not yet melted), and one or more outlet openings (e.g., "tapholes") in the button, along with a corresponding section (usually called the "casting floor").

[0004] A prominent example of such a furnace is a blast furnace, which produces iron (or more specifically, "pig iron"). The blast furnace has further openings, such as tuyere openings. While chemical reactions occur, the outlet opening at the bottom of the furnace is kept closed by clay (also called "mud" or similar). No material leaves the furnace.

[0005] Sometimes operators use drilling machines to remove the clay, a process known as "tapping," which opens tapholes.

[0006] In the next processing step, commonly called "casting" or "casting," the molten material (i.e., metal and slag) flows through the taphole into a runner and then into another vessel. Casting typically takes several hours (to remove, e.g., a 2,500 cubic meter volume of molten material from a furnace).

[0007] During casting, workers can inspect samples of the molten material for chemical composition, temperature, appearance, etc. Workers can take material samples with tools (such as immersion lances), for example, from the runners, or from skimmers (mechanical structures that separate metal from slag).

[0008] The sensors are distributed in various locations in the furnace and can provide measurement data at virtually any time. Operators use the measurement data to fine-tune the operation of the furnace, usually with an optimization goal in mind, such as to minimize energy consumption or to minimize emissions to the environment (such as CO2 emissions). In other words, the data relates to the amount of coal or other fuel in the material, the energy used to heat the furnace, etc.

[0009] Temperature measurements at the taphole (or elsewhere) are particularly important: they relate to the temperature of the molten metal (such as the "hot metal temperature" HMT) and to the temperature of the slag.

[0010] Other measurements target the chemical composition of materials, such as to identify the silicon (Si) content in metals.

[0011] However, the following problems also exist. Workers cannot always take samples during casting. Taking material samples poses a relatively high safety risk. On the other hand, non-contact measurement tools are not always accurate. Two reasons should be mentioned: the taphole may be obscured by fumes or other contaminants, and the materials in the taphole (i.e., metal and slag) have different emissivities (effectiveness at emitting energy as thermal radiation) at different wavelengths, which affects accuracy. Alternative measurement techniques such as laser-induced plasma spectroscopy (LIPS) may pose additional safety risks (due to the application of lasers and the generation of plasma).

[0012] A recent publication by Pauna et al. describes the application of spectrometers during steel processing (Non-Patent Document 1). Summary of the Invention [Means for solving the problem]

[0013] Embodiments of the present invention are directed to a computer system and computer-implemented method for estimating the content of specific chemical elements in molten material available at an opening of a metallurgical vessel.

[0014] A pyrometer is mechanically positioned on the container at a monitoring distance. The pyrometer monitors the opening, and a computer receives radiation data from the pyrometer. The radiation data represents thermal radiation from the material at the opening at at least two wavelengths. A processing module has a regression model for estimating the content of specific chemical elements (i.e., "content data"), the processing module being pre-trained.

[0015] In an embodiment, a pyrometer may measure emissions from a taphole in a blast furnace. To further increase accuracy, a camera monitors the opening (or the path between the opening and the pyrometer). Images are processed to obtain a confidence classification of the emission data.

[0016] A system is provided for estimating the content of a particular chemical element in a molten material available at an opening of a metallurgical vessel.

[0017] A pyrometer is adapted to monitor the aperture and provide radiation data representative of thermal radiation from the molten material at the aperture for at least two wavelengths.

[0018] The camera is adapted to monitor the opening or to monitor a path between the opening and the pyrometer and thus provide one or more images showing the opening or showing the path.

[0019] The computer includes an estimation module and an image classification module. The estimation module is adapted to process the emission data using a regression model to provide an estimation using preliminary content data for specific chemical elements in the molten material. The estimation module is trained by training data collected during a data collection phase. The collected data is a combination of historical emission data and historical content data (obtained by measurements on samples taken from the molten material at an aperture).

[0020] The image classification module is adapted to classify one or more images to identify phenomena in the opening or path, the phenomena selectively including at least one of the following: (i) deposits, fumes, or reflections located in the path, (ii) the degree of alignment of the pyrometer with respect to the opening, and (iii) percentages of metal and slag derivable from the image.

[0021] The image classification module is then further adapted to provide a confidence classification of the radiation data according to the identified phenomenon, such that the computer outputs the preliminary content data as content data only if the confidence classification complies with a predetermined rule.

[0022] Alternatively, the reliability classification can be used to enable or disable the estimation module. Unreliable emission data does not need to be processed. Such an approach may have the advantage of saving computational resources.

[0023] Optionally, the computer further comprises a feature enhancer module adapted to process the emission data into feature data according to the plurality of feature enhancement rules. The estimation module is adapted to further process the feature data together with the emission data, wherein the emission data together with the feature data becomes feature-enhanced emission data. The estimation module has been trained with training data further including feature data identified according to the plurality of feature enhancement rules.

[0024] A computer-implemented method for estimating the content of a particular chemical element in a molten material (available at an opening in a metallurgical vessel) is provided, in which a computer receives, from a pyrometer positioned to monitor the opening, radiation data representing thermal radiation from the molten material at the opening for at least two wavelengths.

[0025] The computer receives one or more images from the camera showing the opening or showing the path between the opening and the pyrometer.

[0026] The computer classifies one or more images to identify phenomena, which selectively include at least one of the following: (i) deposits, fumes, or reflections located in the path; (ii) the degree of alignment of the pyrometer relative to the aperture; and (ii) the percentage of metal and slag on (i.e., derivable from) the image. The computer classifies the emission data from the pyrometer as reliable (trusted) or unreliable (unreliable) emission data according to the identified phenomena.

[0027] The computer operates a processing module that processes the reliable radiation data, the processing module having a pre-trained regression model for estimating the content of a particular chemical element.

[0028] Optionally, operating the processing module comprises operating a module trained with training data collected during the data collection phase, the collected training data being a combination of historical emission data and historical content data (obtained by measurements on samples taken from the molten material at the orifice).

[0029] Optionally, in a further receiving step, the computer receives sensor data from sensors positioned to monitor the container, these sensors being in addition to the camera, and in the step of operating a processing module, the computer processes the emission data in combination with the sensor data.

[0030] Optionally, after receiving the sensor data, the computer performs a step of classifying one or more pieces of data to obtain a reliability classification as intermediate data. Thereby, the computer performs a step of operating the processing module also using the reliability classification as intermediate data. This approach may be advantageous for further using the reliability classification. The classification separates reliable emission data from unreliable emission data (the processing module processes the reliable emission data), while in the case of reliable emission data, the classification itself can be processed by the processing module.

[0031] Optionally, the computer performs the classifying step by an auxiliary machine learning tool, in other words, the classification module may be provided and trained separately from the processing module.

[0032] Optionally, the computer performs the step of classifying by an auxiliary machine learning tool implemented by an autoencoder.

[0033] Optionally, in the receiving step, the computer receives the feature-enhanced emission data, and in the operating step, the computer processes the feature-enhanced emission data. Feature enhancement is also applied during training.

[0034] The molten material may be a composition comprising a molten metal or metal alloy, with the specific chemical elements selected from carbon (C), silicon (Si), iron (Fe), and sulfur (S).

[0035] The method steps may be applied to a metallurgical vessel that is a blast furnace, for which the opening comprises a taphole and runners for transporting material from the blast furnace during casting.

[0036] Optionally, the computer performs the further step of presenting an estimate using the content data to a blast furnace operator.

[0037] The method can be used repeatedly to control the movement of the container.

[0038] A computer program product that, when loaded into the memory of a computer system and executed by at least one processor of the computer system, causes the computer system to perform the steps of a computer-implemented method.

[0039] The computer system comprises a number of modules that perform the steps of the computer-implemented method.

[0040] A system is provided for estimating the content of a particular chemical element in a molten material available at an opening of a metallurgical vessel.

[0041] In other words, a system including a computer estimates the content of a specific chemical element in molten material available at an opening in a metallurgical vessel. In this system, a pyrometer monitors the opening and provides radiation data representing thermal radiation from the molten material at the opening. The computer estimates the content through a pre-trained module that processes the radiation data. Because phenomena at the opening, such as deposits, fumes, or reflections, can interfere with the estimation, the system further includes a camera monitoring the opening and an image classification module that identifies the phenomena. The computer obtains a phenomenon-based confidence classification and improves overall estimation accuracy by filtering out unreliable radiation data from processing or ignoring contents that would be based on unreliable radiation data. [Brief explanation of the drawings]

[0042] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0043] [Figure 1] FIG. 1 illustrates a highly simplified side view of a metallurgical vessel and an operator using content data to control the operation of the vessel.

[0044] [Figure 2] FIG. 1 illustrates an example vessel, a blast furnace, equipped with pyrometers during a data collection phase.

[0045] [Figure 3A] FIG. 1 illustrates a computer with a processing module during a training phase. [Figure 3B] FIG. 1 illustrates a computer with a processing module during a training phase, further illustrating a feature enhancer module.

[0046] [Figure 4] FIG. 1 shows a blast furnace still equipped with pyrometers during the estimation phase.

[0047] [Figure 5A]FIG. 1 illustrates a computer-implemented method for estimating the content of specific chemical elements in molten material available at an opening of a metallurgical vessel, and provides an overview of the data collection, training, and estimation stages. [Figure 5B] FIG. 1 illustrates a computer-implemented method for estimating the content of certain chemical elements in molten material available at an opening of a metallurgical vessel and also illustrates optional steps.

[0048] [Figure 6] FIG. 10 shows the vessel in relation to a pyrometer and a camera to symbolize the sensors that provide data to a computer.

[0049] [Figure 7A] 1A and 1B show time series with emission data and sensor data, and FIG. 1C shows emission data and sensor data. [Figure 7B] FIG. 10 shows a time series with emission data as well as sensor data, illustrating the application of feature enrichment rules to obtain further feature data.

[0050] [Figure 8] FIG. 10 shows an image from a camera in which metal and slag are differentiated by classification, which therefore also provides sensor data.

[0051] [Figure 9] FIG. 1 shows an image from a camera with the phenomenon.

[0052] [Figure 10] FIG. 1 is a diagram showing an overview of a blast furnace and a computer.

[0053] [Figure 11] FIG. 1 illustrates a general-purpose computer. DETAILED DESCRIPTION OF THE INVENTION

[0054] Notational Conventions This specification sometimes refers to human workers fulfilling different roles, such as the role of container worker, the role of sample worker (or rather sample collection worker), etc. The roles can be performed by the same person or by different people.

[0055] This specification occasionally refers to time series with data. The notation {} denotes a univariate time series, and the notation {{}} denotes a multivariate time series. For convenience, a more detailed introduction to time series is provided at the end of this specification in connection with Figures 7A and 7B.

[0056] As used herein, the term "computer" (in the singular, with or without reference) refers to a computing function or a function of a computer-implemented module (such as a processing module). The computing function may be distributed across different physical computers.

[0057] The drawings also illustrate computer programs or computer program products that, when loaded into the memory of a computer and executed by at least one processor of the computer, cause the computer to perform the steps of a computer-implemented method. In other words, the program provides instructions for computer-implemented modules.

[0058] From different perspectives, the figures show modules of a computer system comprising a number of computer-implemented functional modules that, when executed by the computer system, perform steps of a computer-implemented method.

[0059] Although this specification focuses on computers performing computer-implemented methods, in some cases the computers and methods are placed in the context of industrial systems in that the metallurgical vessel is a system component.

[0060] This specification refers to techniques already known in the art, among them: The controller can implement a control loop within the industrial system. This specification describes use cases for the results (such as content data) provided by the computer when executing the method. The control loop does not have to be implemented by the controller; in some situations, it may be sufficient for a container operator to view the content data and interact with the container accordingly. · The auxiliary machine learning (ML) tool provides intermediate data to the computer by using commercially available techniques, such as classifying images to obtain classification data as intermediate data.

[0061] As a result, implementation details for the controller and supporting ML tools are available to those skilled in the art.

[0062] In the context of this patent application, industrial systems (or industrial machines) are not considered to be computer-implemented modules (as they do not perform method steps).

[0063] In this specification, the "sampling interval" used in signal processing is described as the time interval for which data is made available in digital form, but the "sample collection time point" is used for the time at which the sampler collects the physical sample.

[0064] The term "element" refers to a chemical element. overview

[0065] FIG. 1 illustrates a highly simplified, symbolic side view of a metallurgical vessel 100. FIG. 1 also illustrates a vessel operator 190 ("operator" for short) using content data 270 / 370 to control the operation of vessel 100. Vessel 100 is part of an industrial system that includes sensors, actuators, computers, data communication devices, etc., and many other components or equipment. For simplicity, details of such systems have been omitted from the illustration.

[0066] As shown on the left by part A of the figure, the container 100 has an opening 110 for allowing the molten material 300 to exit the container 100. Part B shows a slightly modified example with the container 100' having an opening 110' for receiving the molten material 300.

[0067] The opening 110 can be further differentiated into a hole 120 and a conductor 130 (shown only in part A), and the flow of molten material 300 is shown from left to right.

[0068] Molten material 300 can be further differentiated into metal 350 and slag 360, shown as solid and dotted lines, respectively. This differentiation may be more relevant for vessel 100 in part A than for vessel 100 in part B.

[0069] The functions of vessels 100' and 100 may be combined, for example, into a vessel for transporting materials within a plant.

[0070] A worker 190 (at the vessel) needs to know the content of several chemical elements (ultimately differentiated into metals 350 and slag 360) in the material 300. The content is represented here by content data 270 / 370 symbolized by a pie chart (not scaled).

[0071] This figure uses reference numeral 370 for content data obtained by more traditional methods (such as sample collection with subsequent laboratory analysis) and reference numeral 270 for content data obtained by computer-implemented methods.

[0072] The elemental contents in metals 350 may be given in percentages for elements such as carbon C, silicon Si, iron Fe, and sulfur S. The slag 360 contents may be given in terms of other chemical elements. Operator 190 uses content data 270 / 370 to control the operation of vessel 100 (e.g., by changing feed materials, etc., or by changing other operating parameters).

[0073] Control is not limited to vessel 100, and content data 270 / 370 can be used to control other equipment that further processes the material. For example, molten material 300 from vessel 100 (e.g., iron from a blast furnace) typically passes to another vessel (e.g., vessel 100') that further processes the material (e.g., to produce steel in a steel plant).

[0074] In other words, the content data 270 / 370 is relevant to controlling the container 100 (which provides the material 300), but also to controlling further processing steps (by controlling further equipment such as the container 100').

[0075] Below, the specification describes a computer-implemented technique for estimating content data 270.

[0076] In this specification, a detailed description will be given by using an example in which the vessel 100 is a blast furnace. Those skilled in the art are very familiar with such devices and know various implementations of tapholes (i.e., examples of holes 120) and runners (i.e., examples of conductors 130). A real-world furnace may have two or more tapholes, two or more runners, etc. Further implementations usually include so-called skimmers (to separate the metal 350 from the slag 360). The tapholes and runners transport the molten material 300 (i.e., the metal 350 / slag 360) from the blast furnace during casting.

[0077] Although computer-implemented, this approach also uses data obtained by traditional tools (see Figure 2) and uses a distinction between stages, where the term "stage" is used here to mean "time interval."

[0078] Figure 1 shows the items ** We refer to the items by calling them 0, whereas other figures distinguish items according to stage (see Figure 5A for further overview). Data collection stage **In 1, an industrial system (having a vessel 101) collects data that will be used for training, the details of which are outlined in FIG. Training stage ** In 2, a computer-implemented tool, shown as processing module 252, is trained, and is outlined in more detail in FIGS. 3A and 3B. Estimation stage ** In 3, the computer executes a method for estimating content data, the details of which are outlined in FIG. 4 (apparatus, computer 203) and FIGS. 5A and 5B (method 403).

[0079] In other words, throughout this specification: ** 1 / ** 2 / ** Reference numerals such as 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, ** In 3, the computer estimates the elemental content of the material.

[0080] Again, the focus of this specification is on the computer (and computer-implemented methods), and it does not matter whether vessel 100 is a blast furnace or other metallurgical vessel. assignment

[0081] However, most scenarios have technical challenges, only a few of which are mentioned here. The molten material 300 in the vessel 100 can be dangerous to human workers 190. The molten material 300 has a temperature of the order of 1,000°C or more and must not come into physical contact with the workers 190 at any time. The workers' protective clothing may not function. The openings 110 (especially the tapholes) may open unexpectedly, so humans should not remain in the vicinity, but may need to take samples that can only be obtained from the openings (see Figures 2 and 4, sample workers 171 and 173). Determining the content data 370 typically involves sending a material sample to a laboratory (see Figure 2), but the results with the content data arrive with a waiting period. Such delays make it more difficult to control the vessel (see DELAY in Laboratory 181 in Figure 2). During a single casting process (which can take several hours), samples can only be taken once or twice, making continuous monitoring impossible. However, for real-time control of the operation of the vessel 100 (and other equipment), continuous monitoring (which is relatively short, see δT_COM in Figure 4) is preferred. Inspection of the molten material 300 in the container 100 is complicated for additional reasons (e.g., the opening of the container is usually closed most of the time, and although material can be obtained in different layers within the container, the material composition of the opening matters). Solution Overview

[0082] While the operator continues to send material samples to the laboratory (i.e., the traditional method of measuring content), the computer-implemented technique assists the operator by estimating the content. The computer uses data from a sensor, not from a physical sample. A pyrometer is an example of a sensor that provides such data. To keep the illustration as simple as possible, this specification refers to a pyrometer (and data from a pyrometer) as a sensor that collectively represents multiple sensors (of different types and measuring different properties or items).

[0083] In this specification, measurement data (ie, data from a sensor) is distinguished as "emission data" when the sensor is a pyrometer, and as "sensor data" when the sensor is not a pyrometer.

[0084] Such "non-pyrometer" sensors are described and discussed below (FIGS. 6-9), with cameras being a representative example. Generally, the description and drawings use the reference numeral 140 for the sensor and the reference numeral 150 for the measurement data (from the sensor).

[0085] Reference numbers 140 / 150 may use acronyms such as "P" (for "pyrometer") or "C" (for "camera").

[0086] In some cases, the reference numbers distinguish between stages.

[0087] Figure 2 shows the data collection stage. ** 1 shows a blast furnace 101 (i.e., an example of a vessel) equipped with a pyrometer 141-P. Simply put, a pyrometer measures infrared (IR) radiation at one or more wavelengths λ passing through an optical system. The term "one wavelength" is understood as the range (i.e., bandwidth) between a minimum and a maximum wavelength. An equivalent term to "wavelength" would be the radiation frequency (i.e., calculated using the reciprocal value).

[0088] Pyrometer 141-P is positioned to monitor aperture 111, and pyrometer 141-P provides radiation data 151-P. Radiation can be measured as "spectral energy density," and those skilled in the art are familiar with the basic principle that, very simply, the radiation emitted by an object depends on the temperature of that object. Emissivity is a function of the object's temperature and its elemental composition, and radiation can be described by an equation that has the so-called Planck constant as a component.

[0089] Depending on the implementation, the pyrometer 141-P is located within a monitoring distance, typically within a few meters, from the opening 111. A minimum monitoring distance avoids thermal damage to the pyrometer. An adequate monitoring distance allows for measurement of radiation while limiting external influences on the measurement. As the distance increases, perturbations to the measurement also increase. The pyrometer can be mounted on a holder such that its relative position (with respect to the vessel) remains substantially unchanged (FIG. 10 shows the pyrometer on a holder support or carriage). The support allows the pyrometer to be adjusted to capture radiation from the opening.

[0090] The radiation data 151-P represents the thermal radiation emitted by the molten material 301 at the opening 111. Because the light emission resulting in the radiation is wavelength dependent, the pyrometer 141-P makes the radiation data 151-P available for at least two wavelengths. Thus, the pyrometer 141-P is a multi-wavelength pyrometer. The data is available in a computer-readable format for processing by a computer, and the data represents, for example, radiation intensity at particular wavelengths (such as λ1 and λ2).

[0091] Those skilled in the art of pyrometers may also refer to aperture 111 as the "target" because the primary field of view of pyrometer 141-P is directed toward aperture 111. Optional techniques for aligning pyrometer 141-P with aperture 111 (to expose the pyrometer to radiation from the entire aperture and to maintain a constant relative position) are described in connection with FIG.

[0092] The emission data 151-P that becomes available over time (e.g., λ1 and λ2) may be available in computer-readable form as univariate time series {λ1}1 and {λ2}1, or in other notations, as multivariate time series {{λ}}1, with index i=1 to N. The number "1" on the right indicates the data collection stage. ** Represents 1.

[0093] The number of wavelengths (i.e., "first and second") need not be N=2; Figure 2 shows λ1 having N=3 lines (i.e., three signals or "channels"). It is contemplated to use state-of-the-art pyrometers with N=5 channels (i.e., λ1 through λ5).

[0094] The pyrometer 141-P may provide data values ​​for λ at typically equidistant sampling intervals (such as ΔT). For illustrative purposes, a ΔT of 1 second may be assumed. This is merely an example, and shorter ΔT, such as 125 milliseconds, are also contemplated. General pyrometer

[0095] Pyrometers are typically commercially available for remote temperature measurement applications, for example from Paul Wurth SA, 32, rue d'Alsace, L-1122 Luxembourg, LUXEMBOURG.

[0096] In the context of a blast furnace, US Pat. No. 4,619,533 describes a pyrometer that measures the furnace bath temperature through a tuyere.

[0097] Note that this method of estimating content does not require that the temperature be calculated. Nevertheless, the pyrometer (or better, the electronics within the pyrometer) provides the temperature data, which is presented to the operator (usually on a large display). One skilled in the art can intercept radiation data 151-P (see 153-P in FIG. 4) directly from the pyrometer's optical sensor. Physical Sample

[0098] FIG. 2 further shows a sample operator 171 taking a physical sample 311 (of the material 301 available at the opening 111).

[0099] Sampling is known in the art and will not be discussed in great detail in the description and drawings herein. Those skilled in the art will be able to further distinguish between metal samples (taken from the molten metal) and slag samples (taken from the slag). The sample operator 171 may take the sample from the runner (see 130 in FIG. 1) that carries the material rather than directly from the taphole (from 110 in FIG. 1), or may take the sample after the skimmer (which separates the metal from the slag). One implementation is shown in FIG. 10.

[0100] It is known that samples are sent to a laboratory (here laboratory 181), which delivers content data 371 as measurement results, also shown in time series. Because the time between successive samples is relatively long (i.e., in hours) compared to ΔT of λi (at the pyrometer), the diagram symbolizes the content data 371 by dots (rather than lines).

[0101] The laboratory 181 can typically provide separate content data 371 for different chemical elements, i.e., elements that are expected to occur in the molten material, e.g., Fe, C, S, Si, Ti, Mn, P, Cu, Cr, etc., though not for all of the more than 100 chemical elements. The figure shows the multivariate time series {{content}} (variables corresponding to elements) along with the univariate time series.

[0102] Continuing with the blast furnace example, a single casting process has an average duration of about three hours, but the contents of chemical elements may vary during casting. While a laboratory may provide content data 371 with a relatively high degree of accuracy (i.e., with a relatively narrow tolerance band), the data from the laboratory may no longer correspond to reality when the data arrives from the laboratory. There is a delay from taking physical samples (by operator 171 at t1, t2, etc.) to having the results as content data 371.

[0103] As an example, a multivariate time series with content data 371 is {{Content}}1={{Content-C}1, {Content-Si}1, {Content-Fe}1, {Content-S}1} where again the number "1" may represent the data collection stage. The acronyms are the elemental symbols for carbon, silicon, iron, and sulfur, respectively. The figure shows the {{content}}1 along with the time point (i.e., when the sample 311 was taken).

[0104] It is advantageous to distinguish between metals and slag. In that case, there can be a first multivariate time series for metals and a second multivariate time series for slag. Slag can be distinguished differently for other elements.

[0105] The figure shows content data 371 as a percentage (weight wt%, mass, or volume, as selected by one of skill in the art) for separate sample collection times t1, t2, etc. These times indicate the time at which the physical sample was taken, rather than the availability of delayed results. For ease of illustration, dots represent content values ​​for only one element.

[0106] Although this specification refers to multivariate time series {{contents}}, it is also possible to limit the analysis to a single element within a univariate time series.

[0107] The time series λ1 and the content λ1 (for at least one univariate for one chemical element) are subjected to the following training phase: ** Form the training dataset that will be used in 2. Control Loop

[0108] FIG. 2 also shows a controller 501 as an optional configuration that can use content data 371 as an input to implement the control loop described above.

[0109] The controller 501 can use the content data 371 to control the vessel and to control further equipment. However, DELAY may conflict with efficiency control. For example, by the time the content data 371 becomes available (i.e., sample time t1 or t2 + DELAY), casting may already be finished (vessel empty), so that vessel control may no longer be applicable. Only further equipment control would still be possible. Training the Processing Module

[0110] Figure 3A shows the training phase. **2 shows a computer 202 having two processing modules 252.

[0111] On the left, the diagram repeats {{λ}}1 (emission data 151-P) in combination with {{content}}1 (laboratory content data 371) from Figure 2, but with added detail. The combination of emission data 151-P and content data 371 becomes training data 241.

[0112] As mentioned above, ΔT is the sampling interval (for the radiation data). The content data {{content}}1 are presented in clusters (clustered by temporal proximity during casting, dashed circles) for the following reasons: The opening 111 (of the vessel 101) is not open all the time, but only temporarily (only during casting in the case of a blast furnace). Physical samples 311 are not taken at other times (because they are not available). · The number of samples taken during a single casting process is limited (to one or two, possibly three samples per casting). Therefore, the temporal distance (between t1 and t2, t2 and t3, etc., or "δT_LAB", Figure 2) is not constant. DELAY is not an issue, since t1, t2, etc. can be easily matched to the time points of the emission data.

[0113] The radiation data can be provided without regard to casting and is available even if casting has not yet begun or has been completed. Data not related to casting is excluded from the training data 241.

[0114] Nevertheless, the content data {{content}}1 can be considered ground truth for training. In other words, {{content}}1 can be reference data for training. Although the processing module 252 is shown with a single line in the output, there are several outputs separated by element (see, for example, C, Si, Fe, etc.).

[0115] Data is collected over a time interval called T_historical. As time goes on, this interval gets longer. It is convenient to assume that T_historical is at least greater than one year.

[0116] However, it will be appreciated that the interval T_historical is divided into shorter time frames such that training data 241 is provided to the processing module 252 over multiple iterations M. Staying with the blast furnace example, these shorter time frames "T_cast" may relate to individual casting processes. T_cast may begin when the piercer begins to open the taphole, and T_cast may end when the furnace empties (in the sense that molten material stops flowing from the furnace).

[0117] It is common for each individual casting to be numbered consecutively (by a "unique casting ID") and for further details to be recorded such as the identification of the particular taphole (in the case of furnaces with multiple tapholes), the volume of molten material obtained, and the dimensions of the taphole (perforation length).

[0118] Those skilled in the art can review historical data, for example, by comparing data from individual castings, can ignore data from some castings, or can ignore data that is not related to the casting. For example, confidence scores can be assigned to particular castings. For example, a particular casting process for which, for some reason, sample 311 (see FIG. 2) and its associated content data 371 do not exist is not considered to serve as training data (or is excluded from the training data set). In other words, some sections of T_historical may be ignored, resulting in a larger error in T_cast*M. <T_historicalとなる。

[0119] Applying a confidence score is just one example for disregarding data. Further below, this specification describes the concept of confidence classification RC (see FIG. 6 for an explanation), which can be applied in a similar manner.

[0120] The content data 371 can be considered "annotations" to the emission data 151-P, but does not require expert human oversight. The content data 371 will vary from casting to casting.

[0121] Thus, in a single training run, the module 252 receives training data 241 in M ​​data sets and obtains module internal values ​​accordingly, which will be different for different architectures as will be explained next. Training a processing module using feature-enhanced radiation data.

[0122] FIG. 3B illustrates the training phase with an additional feature enhancer module 282. ** 3B shows a computer 202 with a processing module 252 in FIG. 2. In other words, FIG. 3B is a copy of FIG. 3A that also shows the enhancer module.

[0123] The emission data 151-P can be enhanced with features, so that the feature enhancer module 282 (which may be part of the computer 202 or external to the computer 202) can identify the features according to a number of feature enhancement rules (see FIG. 7B for examples). The emission data 151-P becomes feature-enhanced emission data 151-F. In other words, the combination of the emission data and the feature data is called "feature-enhanced emission data." This is described in more detail below with reference to FIGS. 7A and 7B.

[0124] The feature-enhanced emission data 151-F does not need to be "old", since enhancement can be the immediate previous step before training. However, it is possible to store the feature-enhanced emission data 151-F (and use it, for example, as training data) if desired.

[0125] By referencing the reference number "283" in parentheses, the figure illustrates that feature enhancer 282 / 283 is ** This shows that it can also work with 3.

[0126] The use of the feature enhancer module 282 (and also module 283) is optional, but feature enhancement can increase the estimation accuracy (ie, estimation of the content of chemical elements at a higher resolution).

[0127] Implementation of the feature enhancement rules can include using a software library, building a feature enhancement module with an autoencoder, and other implementations.

[0128] The module can also be implemented as a machine learning module that is trained to select feature reinforcement rules from the above rules, and auto-encoding is also an option. In other words, the reinforcement rules can be obtained by training. Model Architecture

[0129] The processing module 252 uses an underlying model. The model can be based on so-called "deep learning," and those skilled in the art are familiar with selecting an appropriate model. The following paper provides useful information: Amal Mahmud, Ammar Mohammed, "A Survey on Deep Learning for Time-Series Forecasting," in the book "Machine Learning and Big Data Analytics Paradigms: Analysis, Applications and Challenges (pp. 365-392)," January 2021, DOI: 10.1007 / 978-3-030-59338-4_19.

[0130] The paper mentions architectures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memories (LSTMs), gated recurrent units (GRUs), deep autoencoders (AEs), restricted Boltzmann machines (RBMs), and deep belief networks (DBNs).

[0131] Due to the nature of the data being multivariate time series, RNNs, LSTMs, GRUs, or CNNs applied to time series are particularly suitable architectures.

[0132] Optionally, deep autoencoders can be used to identify features that can be used to train architectures such as those described in papers such as Karl-Philipp Kortmann, Moritz Fehsenfeld, and Mark Wielitzka, "Autoencoder-based Representation Learning from Heterogeneous Multivariate Time Series Data of Mechatronic Systems," arXiv:2104.02784.

[0133] Other traditional machine learning algorithms such as neural networks, XGBoost, and random forests are also suitable for training modules where the amount and size of historical data for deep learning architectures is limited.

[0134] The model architecture is ** 3 is conveniently described in terms of "inputs" and "outputs" (e.g., with data flow from left to right, as in Figure 1 of the Mahmoud et al. paper). Module 252 also has its "inputs" and "outputs." Training ** 2, the emission data 151-P is provided at "input" and the content data 371 is provided at "output."

[0135] It is possible to train the processing module 252 separately for different chemical elements.

[0136] Figure 4 shows the estimation stage ** 3 shows a blast furnace 103 equipped with a pyrometer 143-P.

[0137] 2-3 regarding instruments (such as pyrometer 141-P in FIG. 2) and time series are also applicable to the training and estimation phases. In other words, by way of example, pyrometer 143-P in FIG. 4 is positioned to monitor aperture 113 and provide emission data 153-P (for at least two wavelengths λ1 and λ2). The data corresponds to thermal emission of at least two wavelengths from molten material 303 at aperture 113, as described in FIG. 2.

[0138] When executing the computer-implemented method, the computer 203 operates a processing module 253 that processes the radiation data 153-P (at least data available before the time t_current, the current time) (see step 453). The processing module 253 has a regression model (the architecture allows regression) for estimating the content of certain chemical elements.

[0139] As already explained in Figure 3, the processing module 253 is pre-trained. The time required for the computer to execute the method is negligible (compared to δT_COM and also relatively short for ΔT, see Figure 2), so the computer can repeat the method multiple times. As a result, by receiving updated emission data, the computer can update its estimates.

[0140] The computer 203 provides content data 273, but without safety risks to the workers (unless an alternative approach such as LIPS described above is applied). The container worker 193 is not involved in providing the content data 273. The sample worker 171 (who takes the samples) is not involved in providing the content data 273. (Note that every new sample taken and evaluated by the laboratory can potentially serve as training data, see Figures 3A and 3B).

[0141] Vessel 101 / 103 and pyrometer 141-P / 143-P may be the same physical entity.

[0142] Figure 4 also indicates by dashed lines that computer 203 can optionally receive sensor data 153-X, as described in more detail below with reference to Figures 6-10. Figure 4 also shows T_cast (discussed above) and introduces a time interval T_back.

[0143] For ease of illustration, Figure 4 omits the feature enhancer module 283 (see Figure 3B), which may optionally use feature enhancement to improve accuracy, as described above. In such an implementation, the processing module 253 receives feature-enhanced emission data 153-F (shown in parentheses in the figure) instead of data 153-P. Control Loop

[0144] Optionally, the controller 503 controls the operation of the vessel 103. Briefly, the controller 503 can process the content data 273 to control signals for the vessel 103 (i.e., furnace) or an industrial system. It does not matter whether a vessel operator 193 uses the control signals to adjust operating parameters or whether the controller 503 automatically acts on the vessel 103. In either case, the content data 273 can be considered as data describing the state of the technical system.

[0145] 4 symbolizes the controller 503 and control signals with dashed arrows from the output of computer 203 to vessel 103. In other words, a control loop can be established using pyrometer 143-P and computer 203. Content data 273 serves as input data for the loop. Controller 503 processes content data 273 to control signals. Vessel 103 (and optionally other parts of the industrial system) receive control signals (to set operating parameters, etc. for the industrial system).

[0146] Suitable control targets are mentioned only as examples, as those skilled in the art will be able to configure the controller 503 accordingly without further explanation herein. In a first example, the control loop aims to minimize energy consumption (of the vessel 103 or of the entire industrial system). In a second example, the control loop is aimed at a vessel that provides a material with a content that is being optimized for follow-up action (e.g., to provide pig iron with a predetermined elemental content so that the iron is optimized for conversion to steel in a subsequent process step). In the third example, the control loop aims to reduce the environmental impact of the industrial system, for example by reducing emissions of substances (e.g., CO2) into the environment.

[0147] Multiple control loops can be implemented, with content data 373 serving as input to multiple controllers. Stage Considerations

[0148] Although Figures 2-4 show the apparatus for separate stages, these stages are merely a convenient convention for describing operations that should be performed in a sequential order.

[0149] Operating the vessel 101 during the data collection phase (see FIG. 2) is a condition for training the processing module 252 in the training phase (see FIGS. 3A and 3B) using the training data acquired during the data collection phase.

[0150] Training the processing module 252 in a training phase is a prerequisite for estimating content (see FIG. 4, providing content data 273).

[0151] However, in a real-world scenario, collecting data for training purposes (as in the data collection phase) is more or less always possible. Even if the processing module 253 (see FIG. 4) provides content data 273, manual sampling (by a sampler 173) continues to obtain further training data and evaluate the content data 273 (from module 253).

[0152] The operation of the vessel during the data collection phase (as vessel 101 in the continuation of that phase) and during the estimation phase (as vessel 103) continues independently from training (see Figures 3A and 3B).

[0153] The sample operator 173 is still active in taking the sample 313 (see FIG. 4). The time series λ and the content can be considered "historical data", while the continued operation of the vessel 101 / 103 (inferential phase) ** 3), providing data that can be used for a dual purpose: to determine content (see Figure 2, content data from the laboratory 371) and to provide further training data (see Figures 3A and 3B). method

[0154] 5A and 5B show a computer-implemented method 403 (dashed box in FIG. 5A) for estimating the content of a particular chemical element in the molten material (303 in FIG. 4) available at the opening of a metallurgical vessel (113, 103, see FIG. 4). In other words, the method 403 includes an estimation step ** Related to 3. Estimating the content yields content data 273 (see also FIG. 4).

[0155] For convenience, Figure 5A places method 403 in the context of other stages. The diagram shows the stages from left to right in columns: data collection stage ** 1 (leading to training data 241, Figures 3A and 3B), training phase ** 2 (leading to the processing module 252 being trained, also FIGS. 3A and 3B), and the estimation stage **3 (using method 403).

[0156] More specifically, the computer-implemented method 403 includes step 413 "Receive Data" and step 453 "Operate a Processing Module."

[0157] From the pyrometer 143-P positioned to monitor the opening 113, the computer 203 receives radiation data 153-P (i.e., optionally as feature-enhanced radiation data 153-F) representing thermal radiation from the material 303 at the opening 113 for at least two wavelengths λ1, λ2 (step 413).

[0158] The computer 203 operates a processing module 253 that processes the emission data 151-P (step 453). The processing module 253 includes a regression model for estimating the content of a particular chemical element.

[0159] Although the specification refers to the singular "element," the method can also be carried out for multiple elements (see, for example, C, Si, Fe, and S), either separately (sequentially or in parallel) for each element, or in combination.

[0160] The processing module 253 is pre-trained as previously described (see Figures 3A and 3B). Optional Steps

[0161] 5B indicates by dashed boxes that method 403 can optionally include further steps 423, 433, 443, and 463. The optional steps are introduced for optional application of sensor data (detailed below in FIGS. 6-9). Applying optional steps can be advantageous due to timing and accuracy considerations described below.

[0162] The figure shows a method 403 with optional steps (in dashed boxes) on the left, but also shows the application of these steps in a simplified step sequence on the right.

[0163] The computer 203 optionally receives 423 sensor data (150-X in FIG. 7A, or otherwise represented as {{X}}) from sensors positioned to monitor the vessel 103 (and / or monitor the industrial system).

[0164] Within the same dashed box, Figure 5B also shows a further optional step 433, "Adapt sensor data." The computer can perform this adaptation step with an auxiliary ML tool (see adapter 260 in Figure 6 for an example), or otherwise. The adaptation results in intermediate data. An example of adapting image data to identify phenomena and derive classification data will now be described.

[0165] In the "Training-Based Application" column, FIG. 5B shows that when sensor data (150-X) is received (and still optionally adapted), step 453 of operating the processing module can also be performed by processing the sensor data.

[0166] Further steps are step 443 "Evaluate Emission Data" (or "Evaluate Sensor Data") and step 463 "Evaluate Content Data". Evaluation steps 443 and 463 are optionally performed after the data to be evaluated is available. Briefly, the result of the evaluation can allow the sensor data to be used in a subsequent step (such as operation 453) or block the sensor data, or the result of the evaluation can "allow" or "block" the content data.

[0167] In the "Rule-Based Application" column, Figure 5B shows that if sensor data is still received and still optionally adapted, step 443 can still be performed to evaluate the sensor data. An example is described herein below.

[0168] The same column "Rule-based application" indicates further evaluation of content data based on sensor data in step 463. An example of such application is described with reference to FIG. 6, in that preliminary content data is forwarded (allowed) or blocked. Timing and Accuracy

[0169] The computer can repeatedly execute method 403 (FIG. 5A) to estimate "new" content (i.e., provide content data 273) for each iteration of the method. A time diagram for content data 273 is shown in FIG. 4, where each dot corresponds to a method run. For ease of explanation, it can be further assumed that content data 273 is available at substantially equidistant times in the output interval δT_COM (where "COM" stands for "computer").

[0170] The repetition of the method can be synchronized with the availability of data at the input (of the computer 203, see FIG. 4). For example, if emission data 153-P becomes available every ΔT, estimates for content data 273 can be available at intervals δT_COM approximately equal to ΔT.

[0171] Those skilled in the art will configure a computer with sufficient processor and memory capabilities so that the runtime of a single method execution is negligible (with respect to the interval δT_COM).

[0172] In short, running a computer to perform the method or operating a laboratory both transforms input information into output information, but the method does not virtualize a laboratory and uses data that is not derived from physical samples. Furthermore, the "DELAY" of the method is negligible.

[0173] In either case, the interval δT_COM (ie, in seconds) is much shorter than the interval δT_LAB of the content data 371 from the laboratory 181 (see FIG. 2, eg, about 20 minutes).

[0174] Given the optional controller 503 (FIG. 4) introduced above, which can implement a control loop, it may not be necessary to repeat the method 403 at a relatively high rate. As explained above, the content data 273 can be used to control the operation of the vessel (and the operation of the industrial system to which the vessel belongs). It can be assumed that control signals that change the operation of the vessel are required at time intervals longer than δT_COM but shorter than δT_LAB.

[0175] However, despite such advantages in "speed", i.e., availability of content data 273 in δT_COM<δT_LAB and the substantial lack of DELAY, at least the following accuracy aspects should be considered: Tolerance band, and · Reliability (in other words, the question of data reliability).

[0176] With respect to the tolerance band, the content data 273 may be given as a percentage (or other) and the container operator 193 will view the data with an understanding of the tolerance band. As a simplified example, {content-Si} may be a percentage between 0.5 and 1.5 percent, and the tolerance band may be assumed to be plus / minus 0.2 percent (or 0.4 percent from lower limit to upper limit).

[0177] Naturally, the trained processing module 252 / 253 will not be more accurate than its laboratory master (i.e., the training data 241 using the content data 371 of FIGS. 3A and 3B is likely to be more accurate). The tolerance band may be wider (e.g., 0.4 percent as mentioned for content data 273) than the laboratory measurement (e.g., 0.1 percent for content data 371).

[0178] However, obtaining content data 273 in a real-time scenario (from a computer) with relatively low accuracy (i.e., a larger tolerance band) may be more suitable for control purposes than obtaining content data 373 with relatively high accuracy (but a shorter tolerance band from a laboratory).

[0179] Regarding reliability, it should be noted that despite the relatively high output rate (δT_COM in FIG. 4, which is relatively shorter than δT_LAB in FIG. 2), the emission data 153-P is not always accurate.

[0180] The pyrometer may still provide emission data that is within a normal tolerance range between minimum and maximum values ​​(i.e., for a particular wavelength), but the data may not correspond to the emission. As will be explained, a computer can process the sensor data (see 150-X in FIG. 7A) so that the emission data can be differentiated by a reliability classification, or RC for short (see FIG. 6). The RC may be a binary classification, for example, into "reliable" or "unreliable," or may be a classification by degree of reliability.

[0181] This document describes a technique for obtaining the RC by processing an image (ie, the further data is an image). Overview of sensors and sensor data

[0182] Although a method 403 for processing emission data 153-P (training data 151-P, also implemented as feature-enhanced emission data 153-F) has been described, industrial systems have many more sensors. Processing data from such sensors can improve the accuracy of content data 273. Figure 5B already introduces step 423 of receiving sensor data, step 433 of adapting the sensor data, and their application (by processing step 453 and by evaluation steps 443, 463).

[0183] Below, this specification provides a sensor overview (FIG. 6) and a time series overview (FIGS. 7A and 7B) of the measurement data (i.e., data from the sensor). Next, this specification discusses the use of the sensor data to evaluate content data 273 (see optional step 463 in FIG. 5B). Those skilled in the art can carry over these teachings when evaluating emission / sensor data.

[0184] Note again that the use of sensor data is optional.

[0185] While the computer processes the emission data through a processing module (as described for step 453 of FIG. 5A), the computer can use the sensor data in at least two applications: (Rule-Based Application) The computer can use the sensor data (see 150-X in FIG. 7A) to evaluate the content data 273 by applying predefined rules (see step 463). This scenario is illustrated by the phenomenon and confidence classification RC (see FIG. 6). (Training-based approach) The computer can use sensor data (see 150-X in FIG. 7A) in processing module 253 (see step 453 in FIG. 5B). In that case, the sensor data is treated like emission data, and the data collection and training described above is applied. In other words, the sensor data can be used in parallel with or in combination with emission data (see FIG. 7A).

[0186] In both applications, the sensor data can be used to narrow the tolerance band (of the content data 273 since the processing of step 453 uses more data) and increase the confidence in the content data 273 (or at least provide a confidence classification for further evaluation).

[0187] The computer can use some sensor data in both applications. Herein, image data is an example of sensor data used in both rule-based and training-based applications (see FIG. 5B). Given the particular sensor data, these approaches are complementary rather than exclusive.

[0188] Since the distinction between stages is not an issue, Figs. ** 0 notation is used. In other words, the diagram ** 1 (data collection), and ** It can also be applied to 3 (Estimation of content). Auxiliary Machine Learning Tools

[0189] Thus far, this specification has described the processing module 252 / 253 (during training and estimation) in detail as a machine learning tool. In a training-based approach, the historical sensor data is part of the training data 241 of Figures 3A and 3B.

[0190] However, further tools, also referred to below as "auxiliary ML tools", can be used to adapt the sensor data in step 433.

[0191] Traditional stage distinction ** 1. ** 2. ** 3 does not necessarily apply to these auxiliary tools. Some of these auxiliary tools use image processing and can be trained in other ways. Or, the auxiliary tools can use pre-made models. For example, an auxiliary ML tool for image processing can be trained in stages. ** 2 can be trained during the phase where there is no need to respond.

[0192] Figure 6 shows the vessel 100 and (from left to right) the pyrometer 140-P and camera 140-C that provide data to the computer 200. Figure 6 shows the concept; there may be more than one pyrometer or more than one camera. The camera 140-C that provides the image: Monitoring the phenomenon to obtain a confidence classification RC (e.g., see steps 443, 463 for rule-based application of sensor data to evaluate the data); Providing sensor data to be processed by module 253 (see optional step 423 of FIG. 5B as an example of training-based application of sensor data); This is just one example of a dual-purpose image for

[0193] The computer 200 comprises a processing module 250 for processing the emission data 150-P (optionally as feature-enhanced emission data 150-F) into content data 270 (modules 252 / 253 as described above, see method 403 of FIG. 5A). This diagram is stage independent, and therefore: ** Use 0 style reference numbers.

[0194] Container 100 has opening 110 (see FIG. 1), and corresponding to FIGS. 2 and 4, pyrometer 140-P monitors radiation (dashed arrow). Camera 140-C captures image 150-C of opening 110 (or its surroundings) but ignores certain wavelengths. In other words, pyrometer 140-P is monochromatic and camera 140-C is polychromatic. The image capture generates sensor data, which the computer receives in step 423 (see FIG. 5B).

[0195] Image 150-C shows a phenomenon 160 that, in some circumstances, can make the estimation by method 403 (FIG. 5A) unreliable. The phenomenon 160 is detected by fitting (see step 433 in FIG. 5B). Briefly, the presence or absence (or degree) of the phenomenon 160 is given in the intermediate data.

[0196] For example, the following phenomenon 160 is given. Dust (or other particles) may be present in the space (or path 110 / 140) between the opening 110 and the pyrometer 140-P. Such dust attenuates the radiation, which may cause the pyrometer 140-P to detect the radiation erroneously. Or, even worse, the pyrometer 140-P outputs radiation data 150-P that the computer 200 cannot process (or that will result in erroneous content data 270). Such a scenario is particularly expected during casting (hot material may simply interact with the air). In the metallurgical industry, dust or other particles are sometimes referred to as "deposits" or "fumes." The pyrometer 140-P may not be aligned with the aperture 110 and may potentially only partially view the aperture. An ideally aligned pyrometer shows radiation from the aperture as a whole. Pyrometer 140-P may see substantially the slag but not the metal, but the content should be estimated for the metal and not the slag.

[0197] Phenomenon 160 should be understood in a more abstract sense, and the phenomenon need not be located within a path.

[0198] If properly detected, such and other undesirable effects of phenomena 160 can be taken into account and ultimately mitigated. If dust or the like is detected, the content data 270 can be ignored, or if a misalignment is detected, the pyrometer 140-P can simply be realigned. To detect such phenomena 160, one skilled in the art can use an adapter 260 implemented by an auxiliary ML tool.

[0199] However, phenomenon 160 may not be limited to dust and misalignment, phenomenon 160 may include other things that can be converted into data.

[0200] Method 403 (see FIG. 5B ) may optionally include the following: in a receiving step, computer 200 receives one or more images 150-C showing aperture 110 (or showing path 110 / 140 between aperture 110 and pyrometer 140-P). In a classification step, computer 200 classifies one or more images 150-C to identify phenomenon 160. This classification can be performed by adapter 260 implemented by an auxiliary ML tool. The adapter operates within the functionality of an image classification module. In a further classification step, the computer can classify emission data 150-P from pyrometer 140-P as reliable (unreliable) emission data according to the identified phenomenon 160. In other words, adapter 260 provides a reliability classification RC. This classification in the adapter can also be performed by an auxiliary ML tool.

[0201] FIG. 5B already shows optional steps, more generally the receiving step 423, the adapting step 433 (to classify the phenomenon) and the evaluating step 443 (here for evaluating the emission data).

[0202] Again, the stages ** 1 or ** 3 is not a problem. Reliability classification RC is performed in the data collection stage. ** In 1, and in the estimation stage ** 3. For example, unreliable data may not be suitable for training.

[0203] 6 shows RC and content data 270 as separate outputs from computer 200, presentation of content data 270 to operator 193 (see FIG. 4 as content data 273) can be conditional, such as presenting only content data for "trustworthy" RC. In other words, this is just one further example of the optional evaluation in step 463 of FIG. 5B. Data evaluation steps may be related to each other; for example, evaluating the emission data as untrustworthy (step 443) may also mean that the content data is evaluated as untrustworthy (step 463, "switch open," no transfer). Reliability

[0204] In other words, the confidence classification RC can be used as a criterion for using the content data 270 (for control or other purposes) or for ignoring them (for example, if the binary RC is negative).

[0205] Depending on the RC, the operator 193 may apply several corrective actions, among them: applying measures to remove fumes from the path, realigning the pyrometer, and disregarding content data in situations where the slag percentage is relatively high (e.g., more slag than metal at the taphole).

[0206] However, processing an image to obtain a confidence classification RC is just one example for using sensor data (i.e., data that is not radiometric data). This specification now describes several approaches for using sensor data in content estimation (training-based application, steps 423 and 453). Considering the optional step 423 introduced above (see FIG. 5B), the sensor data (with or without adaptation) has been processed by module 253 (the module has been trained on past sensor data). system

[0207] FIG. 6 is also useful for illustrating a system 1000 for estimating the content of specific chemical elements in molten material (available at opening 110 of metallurgical vessel 100).

[0208] Pyrometer 140-P is adapted to monitor aperture 110 and provide emission data 150-P. As already explained above, the emission data represents thermal emission from the molten material at aperture 110 for at least two wavelengths λ1, λ2.

[0209] Camera 140-C is adapted to monitor opening 110 (or to monitor path 110 / 140 between opening 110 and pyrometer 140-P, or to monitor both the opening and the path). Camera 140-C provides one or more images 150-C showing the opening (or the path, or both). With reference to FIG. 5B, computer 200 receives images 150-C as sensor data in step 423.

[0210] The computer 200 comprises an estimation module 250 (i.e., the processing module in the above description) and an image classification module 260 (implemented as an auxiliary ML tool). The estimation module 250 is adapted to process the radiation data 150-P using a regression model and to provide an estimation using preliminary content data 270' for specific chemical elements in the molten material. The estimation module 250 is implemented during the data collection phase, as described in FIG. 3A. ** The system is trained using training data collected during step 1, which is a combination of past radiation data and past content data obtained by measurements on samples taken from the molten material at the opening.

[0211] The image classification module 260 is adapted to classify one or more images 150-C to identify phenomena 160 at the aperture 110 (and / or along the path), and then provide a confidence classification RC of the emission data according to the identified phenomena 160. The computer 200 outputs the preliminary content data 270' as content data 270 only if the confidence classification RC complies with a predetermined rule. FIG. 6 illustrates this conditional data transfer by analogy to a switch controlled by the RC. A rule-compliant RC turns the switch on; otherwise, the switch is opened (to block the preliminary content data 270').

[0212] The switch metaphor can also be seen as dividing time into periods with reliable emission data (or trustworthy data) and periods without (i.e., unreliable periods). Switching can also be applied to the input of module 250 so that unreliable emission data is not processed.

[0213] With respect to Figure 5B, the computer matches the image (step 433, classification as an example of matching) and applies the matching results to evaluate the content data (step 463). Note that in the embodiment of Figure 6, step 433 (generally: Matching sensor data) has a specific and non-optional implementation for classifying image data.

[0214] It should be noted that although phenomenon 160 has been described, vessel 100 does not belong to system 1000 .

[0215] Feature enhancement can optionally be used, and Figure 6 shows, by a dashed box, a feature enhancer module 280 (which enhances emission data 150-P into feature-enhanced emission data 150-F as a combination of 150-P and feature 150-Q, see Figure 7B). Radiation and other data

[0216] FIG. 7A shows a time series with emission data 150-P (also denoted as {λ1}, {λ2}, etc.) as well as sensor data 150-X.

[0217] As mentioned above, a pyrometer is not necessarily the only sensor, and by way of example only, the sensors may be one or more cameras (see FIG. 6 for a rule-based application of deriving RC from images), one or more microphones, sensors for measuring the viscosity of the material, the speed at which the material moves as in a runner, etc., and sensors for measuring certain gaseous substances in the vicinity of the opening.

[0218] Temperature data may be among the sensor data, noting that the primary purpose of a pyrometer is to measure temperature. Thus, a pyrometer may also provide temperature data.

[0219] For example, {X1}, {X2}, and {X3} are univariate time series representing measurement data available through scalars. Two or more variables can be combined into vectors. For example, {{X1},{X2}} is a two-scalar vector.

[0220] As an example, {X4} represents data obtained from a microphone, which may represent noise intensity (e.g., as sound pressure level), sound frequency, or other sound-related observations (e.g., characteristic sounds occurring during casting).

[0221] As an example, {X5} represents data acquired from a camera (see camera 140-C in FIG. 6), such as an image. As explained, an image is a prominent example that is matched to intermediate data, for example, by classification (see step 433 in FIG. 5B).

[0222] Those skilled in the art are familiar with the installation of sensors on industrial equipment, so the description and drawings will focus on adapting data from the sensors, but need not show the sensors in great detail.

[0223] Training Phase** 2. Intermediate and estimated stages ** By using sensor data in both methods, the accuracy of the estimation (of the content) can be improved.

[0224] The computer 200 (having processing modules 252 / 253, 250 in FIG. 6) receives the radiation data 150-P (i.e., {{λ}}) having variations corresponding to different wavelengths, as previously described for method 403 (see FIG. 5A).

[0225] By receiving (and processing) sensor data 150-X (or {{X}} here denoted by {X1} through {X5}) in step 423, the accuracy of content data 273 can be increased (e.g., by narrowing the tolerance band).

[0226] From a different perspective, the computer processes the radiation data 150-P (from the pyrometer in step 453), but in one context, e.g., (Non-)availability of data that may indicate content, see evaluation of radiation or sensor data (step 443) that may block data from processing, the presence of phenomena that may interfere with the operation of the pyrometer (see FIGS. 6 and 9, phenomena 160 such as dust or misalignment), the phenomenon data of which can be input to the processing module 253 or used for evaluation as illustrated by the image example of FIG. 6; and Possibility to detect inconsistencies (openings are closed while the computer receives data from the pyrometer) Process.

[0227] The sensor data 150-X does not have to be obtained directly from the sensor, but can be obtained by adaptation.

[0228] It is also possible for the RC to be a further input data, in which case the RC is not only applied for evaluation (see steps 443, 463) but is also processed in step 453. Feature-enhanced radiation data

[0229] In FIG. 7A, the dashed box indicates feature-enhanced emission data 150-F that is used in optional implementations, but for ease of explanation will be primarily discussed below with reference to FIG. 7B. image

[0230] 8-9 show image 150-C from camera 140-C. Image processing can be applied (by processing modules 252 / 253) or by auxiliary ML tools (such as adapter 260 in FIG. 6). The results of image processing can be used as sensor data (see FIG. 7A), as confidence discriminators (see FIGS. 6 and 9, evaluation), as control signals to pyrometers (see FIG. 8 for metal / slag ratio, see FIG. 9 for controlling alignment), or in other manners.

[0231] When camera 140-C is pointed at opening 110 (i.e., a taphole, runner, etc. for the example of a blast furnace, see also FIG. 1), image 150-C shows opening 110 and the path between the opening and the camera, at least in many circumstances.

[0232] Image processing can be advantageous for additional reasons. While images (received for one ΔT) typically contain pixels counted in the millions, pyrometer radiation data contains several scalar values ​​for each wavelength (e.g., two values ​​for λ1 and λ2, optionally longer values ​​such as λ3, λ4, λ5, etc., but not millions). Sensor data (see FIG. 7A) is also relatively sparse. This results in an imbalance in the amount of data available. Image processing can reduce the dimensionality (i.e., by fitting, step 433), so that a million-pixel image becomes a "short" data vector, but still informative enough to contribute to the estimation of the original content. (Metaphorically speaking, the image data is fitted to operate on the same granularity as the pyrometer data.) Metal and slag ratio

[0233] 8 shows images from a camera in which metal and slag are differentiated by classification using an auxiliary ML tool that has also been appropriately trained. Training does not necessarily have to be done on a particular vessel 100, but can potentially be done on any image showing metal and slag, typically supervised training.

[0234] Figure 8 shows a series of images 150-C taken at successive time points, similar to {X5} in Figure 7A. This figure shows that the auxiliary ML tool distinguishes material 300 into metal 350 (black diamond symbols) and slag 360 (black circles). Image processing allows for the differentiation of image regions (or regions of interest, see dashed lines). Image processing converts image data into intermediate data.

[0235] Such distinctions for single images can separately be the basis for classification (an example of matching in step 433), as explained in the following example. The images can be classified according to the percentage or ratio by which slag and metal are represented. The figures show slag / metal ratios of 50 / 50, 20 / 80, 60 / 40, and 0 / 100 (percent). The images can be classified according to the presence or absence of a "boundary" between the slag and the metal. From left to right, there is a single boundary between the slag and the metal (e.g., "one" and "one" for the data), multiple boundaries, or no boundaries at all (in the metal-only example). The information about the number of boundaries is intermediate data {X7}, which can be processed by the processing module 250 (in step 453). Image sequences can be classified by changes, for example by the change in the slag / metal ratio. The metal ratio increases from the first image to the second (from 50 to 80), the metal ratio decreases from the second image to the third image (from 80 to 40), and increases again in the fourth image. In other words, intermediate data can be obtained by operations similar to mathematical deviations (which will be explained here with the example of ratios). Image sequences can be classified in other ways. For example, a computer can detect that the number of regions has changed. The first and second images have two regions (one for slag and one for metal), while the third image has multiple slag regions, and in the fourth image the slag has disappeared. The change in regions can also be coded in the intermediate data.

[0236] Classifying the images is one example of matching data in optional step 433 (of method 403, FIG. 5B). Depending on the application, (Training-based application) The intermediate data {X6}, {X7}, etc. can then be processed by the processing module 250 (step 453, in parallel with and in combination with the emission data), (Rule-based application) The intermediate data can then be used for evaluation (e.g. rules give threshold conditions for slag / metal ratios).

[0237] Those skilled in the art can use further image processing techniques, such as to obtain histograms (color distributions), etc. The color distribution can even be modified by filtering out certain colors. Again, image classification is just one example of matching in step 433.

[0238] Image classification can also be part of determining a confidence classification RC. For RC, the computer applies rules. For example, the rules can change RC to "negative" meaning that the preliminary content data does not become content data (see FIG. 6 with reference numbers 270' and 270). As an example, for a slag / metal ratio classified as 90 percent slag (or more), an estimation of the element content in the metal will not be possible. Further phenomena

[0239] FIG. 9 shows image 150-C in a different deformation with additional phenomena.

[0240] It can be assumed that pyrometer 140-P and camera 140-C can be co-aligned so that any repositioning of the pyrometer relative to the aperture also repositions the camera. In that case, the image also shows the pyrometer's field of view (one skilled in the art can take into account parallax, if present). To symbolize the alignment, the figure shows image 150-C with crosshairs (dashed lines).

[0241] In this example, pyrometer 140-P should be aimed at taphole 120 (but could also be aimed at a runner or other part, for example).

[0242] In case (A), image processing (by the auxiliary ML tool) results in the category "OPTIMAL FIT", in case (B) the category "OFF", and in case (C) the result is "DUST" (symbolized by an S-shaped line, obscuring the taphole). In cases (B) and (C), the taphole is not visible (to the pyrometer) and therefore the measurement result is unreliable (RC "negative").

[0243] Again, auxiliary ML tools can be used here, and it doesn't matter where they were trained. In theory, detection of misalignment of objects in images (i.e., specific classification) could be trained by using "synthetic" images or by images taken from a mock-up installation.

[0244] An appropriate warning can be given to the operator, such as to readjust the pyrometer (the circle moves to a crosshair) in (B) or to wait until the dust has disappeared in (C). Such a result can be used as a control signal to the processing module 252 / 253, in which case training with the data for processing content values ​​etc. is not allowed.

[0245] 7-8, misalignment detection may result in further data to be processed by the computer (e.g., as binary data OPTIMAL FIT / OFF). Module 252 / 253 will eventually learn that it may have to ignore the data from the misaligned pyrometer, but it may be appropriate to issue a warning (leading to a readjustment).

[0246] Note that the pyrometer is automatically aligned based on image detection.

[0247] If the pyrometer and camera are not co-aligned (as they may be independently supported), OPTIMAL FIT may be a "false positive" and OFF may be a false negative, but one skilled in the art can take such a situation into account. radiation

[0248] As mentioned above, pyrometers measure infrared (IR) radiation at one or more wavelengths that pass through their optics. Typically, IR has wavelengths between 700 nanometers (nm) and 1 millimeter (mm), although radiation outside the IR range is contemplated.

[0249] One criterion for selecting an appropriate wavelength is the effectiveness of the emissivity of the surface of the material (ie, material 300) in emitting energy as thermal radiation.

[0250] Therefore, it seems reasonable to use longer wavelengths (greater than 1 millimeter) or shorter wavelengths (which would make it visible).

[0251] Note that using an image (such as image 150-C from camera 140-C in FIG. 6) means that a computer (202, 203 in FIGS. 3-4) processes data related to radiation by visible light. Autoencoder

[0252] As explained, the step of fitting the sensor data (433 in FIG. 5B) can be performed by an auxiliary ML tool.

[0253] A further example for the implementation of such tools is an autoencoder. The use of an autoencoder can be advantageous as it can reduce the effort required for a human expert to annotate relevant features. Autoencoders have been discussed in papers such as the above-mentioned paper by Kortmann et al. Example

[0254] 10 shows a schematic of a blast furnace 600 for a computer 690 to estimate the content of specific chemical elements, which allows the computer to use image data (alternatively or in combination) for the following functions: As described above for system 1000 in conjunction with FIG. 7A, computer 690 may use image data to obtain and apply RC. 5B, optional steps 423 and 433, the computer 690 can use the image data to provide intermediate data for processing (step 453). For example, the intermediate data can indicate the slag / metal ratio.

[0255] On the left, the diagram shows a blast furnace 600 with a taphole 610 (an example of a metallurgical vessel with an opening) during casting. Material flow, from left to right, from the furnace to several runners. Molten material (slag 660 and iron 650) moves from the furnace through the taphole and into the runners. As shown in the center of the diagram, the slag and cast iron are mixed in the initial runner but are quickly separated (by a skimmer) into slag and cast iron (which flow into separate runners). The diagram also shows (by vertical arrow 670) that manual sampling (see operator 371 in Figure 2) is performed on the cast iron runner.

[0256] Similar diagrams with furnace, taphole, runner, and skimmer can be found in textbooks, but FIG. 10 adds a pyrometer 640 (see pyrometers 151 / 153 and processing unit 690; see computer 200 executing method 403). A camera can be mounted on the same support as the pyrometer. Image 650 is symbolized above an arrow pointing from the pyrometer to the taphole. In other words, if the pyrometer is properly aligned with the taphole, the image will show slag and iron in various proportions. The image may also show phenomena such as dust. Fitting images to intermediate data and applying the intermediate data for evaluation or other purposes has been described in detail above.

[0257] If radiation transfers relatively high temperatures from the taphole to its vicinity, the pyrometer 640, and optionally the camera and other sensors, are adapted to operate under such conditions. One skilled in the art can provide cooling for the electronics, for example, by water cooling, or can provide appropriate shielding. Figure 10 symbolizes such a protective measure by a protective shield attached to the left of the pyrometer / camera. Time series

[0258] Herein, we sometimes refer to multivariate time series that represent multiple measurement data. An example is shown in Figure 7A. In general, {{X}} represents a multivariate time series, with i = 1 to N variables {Xi}. The index i is the variable index, and N is the number of variables. {{X}} contains multiple univariate time series. The variables {Xi} are given by the univariate time series. In an alternative notation, a univariate time series can be given as a sequence of data samples.

[0259] M is the number of samples in the observation interval WINDOW (i.e., the time length of the time series), and individual samples are identified by index m. The sampling interval has a duration of WINDOW = ΔT * M. ΔT represents the sampling interval. ΔT may be the same for all i. This is convenient for illustration purposes, but is not required in practice. Different univariate time series can use different sampling intervals. For example, temperature may be measured every minute, Δt = 60 seconds, while radiation (as explained) may be measured every second, ΔT = 1 second.

[0260] At each time point tm, data representing the value of each variable Xi may be available. Since semantics do not matter, the values ​​can be normalized. For example, in the temperature range of 100°C to 1,100°C, the normalized values ​​can be treated as "0" and "1". One skilled in the art can apply preprocessing to filter out infeasible variable values. For example, a faulty sensor may output negligible excess values.

[0261] A time point tm is given for the end of each sampling interval ΔT. This is simply a convenient convention, in other words tm identifies the sampling interval ΔT that ends at tm.

[0262] In this specification, we distinguish between multivariate and univariate time series by notating them as {{}} or {}, and therefore "multivariate / univariate" may be omitted herein, or they may be written in parentheses, respectively. Radiation data provides further information

[0263] As mentioned above, emissivity is a function of the temperature of an object and of its elemental composition. Although radiation data from pyrometers (see 151-P in Figure 2, 153-P in Figure 4, and 150-P in Figure 6) is related to emissivity, it is not necessary to calculate emissivity.

[0264] Because the radiation data is wavelength specific, the data can be thought of as a spectrum with N spectral lines (see N wavelengths, N frequencies). Because the radiation intensity varies over time, a computer can analyze the radiation data as a multivariate time series λ (step ** 1 and stage ** 3) Process.

[0265] When radiation carries energy, the energy has two well-known aspects that are simplified here: First, radiation at shorter λ will, in principle, carry more energy than radiation at longer λ; Second, each λ value also corresponds to energy, and this energy can change over time, with lower values ​​corresponding to lower energy and higher values ​​corresponding to higher energy.

[0266] The second aspect is shown in Figure 2 for {λ1}, {λ2}, etc., with values ​​(vertical axis) changing over time. Note that while the change over time is conveniently examined with a curve or graph, the computer need not present such a visualization to the user.

[0267] Figure 7B shows the radiation date 153-P in two univariate time series, (first and second) time series {λ} and {λ}, as well as the feature data 153-Q, also as a time series. For convenience, a ΔT interval is applied to all data 153-P and 153-Q. The interval is given as tm, from tl to tl7.

[0268] This figure provides several examples of feature enhancement according to multiple feature enhancement rules. A computer (see enhancer module 282 / 283 in FIG. 3B, module 280 in FIG. 6) applies the rules by processing the emission data 153-P. The rules (and their corresponding features) can be differentiated according to whether the computer processes (i) one univariate time series or (ii) two or more univariate time series.

[0269] In this example, {λ1} has rising and falling values ​​(thick line), which are sometimes higher and sometimes lower than a reference value. The reference value can be a threshold (here given by a horizontal line). It is also possible to normalize the data to such or other thresholds. For example, the reference can be an average, so that values ​​are sometimes above ("positive") or below ("negative") the average. {λ2} has similar rising and falling values, but with different timing (the thick line is smoother in this example).

[0270] For example, the computer applies the first rule, where the first feature must be the sign of the derivative. The rule can be implemented by comparing successive values. The derivative is positive if value(tm+1) > value(tm), or negative if value(tm+1) < value(tm).

[0271] The computer can apply a second rule, where the second feature should be the identification of threshold crossings, shown here for time intervals t, t, t, and t during which the value of {λ} crosses the threshold (in either direction).

[0272] The first and second features can be derived from {λ1} alone (case (i)), while the third feature should be determined by processing both {λ1} and {λ2} (case (ii)). By applying the third rule, the feature enhancer will identify the third feature. In this example, the computer identifies time intervals, here t1, t2, t5, t6, and t14 onward, where the values ​​of both {λ1} and {λ2} are elevated.

[0273] The first, second, and third rules above are merely examples. Identifying features for time series data does not need to be illustrated, and this specification continues by discussing features without necessarily referring to figures.

[0274] The emission data can be preprocessed to identify features that are local characteristics and may be specific to individual wavelengths. In other words, the changes in the λ values ​​may follow detectable patterns of curvature. By way of example, such a pattern may be as follows for each univariate time series {λi}: The values ​​of a time series may increase or decrease (i.e., first derivative) over the interval of time points, or may remain constant (first derivative at 0). · The time series can show changes in such increases or decreases (i.e., second derivatives). · Time series may have inflection points (which is just another way of looking at the second derivative). The value can reach 0 (and remain 0 for a detectable number of time points), and the same is true for reaching a threshold. The value can reach a maximum (or minimum) value with or without consideration of absolute value etc.

[0275] Furthermore, the emission data can be pre-processed to identify features that are still local characteristics but may be specific to two or more individual wavelengths. By way of example only, features can be detected as follows:

[0276] Over the interval of time tm, the value of the first wavelength λ1 increases while the value of the second wavelength λ2 decreases (and vice versa), in the sense that the first derivatives of both values ​​will have different signs (positive and negative, or vice versa).

[0277] For a pair of wavelengths, there may be a second derivative with signs such as "++", "+-", "-+", "--", etc.

[0278] The values ​​may reach 0 in a particular order (e.g., first λ1, second λ2).

[0279] The values ​​of λ1 and λ2 may reach their maximum values ​​at substantially the same time, at different times, etc.

[0280] Such features can be detected by a computer (e.g., by a feature detector or feature enhancer), and the detected features can be considered enhancements to the emission data. The feature-enhanced emission data 150-F (see FIGS. 7A and 7B) can be processed as if it were emission data (at all stages ** 1. ** 2, and ** In other words, the feature-enhanced emission data 150-F represents the thermal emission from the molten material (similar to emission data 150-P), but represents the emission with greater precision, allowing estimates (i.e., elemental content) to be derived with greater precision.

[0281] Thus, the computer performs the method steps accordingly (see Figures 5A and 5B), where in receiving step 413 the computer receives the feature-enhanced emission data. In acting step 453 the computer processes the feature-enhanced emission data. The same principles apply to training (see discussion of Figure 3B).

[0282] In general, the feature-enhanced emission data (150-F in FIG. 7A and the example in FIG. 7B serve as examples only) are data that represent local characteristics of the emission data. The feature-enhanced emission data are wavelength-specific data (150-F is shown only for λ2). The term "local" refers to a time interval within a time series.

[0283] Further examples include: The value of a particular wavelength λi may have a local minimum or a local maximum (with respect to a "local" time interval). The feature-enhanced emission data indicates the time of that minimum or maximum. The value of a particular wavelength λi may oscillate, and the oscillations can be detected over a particular time interval. Figure 7A shows 150-F for such a temporal oscillation of λ2. The feature-enhanced emission data indicates the start and end times of the oscillation, and optionally indicates a class (such as a relatively low or high oscillation). Since the value of a particular wavelength λi usually changes over time, it is possible to calculate the first (or even second) derivative. For example, an increase in value would be a feature and would correspond to a positive but more or less stable (first) derivative. The computer can use that feature (i.e., the indication of the derivative) as processing input. The change over time can also be seen as a slope (of a curve or graph), and those skilled in the art are familiar with classifying them (e.g., steepness). Changes over time do not necessarily need to be processed by specific wavelength λi values, but it may be advantageous to apply local curve fitting techniques. Such curve fitting may ignore outliers, etc. This is just another example in that feature enhancement can improve accuracy; outliers are simply ignored. While a first univariate time series varies over time in a particular pattern (e.g., with a steep gradient), a second univariate time series may vary in a similar pattern (e.g., with a similarly steep gradient) or in the opposite pattern (in the other gradient direction). Such pattern similarities can also be detected as features. Pattern similarity may include time differences (similar to the phase shift between voltage and current in electrical engineering), which may also serve as features. Since the univariate time series (here, lambda) represents radiant energy, the ratio of values ​​between two time series can also be characterized (e.g., value at λ1(tm) / value at λ1(tm) for equal tm). Time windows and further statistics

[0284] In this document, features are described as examples using a temporal aspect, considering one or two (or even more) time series, considering value comparisons, etc. In general, feature identification can be applied to statistics. Therefore, the following is only an overview of further options. Features can develop over time. An example of such a temporal evolution of a feature is a change over a period longer than ΔT. It is possible to apply moving time windows. Statistics can be calculated from the distribution of values ​​within such windows (e.g., mean or average, median, standard deviation, quantiles, skewness, polynomial fit parameters, or any curve fit). It is also possible to apply a moving window to obtain the signal spectrum. For example, the following transformations can be applied: Fast Fourier Transform (FFT) or Wavelet Transform. Applying a moving window has additional aspects in computing the signal autocorrelation and related properties (e.g., the time shift at which the autocorrelation falls below a predetermined threshold). Because the autocorrelation is applied to a single time series, the correlation can also be computed for multiple time series (see example in Figure 7B for {λ1} and {λ2}). Furthermore, it is possible to use pre-trained neural networks or rules (i.e., further auxiliary networks) to detect time series patterns using similarity metrics such as dynamic time wrapping. Furthermore, the correlation between one emitted signal and another signal (for a moving window) can be recorded. Comparisons with specific historical data can be classified according to specific historical castings (e.g., "high quality castings" with the best performance, or even castings with worse performance). To limit the computational effort (in terms of memory, CPU, bandwidth, and other resources), it is possible to apply feature identification through dimensionality reduction, such as principal component analysis or autoencoder machine learning. similarity

[0285] Those skilled in the art can see some similarities with enhanced emission data by talking to astronomers. Stars can periodically change color, brightness, etc. Such and other variations can indicate stellar properties. However, since astronomers do not take physical samples (the moon and comet examples do not apply), the similarities end there. Data Time Interval

[0286] 4 shows the time interval T_cast, which, as already explained, can start when the drill starts to open the taphole. In other words, T_cast can mark the start of each individual casting process. It is advantageous for the computer 203 to receive the emission data for the data from the start of each casting (and process them in step 413, FIGS. 5A, 5B, and step 453). The same principle applies to training. The (optional) separation of historical data by T_cast has already been explained as being advantageous.

[0287] However, it is not required to follow that approach. Figure 4 shows the time interval T_back from the point when casting is already underway to t_current. T_back can therefore mark the time length of the time series to be processed. In many situations, T_back will correspond to a WINDOW (see Figure 7A).

[0288] T_back (i.e., a value representing its duration) may also be input to the processing module 253 of the computer 203. (The same principle applies to training.) Although T_back does not come from a sensor, it can be treated like sensor data. general-purpose computer

[0289] FIG. 11 illustrates an example of a general-purpose computing device that can be used with the techniques described herein. FIG. 11 illustrates an example of a general-purpose computing device 900 and a general-purpose mobile computing device 950 that can be used with the techniques described herein. Computing device 900 is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers. General-purpose computing device 900 may correspond to computers illustrated by other figures. Computing device 950 is intended to represent various forms of mobile devices, such as personal digital assistants, mobile phones, smartphones, driver assistance systems or vehicle on-board computers, and other similar computing devices. For example, computing device 950 can be used as a front end by a user (e.g., a blast furnace operator) to interact with computing device 900. The components, their connections and relationships, and their functions illustrated herein are merely exemplary and do not limit the implementation of the invention(s) described and / or claimed herein.

[0290] Computing device 900 includes a processor 902, memory 904, a storage device 906, a high-speed interface 908 connecting to memory 904 and a high-speed expansion port 910, and a low-speed interface 912 connecting to a low-speed bus 914 and storage device 906. Each of the components 902, 904, 906, 908, 910, and 912 are interconnected using various buses and may be mounted on a common motherboard or otherwise as needed. Processor 902 can process instructions for execution within computing device 900, including instructions stored in memory 904 or on storage device 906 to display graphical information for a GUI on an external input / output device, such as a display 916 coupled to the high-speed interface 908. In other implementations, multiple processors and / or multiple buses may be used, along with multiple memories and multiple types of memory, as needed. Multiple computing devices 900 may also be connected, each providing a portion of the required operations (e.g., as a server bank, a group of blade servers, or a multiprocessor system).

[0291] The memory 904 stores information within the computing device 900. In one implementation, the memory 904 is one or more volatile memory units. In another implementation, the memory 904 is one or more non-volatile memory units. The memory 904 may also be another form of computer-readable medium, such as a magnetic or optical disk.

[0292] The storage device 906 is capable of providing mass storage for the computing device 900. In one implementation, the storage device 906 can be or include a computer-readable medium such as a floppy disk device, a hard disk device, an optical disk device, or a tape device, a flash memory or other similar solid-state memory device, or an array of devices including devices in a storage area network or other configuration. A computer program product can be tangibly embodied on an information carrier. The computer program product can also include instructions that, when executed, perform one or more methods, such as those described above. The information carrier is a computer- or machine-readable medium, such as memory 904, the storage device 906, or memory on the processor 902.

[0293] The high-speed controller 908 manages bandwidth-intensive operations of the computing device 900, while the low-speed controller 912 manages less bandwidth-intensive operations. This allocation of functionality is merely exemplary. In one implementation, the high-speed controller 908 is coupled to the memory 904, the display 916 (e.g., via a graphics processor or accelerator), and a high-speed expansion port 910 that can accept various expansion cards (not shown). In this implementation, the low-speed controller 912 is coupled to the storage device 906 and the low-speed expansion port 914. The low-speed expansion port, which may include various communication ports (e.g., USB, Bluetooth, Ethernet, wireless Ethernet), can be coupled, for example, via a network adapter, to one or more input / output devices, such as a keyboard, a pointing device, a scanner, or a networking device such as a switch or router.

[0294] Computing device 900, as shown, can be implemented in several different forms. For example, it may be implemented as a standard server 920, or multiple times in a group of such servers. It may also be implemented as part of a rack server system 924. Additionally, it may be implemented in a personal computer, such as a laptop computer 922. Alternatively, components from computing device 900 can be combined with other components in a mobile device (not shown), such as device 950. Each such device can include one or more of computing devices 900, 950, and the entire system can be composed of multiple computing devices 900, 950 communicating with each other.

[0295] Computing device 950 includes, among other components, a processor 952, memory 964, input / output devices such as a display 954, a communications interface 966, and a transceiver 968. Device 950 may also include a storage device, such as a microdrive or other device, to provide additional storage. Each of components 950, 952, 964, 954, 966, and 968 are interconnected using various buses, and some of the components may be mounted on a common motherboard or otherwise as desired.

[0296] The processor 952 can execute instructions within the computing device 950, including instructions stored in the memory 964. The processor may be implemented as a chipset of chips including separate analog and digital processors. The processor can provide coordination of other components of the device 950, such as control of a user interface, applications run by the device 950, and wireless communications by the device 950.

[0297] The processor 952 can communicate with a user via a control interface 958 and a display interface 956 coupled to a display 954. The display 954 can be, for example, a TFT LCD (thin film transistor liquid crystal display) or an OLED (organic light emitting diode) display, or other suitable display technology. The display interface 956 can comprise appropriate circuitry for driving the display 954 to present graphics and other information to the user. The control interface 958 can receive commands from the user and convert them for submission to the processor 952. Additionally, an external interface 962 can be provided in communication with the processor 952 to enable near-field communication of the device 950 with other devices. The external interface 962 can provide, for example, wired communication in some implementations or wireless communication in other implementations, and multiple interfaces may also be used.

[0298] Memory 964 stores information within computing device 950. Memory 964 may be implemented as one or more computer-readable media, one or more volatile memory units, or one or more non-volatile memory units. Expansion memory 984 may also be provided and connected to device 950 via expansion interface 982, which may include, for example, a SIMM (single in-line memory module) card interface. Such expansion memory 984 may provide additional storage space for device 950 or may store applications or other information for device 950. Specifically, expansion memory 984 may include instructions for performing or supplementing the processes described above and may also include secure information. Thus, for example, expansion memory 984 may function as a security module for device 950 and be programmed with instructions that enable secure use of device 950. Additionally, secure applications may be provided via a SIMM card along with additional information, such as placing identifying information on the SIMM card in an unhackable manner.

[0299] The memory may include, for example, flash memory and / or NVRAM memory, as discussed below. In one implementation, a computer program product is tangibly embodied on an information carrier. The computer program product includes instructions that, when executed, perform one or more methods, such as those described above. The information carrier is a computer- or machine-readable medium, such as memory 964, expansion memory 984, or memory on processor 952, which may be received, for example, via transceiver 968 or external interface 962.

[0300] Device 950 can communicate wirelessly via communication interface 966, which may include digital signal processing circuitry as needed. Communication interface 966 can provide for communication under various modes or protocols, such as GSM voice calls, SMS, EMS, or MMS messaging, CDMA, TDMA, PDC, WCDMA, CDMA2000, or GPRS, among others. Such communication may occur, for example, via radio frequency transceiver 968. Additionally, short-range communication may occur, such as using Bluetooth, WiFi, or other such transceivers (not shown). Additionally, GPS (Global Positioning System) receiver module 980 may provide device 950 with additional navigation- and location-related wireless data that may be used as needed by applications executing on device 950.

[0301] Device 950 can also communicate audibly using audio codec 960, which can receive verbal information from a user and convert it into usable digital information. Audio codec 960 can likewise generate audible sounds for the user, such as through a speaker in the handset of device 950. Such sounds can include sounds from voice calls, recorded sounds (e.g., voice messages, music files, etc.), and sounds generated by applications running on device 950.

[0302] The computing device 950, as shown, can be implemented in several different forms. For example, it can be implemented as a mobile phone 980. It can also be implemented as part of a smartphone 982, personal digital assistant, or other similar mobile device.

[0303] Various implementations of the systems and techniques described herein can be realized in digital electronic circuitry, integrated circuits, specially designed ASICs (application-specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs executable and / or interpretable on a programmable system including at least one programmable processor, which can be special purpose or general purpose, coupled to receive data and instructions from, and send data and instructions to, a storage system, at least one input device, and at least one output device.

[0304] These computer programs (also known as programs, software, software applications, or code) include machine instructions for a programmable processor and may be implemented in a high-level procedural and / or object-oriented programming language and / or in an assembly / machine language. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and / or device (e.g., magnetic disk, optical disk, memory, programmable logic device (PLD)) used to provide machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0305] To provide for user interaction, the systems and techniques described herein can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user, and a keyboard and pointing device (e.g., a mouse or trackball) by which the user can provide input to the computer. Other types of devices can also be used to provide for user interaction; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback), and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0306] The systems and techniques described herein can be implemented on computing devices that include back-end components (e.g., as data servers), or that include middleware components (e.g., application servers), or that include front-end components (e.g., client computers having a graphical user interface or web browser through which a user can interact with an implementation of the systems and techniques described herein), or any combination of such back-end, middleware, or front-end components. The components of the systems can be interconnected by any form or medium of digital data communication (e.g., a communications network). Examples of communications networks include a local area network ("LAN"), a wide area network ("WAN"), and the Internet.

[0307] Computing devices may include clients and servers. Clients and servers are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

[0308] Although several embodiments have been described, it will be understood that various modifications can be made without departing from the spirit and scope of the invention.

[0309] Additionally, the logic flows depicted in the figures do not require the particular order shown, or sequential order, to achieve desirable results. Additionally, other steps may be provided or steps may be eliminated from the described flows, and other components may be added to or removed from the described systems. Accordingly, other embodiments are within the scope of the claims. [Explanation of symbols]

[0310] The following summary considers the stages and cites the reference numbers where the items are indicated.

[0311] 100 / 101 / 103 Metallurgical containers 110 / 111 / 113 opening 120 holes or holes on the image 130 Conductor 141-P / 143-P Pyrometer 140-C Camera 150-P / 151-P / 153-P Radiation Data 150-X sensor data 150-C Image (Example of sensor data) 150-F Feature Enhanced Emission Data 150-Q, 153-Q feature data 160 Phenomenon 171 / 173 Sample Worker 181 Laboratory 190 / 193 Container Worker

[0312] 200 / 202 / 203 Computer 241 training data 252 / 253 Processing Module 260 adapter 270,273 Computer-estimated content data 270´ Preliminary Content Data 280 / 282 / 283 Feature Enhancer Module

[0313] 300 / 301 / 303 molten material 311 / 313 Molten material sample 350 metal 360 Slug 370,371 Laboratory-measured content data

[0314] 4xx Method, Method Step

[0315] 501,503 Controller

[0316] 6xx Blast Furnace and Example System Components

[0317] 9xx General-purpose computers and their components

[0318] 1000 systems

[0319] δT_COM The interval at which content data is available at the output of the processing module δT_LAB Interval when content data from the laboratory is available ΔT is the sampling interval at the input of the processing module M Repetition of training RC Reliability Classification

[0320] As mentioned above, the radiation data λ1, λ2 are given without scale. [Prior art documents] [Non-patent literature]

[0321] [Non-Patent Document 1] 「Optical emission spectroscopy as a method to improve the process automation of electric arc furnaces and ladle furnaces」ScienceDirect、IFAC PapersOnLine 55-2(2022)78~83

Claims

1. 1. A system (1000) for estimating the content of a particular chemical element in a molten material available at an opening (110) of a metallurgical vessel (100), comprising: a pyrometer (140-P) adapted to monitor the aperture (110) and provide radiation data (150-P) representative of thermal radiation from the molten material (300) at the aperture (110) for at least two wavelengths (λ1, λ2); a camera (140-C) adapted to monitor the opening (110) or a path (110 / 140) between the opening (110) and the pyrometer (140-P) and provide one or more images (150-C) indicative of the opening or the path; a computer (200) having an estimation module (250) and having an image classification module (260); Equipped with the estimation module (250) is adapted to process the emission data (150-P) using a regression model to provide an estimation using preliminary content data (270') for the specific chemical element in the molten material, the estimation module (250) being trained by training data (241) collected during a data collection phase (**1), the collected data being a combination of past emission data (151-P) and past content data (371) obtained by measurements on samples (311) taken from the molten material (301) at the opening (111); The image classification module (260) is adapted to classify (433) the one or more images (150-C) to identify phenomena (160) in the opening (110) or the path (110 / 140), the phenomena (160) being one of the following: (i) deposits, fumes, or reflections located within said path (110 / 140); and (ii) the degree of alignment of the pyrometer with respect to the opening (110); (iii) the percentage of metal and slag on the image; and and optionally including at least one of: The system (1000) is further adapted such that the image classification module (260) subsequently provides a confidence classification (RC) of the radiation data according to the identified phenomenon (160), such that the computer outputs preliminary content data (270') as content data (270) only if the confidence classification (RC) complies with predetermined rules.

2. the computer (200) further comprises a feature enhancer module (280) adapted to process the emission data (150-P) into feature data (150-Q) according to a plurality of feature enhancement rules; the estimation module (250) is adapted to further process the feature data (150-Q) together with the emission data (150-P), so that the emission data (153-P) together with the feature data (150-Q) become feature-enhanced emission data (150-F); The system of claim 1 , wherein the estimation module is trained with training data further comprising feature data identified according to the plurality of feature enrichment rules.

3. 1. A computer-implemented method (403) for estimating the content (273) of a particular chemical element in a molten material (303) available at an opening (113) of a metallurgical vessel (103), the method (403) comprising the following steps: receiving (413) radiation data (153-P) representing thermal radiation from the molten material (303) at the opening (113) for at least two wavelengths (λ1, λ2) from a pyrometer (143-P) positioned to monitor the opening (113); receiving, by the computer (200) from a camera (140-C), one or more images (150-C) showing the opening (110) or showing a path (110 / 140) between the opening (110) and the pyrometer (140-P); The computer (200) classifies (433) the one or more images (150-C) to identify a phenomenon (160), wherein the phenomenon (160) is one of the following: (i) deposits, fumes, or reflections located within said path (110 / 140); and (ii) the degree of alignment of the pyrometer with respect to the opening (110); (iii) the proportion of metal and slag on said image (150-C); and Step (433), which selectively includes at least one of: classifying the radiation data (153-P) from the pyrometer as reliable radiation data or unreliable radiation data according to the identified phenomenon (160); a step (453) in which the computer (203) operates a processing module (253) for processing the reliable radiation data (153-P), the processing module (253) having a regression model for estimating the content (273) of the specific chemical element, the processing module (253) having been pre-trained; A method (403) comprising:

4. 4. The method (403) of claim 3, wherein the step (453) of operating the processing module (253) comprises operating a module (253 / 252) that has been trained by training data (241) collected during a data collection phase (**1), the collected training data (241) being a combination of past radiation data (151-P) and past content data (371) obtained by measurements on samples (311) taken from the molten material (301) at the opening (111).

5. In a further receiving step (423), the computer (203) receives sensor data (150-X, 153-X) from sensors positioned to monitor the container (103); 5. The method (400) according to claim 3, wherein in the step (453) of operating the processing module (253), the computer (203) processes the radiation data (153-P) in combination with the sensor data (150-X, 153-X).

6. 6. The method (400) of claim 5, wherein after receiving (423) sensor data (150-X, 153-X), the computer performs a step (433) of classifying the one or more images to obtain the confidence classification (RC) as intermediate data, and the computer performs the step (453) of operating the processing module (253) also using the confidence classification (RC) as the intermediate data.

7. 7. The method (400) of claim 6, wherein the computer performs the step of classifying (433) with an auxiliary machine learning tool (260).

8. 8. The method (400) of claim 7, wherein the computer performs the classifying (433) step with an auxiliary machine learning tool (260) that is an autoencoder.

9. 10. The method (400) of any one of claims 1 to 9, wherein in the receiving step (413), the computer receives feature-enhanced radiation data (150-F), and in the operating step (453), the computer processes the feature-enhanced radiation data (150-F).

10. 10. The method (403) of any one of claims 1 to 9, wherein the molten material (123) is a composition ({C, Fe, S, Si}) comprising a molten metal or metal alloy, and the specific chemical elements are selected from carbon (C), silicon (Si), iron (Fe), and sulfur (S).

11. 11. The method (400) of claim 10, wherein the method steps (413, 423, 433, 453) are applied to the metallurgical vessel (101 / 103) being a blast furnace, such that the opening (123) comprises a taphole and runners for transporting the material from the blast furnace during casting.

12. 12. The method (400) of claim 11, comprising the further step of presenting the estimation using the content data (273) to an operator (193) of the blast furnace (100).

13. Repeatedly using the method (403) of any one of claims 1 to 12 to control the movement of the vessel (103).

14. A computer program product which, when loaded into a memory of a computer system and executed by at least one processor of said computer system, causes said computer system to perform the steps of the computer-implemented method of any one of claims 1 to 12.

15. A computer system comprising a plurality of modules for performing the steps of the computer-implemented method of any one of claims 1 to 12.

Images