Characterization of direct reduction

Abstract

The present invention relates to a method (100) and a system (10) for characterizing a direct reduction process, to a direct reduction plant (1) and to methods (200, 300) for machine learning of models (16, 24) usable in such a method (100) and system (10). For a reduction process currently taking place in at least one reactor (3) of a direct reduction plant (1), parameter values (A, B, C, D, E) of a predetermined number of different process parameters are determined (S1). On the basis of the parameter values (A, B, C, D, E) determined, a machine-learned model (16) then classifies a reactor operating state (30) defined by the parameter values (A, B, C, D, E) under an operating category (X, Y, Z) (S2).

Description

[0001] Description

[0002] Direct reduction characterization

[0003] field of technology

[0004] The present invention relates to a method and a system for characterizing a direct reduction process, a direct reduction plant and methods for machine learning models that can be used in such a method and system.

[0005] State of the art

[0006] In direct reduction plants, metal oxides are directly reduced using reducing gas, such as carbon monoxide (CO) or hydrogen (H2). These reduction processes are typically closely monitored. A variety of different sensors can be used for automation and quality assurance. These sensors can be used to record a wide range of different process parameters, which in principle allow statements and / or predictions to be made, for example, regarding reaction processes and rates.

[0007] However, evaluating these parameter values ​​is very complex and time-consuming, as it is generally not sufficient to simply examine individual process parameters to identify sources of error, for example. Rather, a combination of numerous process parameters must usually be considered. Therefore, undesirable process behaviors are often only detected late. Furthermore, it is often not possible, or only with considerable effort and / or experience, to determine the cause of the undesirable process behavior and initiate appropriate countermeasures based solely on the multitude of parameter values.

[0008] Summary of the invention

[0009] Against this background, it is an object of the present invention to further improve the characterization of direct reduction processes, in particular to simplify and accelerate them.

[0010] This object is achieved by a method and a system for characterizing a direct reduction process and a direct reduction plant according to the independent claims.

[0011] A further object is to provide methods by which models that can be used for the improved characterization of a direct reduction process can be machine-learned. This object is achieved by methods for machine-learning models that can be used in a method and system described here, according to the independent claims.

[0012] Preferred embodiments are subject of the dependent claims and the following description.

[0013] According to a first aspect of the invention, in the method for characterizing a direct reduction process, parameter values ​​of a predetermined number of different process parameters are determined for a reduction process currently taking place in at least one reactor of a direct reduction plant. A machine-learned model then assigns a reactor operating state defined by the parameter values ​​to an operating category based on the determined parameter values.

[0014] A machine-learned model within the meaning of the invention is preferably a statistical model generated by one or more algorithms based on training data. The machine-learned model can therefore also be referred to as a trained model. The machine-learned model is expediently the result of machine learning. Such a machine-learned model can also be commonly referred to as artificial intelligence, which has recognized patterns and regularities in the training data and can thus also assess or evaluate unknown data according to these patterns and regularities. The machine-learned model can be based, for example, on a neural network or a random forest.

[0015] One aspect of the invention is based on the approach of assessing an operating state of a reactor of a direct reduction plant or a state of the direct reduction process taking place therein, i.e. a reduction process based on direct reduction, using artificial intelligence. This assessment is expediently based on a limited, for example single-digit, number of process parameters. Based on current parameter values ​​of these process parameters, also referred to as process variables, the artificial intelligence can assign the current operating state of the reactor or the state of the direct reduction process taking place therein to one of several operating categories. For example, the operating or process state can be assigned to the categories "good", "neutral" or "poor". In principle, however, other categories are also conceivable.This allows for timely detection when the reduction process deteriorates or the operating condition deteriorates. Consequently, this deterioration can be responded to promptly, saving resources, especially energy. For example, improved quality feedstocks can be used or the feedstock mix can be changed. Furthermore, many process parameters no longer need to be manually monitored and evaluated by experienced operators, which can further reduce effort and costs.

[0016] The artificial intelligence is expediently represented by an appropriately trained model, also referred to as a machine-learned model. The machine-learned model is preferably trained to categorize the operating or process state based on current parameter values ​​of a predetermined number of (a few) process parameters. The operating or process state can thus be assessed effectively, if necessary even without in-depth prior knowledge on the part of the operating personnel, in particular automatically and / or continuously.

[0017] The machine-learned model outputs a measure of the quality of the reduction process to assign the operating or process state to the operating category. The parameter values ​​characterizing the operating or process state are thus effectively and automatically summarized by the machine-learned model into an easy-to-interpret key figure that is meaningful even for inexperienced operators.

[0018] Preferred embodiments of the invention and their further developments are described below. These embodiments can be combined with each other and with the aspects of the invention described below, unless expressly excluded.

[0019] Using the machine-learned model, it is not only possible to predict the current operating state of the reactor. Such a model can also be advantageously used to predict future operating states. In order to be able to conduct a particularly comprehensive assessment of plant operation, at least one future operating and / or product parameter, for example a process temperature, a process pressure, a product property such as a content of a specific substance, a production volume, and / or the like, is preferably predicted using the machine-learned model based on the determined parameter values.

[0020] It has been shown that a reliable assessment of the reactor operating state or the reduction process does not necessarily require the consideration of a large number of process parameters. Rather, a reliable and rapid categorization of the state can be achieved even if only parameter values ​​of fifteen or fewer, preferably ten or fewer, or even just eight or fewer, process parameters are determined and used as the basis for assigning the reactor operating state to an operating category and / or predicting at least one future operating and / or product parameter. This can have advantages not only with regard to the required infrastructure or hardware (e.g., a reduced number of sensors for determining the parameter values ​​on and / or in the reactor or other components of the direct reduction plant, such as a reformer), but also with regard to the machine learning of the model.In particular, this allows the size of a training dataset to be significantly reduced and model training to be accelerated.

[0021] Examples of process parameters that have proven particularly relevant for assignment to an operating category by the machine-learned model are process parameters that relate to the properties and / or handling of at least one reducing gas, also referred to as a "bustle gas," passed through the reactor. Properties and / or handling of feedstocks fed to the reactor have also shown equal relevance. In addition, pressure or temperature can also play a significant role in assessing the operating state of the reactor or the direct reduction-based process. Accordingly, it is preferred that the predetermined number of different process parameters comprise at least one reducing gas-related process parameter, at least one pressure-related process parameter, at least one temperature-related process parameter, and / or at least one feedstock-related process parameter.

[0022] Properties of the reducing gas can be taken into account, for example, by determining a parameter value that characterizes a reducing gas composition. For example, it is conceivable that a stoichiometry of the molecules of gases, for example the ratio of carbon monoxide to hydrogen in the gas mixture introduced into the reactor as reducing gas, is determined. Certain reducing gas compositions, for example a significant imbalance of carbon monoxide and hydrogen in the reducing gas, can have a detrimental effect on the direct reduction process and thus also on the final product. Alternatively or additionally, the handling of the reducing gas can be taken into account, for example by determining a parameter value that characterizes an absolute and / or specific reducing gas flow in the reactor and / or a parameter value that characterizes a reducing gas exit velocity from a reactor base, e.g. a nozzle base.

[0023] Properties of the feedstock—where multiple feedstocks or mixtures thereof can also be considered—can be taken into account by determining a parameter value that characterizes the feedstock quality. The feedstock quality can, for example, relate to the homogeneity and / or pellet size of the iron ore pellets used. Quality or homogeneity can refer to mechanical strength (CCS - Cold Crushing Strength), abrasion resistance, reducibility, porosity, and / or the like. The use of lower-quality pellets, for example, can lead to a lower degree of reduction, while very large pellets may not achieve sufficient, continuous reduction of the iron ore.Alternatively or additionally, feedstock handling can be taken into account by determining i) a parameter value that characterizes the feed rate of at least one feedstock into the reactor, ii) a parameter value that characterizes the feedstock bed height in the reactor, and / or iii) a parameter value that characterizes the feedstock residence time in the reactor. Insufficient feedstock feed, a low feedstock bed height, or a short feedstock residence time can also lead to inadequate reduction and, consequently, a low-quality end product.

[0024] A particularly reliable assignment of the reactor operating state to an operating category can be achieved by determining a reducing gas temperature, a reactor temperature (e.g., in a reactor bed formed from the feedstock), and / or a temperature in the reformer or gas heater. Parameter values ​​characterizing such temperatures can provide information about whether the conditions for efficient direct reduction in the reactor or, more generally, for the efficient operation of the direct reduction plant as a whole are met.

[0025] Alternatively or additionally, a particularly reliable assignment of the reactor operating state to an operating category can also be achieved by determining a reactor pressure, e.g., a gas pressure in the (closed) reactor, a hydraulic pressure, e.g., the pressure of a hydraulic medium required to move a transport or circulation device in the reactor, such as a burden feeder, and / or a differential pressure between an upper reactor region and a lower reactor region, particularly in a fluidized-bed reactor. Parameter values ​​characterizing such pressures can also provide information about whether the conditions for efficient direct reduction are met at all or whether the functionality of the direct reduction plant remains unimpaired.

[0026] If necessary, other process parameters can also be considered in addition or as an alternative. This can, in some cases, at least slightly further increase the robustness of the categorization by the machine-learned model.

[0027] After determining the previously mentioned parameter values, these are expediently used as the basis for assigning the reactor operating state - defined by them - to the operating category.

[0028] In principle, it is possible to estimate the development of the reactor operating state or the reduction process based on the predetermined number of process parameters or the corresponding parameter values, in particular a time profile of these parameter values. For this purpose, the current parameter value of at least one of the process parameters is preferably estimated using at least one further machine-learned model. This estimation is expediently carried out on the basis of the current parameter values ​​of the remaining process parameters.For example, a parameter value characterizing the current reducing gas flow through the reactor can be estimated based on current parameter values ​​for the exit velocity of the reducing gas, the reactor pressure, the differential pressure between an upper and lower reactor region, the reducing gas temperature, the reactor temperature, the feed rate of at least one feedstock, the feedstock bed height, and / or the like. Based on an error output by at least one further machine-learned model for this estimation, it can then be checked whether there is a disturbance in the reactor operating state or reduction process defined by the current parameter values. The error is expediently taken into account over a predetermined time period. For example, it can be checked whether the error reaches or exceeds a predetermined threshold within this time period.If this is the case, it may indicate a disruption of the reduction process taking place in the reactor and the operating personnel will be informed accordingly.

[0029] In this context, a predetermined time period is preferably a maximum of 4 to 6 hours, particularly preferably 2 to 4 hours, and in particular a minimum of 30 minutes. Depending on the length of the predetermined time period, different types of disturbances can be identified.

[0030] If the error output by the further machine-learned model or a parameter derived from the error only reaches or exceeds the threshold value for a short time within the predetermined time range, i.e., within a short period of time, for example, a few minutes, up to a maximum of 30 minutes, the disturbance may be a so-called spontaneous state change (point anomaly). The threshold value can therefore also be referred to as a "spike limit." Such spontaneous state changes are detected, for example, when a sensor is defective and accordingly delivers incorrect parameter values. In this sense, a spontaneous state change can be a virtual state change, and the associated disturbance of the reactor operating state or reduction process can be a virtual disturbance.On the other hand, such a spontaneous change of state can also occur due to a temporary disturbance in the direct reduction process, such as a disturbance in the material flow or gas flow. Such a disturbance can lead to the collapse of the fluidized bed in a fluidized bed reactor (defluidization). The type of spontaneous change of state occurring can be determined, if necessary, by further analysis of the process parameters, in particular the temporal development of the other process parameters and / or the other errors output by the at least one machine-learned model.

[0031] However, if the error output by the additional machine-learned model or the parameter derived from it reaches or exceeds the threshold within the predetermined time range over a longer period, i.e., within a long period of time of, for example, more than 1 hour, the fault may be a state change (i.e., a change to a different operating state). Such state changes can occur, for example, due to a change in the operation of the direct reduction plant, in particular the reactor, or other phenomena, such as buildup, agglomeration of feed materials, or unfavorable feed material mixtures (proportion of lump or fines in the feed material) in a shaft reactor.Here too, a further analysis of the process parameters, in particular the temporal development of the other process parameters and / or the other errors output by the at least one further machine-learned model, may provide additional information about the state change, in particular its cause.

[0032] Particularly preferably, therefore, not only is a current parameter value estimated for one process parameter and the error determined in this way (or a characteristic value derived therefrom) used as the basis for checking whether a fault exists. Rather, using a corresponding number of further machine-learned models, the current parameter values ​​for each process parameter are preferably estimated on the basis of current parameter values ​​of the respective other process parameters. On the basis of the errors output by the further machine-learned models for these estimates, an overall error is expediently determined, which forms the basis for checking whether a fault exists in the reactor operating state or direct reduction process defined by the parameter values ​​of the process parameters. This overall error or a characteristic value derived therefrom, e.g.A smoothed total error is also checked to determine whether it reaches or exceeds a specified threshold, or, if applicable, how long this threshold is reached or exceeded. The evaluation of the total error enables a particularly reliable assessment of the operating or process state or its development, in particular whether a spontaneous change in state or a change in state is occurring or at least imminent.

[0033] To determine the cause of the disruption to the reduction process or the operating state, the individual errors determined for the estimates and / or the determined total error are preferably considered. This should also include the consideration of derived parameters, such as a smoothed total error.

[0034] For example, it can be checked for which process parameters the detected errors are increasing, for example whether they reach or exceed the corresponding threshold values. Preferably, the temporal development, i.e. the temporal progression, of the parameter values ​​and / or the corresponding errors is also evaluated. For example, if a poor or at least only neutral reactor operating state is suddenly determined using the machine-learned model, those process parameters whose corresponding error reaches or exceeds the corresponding threshold value can provide an indication of the cause of the change in state or the now poor or only neutral state. In particular, on this basis, for example using explainable Kl (explainable AI algorithms), the operating personnel can be given indications as to which of the process parameters is “conspicuous” and thus primarily responsible for a “poor” operating state.

[0035] If necessary, it is also possible to estimate a further development of the reactor operating state or reduction process based on the temporal progression of the parameter values ​​of at least one of the process parameters and a parallel temporal progression of the assigned operating categories. For this purpose, for example, a trajectory of the reactor operating state can be determined and visualized. This allows the operating personnel to take countermeasures at an early stage and thus, if necessary, avert a transition to a poor operating state. It is also conceivable to have the temporal progression of the parameter values ​​and / or the parallel assigned operating category, in particular the trajectory, analyzed by an algorithm and to have an alarm issued automatically or at least semi-automatically if the algorithm detects that the reduction process is deteriorating.

[0036] According to a second aspect of the invention, the system for characterizing a direct reduction process comprises: i) a sensor device with which parameter values ​​of a predetermined number of different process parameters can be determined in a reduction process taking place in at least one reactor of a direct reduction plant; and ii) a categorization device comprising a first machine-learned model and configured to assign a reactor operating state defined by the parameter values ​​to an operating category based on the determined parameter values, in particular by means of the machine-learned model.

[0037] With such a system, the reactor operating status and the corresponding reduction process taking place in the reactor can be assessed effectively, possibly even without in-depth prior knowledge on the part of the operating personnel. This eliminates the need to manually monitor and evaluate a multitude of process parameters based on experience, reducing effort and costs. Furthermore, it can also be detected promptly if the reduction process is deteriorating. This allows for a timely response to this deterioration, saving resources, especially energy, used to operate the direct reduction plant.

[0038] The sensor device expediently comprises various sensor types in order to be able to detect the parameter values ​​of the predetermined number of process parameters. The number of sensor types expediently corresponds to the predetermined number of process parameters. One or more sensors can be provided per sensor type in order to be able to detect the parameter values ​​of the corresponding process parameter precisely and comprehensively, for example, with spatial resolution. Accordingly, at least one sensor is preferably provided for monitoring a reducing gas-related process parameter, e.g., at least one sensor for detecting a reducing gas composition, a reducing gas flow in the reactor, and / or a reducing gas exit velocity from a reactor bottom.

[0039] Alternatively or additionally, at least one sensor is provided for monitoring a pressure-related process parameter, e.g., at least one sensor for detecting a reactor pressure, a hydraulic pressure, and / or a differential pressure between an upper and a lower reactor region.

[0040] Alternatively or additionally, at least one sensor is provided for monitoring a temperature-related process parameter, e.g., at least one sensor for detecting a reducing gas temperature, a reactor temperature, and / or a temperature in a process gas reformer.

[0041] Alternatively or additionally, at least one sensor is provided for monitoring a feedstock-related process parameter, e.g., at least one sensor for detecting a feed rate of at least one feedstock into the reactor, a feedstock quality, a feedstock bed height in the reactor, and / or a feedstock residence time in the reactor.

[0042] The categorization device can be implemented in software and, if appropriate, additionally, at least partially, in hardware. The categorization device can, in particular, comprise a processing unit, preferably connected to a memory and / or bus system for data or signals. For example, the categorization device can comprise a microprocessor unit (CPU) or a module thereof and / or one or more programs or program modules. In particular, the machine-learned model can form a program or software module of the categorization device. The categorization device can be configured to process instructions implemented as a program stored in a memory system or as a machine-learned model, to acquire input signals from a data bus, and / or to output output signals to a data bus.A storage system may comprise one or more, in particular different, storage media, in particular optical, magnetic, solid-state, and / or other non-volatile media. The program or the machine-learned model may be configured such that it embodies or is capable of executing the method described here or at least parts thereof, so that a process based on direct reduction can be characterized.

[0043] According to a third aspect of the invention, the direct reduction plant for reducing at least one feedstock, in particular a metal ore, has at least one reactor and a system for characterizing a direct reduction process according to the second aspect of the invention. At least part of the sensor device of the system is preferably arranged in the region of the reactor. For example, at least one sensor of the sensor device can be arranged in and / or on the reactor and configured to determine characteristic process parameters for sub-processes of the reduction process taking place in and / or on the reactor - such as, for example, the reduction of the feedstock, the swirling of the feedstock layered in the reactor, i.e. the reactor bed, the removal of thermal energy, a resulting pressure difference in the reactor and / or the like. In this sense, a sensor device orA sensor can also be understood as an interface through which corresponding parameter values ​​can be accessed. For example, parameter values ​​such as those for controlling the temperature of the process gas heated in a reformer and / or hydraulic pressure in a hydraulic system for moving transport or circulation elements in the reactor can be provided by a plant or reactor control system, e.g., via an interface. This allows process-critical process parameters, in particular, to be determined directly and with exceptional reliability.

[0044] According to a fourth aspect of the invention, in the method for machine learning a model, which can be used in a method according to the first aspect of the invention, i) parameter values ​​of a predetermined number of different process parameters are provided, which were determined in a reduction process taking place in at least one reactor of a direct reduction plant over a predetermined period of time, ii) reactor operating states defined in each case by a set of parameter values ​​of the process parameters are assigned to an operating category at different times within the predetermined period of time, and iii) a model is machine learned by means of an algorithm on the basis of the parameter values ​​defining the reactor operating states and the respectively assigned operating categories.

[0045] Using such a machine-learned model, a current reactor operating state can be quickly and reliably assigned to an operating category. This does not require a large, e.g., double-digit, number of different process parameters. Rather, a small number of process parameters, such as fifteen or even just ten, is sufficient to perform a robust categorization of a current reactor operating state.

[0046] A predetermined period of time within the meaning of the invention is preferably a period of at least two years, in particular essentially three years. Within this period, a reactor of a direct reduction plant typically undergoes a sufficient number of different operating states to allow robust training of the model. To enable robust operation with reproducible results from the machine-learned model, the process parameters and / or time ranges most relevant to reactor operating states are expediently selected against the background of process engineering experience, for example, by experienced operating personnel. The parameter values ​​of these selected process parameters can be taken, for example, from a database in which historical operating data of the reactor or, more generally, of the direct reduction plant is stored.

[0047] Experiments have shown that the most relevant process parameters are the reducing gas composition, the reducing gas flow, the reducing gas outlet velocity from the reactor bottom, the reactor pressure, the hydraulic pressure, the differential pressure between an upper and a lower reactor region, the reducing gas temperature, the reactor temperature, the temperature in a process gas reformer, the feedstock quality, the feedstock feed rate, the feedstock bed height and / or the feedstock residence time.

[0048] For practical purposes, experienced operating personnel manually assign each set of process parameter values, or the reactor operating state defined thereby, to a process category. For example, the respective reactor operating state can be assigned to one of three states: "good," "neutral," or "poor" for a variety of time points. The resulting data set can then be analyzed by artificial intelligence, i.e., the model can be machine-learned using this data set.

[0049] According to a fifth aspect of the invention, in the method for machine learning a further model, which can be used in a method according to the first aspect of the invention, i) parameter values ​​of a predetermined number of different process parameters are provided, which were determined in a reduction process running in at least one reactor of a direct reduction plant over a predetermined period of time, and ii) by means of an algorithm on the basis of parameter values ​​of the process parameters from a plurality of predetermined process sections, a further model is machine-learned, with which a current parameter value can be estimated for at least one process parameter on the basis of the current parameter values ​​of the other process parameters.

[0050] A predetermined process section in the sense of the invention is preferably a predetermined time range or is defined by the predetermined time range, i.e. a predetermined period of time.

[0051] Using such an additional machine-learned model, the current parameter value of an additional process parameter can be reliably estimated based on the current parameter values ​​of various process parameters. This does not require the use of a large number of different process parameters. Rather, current parameter values ​​of a small number of process parameters, for example, fifteen or even just ten process parameters, are sufficient to enable a robust estimation of the parameter value of the additional process parameter.

[0052] Here, too, process parameters and time ranges in which the parameter values ​​of the selected process parameters are taken into account are expediently selected on the basis of process engineering experience, for example by experienced operating personnel. The specific reactor operating states corresponding to this selection can serve as training data for one or more further models. By selecting in particular such time ranges - i.e. corresponding process sections of the reduction process - in which different reactor operating states occur, the model or models can be trained in particular to recognize deviations from a current reactor operating state and / or changes to another operating state, i.e. a different operating category. The error output by the trained model can also be used as an indicator of such deviations or changes in operating state.

[0053] Short description of the drawings

[0054] The above-described properties, features, and advantages of this invention, as well as the manner in which they are achieved, will become clearer and more readily understood in connection with the following description of an embodiment, which is explained in more detail in conjunction with the drawings.

[0055] FIG 1 shows an example of a direct reduction plant for reducing a feedstock;

[0056] FIG 2 shows an example of a method for characterizing a direct reduction process;

[0057] FIG 3 shows an example of a trajectory of a reactor state;

[0058] FIG 4 shows an example of a time course of a total error determined when estimating current parameter values;

[0059] FIG 5 shows an example of a method for machine learning a model; and

[0060] FIG 6 shows an example of a method for machine learning another model. Where appropriate, the same reference numerals are used in the figures for the same or corresponding elements of the invention.

[0061] Description of the embodiments

[0062] FIG 1 shows an example of a direct reduction plant 1 for reducing a feedstock 2, in particular a metal ore, with a reactor 3, a gas reformer or gas heater 4, and a system 10 for characterizing a direct reduction process taking place in the reactor 3. The reactor 3, designed as a shaft furnace in the present example, can be charged from above with the feedstock 2, which forms a bed 5 in the reactor 3, also referred to as a reactor bed. The bed 5 can be flowed through from below, i.e. effectively in a countercurrent process, with a reducing gas 6, whereby the feedstock 2 is reduced and can ultimately be discharged as the end product at a lower end of the reactor 3.

[0063] The reducing gas 6, which can also be a gas mixture of several gases that reduce the feedstock 2, such as hydrogen and carbon monoxide, is provided by the reformation of a process gas 7, for example methane, in the gas reformer or gas heater 4. The exhaust gas produced by the reduction of the feedstock 2 as it flows through the bed 5 is scrubbed in a gas scrubber 8. Part of the exhaust gas can be fed into the gas reformer or gas heater 4 as fuel gas. Another part of the exhaust gas is compressed as a so-called carrier gas in a compressor 9 and, if appropriate after the addition of additional gas (not shown), in particular hydrogen, serves again as process gas 7. If appropriate, further additional gases 7a, for example natural gas and / or oxygen, can be admixed with the reducing gas 6 before it is introduced into the reactor 3.Alternatively or additionally, these additional gases 7a can also be introduced directly into the reactor 3.

[0064] The system 10 for characterizing a reduction process taking place in the reactor 3 comprises a sensor device 12 and a categorization device 14 with a machine-learned model 16. One part of the sensor device 12 is arranged in the region of the reactor 3, in particular in and / or on the reactor 3, another part is arranged in the region of the reformer or gas heater 4, in particular in and / or on the reformer or gas heater 4. The sensor device 12 comprises a plurality of sensors 18 in order to determine parameter values ​​of a predetermined number of process parameters. The sensors 18 can represent different sensor types with which, for example, a composition of the reducing gas 6, a flow of the reducing gas 6 in the reactor 3, an exit velocity of the reducing gas 6 into the lower part of the reactor 3, e.g.from a reactor bottom, a pressure in the reactor 3, a pressure in a hydraulic system in and / or on the reactor 3, a differential pressure between an upper and lower reactor region, a temperature of the reducing gas 6, a temperature in the reactor 3, a temperature in the gas reformer or gas heater 4, a feed rate of the feedstock 2, a quality of the feedstock 2, the height of the bed 5 and / or a residence time of the feedstock 2 in the reactor 3 can be detected. Alternatively or additionally, the sensor device 12 can also have an interface to be understood as a sensor 18 for this purpose, via which parameter values ​​for such process parameters can be retrieved, for example, by a plant control 20.

[0065] The categorization device 14 is configured to assign a reactor operating state defined by the parameter values ​​to one of several operating categories based on the determined parameter values. For this purpose, the determined parameter values ​​can, for example, be fed to the machine-learned model 16 as input values. In other words, the machine-learned model 16 can evaluate the determined parameter values ​​to determine the reactor operating state and assign it to an operating category. The result of this assignment, for example, a code characterizing the operating category, can expediently be output via an output interface 22, such as a screen, a data bus, a printer, and / or the like.

[0066] For example, the MIDREX® process can be carried out using the direct reduction plant 1 shown in FIG. A configuration with a reactor 3 designed as a shaft furnace, which can be used to carry out the Energiron® process, is equally conceivable. It is also possible to implement the direct reduction process with a reactor 3 designed as a fluidized-bed reactor, in which the direct reduction of the feedstock 2 takes place.

[0067] FIG. 2 shows an example of a method 100 for characterizing a direct reduction process. In a method step S1, parameter values ​​A, B, C, D, and E of a predetermined number of different process parameters are determined for a reduction process currently taking place in a reactor of a direct reduction plant. Five process parameters are considered here purely as examples. The parameter values ​​A to E of the five process parameters are expediently characteristic of the current reactor operating state.They can, for example, represent a composition of a reducing gas, a flow of the reducing gas in the reactor, an exit velocity of the reducing gas from a reactor bottom, a pressure in the reactor, a pressure in a hydraulic system in and / or on the reactor, a differential pressure between an upper and lower reactor region, a temperature of the reducing gas, a temperature in the reactor, a temperature in a reformer, a feed rate of a feedstock, a quality of the feedstock, a feedstock bed height, and / or a residence time of the feedstock in the reactor. The parameter values ​​A to E can, for example, be recorded and / or retrieved by a sensor device. In a further method step S2, a machine-learned model 16 assigns a reactor operating state defined by the parameter values ​​A to E to an operating category X, Y, Z on the basis of the determined parameter values ​​A to E.In particular, the reactor operating state, characterized by the parameter values ​​A to E, is compressed into an easily interpretable key figure. The machine-learned model 16 can thus represent the several years of experience of a large number of operators or the corresponding expert knowledge.

[0068] One of the operating categories, operating category X, can, for example, represent "good," i.e., desired, reactor operating conditions in which the reduction process proceeds efficiently and essentially without disruption. Reactor operating conditions of operating category X require a relatively low use of resources such as energy, reducing gas, optionally additional gas, and / or the like to operate the direct reduction plant, while achieving good and / or consistent product quality. Operating category X is indicated in FIG. 2 by the absence of hatching.

[0069] Another operating category, operating category Y, can represent "neutral" reactor operating conditions that are tolerable but not optimal. In reactor operating conditions of operating category Y, the reduction process may be comparatively less efficient and may be subject to minor disruptions. For example, this may result in temporarily increased resource consumption and / or a failure or defect of individual sensors in the sensor device used to determine parameter values ​​A to E. Operating category Y is indicated in FIG. 2 by the vertical hatching.

[0070] Another operating category, operating category Z, can represent "poor," i.e., undesirable, reactor operating conditions in which the reduction process is inefficient and / or permanently disrupted. Reactor operating conditions in operating category Z accordingly require a relatively high input of resources for plant operation and / or include problematic or even critical process developments and consequences, such as the formation of deposits on the reactor shaft walls or agglomeration of feed materials. Operating category Z is indicated in FIG. 2 by horizontal hatching.

[0071] In a further, preferred method step S3, the operating category X to Z assigned to the current reactor operating state is output, for example via an interface such as a display or at least a data bus. The output can be conceivable as a type of "status traffic light", whereby the reactor operating state can be reliably identified by the operating personnel even without complex and time-consuming manual evaluation of a large number of operating parameters or without extensive process engineering experience. If necessary, in method step S3, not only the operating category X to Z but also a status forecast or at least a trend of how the reactor operating state is developing can be output. This status forecast orThe trend can be visualized in one of several ways, for example, as a 2D plot, in particular as a projection of selected process parameters into a two-dimensional representation, or as a trend over a predetermined period of time, such as a time trend over the last week. A possible visual representation is shown in FIG 3.

[0072] FIG. 3 shows an example of a trajectory 110 of a reactor operating state 30. The reactor operating state 30 can fundamentally be defined by the parameter values ​​of a predetermined number of process parameters. To illustrate the trajectory 110 in FIG. 3, operating categories assigned to the reactor operating state 30 at different times, which are indicated by different hatching, are plotted against two of these process parameters with the parameter values ​​A, B. Temporarily consecutive times are connected by lines. For reasons of clarity, the reactor operating state 30 is provided with a reference symbol only at the start time.

[0073] FIG 3 shows that the reactor in which the direct reduction process takes place is initially in a "good" state. This reactor operating state 30 is indicated in FIG 3 by the absence of hatching.

[0074] As time progresses, the process parameter plotted on the abscissa increases, ie, the corresponding parameter value A increases. The reactor then transitions to a "neutral" state, which is indicated by vertical hatching in FIG. 3.

[0075] Eventually, both process parameters shrink, i.e., the corresponding parameter values ​​A and B decrease, and the reactor enters a "poor" state. This reactor operating state 30 is indicated by horizontal hatching in FIG. 3. Towards the end of trajectory 110, parameter value B increases again, and the "good" operating state is restored.

[0076] Such a trajectory 110 can be determined and, as shown by way of example in FIG 3, visualized by assigning the reactor operating state 30 to one of the operating categories by means of a correspondingly machine-learned model using the method described in connection with FIG 2 on the basis of the parameter values ​​determined at several successive points in time.

[0077] Advantageously, not just two process parameters are considered, but several, e.g. a composition of a reducing gas, a flow of the reducing gas in the reactor, an exit velocity of the reducing gas from a reactor bottom, a pressure in the reactor, a pressure in a hydraulic system in and / or on the reactor, a differential pressure between an upper and lower reactor region, a temperature of the reducing gas, a temperature in the reactor, a temperature in a reformer, a feed rate of a feedstock, a quality of the feedstock, a feedstock bed height and / or a residence time of the feedstock in the reactor. In this respect, the trajectory can also be multidimensional, i.e. extend in more than two dimensions. For visualization, such a multidimensional trajectory of the reactor state can be projected onto a plane, in particular onto main components, and consequently represented as a two-dimensional plot.This allows a holistic overview of the operating status to be created.

[0078] In principle, it is also conceivable to additionally display "historical" operating data of the reactor, i.e., older pairs and / or projections of parameter values ​​A, B, in such a plot, e.g., in the trajectory 110 shown in FIG. 3. These operating data can, for example, be displayed transparently in the background of the trajectory 110 to enable a comparison with past operating states.

[0079] FIG 4 shows an example of a time course 120 of a total error G determined during the estimation of current parameter values, for example by means of further machine-learned models. For this purpose, the total error G is plotted against time t in a process section T, ie within a certain time range.

[0080] Such a total error G can be determined, for example, by determining a parameter value for a predetermined number of process parameters at a time while a reduction process is taking place in a reactor of a direct reduction plant. For each of the process parameters, the parameter value can then be additionally estimated using the additional machine-learned model specifically for this process parameter, based on the parameter values ​​of the remaining process parameters. The additional machine-learned model expediently outputs an error, for example as the difference between the estimated parameter value and the actually determined, e.g. measured, parameter value. Using the corresponding additional machine-learned models, this can be done for each process parameter, so that an error is determined for each process parameter.The total error G can then be calculated based on the errors determined for each process parameter, for example, by averaging, weighted averaging, and / or another suitable statistical method. If this is performed for several consecutive points in time, a time course 120 of the total error G is obtained.

[0081] Due to the strong fluctuation of the determined total error G, smoothing may be considered. A suitable smoothing method, such as a regression analysis or a regression analysis, can be applied for this purpose. The result of such a smoothing, the smoothed total error 122, is shown as a dashed line in FIG. 4.

[0082] In principle, the total error G, in particular the time profile 120, can be used to check whether a disturbance S in the reactor operating state is present. For example, it can be checked whether the smoothed total error 122 reaches or exceeds a first predetermined threshold value 124, which can also be referred to as the "average limit" and is shown in FIG 4 as a horizontal, dash-dotted line. In FIG 4, this is the case in the period marked by the dashed vertical lines. Such a disturbance S, present over a longer period, can indicate a change in state, also referred to as "drift", such as can occur, for example, due to the formation of deposits on a shaft wall or the agglomeration of feed materials.

[0083] Alternatively or additionally, it is also possible to check whether the total error G reaches or exceeds a second predetermined threshold value 126, which can also be referred to as the "spike limit" and is also shown in FIG. 4 as a horizontal, dash-dotted line. In FIG. 4, this is the case for a short period of time. Such a short disturbance S is also referred to as a spontaneous change of state or point anomaly and can be triggered, for example, by a defect in a measuring instrument or by a sudden, severe subsidence of a charge material in a reactor designed as a shaft furnace or by defluidization in a fluidized bed reactor.

[0084] If necessary, the cause of the fault S can be determined more precisely by also considering the temporal progression 120 of the individual errors determined for each process parameter and / or the temporal progression 120 of the corresponding parameter values, i.e., the temporal development of these errors or values. In particular, just as for the overall error G, it can be checked for each individual error determined for each process parameter whether a first predetermined threshold value 124 and / or a second predetermined threshold value 126 is reached or exceeded and, if applicable, also whether this correlates in time with the fault S identified on the basis of the overall error G. In this way, it can be deduced, if applicable, which of the process parameters are involved in the fault S or even cause it.

[0085] FIG. 5 shows an example of a method 200 for machine learning a model 16, which assigns a reactor operating state defined by determined parameter values ​​A, B, C, D, E of a predetermined number of process parameters to an operating category X, Y, Z. The operating categories X to Z are indicated by hatching, as in FIG. 2.

[0086] In a method step V1, the parameter values ​​A to E determined over a predetermined period of time during a reduction process taking place in a reactor of a direct reduction plant are provided. Providing these values ​​may include determining them, for example, by sensory detection or retrieval. For example, in method step V1, a database in which the parameter values ​​A to E are stored over a period of several years, for example, up to three years, may be read out.

[0087] In a further process step V2, reactor operating states defined by a set of parameter values ​​A to E are assigned to one of the operating categories X to Z at different times within the predetermined period. This assignment can be performed manually, for example, by experienced operators. The assignment is therefore based on expert knowledge. This assignment creates a new data set that can be used as a training data set in a further process step V3.

[0088] In the subsequent process step V3, model 16 is machine-learned using an algorithm based on the parameter values ​​A to E defining the reactor operating states and the respective associated operating categories X to Z. In other words, model 16 is trained based on the training data set generated in process step V2, so that it can automatically assign one of the operating categories X to Z to a reactor operating state defined by new current parameter values ​​A to E.

[0089] FIG 6 shows an example of a method 300 for machine learning of a further model 24, which estimates a current parameter value A of a process parameter from a predetermined number of process parameters on the basis of current parameter values ​​B, C, D, E of further process parameters.

[0090] In a method step W1, the parameter values ​​A to E of the predetermined number of different process parameters are provided, which were determined over a predetermined period of time during a reduction process taking place in a reactor of a direct reduction plant. Providing these values ​​can include determining them, for example, by sensory detection or retrieval. For example, in method step W1, a database in which the parameter values ​​A to E are stored over a period of several years, for example, up to three years, can be read out.

[0091] In a further method step W2, the further model 24 is machine-learned by means of an algorithm on the basis of parameter values ​​A to E of the process parameters from a plurality of predetermined process sections T within the predetermined period of time. For example, the further model 24 can be trained under the proviso that the parameter value A is estimated on the basis of the parameter values ​​B to E. In this case, in particular for each of the predetermined process sections T, the time profile 120 of the parameter value A and the corresponding parameter values ​​B to E can also be taken into account, ie the learning of the estimate of the parameter value A can be used as a basis. These time profiles 120 are shown in FIG 6 for the parameter values ​​A to C during a process section T by plotting the parameter values ​​A to C against time t.

[0092] Although the invention has been illustrated and described in detail by the preferred embodiments, the invention is not limited by the disclosed examples and other variations may be derived therefrom by those skilled in the art without departing from the scope of the invention.

[0093] List of reference symbols

[0094] 1 direct reduction plant

[0095] 2 Input material

[0096] 3 reactor

[0097] 4 gas reformers / gas heaters

[0098] 5 Bulk

[0099] 6 Reducing gas

[0100] 7 Process gas

[0101] 7a Additional gas

[0102] 8 gas scrubbers

[0103] 9 Compressor

[0104] 10 systems

[0105] 12 Sensor device

[0106] 14 Categorization device

[0107] 16 Model

[0108] 18 Sensor

[0109] 20 Plant control

[0110] 22 Output interface

[0111] 24 additional models

[0112] 30 Reactor operating condition

[0113] 100 methods for characterizing a direct reduction process

[0114] S1 Determination of parameter values

[0115] S2 Assigning an operating category

[0116] S3 Output of the assigned operating category

[0117] 200 methods for machine learning a model

[0118] V1 Providing parameter values

[0119] V2 Assigning operating categories

[0120] V3 Learning the model

[0121] 300 methods for machine learning another model

[0122] W1 Providing parameter values

[0123] W2 Learning the further model

[0124] 110 T rajectory

[0125] 120 time course 122 smoothed total error

[0126] 124 first threshold

[0127] 126 second threshold

[0128] A, B, C, D, E parameter value

[0129] X, Y, Z operating category

[0130] G Total error t Time

[0131] S disturbance

[0132] T Process section

Claims

Claims 1. Method (100) for characterizing a direct reduction process, wherein - parameter values ​​(A, B, C, D, E) of a predetermined number of different process parameters are determined (S1) for a reduction process taking place in at least one reactor (3) of a direct reduction plant (1), and - a machine-learned model (16) assigns a reactor operating state (30) defined by the parameter values ​​(A, B, C, D, E) to an operating category (X, Y, Z) on the basis of the determined parameter values ​​(A, B, C, D, E) and outputs a measure of the quality of the reduction process, so that the parameter values ​​(A, B, C, D, E) characterizing the reactor operating state (30) are combined by means of the machine-learned model (16) to form an easily interpretable key figure.

2. The method (100) according to claim 1, wherein at least one future operating and / or product parameter is predicted by means of the machine-learned model (16) on the basis of the determined parameter values ​​(A, B, C, D, E).

3. Method (100) according to claim 1 or 2, wherein the parameter values ​​(A, B, C, D, E) of fifteen or fewer, preferably ten or fewer, in particular eight or fewer, process parameters are determined and used as the basis for assigning the reactor operating state (30) to an operating category (X, Y; Z) and / or for predicting at least one future operating and / or product parameter.

4. The method (100) according to claim 1 or 2, wherein the predetermined number of different process parameters comprises at least one process gas-related process parameter, at least one pressure-related process parameter, at least one temperature-related process parameter and / or at least one feedstock-related process parameter.

5. The method (100) according to claim 3, wherein - a parameter value (A, B, C, D, E) characterising a reducing gas composition, - a parameter value (A, B, C, D, E) characterising an absolute and / or specific reducing gas flow in the reactor (3), and / or - a parameter value (A, B, C, D, E) which characterizes a differential pressure between the reactor inlet region and the reactor outlet region is determined and used as the basis for assigning the reactor operating state (30) to the operating category (X, Y, Z) and / or for predicting at least one future operating and / or product parameter.

6. Method (100) according to one of claims 3 to 5, wherein - a parameter value (A, B, C, D, E) characterising a reducing gas temperature, - a parameter value (A, B, C, D, E) characterising a reactor temperature, and / or - a parameter value (A, B, C, D, E) which characterizes a reformer temperature or gas heater temperature (4) is determined and used as the basis for assigning the reactor operating state (30) to the operating category (X, Y, Z) and / or for predicting at least one future operating and / or product parameter.

7. Method (100) according to one of claims 3 to 6, wherein - a parameter value (A, B, C, D, E) which characterises a feed rate of at least one feedstock (2) into the reactor (3) and / or a discharge rate from the reactor, - a parameter value (A, B, C, D, E) that characterizes a feedstock quality, - a parameter value (A, B, C, D, E) characterising a feedstock bed height in the reactor (3), and / or - a parameter value (A, B, C, D, E) which characterizes a feedstock residence time in the reactor (3) is determined and used as the basis for assigning the reactor operating state (30) to the operating category (X, Y, Z) and / or for predicting at least one operating and / or product parameter.

8. Method (100) according to one of the preceding claims, wherein - by means of at least one further machine-learned model (24), the current parameter value (A) of at least one of the process parameters is estimated on the basis of current parameter values ​​(B, C, D, E) of the remaining process parameters and - on the basis of an error output by at least one further machine-learned model (24) for this estimation, it is checked whether a disturbance (S) of the reactor operating state (30) defined by the current parameter values ​​(A, B, C, D, E) is present.

9. Method (100) according to one of the preceding claims, wherein - by means of further machine-learned models (24), the current parameter values ​​(A, B, C, D, E) for each process parameter are estimated on the basis of current parameter values ​​(A, B, C, D, E) of the remaining process parameters, and - on the basis of errors output by the further machine-learned models (24) for these estimates, a total error (G) is determined which is used as the basis for a check as to whether there is a disturbance (S) in the reactor operating state (30) defined by the parameter values ​​(A, B, C, D, E) of the process parameters.

10. The method (100) according to claim 9, wherein a cause for the disturbance (S) of the reactor operating state (30) defined by the parameter values ​​(A, B, C, D, E) of the process parameters is determined on the basis of the errors determined for the estimates and / or the total error (G).

11. Method (100) according to one of the preceding claims, wherein a further development of the reactor operating state (30) is estimated on the basis of a time profile (120) of the parameter values ​​(A, B, C, D, E) of at least one of the process parameters and a parallel time profile (120) of the assigned operating categories (X, Y, Z).

12. System (10) for characterizing a direct reduction process, comprising - a sensor device (12) with which current parameter values ​​(A, B, C, D, E) of a predetermined number of different process parameters can be determined in a reduction process taking place in at least one reactor (3) of a direct reduction plant (1), and - a categorisation device (14) which has a machine-learned model (16) and is designed to, on the basis of the determined parameter values ​​(A, B, C, D, E), create a reactor defined by the parameter values ​​(A, B, C, D, E) of the process parameters. to assign an operating state (30) to an operating category (X, Y, Z) and to output a measure of the quality of the reduction process, so that the parameter values ​​(A, B, C, D, E) characterizing the reactor operating state (30) are summarized by means of the machine-learned model (16) to form an easily interpretable key figure.

13. Direct reduction plant (1) for reducing at least one feedstock (2), in particular a metal ore, with at least one reactor (3) and a system (10) for characterizing a direct reduction process according to claim 12.

14. A method (200) for machine learning a model (16) which can be used in a method (100) according to one of claims 1 to 11, wherein - parameter values ​​(A, B, C, D, E) of a predetermined number of different process parameters are provided (V1), which were determined in a reduction process taking place in at least one reactor (3) of a direct reduction plant (1) over a predetermined period of time, - reactor operating states (30) defined by a set of parameter values ​​(A, B, C, D, E) of the process parameters are assigned to an operating category (X, Y, Z) at different times within the predetermined period (V2) and - a model (16) is machine-learned (V3) by means of an algorithm based on the parameter values ​​(A, B, C, D, E) defining the reactor operating states (30) and the respectively assigned operating categories (X, Y, Z).

15. A method (300) for machine learning a further model (24) which can be used in a method (100) according to one of claims 8 to 10, wherein - parameter values ​​(A, B, C, D, E) of a predetermined number of different process parameters are provided (W1) which are determined in a direct reduction process in at least one reactor (3). reduction process taking place in the plant (1) over a predetermined period of time, and - by means of an algorithm on the basis of parameter values ​​(A, B, C, D, E) of the process parameters from a plurality of predetermined process sections (T), a further model (24) is machine-learned (W2), with which a current parameter value (A) can be estimated for at least one process parameter on the basis of the current parameter values ​​(B, C, D, E) of the remaining process parameters.

Images