Identifying a value sequence for controlling the pressure in the gas-collecting main of COKE-ovens

Abstract

A time-variant gas-flow parameter ({P}|13) is applied to a pressure-control device (163) in a gas-collecting main (183) of an offtake piping system (143, 143', 183) of at least one stamp-charging coke-oven (103) that is adapted to process a coal-cake (113). The parameter ({P}| 3) is a parameter value sequence (503) that is provided by a computer (203) with a pre-trained neural network (253). The computer (203) receives a set (303) of input parameters ({X}|3), with a modality parameter set ({X0}) that represents a charging modality by that the coal-cake (113) is to be inserted into the stamp-charging coke-oven (103), and a coal-cake property parameter set ({X1}) that represents a property of the coal-cake (113) before insertion. The neural network (253) has been trained with historical data-sets of input parameters from reference operations and with historical sequences of time-variant gas-flow parameters as the ground-truth.

Description

IDENTIFYING A VALUE SEQUENCE FOR CONTROLLING THE PRESSURE IN THE GAS¬COLLECTING MAIN OF COKE-OVENSTechnical Field

[0001] In general, the disclosure relates to industrial production processes and to equipment that performs such processes. More particularly, the disclosure relates to computer systems, methods and computer-program products that support controlling coke-ovens.Background

[0002] Much simplified, industrial ovens process materials at high temperatures. The materials are solid-state materials, such as coke, and the processes in the oven change the properties of the materials. To name an example, some ovens are designed to convert coal into coke that is suitable for use in iron-making furnaces.

[0003] Ovens have openings - such as doors - that are open on a regular basis on purpose, for example, to charge the oven with material and to discharge the oven from processed material. For example, stamp-charging is a technology by that coal-cakes are pressed into the oven. After processing, the coke is removed from the oven. Stamp-charging is known in the art for decades. A reference that explains details is a paper by Jorge Madias et al "A review on stamped charging of coals", September 2013 Conference: 43rd Ironmaking and Raw Materials Seminar, 12th Brazilian Symposium on Iron Ore and 1st Brazilian Symposium on Agglomeration of Iron Ore, September 1st to 4th, 2013, Belo Horizonte, MG, Brazil.

[0004] Processing involves the presence of gases within the oven (so-called "coke-oven gases"). Simplified, the processing ovens are the "ovens under distillation", and the oven gases can be distillation gases and vapors. In other words, a gas-free oven would not work. In principle, the doors are no obstacles to the gases, especially when they are open during charging and discharging. Similarly, leakages (at the doors or elsewhere) would allow gases to leave the oven as well.

[0005] For the gases, at least two aspects should be considered: (i) oxygen from the environmental air may change the chemical composition of the oven gases, and (ii) oven gases should not turn into emissions to the environment.

[0006] The gas pressure inside the oven is controlled to be a slight over-pressure in relation to the air pressure outside the oven. The oven gases flow in a pre-defined direction from the oven to one or more gas-collecting pipes (such as to an oven offtake piping system). There is a pressure gradient from the oven to the pipe.

[0007] Multiple ovens are usually arranged in batteries, and the ovens of a battery have the gas-collecting pipe in common. The skilled person is familiar with adjusting the pressure. While it would be possible to install valves separately for each oven in the battery, usual implementations use a common pressure-control device in the gascollecting pipe.

[0008] Coal-cakes are being inserted to individual ovens separately. During insertion, the doors of the other ovens remain closed. In other words, inserting two cakes into two ovens of the same battery at the same time is not an option. This separation in time allows that the same pressure-control device can control the operation of the multiple ovens in the battery. From an overall perspective, a common pressurecontrol device is a shared resource in multiplex use.

[0009] However, controlling the pressure-control device is a relatively complex task. Usually, a first pressure set-point is used when one of the ovens of the battery is being charged, and a second pressure set-point is used when no oven is being charged. Moreover, during charging, the pressure set-point could be modulated over time to minimize both air intake and gas emissions through the open door. Such modulation would be required to accommodate different coal-cake parameters, such as humidity and the share of volatile matters in the cake, but also different charging modalities.

[0010] Regarding the process states, there are also constraints in the properties of the coal-cake, because every cake would be different. For example, individual cakes differ by volatile matters or humidity.

[0011] Regarding the chemical composition of the gases in the gas-collecting main, there are also constraints that are mainly driven by safety concerns. For example, the share of oxygen O2 should not exceed a certain threshold. Otherwise, equipment that processes the gases from the gas-collecting main may see unwanted reactions with O2.

[0012] Complexity is further caused by certain delays (or inaccuracies) in obtaining parameter values and in actually controlling the valve. For example, detecting environmental-critical emissions from closed doors (i.e., from leakages) takes a couple of minutes, and may not be accurate. Further, when a control signal to change the pressure arrives at the device, it can not change the pressure immediately.

[0013] There is a need to calculate the pressure set-point over time but an analytical solution (in terms of an equation system with multiple parameter values) is not available.

[0014] A stamp-charging coke-oven is adapted to process a coal-cake. The coke-oven comprises a pressure-control device in a gas-collecting pipe. The coke-oven is either a stand-alone oven (i.e., at least one oven) or an oven in a battery (i.e., multiple ovens). In battery implementations, the coke-ovens share the gas-collecting pipe with each other, so that the gas-collecting pipe is a common gas-collecting pipe. To apply commonly used terminology, the description uses the term "gas-collecting main", or GCM.

[0015] The pressure-control device receives control commands from a computer. The computer executes a computer-implemented method to identify a parameter value sequence for a time-variant gas-flow parameter. The computer applies the parameter with the value sequence to the pressure-control device of the at least one coke-oven (or of the GCM of the battery).

[0016] The computer comprises a pre-trained neural network. The neural network is a classifier and it has been trained in advance with training data. During training, theinput of the neural network (under training) has received a plurality of historical input parameter values from reference operations of a reference stamp-charging coke-oven, and the output of the neural network (under training) has received a plurality of historical sequences of time-variant gas-flow parameters as the groundtruth.Training further comprises to identify network weights at network nodes. Once training has been completed, the pre-trained neural network uses these identified network weights.

[0017] The computer performs the method to identify the value sequence while the coke- oven is in operation at the occasion that a coal-cake is to be inserted to the coke- oven.

[0018] In a receiving step, the computer receives a set of input parameters. The input parameters have a modality parameter set that represents a charging modality by that the coal-cake is to be inserted into the stamp-charging coke-oven. The input parameters further comprise a coal-cake property parameter set that represents one or more properties of the coal-cake before insertion.

[0019] In a data-processing step, the computer processes the received set of input parameters by the neural network. By that data-processing the computer provides a preliminary value sequence (as the result of the classifier). The preliminary value sequence is then applied to the pressure-control device in the gas-collecting main (of the at least one stamp-charging coke-oven) during cake insertion.

[0020] Optionally, the modality parameter set comprises modality parameters that represent how the coal-cake is to be inserted into the coke-oven, such parameters are actuator parameters for the oven.

[0021] Optionally, the coal-cake property parameter set comprises one or more parameters that describe properties of the coal-cake. The properties are known before the coal-cake is being inserted.

[0022] There can be no guarantee that the neural network always provides a value sequence that is suitable, such as in view of oxygen and of emissions. Therefore,optiona I ly, the step applying the value sequence is performed conditionally. In other words, the parameter value sequence turns from a preliminary parameter value sequence to a to-be-applied parameter value sequence following a predefined check.

[0023] Optionally, applying the preliminary value sequence is performed conditionally after testing to confirm that that the sequence has its start values within tolerances given by a control loop. The control loop belongs to the pressure-control device.

[0024] But even if the test confirms suitability, and the coal-cake insertion starts with the to-be-applied sequence, the above-mentioned two aspects (oxygen and gas emissions) are considered: optionally, the step applying is performed simultaneously with controlling the pressure-control device by a control loop with at least the following set-points: (i) controlling the pressure-control device such that the oxygen amount of the gas in downstream direction of the gas-collecting main remains below a pre-defined threshold; and (ii) controlling the pressure-control device such that amount of gas emissions from the oven remains within a predefined range.

[0025] While measuring the oxygen amount of the gas can be performed with data from gas sensors, the amount of gas emissions would have to be measured indirectly. An auxiliary neural network that processes camara images is presented as a data provider for emission data (emission classifier network).

[0026] Emission data and oxygen data can be used to retrain the neural network, in reinforcement learning scenarios.

[0027] A computer-implemented method is presented to identify a parameter value sequence for a time-variant gas-flow parameter to be applied to a pressure-control device in a gas collecting main. The main belongs to an offtake piping system of at least one stamp-charging coke-oven that is adapted to process a coal-cake.

[0028] In a receiving step, the computer receives a set of input parameters, with at least a modality parameter set that represents a charging modality by that the coal-cake is to be inserted into the stamp-charging coke-oven, and receives a coal-cake propertyparameter set that represents a property of the coal-cake before insertion.

[0029] In a data-processing step, the computer processes the received set of input parameters by a neural network that is a classifier and that has been trained in advance with training data. During training, the input of the neural network has received a plurality of historical data-sets of input parameters from reference operations of a reference stamp-charging coke-oven; and the output of the neural network has received a plurality of historical sequences of time-variant gas-flow parameters as the ground-truth.

[0030] By data-processing the computer provides a preliminary value sequence. In a providing step, the computer provides the preliminary value sequence to the pressure-control device of the stamp-charging coke-oven for application during cake insertion.

[0031] Optionally, the modality parameter set comprises one or more of the following parameters: a first modality parameter that represents a speed increase in an acceleration phase; a second modality parameter that represents the duration of the acceleration phase; a third modality parameter that represents a speed of insertion; a fourth modality parameter that represents a duration of insertion at regime speed; a fifth modality parameter that represents a speed decrease in a deceleration phase; a sixth modality parameter that represents a speed decrease in a deceleration phase; and a seventh modality parameter that represents the duration of the deceleration phase.

[0032] Optionally, the coal-cake property parameter set comprises one or more of the following parameters: a first property parameter that represents the share of volatile matters in the coal-cake; a second property parameter that represents the humidity in the coal-cake; a third property that represents the density of the coalcake; and a fourth property parameter that represents the height of the coal-cake.

[0033] Optionally, providing the preliminary value sequence is performed conditionally after testing to confirm that the preliminary value sequence has its start values within tolerances given by a control loop.

[0034] Optionally, the computer performs the providing step simultaneously with controlling the pressure-control device by a control loop with controlling the pressure-control device such that the oxygen amount of the gas in downstream direction of the gas collecting main remains below a pre-defined threshold.

[0035] Optionally, the computer performs the providing step simultaneously with controlling the pressure-control device by a control loop with controlling the pressure-control device such that the amount of gas emissions from the oven remains within a pre-defined range.

[0036] Optionally, the time-variant gas-flow parameter for the pressure-control device of the oven is selected from the following: a pressure parameter, a water level parameter, an angular or linear location parameter of a mechanical element, and a parameter to control a plurality of ammonia water jets.

[0037] Optionally, - during training - the neural network has received the plurality of historical sequences of time-variant gas-flow parameters as the ground-truth together with historical emission data so that in the data-processing step, the neural network provides the preliminary value sequence such that it is optimized for minimal emissions.

[0038] Optionally, data-processing the received set of input parameters by the neural network involves the use of a neural network that has also been trained with historical sets of emission data from reference operations of a reference stampcharging coke-oven as a ground-truth for gas emissions, so that data-processing not only provides the preliminary value sequence but also provides an estimation of emission data.

[0039] Optionally, the neural network has been trained either simultaneously with historical data-sets of input parameters and historical sets of emission data, or consecutively with the historical data-sets that are the sequences of time-variant gas-flow parameters as the ground-truth, and subsequently - during re-training - with historical data-sets of emission data as further ground-truth.

[0040] Optionally, the historical data-sets of emission data have been obtained by applyingan emission detecting technique that uses a camera and that provides emission classification by an emission classifier network.

[0041] Optionally, data-processing the received set of input parameters by the neural network involves the use of a neural network that has also been trained with historical sets of oxygen data from reference operations of a reference stampcharging coke-oven as a ground-truth for oxygen data, so that data-processing not only provides the preliminary value sequence but also provides an estimation of oxygen data.

[0042] Optionally, the neural network has been trained simultaneously, with historical data-sets of input parameters and of oxygen data, or has been trained consecutively with historical data-sets that are the sequences of time-variant gas-flow parameters as the ground-truth, and subsequently - during re-training - with historical data-sets of oxygen data as further ground-truth.

[0043] Optionally, the step applying the preliminary value sequence to the pressurecontrol device of the stamp-charging coke-oven during cake insertion is followed by: collecting feedback with emission data that represents emissions from the stamp-charging coke-oven, and / or with oxygen data from a sensor in the gas collecting main, and rewarding the neural network in a reinforced learning approach by iteratively re-defining and re-adjusting the weights of the network.

[0044] Optionally, collecting feedback with emission data applies an emission detecting technique that uses a camera and that provides emission classification by an emission classifier network.

[0045] The disclosure further relates to the use of the computer-implemented method that is being applied to a plurality of coke-ovens, wherein the computer is located remotely to the controller of the coke-ovens. This approach allows offering the method in application-as-a-service or software-as-a-service settings.

[0046] A computer system is presented to identify a parameter value sequence for a timevariant gas-flow parameter to be applied to a pressure-control device in a gas collecting main of an offtake piping system of at least one stamp-charging coke-oven that is adapted to process a coal-cake, the computer system being adapted by comprising modules to perform the computer-implemented method.

[0047] A computer program (or a computer program product) is also disclosed. The computer program - when loaded into a memory of a computer and being executed by at least one processor of the computer - causes the computer to perform the steps of the computer-implemented method.

[0048] Further disclosed is a computer-implemented method for training a neural network to enable the neural network to identify a parameter value sequence for a timevariant gas-flow parameter to be applied to a pressure-control device in a gas collecting main of an offtake piping system of at least one stamp-charging coke- oven that is adapted to process a coal-cake. At the input of the neural network under training, the method-executing computer is applying a plurality of historical data-sets of input parameters from reference operations of a reference stampcharging coke-oven, the input parameters with at least a modality parameter set that represents a charging modality by that coal-cake had been inserted during the reference operations, and a coal-cake property parameter set that represents a property of the coal-cake before insertion. At the output of the neural network under training, the method-executing computer is applying a plurality of historical sequences of time-variant gas-flow parameters as the ground-truth.Brief Description of the Drawings

[0049] FIG. 1 illustrates an overview to a stamp-charging coke-oven and to a computer, in an operating phase;

[0050] FIG. 2 illustrates a neural network that has been trained;

[0051] FIG. 3 illustrates a modality parameter set, and illustrates a coal-cake property parameter set;

[0052] FIG. 4 illustrates a flow-chart overview, from left to right with, a collecting phase, a training phase with the neural network under training, and the operation phase;

[0053] FIG. 5 illustrates re-training as reinforced learning;

[0054] FIG. 7 illustrates a single oven, a closure of that oven, and visible gas emissions at a leakage-area, and also introduces emission degrees;

[0055] FIG. 8 illustrates a training phase with a neural network being trained;

[0056] FIG. 9 illustrates two control loops applied to a single oven, with controllers as well as a camera and a pre-trained network;

[0057] FIG. 10 illustrates a flow-chart diagram of a computer-implemented method to obtain a pressure set-point for a programmable controller that is associated with an oven;

[0058] FIG. 11 illustrates a flow-chart diagram for an approach in that some steps of the method of FIG. 10 are performed in multiple instances;

[0059] FIG. 12 illustrates images in various situations to describe further approaches to increase the classification accuracy; and

[0060] FIG. 13 illustrates a generic computer.Detailed DescriptionWording

[0061] Usually, the terms "set" and "plurality" are used as synonyms, but the description here writes "data-set" for data collections that can be processed together at the same time (such as, for example, data-sets at the input or at the output of a neural network). The term "plurality" is used for multiple data-sets that are processed serially (such as the training data during training a network). For example, FIG. 4 shows data-sets 301-m and 501-m (with m = 1 to M) and shows pluralities 301 and 501 (each with M data-sets). Data-sets can also be noted by { }. Data-sets can stand for sequences for values that change with the progress of time (e.g., {P} in FIG. 1), and data-set can be collections of parameters. The description occasionally writes "set" instead of "data-set" (e.g. in "parameter set" or the like).

[0062] The term "plurality" is occasionally used for multiple things, such as "plurality of water jets", "plurality of ovens in a battery" and so on.

[0063] The description occasionally simplifies the wording. Receiving, processing, providing, etc. a data-set comprises that the computer receives, processes etc. numerical values for the elements in the data-set.Overview, first scenario

[0064] Referring to FIGS. 1-5, the description explains computer-implemented method 403 to identify a parameter value sequence {P} for a time-variant gas-flow parameter to be applied to gas-flow control device 163. Device 163 belongs to the gas -collecting main of (at least one) stamp-charging coke-oven 103 that is adapted to process coal-cake 113. The parameter value sequence {P} is applicable during cake insertion (starting with opening the door and ending with closing the door).

[0065] Device 163 allows changing the gas-flow over time. Gas-flows are usually associated with the physical phenomena "pressure". The shares of particular chemical elements in the gas remain substantially constant (at least during insertion). To enhance readability, the description therefore uses the attribute "pressure-control".

[0066] In terms of controlling industrial equipment (such as the oven, the device in the gascollecting main, etc.), pressure-control device 163 can be regarded as an actuator, and the parameter value sequence {P} would be the sequence of an actuator parameter.

[0067] FIGS. 1-5 thereby show a first scenario, in that the parameter value sequence {P} depends• on modality parameters that represent how the coal-cake is to be inserted into the coke-oven (i.e., also actuator parameters, but pre-defined), and• on property parameters that represent properties of coal-cake 113 that are known before inserting.

[0068] Method-executing computer 203 runs pre-trained neural network 253 to process data. Description and drawings differentiate data-set 303 / 301-m at the INPUT and data-set 503 / 501-m the OUTPUT of network 253 (cf. FIG. 1, FIG. 4).

[0069] In the first scenario, modality parameters and property parameters are received at the INPUT of the network and the parameter value sequence {P} is provided at theoutput of the network.

[0070] Further scenarios may use different or additional INPUT / OUTPUT relations.Training options

[0071] For the first scenario (and for any further scenario) the neural network has to be trained. Regarding the training for the neural network, there are in principle two training options:• The first training option is characterized by the neural network being trained with training data that comprise historical data. In other words, the existence of previously collected data is required. The first training option is assumed for FIGS. 1-5, and the trained network (i.e. initial training completed) is given by reference 253. In implementations of this first training option, obtaining the training data would usually imply some supervision (e.g., to follow the well-known supervised learning paradigm). For example, a human expert would split historical data (among them values sequences) into (positive) data associated with standard-compliant coke (that the oven produces) and (negative) data associated with failures (e.g., the oven producing non-suitable coke, the observation of threshold exceeding emissions, the detection of safety hazards or other failures).• The second training option is characterized by the neural network being trained by experiment-based training, following the known reinforced learning paradigm. Historical data is not yet available and can therefore not be used. The second training option will be explained below with FIG. 6. The network continues to operate: in the first scenario, the network provides parameter value sequence {P} and that sequence is actually be applied to the oven. Feedback from the oven is used in subsequent operations (i.e., with further coal-cakes) and over time, the network is being trained. The reference for the network will be 254.

[0072] It is expected that data is being accumulated over time (i.e., for multiple insertion operations) so that training can continue with that data being historical data. In other words, both training options can be combined. For example, the operatorscan apply an initial value sequence that can be based on operator experience. Such initial value sequence can be a test sequence, or tentative sequence (in a trial-and- error approach). Over time, such initial values sequence could be perfectionated. Values sequences may turn into historical data (first training option) and value sequences may trigger feedback (second training option).Data granularity in sets

[0073] The following applies for all scenarios and for both training options. As data are usually available in data-sets (with multiple values, not necessarily single values), description and drawings refer to data-sets in { } notation. The type of data is noted by a letter inside the { }.

[0074] As used herein, { } can stand for data-sets with elements with different types, but can also stand for elements of the same type that may vary over time.

[0075] In general, the notation {X} symbolizes data-sets with parameters (parameter sets) that are related to inserting the coal-cake into the coke-oven. The data-sets can be differentiated in modality parameter set {X0} and property parameter set {XI}.

[0076] Modality parameter set {X0} comprises one or more parameters that represent a modality by that the coal-cake 113 is to be inserted into stamp-charging coke-oven 103. Modality parameters can represent, for example, the time it takes to insert the cake, the speed by that the cake is being inserted, and so on. Details are explained with examples in FIG. 3.

[0077] Property parameter set {XI} comprises one or more coal-cake property parameters that represents one or more properties of the coal-cake before insertion. Properties are known before insertion. Property parameters can represent, for example, the humidity of the cake, its external dimensions and so on. Details are explained with FIG. 3 as well.

[0078] In relation to the coal-cake, it is possible to consider modality parameter set {X0} as a set with extrinsic parameters, and to consider property parameter set {XI} as a set with intrinsic parameters.

[0079] Parameter value sequence {P} symbolizes a data-set with the value sequence {Pl,P2, ...} for the gas-flow parameter, the parameter changes over time.

[0080] Emission data-set {E} symbolizes a data-set that represents gas emissions that occur during inserting or after inserting. It is noted that emission data is available for other times as well.

[0081] Gas data-set {02} symbolizes a data-set that represents the share of a particular chemical element in the oven-gas, such as the share of oxygen. Data that represents other gases can be processed as well, but for simplicity of explanation, the description writes 02.

[0082] It is also possible to consider the notations to stand for a vector or even for a matrix. Individual set elements can be identified by "dot notation" (for vectors). For example, modality parameter set {XO} can have elements X0.1 and X0.2 that are particular modality parameters. Likewise, property parameter set {XI} can have XI.1, XI.2 and so on that are particular property parameters.

[0083] Data-processing by network 253 can be summarized in an arrow notation that is inspired by chemistry. The left side shows data at a network INPUT, and the right side gives data at the network OUTPUT: data-set 303 -> data-set 503.

[0084] In other words, one or more data-sets at the INPUT are processed to one or more data-sets at the OUTPUT.

[0085] For example, the first scenario can be summarized by the following function: {X} -> {P} (generally, writing {X} for both {XO}, {XI}

[0086] The function can be seen as mapping between sets { }. Mathematical formulas between set elements (such as parameters and / or time sequences) are not necessarily available. But the pre-trained neural network steps in.

[0087] In that first scenario, neural network 253 (i.e., the pre-trained network) receives parameter sets {XO} and {XI} at the network INPUT and provides the value sequence {P} at the network OUTPUT: 303 = {X} | 3 -> 503 = {P} | 3.

[0088] In that first scenario, the network has the function to identify {P}. It is therefore occasionally called the "value-sequence network".Further scenarios

[0089] Further scenarios with other data configurations are to be explained below, so that FIGS. 1-5 remain focused on the first scenario.

[0090] It is possible to provide multiple data-sets to the INPUT and to obtain multiple datasets at the OUTPUT.Phases

[0091] In order to explain how network 253 is being trained and how it operates, it is convenient to differentiate phases.

[0092] For the first training option, the phases are noted as **1, **2, and **3 in the references, and noted { } 11, { } 13 for the data-sets, { } 12 would stand for intermediate data that occurs during training, but that is not relevant here.• Collecting phase **1 stands for the time to obtain historical data { } 11 to train the network (cf. FIG. 4, item 401).• Training phase **2 stands for the time in that the network is being trained with a training set (that comprises the historical data { } 11, cf. FIG. 4, items 402).• Operating phase **3 stands for the time to execute method 403 (FIG. 4). Phase **3 also stands also for the time to operate the stamp-charging coke- oven at least when the oven-processes starts.

[0093] For the second training option, trial phase **4 stands for the time to execute method 404 (FIG. 6), and data-sets are identified as { }|4. Method 404 comprises to operate the coke-oven with an initial sequence {P} 14.0, to monitor the operation (and to track {X} 14.0). The initial sequence {P} 14.0 can be obtained from an operator (as mentioned, based on operator experience), or from a network that has been trained with a relatively low quantity of training data, etc. Method 404 comprise to repeat the operation (with modified sequences {P} 14.1, {P} 14.2, {P} 14.3, and so on.

[0094] During the repetitions in method 404, data turns into historical data so that the first training option (with phases **1 and **2) can be applied.Phases in combination

[0095] In other words, both training options can be combined in a cycle: Starting with method 404 (trial phase **4) allows to acquire training data. In that sense, trial phase **4 can serve as collecting phase **1.Notation

[0096] The first scenario (first training option) can be noted as:{X}| 1 ^ {P} | 1Training with {X} | 1 at network input and with {P} as ground-truth at the network output.{X}| 3 {P} | 3 during operation, from input to output, the network identifies the value sequence {P}.

[0097] The first scenario in the second training option can be noted likewise but with a restriction:{X} 14.i (- ) {P} | 4.i (index i for repetitions) in that the arrow is placed in parenthesis because the network is not yet working. The computer keeps track but does not yet operate the network.Responses to {E} and {02} with two approaches

[0098] This notation with a single input and a single output of the neural network is simplified because emission data {E} and oxygen data {02} have influence as well. In other words, the arrow -> can also lead to {E} and can lead to {02}.

[0099] There are - in principle - two approaches:

[0100] In the first approach, {E} or {02} (alone or in combination) are processed at the INPUT or the result at the OUTPUT. This leads to further scenarios.

[0101] In the second approach, the OUTPUT (of network 253 of FIG. 1, or network 254 of FIG. 6) is considered as candidate, and a further processing unit (that is not the same at the neural network), in FIG. 1 as shown as test unit 283 checks if theOUTPUT can be used or not. The second approach is best explained for the first scenario {X} -> {P}. The sequence {P} may be associated with "non-allowable" emission data {E} or oxygen data {02}. The test may fail.Data origin

[0102] The origin of data is applicable to all scenarios (to both training options, and to both {E} and {02} approaches). For {X} and {P}, details will be explained with FIG. 1.

[0103] For data of type {E} - emission data - the skilled person is able to detect and to classify emissions to the environment. For example, human operators can regularly look for emissions, by visually inspecting the ovens and their closures. The operators estimate the intensity of emissions, the duration of the emissions, and other phenomena. However, it is also possible to apply an emission detecting technique that uses a camera and that provides emission classification by an emission classifier network (ECN). FIGS. 7-12 are provided to introduce that ECN. For the emission classifier network, phase for collecting, training and operation would be differentiated as well. Using the ECN is convenient because it makes data collecting easier.

[0104] For data of type {02} - the presence of particular chemical elements in the gas such as oxygen - the skilled person is able to measure characteristics of oven gases, and can - for example - use oxygen sensors to measure the share of oxygen.

[0105] In other words, automatically obtaining data {E} and {02} via sensors allows the data to be used during training **2, and / or during trial **4.

[0106] It is noted that in some scenarios, data {E} and {02} can be estimated by the neural network (appropriately trained), and that the notation would be {E}| 3 and {02} 13.

[0107] FIG. 1 illustrates an overview to stamp-charging coke-oven 103 ("coke-oven") and to computer 203, in operating phase **3. Coke-oven 103 is adapted to process coalcake 113 and the figure illustrates coal-cake 113 just before it is being inserted into coke-oven 103.

[0108] FIG. 1 illustrates oven 103 as the particular oven in that the cake is being inserted, and illustrates oven 103' as one of many other ovens in a battery, with closed doors.Of course, the battery does not have only two ovens, but many more. Eventually, a cake will be inserted to oven 103' and operation phase **3 will be repeated.

[0109] One or more human operators 193 supervise the operation of coke-oven 103.Operators 193 can decide when and how to insert coal-cake 113. Computer 203 assists them. For example, in the first scenario, computer 203 identifies data-set 503 that is the parameter value sequence {P} 13 for the time-variant gas-flow parameter. Once identified (and checked for compliance with safety rule or the like), computer 203 could also control pressure-control device 163 directly.

[0110] The same principle applies for the other ovens in the battery as well.Input- and output parameters for the first scenario

[0111] Computer 203 (with value-sequence network 253) receives the values of the input parameter sets {X0}, {XI} and provides the gas-flow parameter as output parameter (i.e., value sequence {P}).

[0112] As mentioned, parameters set {X} = modality parameter set {X0} and property parameter set {XI} are parameters sets that are related to inserting. Collectively, the input parameters form data-set 303 of input parameters to network 253. For each parameter set, the values can be single values, or can be vector values.

[0113] The description will discuss further details for the parameter sets in connection with FIG. 3.

[0114] Regarding the parameter value sequence {P}, the description takes a short excurse to the oven hardware in view of FIG. 1: Coke-ovens 103 and 103' are coupled to gascollecting main (GCM) 183 (of the offtake piping system) that has pressure-control device 163.

[0115] Coke-oven 103 can belong to a battery of coke-ovens (103, 103' and other that are omitted here from illustration), having individual pipes 143, 143' that lead to GCM 183 (of the offtake piping system that is common to all ovens). FIG. 1 is simplified, individual pipes 143, 143' can have individual pressure valves, but they are not illustrated.

[0116] Taking the connections of individual pipes 143, 143' to GCM 183 (i.e., collectively (common) offtake piping system) as reference, the skilled person would differentiate individual "upstream" pipes and a common "down-stream" pipe (i.e., GCM 183).

[0117] In such a convention, 02 sensor 133 (to obtain {02}) is located "down-stream", and oxygen data {02} stand for the share of chemical elements in GCM 183.

[0118] Pressure-control device 163 (in GCM 183) is illustrated only symbolically. Any system that allows controlling the pressure (i.e., the gas-flow in general) would correspond to pressure-control device 163. The skilled person can select the system that fits best.

[0119] Pressure-control device 163 can operate to introduce negative pressure (to have a suction effect) to remove gases from GCM 183 (and from pipes 143, 143'). For example, device 163 can be implemented with a fan. Much simplified, a relatively fast rotating fan causes more suction that a relatively slowly rotating fan. Other implementations comprise devices in that water flow cause negative gas pressure (i.e., using the well-known Bernoulli and / or Venturi effects). Devices are commercially available, for example, from Paul Wurth S.A. as well.

[0120] Identifying the parameter value sequence (to be applied to pressure-control device 163) is also compatible with different implementations of batteries.

[0121] Some batteries may have a single oven pressure control installed, so that the inoven pressure of the individual ovens can be controlled individually (commercial availability from Paul Wurth S.A. as well, under the trademark SOPRECO, cf. also in EP 2 160449 Bl). While one oven is being charged with the coal-cake, the gascollecting main pressure (GCM pressure) is controlled to be negative (i.e., oven-to- pipe gradient as mentioned above, modulated by the value sequence) through pressure-control device 163. The other ovens are the ovens under distillation, but they are not affected by such negative pressure, as their individual pressure is controlled.

[0122] Some more traditional batteries may not have such as single oven pressure controlinstalled. Pressure control device 163 (that is used during charging) can be implemented by one or more high pressure ammonia waterjets. The jets are injected to exploit the Venturi effect and the jets create a local suction in the particular gas-collecting pipe that corresponds to the oven being charged.

[0123] The time-variant gas-flow parameter {P} for pressure-control device 163 can therefore be a parameter selected from the following:• a pressure parameter (i.e., its numerical value corresponds to a particular pressure),• a water level parameter (i.e., to set water to an amount that correspond to the desired gas-flow),• an angular or linear location parameter of a mechanical element (e.g., the position of a handle or crank),• a plurality of pressure ammonia water jets.

[0124] The parameters are merely given as examples.

[0125] Although the description just writes P, it is contemplated to have two (or even more) sub-parameters. The reader is familiar with car driving: to set the speed, the driver operates multiple controls - gear stick and pedals. The gear number would be the gear (sub)-parameter and the pedal position would be the pedal (sub-) parameter.Value sequence

[0126] Simplified, depending on the parameter value sequence {P}, computer 203 controls device 163 with appropriate set-points for the gas-flow (of the coke-oven gases) through pressure-control device 163. Computer 253 thereby uses pre-trained neural network 253.

[0127] More in detail, the set-points for the gas-flow can be given as value sequence that can be written as a time-series {P} = {Pl, P2, P3, ..., Pn,..., PK}. The values Pk can be applicable for equidistant time-slots At, but that is not required. The overall duration At*K approximately corresponds to the time it takes to insert coal-cake 113 into coke-oven 103 (i.e., opening the door, moving the cake, closing the door,etc. TJNSERT). The duration of time-slots At can be regarded as an accuracy indicator.

[0128] The letter "P" does not necessarily stand for a "pressure", but more in general for time-variant parameters that represent the gas-flow (e.g., pressure, mechanical position of device components, control instructions to the valve to lead to particular gas-flow etc.).

[0129] In the above-mentioned implementation with the fan, the letter P can stand for the rotation frequency (rotation per minute, rpm).

[0130] The time-series {P} can have K values. The number K can be selected according to the physics by that device 163 reacts to a new set-points T_REACT and according to the overall duration to insert the cake (TJNSERT, cf. FIG. 3). For example, it can take about five seconds (i.e., (T_REACT ) from receiving Pk and actually establishing the pressure according to that. TJNSERT could be approximately 100 seconds. K would be an integer that is estimated to be less than TJNSERT / T_REACT (e.g., K = 10 that is less than 100 / 5).

[0131] The skilled person knows that the control loops or the like have time constants and that changing set-points must not be too fast.

[0132] It is also possible to indicate the sequence otherwise, e.g., a first value for a first duration (in seconds), a second value for a second duration (also in second, but not necessarily as long as the first duration), etc. The notation as a time-series is convenient for illustration. If pressure device 163 is implemented with valve control circuitry, the circuitry would receive control instructions that could be expressed otherwise.

[0133] In principle, it is also possible to let operators 193 change the pressure set-point and to measure {P} over time. Such manual operation of device 163 (with appropriate data tracking) is related to the above-mentioned training phase **1 (and also to the trial phase **4, if used)

[0134] FIG. 1 illustrates {P} only symbolically with 3 values Pl, P2 and P3 in K = 3 time-slots At, for {P} being provided by computer 203 as sequence 503.

[0135] Computer 203 receives (cf. step 413 in FIG. 4) set 303 (of input parameters sets{XO}, {XI}) so that neural network 252 processes set 303. Data-processing (step 423) by neural network 253 provides a value sequence 503. Computer 203 that applies preliminary value sequence 503 to control pressure-control device 263 (of coke- oven 103).

[0136] It is noted that the description reduces the parameter value sequence {P} to have values of the same type. (In terms of the driving analogy, this would be similar to telling the driver a particular target speed but to let him or her select the gear etc. individually.)Emissions

[0137] There is an overall goal to keep gas emissions to the environment low. This can be accomplished, for example by the following:• In further scenarios, emission data {E} can be considered as data for the network (network 253) to process, during the training phase **2 and during the operation phase **3.• Optionally, emission data {E} can be determined can be while the coal-cake is being inserted. Eventually, the computer can overrule the value sequence {P} from the neural network if the emissions turn out to excess pre-defined thresholds. The computer could switch to an alternative value sequence for that emissions are expected within allowable ranges. In that sense, emission data {E} is measured as a control input for a control loop (that does not have to be implemented by a network).• Optionally, the neural network can provide two or more values sequences, and the computer can select the sequence for that the emissions are expected to be lowest.• It is also possible to re-train the network, based on {P} 13 that network 253 provides during operation. For reinforced learning, positive feedback is provided by sequences with relatively low {E}, and negative feedback is provided by sequences with relatives high {E}.• If available, human experts can annotate the training data with negative linksso that {X} and {P} relations with undesired emissions can be marked as "not good" already for the training.

[0138] To summarize this section, emission data {E} can be used (i) during the operation in a control function (i.e., to modify charging of excess emissions), (ii) can be processed by network 253 (in some scenarios), and can even be applied for training and reinforced learning (cf. FIG. 5, step 443). There is - however - a requirement: to obtain emission data {E} with minimal efforts and with minimal delays. As already mentioned, the description explains an approach with FIGS. 7-12.Value sequence from the trained network

[0139] FIG. 2 illustrates - simplified - neural network 253, that is the network after training (i.e. the trained network). FIG. 2 symbolizes network nodes by circles and symbolizes weighted connections by lines between the nodes.

[0140] Network 253 is able to receive parameter set 303 (cf. FIG. 1) at the INPUT and to provide a classification by set 503. z (cf. FIG. 1, item 503) at the OUTPUT. Simplified there are Z possible output sets available, from 503.1 to 503.Z. Output set 503-z stands for the particular output set that matches the input). Optionally, there are two or more output sets.

[0141] Input (as parameter set 302) and output (as sequence 502-z) are shown symbolically for the first scenario with set 302 = {X} = {X0}, {XI} going to different input nodes, and by set 503. z with a particular sequence {P} at one of the output nodes. As already mentioned, the neural network is a classifier and 503. z with the particular sequence {P} corresponds to one of the classes.

[0142] FIG. 2 also introduces a convenient notation: the data-sets (rectangles with "round" corners) 303 and 503. z are shown with the type identifiers, here {X} and {P} for the first scenario. For the further scenarios, the data-sets are of different types.

[0143] The description shows a "many-to-few" classification, with a single output set 503. z Since a technical system such as coke-oven 103 is relatively complex, it can happen that different value sequences could be applied (at least in the first scenario). Provided that suitable training data is available (i.e., the network has been trainedwith alternatives), it is contemplated to optionally let network 253 output two or more data-sets 503-za and 5O3-z|3.

[0144] FIG. 3 illustrates set {X0} and set {XI} with more detail. The following is a discussion of parameters that have an impact on the process.

[0145] As illustrated by a WAY-TIME diagram, the modality parameter set {XO} represents a charging modality by that the coal-cake 113 is to be inserted into the stampcharging coke-oven 103.

[0146] From a high-level perspective, inserting the cake into the oven appears similar to driving a car into a residential garage. But there are differences: the car must not change its shape, the car is supposed to return from the garage, and air flow is negligible.

[0147] But likely more important is the ability to control the insertion and to schedule it with parameters to be applied. (For the car, the driver decides, not a garage operator or the like).

[0148] As the coal-cake is inserted by well-know equipment (i.e., with stamp etc.), the insertion can be controlled by controlling its motor, its hydraulics or pneumatics etc. The description therefore discusses the insertion with a more abstract terms, such as position (WAY) and time (TIME).

[0149] During insertion, the cake would be located along a WAY during a time interval TIME (from opening the door to closing the door, approximately At*K). In an acceleration phase (1), the cake would accelerate. In the sub-sequent regime phase (2), the cake would be moved at substantially constant speed. Finally, in the deceleration phase (3), the cake would reach its final position. The insertion curve is indicated by parameter set {X0}.

[0150] For example, the following modality parameters in set {X0} can be applied.• Modality parameter X0.1 can represent a speed increase (in phase (1)).• Modality parameter X0.2 can represent the duration of phase (1);• Modality parameter X0.3 can represent the speed of insertion, such as, for example, the average speed (WAY / TIME, as illustrated by a dashed line) or thespeed in phase (2), cf. the slope of the curve in that phase (cf. the dashed curve approximating the {XO} curve).• Modality parameter X0.4 can represent the duration of insertion at regime speed (i.e., phase (2) with substantially constant speed).• Modality parameter X0.5 can represent the speed decrease in phase (3).• Modality parameter X0.6 can represent the duration of phase (3).• Modality parameter X0.7 can represent the overall duration of inserting (i.e., all phases together, TIME, corresponding approximately to At*K). For example, it may take TIME = 2 minutes.

[0151] The description presents the list of parameters X0.1 to X0.7 merely as an example. Modality parameters will to be processed (in data-processing 423, cf. FIG. 4) by the neural network, but not all parameters have to be processed.

[0152] For processing by the neural network, the modality parameters can be normalized (e.g., to have value ranges from 0 to 1).Property parameters

[0153] By way of example, the description explains some property parameters.• A first property parameter XI.1 can represent the share of volatile matters in coal-cake 113. XI.1. can be available as mass percentage.• A second property parameter XI.2 can represent the humidity in the coalcake 113. In other words, XI.2 would be the water content of the coal-cake.• A third property parameter XI.3 can represent the density of the coal-cake (i.e., its mass over its volume).• A fourth property parameter XI.4 can represent the height of the coal-cake. FIG. 3 shows that symbolically. The cakes are - of course - dimensioned such that they fit into the oven, but the cake does not have to take the available space completely. Looking through an open door, there can be free space above the cake. The space has a rectangular section. During insertion, gas would be displaced: coke-oven gas would flow towards the gas-collecting main (cf. item 183 in FIG. 1). For a larger space, less gas would flow, for asmaller space, more gas would flow. As the height of the coal-cake is related to that free space, lower cakes or higher cakes have a different impact for the gas to flow.• A fifth property parameter XI.5 can be the coal-cake width (i.e., a further dimension with an impact to gas displacement as well)• A sixth property parameter XI.6 can represent the ash content of the coalcake.

[0154] The indices are just given for convenience, but they do not convey importance.

[0155] The list of property parameters can not be complete, and property parameters can be more complex. For examples, a granulometric curve indicates particle sizes in different shares.

[0156] Other properties can be represented by indicators that combine several property parameters. For example, coal-cake can have so-called blends. A cake of the (fictitious) blend ALPHA has certain guaranteed properties, but a BETA cake has different properties. Technical classifications in property groups can also be applied.

[0157] For processing by the neural network, the property parameters can be normalized as well (e.g., to have value ranges from 0 to 1).Selecting the modality parameters and the property parameters

[0158] The parameters can be selected by the network as well. For example, training can initially be performed with a first selection of parameters, and can be repeated with a second selection of parameters (that neglecting a parameter, or that adds a parameter). Some parameters may not have an impact.Phases and method

[0159] FIG. 4 illustrates a flow-chart overview, from left to right with• the collecting phase **1• the training phase **2 with network 252 under training• the operation phase **3 (cf. FIG. 1) with method 403 (to identify the value sequence, in the first scenario).

[0160] Training 402 (in phase **2) is a method in that the INPUT of network 252 receives a plurality 301 of M historical data-sets 301-1 to 301-M from reference operations (phase **1) and the OUTPUT of network 252 receives a plurality 501 of M data-sets 501-1 to 501-M also from reference operations (as ground-truth).

[0161] The number M stands, for example, for the number of historical insertion operations that had been performed for the coke-oven in the past. It is also possible to consider the coke-ovens in the battery and also possible to use historical data from other coke-ovens.

[0162] There are inherent matches: reference operation (with index m) provides both data-set 301-m and data-set 501-m.

[0163] For the first scenario {X} -> {P} the following applies:• The plurality 301 of M data-sets 301-m (i.e., 301-1 to 301-M) are the data-sets {X} 11 with values of the reference operations (X0, XI), of modality parameter set {X0} coal-cake property parameter set {XI}.• The plurality 501 of M data-sets 501-m (i.e., 501-1 to 501-M) are the historical value sequences {P}| 1 that of the time-variant gas-flow parameters as ground-truth.

[0164] The figure also symbolizes that pluralities 301 and 501 can comprise historical datasets such as {X}| 1, {P}| 1, {E}| 1, {02} 11, alone or in combinations. Details will be explained at the end of the description for different further scenarios.Method in general

[0165] As illustrated on the right side of FIG. 4, computer-implemented method 403 is a method to identify parameter value sequence 503, {P} 13, i.e, {Pl, P2, P3, ...} for a time-variant gas-flow parameter to be applied to pressure-control device 163 in gas-collecting main 183 of an offtake piping system of at least one stamp-charging coke-oven 103 that is adapted to process coal-cake 113.

[0166] In receiving step 413, the computer receives set 303 or {X}| 3 of input parameters, with at least• a modality parameter set {X0} 13 that represents a charging modality by thatcoal-cake 113 is to be inserted into stamp-charging coke-oven 103, and a coal-cake property parameter set {XI} 13 that represents a property of the coal-cake before insertion.

[0167] In step data-processing 423, the computer processes received set 303 of the input parameters by neural network 253 (cf. FIG. 1) that is a classifier and that has been trained (cf. 402 in the center of the figure) in advance with training data 301- m / 501-m. During training, the input of neural network 252 has received plurality 301 of historical data-sets 301-m of input parameter values from reference operations (phase **1) of a reference stamp-charging coke-oven (not illustrated). Also, during training, the output of neural network 252 has received plurality 501 of historical sequences 501-m or {P} 11 of time-variant gas-flow parameters as the ground-truth.

[0168] In data-processing - step 423 - the computer provides preliminary value sequence 503 (i.e., {P}| 3); and in step providing 443, the computer provides the preliminary value sequence 503 to pressure-control device 163 of the stamp-charging coke- oven 103 for application during cake insertion. Applying the values sequence is not necessarily a task for the method-executing computer, but rather a task for the controller of device 163.Further aspects of the method

[0169] The computer can provide (step 443) preliminary value sequence 503 conditionally after testing 433 to confirm that preliminary sequence 503 has its start values within tolerances given by a control loop. Such optional testing takes into account that neural networks occasionally may output data that is not suitable.

[0170] Providing 443 can be performed simultaneously with controlling pressure-control device 163 by a control loop with controlling the pressure-control device 163 such that the oxygen amount of the gas in downstream direction of the gas -collecting main 183 remains below a pre-defined threshold.

[0171] Providing 443 can be performed simultaneously with controlling pressure-control device 163 by a control loop with controlling pressure-control device 163 such thatthe amount of gas emissions {E} from oven 103 remains within a pre-defined range.

[0172] As explained in connection with pressure-control device 163 already, there are several parameters available. Time-variant gas-flow parameter {P} for pressurecontrol device 163 of oven 103 can therefore be selected from the following: a pressure parameter, a water level parameter, an angular or linear location parameter of a mechanical element, and a parameter to control a plurality of ammonia waterjets.Reinforced learning

[0173] FIG. 5 illustrates re-training as reinforced learning. The figure repeats method 403 from FIG. 4 partially (as method 403'). The last step of method 403' is applying 443.

[0174] As mentioned above, the (trained) network has the function data-set 303 -> data-set 503, {X} 13 — > {P}| 3 (for example, in the first scenario).

[0175] In an updated notation, the function is expanded, data-set 303 -> data-set 503 (- ) data-set 603

[0176] Data-set 603 stands for data that has been obtained during the operation of coke- oven 103 (cf. FIG. 1) with the assumption that data-set 503 (at the output of network 253) has been used to control the oven. The arrow is given in parenthesis because data-set 603 is not necessarily obtained from the neural network.

[0177] Data-set 603 can be used to re-train the network (in a repetition of phase **2, here as 402')

[0178] To stay with the first scenario, emission data {E} 13 can be collected during operation and can be used as feedback for re-training.{X}| 3 -> {P}| 3 (— >) {E}| 3

[0179] In other words, network 253 has identified (during multiple insertions) value sequences, and the emission sensor (that with the camera, see below, FIGS. 5-12) has provided {E} 13. Some of the value sequences may have resulted in relatively high emissions {E} 13. Re-training would create a "penalty" for the network topropose such {P} again (at least for the corresponding {X}.

[0180] It is also possible to apply positive feedback.

[0181] More in general, collected feedback (e.g., {E} and / or {02}) can be used to calculate a so-called "reward". The figure illustrates that some {X} and {P} combinations are rewarded (bold arrow for rewarded combinations).

[0182] In terms of the method, step providing 443 preliminary value sequence 503 to pressure-control device 163 (of stamp-charging coke-oven 103 during cake insertion) is followed by: collecting 453 feedback with emission data {E} that represents emissions from the stamp-charging coke-oven 103, and / or with oxygen data {02} from a sensor 133 in the gas-collecting main 183, and rewarding 463 neural network 253 during reinforced learning (i.e., by iteratively re-defining and readjusting the weights of neural network 253).

[0183] Collecting 453 feedback with emission data {E} can apply an emission detecting technique that uses a camera and that provides emission classification by an emission classifier network. Emission-data and {E} oxygen data {02} is written here without 13 because that date arrives from measurement equipment (e.g., the camera approach with FIGS. 7-12, and from sensor 133.Negative (initial) training

[0184] As explained, the first scenario is based on historical data-sets, and there is a link (index m) from {X} to {P}. This link is a "positive link" in the sense that particular insertions (described by historical {X}) had been responded by particular value sequences {P}. A trained network would propose a particular {P}. It may happen that historical operations (of the oven) occurred with experiences that should not be repeated. For example, a particular combination of {X}| 1 and {P} | 1 may have caused undesired emissions {E} or undesired {02} measurements (or even failures).

[0185] It is possible to semi-supervise the training by - at least - annotating the combinations that have failed. The annotation would be an expert annotation (by an operator), in the sense of: for certain {X} never provide certain {P}.

[0186] FIG. 5 symbolizes this "negative" {X}-{P} combination by a "broken" arrow.

[0187] However, the collection phase **1 would have resulted in data for {E}| 1 and for {02} 11. In case that {E}| 1 has been collected by the above-mentioned cameratechniques, human involvement would have been minimal.

[0188] Such further data can be processed in further scenarios.Alternative: second training option

[0189] The second training option is characterized by the neural network being trained by experiment-based training.

[0190] FIG. 6 illustrates an alternative flow-chart for method 404 to operate a computer and the coke-oven gas-collecting main to identify parameter value sequence {P} (cf. item 503) for a time-variant gas-flow parameter to be applied to a pressure-control device 163.

[0191] As explained above, the first training option relies on historical data, but historical data may not yet be available and could therefore not be used.

[0192] Method 404 that is presented in FIG. 6 is therefore a method with two participating entities:• a computer that is similar to computer 203 of FIG. 1. and• the coke-oven that can be oven 103 in FIG. 1.

[0193] Method 404 will be presented for a single oven (i.e. oven 103), but there are step repetitions (such inserting) that can be performed for multiple ovens of the battery (i.e. oven 103', cf. FIG. 1).

[0194] Initially, the neural network is not yet trained (or trained with a relatively low amount of training data).

[0195] In step 1, a first coal-cake is being inserted (e.g., cake 113), with arbitrary parameter sets, modality parameter set {X0}, and property parameter set {XI}. The parameter set can be provided by the operator (cf. operator 190 in FIG. 1), for example, based on experience.

[0196] In step 2, both sets {X0} and {XI} are provided to the INPUT of the neural network (e.g., network 253).

[0197] In step 3, the neural network provides a preliminary parameter value sequence {P}.

[0198] In optional step 4, the computer calculates a reward (cf. FIG. 5, item 463, if sufficient data is available and at least some initial training has been performed).

[0199] In step 5, the preliminary parameter value sequence {P} is being used during cake insertion as a to-be-applied sequence, with modality parameter set {X0}, and property parameter set {XI} of step 1. (For simplicity, the description leaves out the suitability tests, cf. item 433 in FIG. 4). Inserting the cake provides results, such as data that indicates failures (e.g., excess emissions {E} that are not tolerable, or gas parameters {02} that are not tolerable), or indicates success.

[0200] In step 6, the network is being trained. The drawing is much simplified here, steps 1 and 5 would have to be performed multiple times, to obtain training data. For example, certain failure combinations ({XO}, {XI}, {E}, {02} etc.) or success combinations can be used to train the network. In other words, the network will take the combinations as (negative or positive) training data. In that sense, inserting the cake and operating the pressure-control device provide historical data. The skilled person knows that training and re-training can involve updating the weights of the network.

[0201] It is noted that step 6 would be performed for multiple ovens of the battery, serially, not at the same time.

[0202] In step 7, the network calculates a further preliminary parameter value sequence {P}, similar as in step 3, but the neural network that has been trained (at least initially).

[0203] In optional step 8, the computer would calculate (i.e., predict) further rewards.

[0204] As illustrated by dashed boxes 9 and 10, the steps would be repeated.

[0205] In other words, there is learning by doing, with accepting mistakes (i.e., failures).Training in parallel to operating the oven

[0206] Although FIG. 6 has been described as a trial-and-error approach, it is possible to run the computer in parallel with the operation of the oven. Some steps 7 can beperformed otherwise, for example, not be using the result from step 6 ({P} from the network), but just starting as step 1.Emission classifier network

[0207] The description now explains the use of an emission classifier network (ECN in short) that processes camera images to obtain emission data {E}. Simplified, a camera is located external to the oven, and the camera obtains a leakage-area image that shows an area of the external surface of the oven where gas emissions can be present (the leakage-area). The ECN processes the leakage-area images and provides emission data {E}.

[0208] In principle it does not matter if the oven is a coke-oven or not, the ECN approach is suitable for ovens in general. The description now mainly uses references with 4- digits, but correspondences to FIGS. 1-6 are easy to see: oven 1100 corresponds can correspond to oven 103 (of FIG. 1).

[0209] FIG. 7 illustrates oven 1100, closure 1110, visible emissions 1120 (i.e., camera- visible emissions) as well as introduces emission degrees d(t). The figure also shows location coordinates Z and X. Coordinate X is conveniently further divided into "left", "mid", and "right", but that simplified scale is just convenient for explanation.

[0210] On its left side, FIG. 7 illustrates oven 1100 (such as, for example, a coke oven 103 that belongs to a battery 103 / 103', cf. FIG. 1), with closure 1110. In the example, closure 1110 is a door in the XZ-plane (i.e., at the extraction side or at the push side of an oven battery). But it does not matter where closure 1110 is located at oven 1100, during operation it should remain closed.

[0211] However, there are camera-visible emissions 1120, such as the emission of gas from oven 1100. The figure symbolizes the emissions by bold dashed lines. Emissions may occur in leakage-areas 1130 (or better: occur from leakage-areas). In the example, leakage-area 1130 is shown to include the door and to include the part at the top of oven 1100.

[0212] The term "leakage-area" stands for any area on the external surface of the oven where camera-visible gas emissions can be present. The following furtherdifferentiation of the leakage-area is useful to make:• The external surface of the oven where emissions can leak, such as at the perimeter of a closure, can be labeled "emission origin" area.• The external surface of the oven where emissions - after having been leaked - would still be visible, can be labeled "emission distribution" area.

[0213] For convenience of explanation, it can be assumed that - simplified - the "emission origin" and the "emission distribution" areas are disjunct. Computer-vision techniques could differentiate them (if properly trained, cf. FIG. 12 for a discussion), and the inherent property of the emissions to move away from their origin (emissions are fugitive) can be used to determine to have emissions.

[0214] As leakages are rather exceptions to be prevented, leakage-areas could be attributed to be "potential leakage-areas". But for simplicity of explanation, the description uses the label "leakage-areas" only.

[0215] In the example, emissions 1120 should occur in the right part (X-coordinate with letter R) of leakage-area 1130 near the door. (The figure is simplified in not further differentiating origin and distribution areas). Simplified, it can be assumed that emissions 1120 are camera-visible where they occur.

[0216] On its center part, FIG. 7 illustrates leakage-area image 1230, showing the door (at least partially) and showing emissions 1230. The reference changes from 1120 to 1230 because the emissions are on the image. The figure is simplified here in that images capturing is showed for ideal situations, substantially all emissions would be camera-visible in the image. The description will discuss accuracy (for non-ideal situations) below with FIG. 11 (and partially with FIG. 12).

[0217] On its right side, FIG. 7 illustrates a collection of images 1232-1, 1232-2, 1232-m, 1232-M. The images had been taken during a preparation phase (as historical data) and had been annotated during that preparation phase. Annotations 1222-1, 1222- 2, 1222-m and 1222-M are human-made annotations, here given in the notation R, L, M (corresponding to the X-coordinates), and "absence of emissions" or "presence of emissions". The # symbols further differentiate presences as "# presence as low emissions", "# # presence as medium emissions", and "# # # presence as highemissions".

[0218] Annotating the images by two main categories (absence, presences) and by subcategories (low, medium, high for presence) is just convenient for illustrations. The number of categories correspond to the number of categories the ECN in operation (ECN 1253 in FIG. 9) will be able to differentiate.

[0219] The collection of leakage-area images 1232-m with annotations 1222-m serve as training data.

[0220] Inspection procedures are usually standardized. For example, the United States Environmental Protection Agency (EPA) has defined an air pollution test method to determine the visible emissions (VE) from coke ovens ("Method 303 - By-product Coke Oven Batteries"). For Europe, a convenient reference is provided by the Umweltbundesamt (German Federal Environment Agency) as the "Merkblatt uber die besten verfugbaren Techniken in der Eisen- und Stahlerzeugung nach der Industrie-Emissionen-Richtlinie 2010 / 75 / EU, Marz 2012 [Directive 2010 / 75 / EU of the European Parliament and of the Council of 24 November 2010 on industrial emissions]". The document refers to method 303 as well. Ghosh at al. mention further standards.

[0221] It is noted that the result of such inspections can be used as annotations. The annotating expert may not have to look at images, but images to be annotated should be captured by cameras during inspection. The skilled person can apply data-binding techniques to relate particular annotations to particular images. For example, the expert inspects a particular oven, enters a particular emission degree (and other data) into a database, and makes sure that a camera takes an image. It is however not required that the expert looks at this image again.

[0222] It is further noted that image 1232-m and annotations 1222-m do not have to be made for a particular oven for that the ECN operates.

[0223] Within a battery, the ovens look almost identical. Each oven may have individual labels (such as identification numbers), and occasionally, the name of the company that runs (or made) the battery may be written on surfaces as well. Betweenbatteries, the style of such labels or name plates may be different.

[0224] Such relatively minor differences in the external view of an oven (or battery) do not influence the emissions. The annotating expert would ignore such differences.

[0225] In order to avoid overfitting during subsequent training it is possible to harmonize the locations with individual labels (or the like) by image pre-processing. For example, numbers - if visible on an image - could be removed.Training for the emission classifier network (ECN)

[0226] FIG. 8 illustrates the training phase in that ECN 1252 (i.e. a neural network) that is being trained to become emission classifier network (ECN) 1253. The annotations serve as the ground-truth.

[0227] It is in the expertise of the skilled person to validate the network once training has been completed. Details regarding validation are therefore omitted for simplicity.Obtaining further training data

[0228] Every time an emission is detected during operation, such emissions can be classified (by the operators) and can be linked to images taken during operation. Such annotated images 1232-m / 1222-m could be used for follow-up training. Over time, the accuracy of emission detection would rise.

[0229] It can happen that for a particular battery, training data (cf. FIG. 8 image 1232- m / 1222-m) is not yet available in sufficient quantities so that the ECN could classify the degrees. It is contemplated for use an initially trained ECN (training based on historical data from other batteries) and to continuously re-train the ECN (if available with historical data from the oven for that ECN 1253 is being applied).

[0230] Emissions (to be used in the training data) may occur only from time to time. That can be expected for relatively new batteries and / or for ovens with relatively new doors. New door sealings would prevent emissions better than old door sealings.

[0231] From the view of operation this is beneficial. To collect less historical data for training might be less optimal to develop the network (by trainings). In other words, the lack of battery-specific training data prevents the ECN to sharpen its accuracyfor this particular battery.

[0232] But there is simple approach by that training data can be enhanced. For batteries with individual pressure control, pressure can be changed - at least temporarily - until emissions show up (least camera-visible emission). Images would show the emissions. Annotating does not have to rely on the images: the experts do not have to look at images, because the experts already know that emissions occur. The pressure difference that leads to such purpose-created sample emissions are known. Data that describes the pressure difference could serve as an emission degree, but it is noted that individual ovens may show different emission degrees for equal pressure difference.Two control loops

[0233] FIG. 9 illustrates that two control loops applied to a single oven 1103, with controllers 1173 and 1263 as well as camera 1143 and pre-trained ECN 1253. ECN 1253 is illustrated here in a main function support a pressure control loop (on the right sight) and ECN 1253 provides d(t), D(t) that go to controller 1173.

[0234] But that is not required, ECN 1253 can also be used to merely provide emission data {E} (noted without 13 because network 253 is not involved).

[0235] ECN 1253 and auxiliary controller 1263 (if used) can be implemented by a computer. ECN 1253 implements the emission classifier, and controller 1263 can implement functions to control the oven.

[0236] Oven 1103 is illustrated with door 1113 (being an example for a closure) in a sideview looking to the YZ-plane. Oven 1103 is much simplified, by showing a single door only.

[0237] Oven 1103 has pressure sensor 1153 that provides a numerical value that represents the inside pressure inside p(t), at any time point t. Programmable controller 1173 (or "programmable logic controller PLC") receives pressure set-point p_set. Programmable controller 1173 instructs pressure valve 1163 to increase or to decrease the pressure in oven 1103. Arrangements with sensor 1153, programmable controller 1173 and valve 1163 are known in the art.

[0238] For the operation of an ECN with the only function to provide emission data {E}, such sensors and PLCs are not required. But the components can be used in synergy.

[0239] In embodiments, controller 1173 does not receive set point p_set all the time from controller 1263, but only when emissions are detected (i.e., when a trigger has been generated).

[0240] Gas emissions 1123 leave the oven 1103 through leakages (of the closed door or otherwise), and it is expected that emissions 1123 occur mainly in leakage-area 1133 at the external surface of oven 1103. A minor fraction of the emissions may also leave through other areas, but for classifying the degree that is not relevant. Emissions 1123 are camera-visible.

[0241] As camera 1143 captures image 1233 from area 1133, most of the emissions would be shown on the image (at least data would represent emissions). The figure illustrates a single camera, but other cameras could be place to other leakage-areas.

[0242] Image 1233 arrives at pretrained ECN 1253 substantially at the same time when it was taken. Signal propagation delay on its way from camera 1143 can be neglected. Of course, image 1233 arrives without any annotations. Pre-trained ECN 1253 provides degree d(t), with categories or classes. Regarding accuracy it is noted that ideally the degrees would be the same that a human expert would determine. (The description provides an example to achieve higher accuracy, by using vector D(t) instead of d(t), cf. FIG. 11).

[0243] Simplified, auxiliary controller 1263 processes d(t) and provides a new set-pressure point (if needed). Instead of receiving d(t), controller 1263 can receive a trigger, cf. FIG. 10. Optionally, controller 1263 can be implemented by a pre-trained neural network as well. For obtaining emission data {E} alone, adapting a set-point is not required, but the NC network provides emission data {E} as a "side-product" (synergistic effect).

[0244] FIG. 9 is simplified by not illustrating meta-data, and for using the approach with multiple ovens (such as for a battery), the skilled person can use data binding andother techniques to ensure that p_set is applied to the oven from that the image was taken. In other words, data binding ensures to have p_set_n, that images are related to particular oven index n, and so on.

[0245] As explained already, programmable controller 1173 (cf. FIG. 9) receives pressure data p(t) from at least one pressure sensor 1153 and interacts with pressure valve 1163. The pressure is thereby maintained to a pressure set-point p_set. Controller 1173 is active when oven 1103 is under pressure.

[0246] FIG. 10 illustrates a flow-chart diagram of computer-implemented method 1403 to obtain a pressure set-point p_set for programmable controller 1173 that is associated with oven 103 and that controls the gas pressure p inside oven 1103.

[0247] The figure conveniently also illustrates that the NC network is primarily a part to control the pressure, but emission data {E} is a "side-product".

[0248] The figure presents method 1403 within a dashed rectangle. Double-lines indicate that the execution of method 1403 would be repeated periodically.

[0249] Method 1403 is illustrated with a single start at line 1499. Method 1403 is illustrated with two alternative ends. On the right side, method 1403 ends by step 1443 changing the set-point p_set for programmable controller 1173. In many situations, the new set-point would be selected such that the pressure decreases gradually. Having a new set-point may however comprise to keep p_set unchanged.

[0250] For obtaining {E}, steps 1413 and 1423 are relevant.

[0251] On the left side, method 1403 is repeated in case that the detected emission degree is below pre-defined thresholds (or other discriminators) so that step 1443 is not required. Or in other words, method 1403 ends with the follow-up activities (step 1443) if a trigger has been generated, or ends when no trigger has been generated.

[0252] As mentioned, when method 1403 has ended, it will be repeated (from point 1499) again. In view of computation efficiency, there can be a waiting time (from end at 1433 / 1443 to start 1499). The waiting time can be related to T2.

[0253] FIG. 10 illustrates method 1403 for a single oven, but method 1403 can be performed for the multiple ovens in batteries, in serial repetitions (in parallel, or incombinations thereof). Method 1403 can be multiplexed for the ovens of a battery, and the execution time T1 of step sequence 1413 / 1423 to obtain the classification (here: {E} would still be sufficiently short.

[0254] Camera 1143 that is located external to the oven 1103 obtains (step 1413) leakagearea image 1233. Image 1233 shows an area 1130, 1133 of the external surfaces of the oven 1100, 1103 where camera-visible gas emissions 1120, 1123 can be present. As explained, area 1130, 1133 is the leakage-area.

[0255] The computer - with emission classifier by ECN 1253 - processes (step 1423) leakage-area image 1233 by pre-trained ECN network 1253 to classify degree 1223 (or d(t), cf. FIG. 9, also {E}) of emissions 1123 on the leakage-area image 1233.

[0256] As explained with FIG. 8, ECN 1253 has been trained from training data 1232-m, 1222-m that comprise historical images 1232-m that had been taken as reference and that comprise human-made degree annotations 1222-m paired to the historical images 1232-m.

[0257] Step sequence 1413 / 1423 is executed during Tl.

[0258] By applying pre-defined rules in step 1433 and depending on classified degree 1223 (d(t), also {E}) of visible emissions 1123, the computer changes (step 1443) the setpoint (p_set) for programmable controller 1173. As illustrated in FIG. 9, this setpoint changes can be implemented outside the ECN, in controller 1263.

[0259] In the example, the (new) set-point is set such that the pressure p(t) in oven 1103 can gradually decrease.Becoming more accurate

[0260] Having explained method 1403 in the context of the control loops for one or more ovens in FIG. 10, the description now discusses optional approaches to change the pressure set-point with more precision. These optional approaches require to obtain the degree (in step 1423) with higher accuracy. Emission data {E} becomes available.

[0261] The approaches will be explained with FIG. 11 (multiple step instances, ruleadaptations) and FIG. 12 (oven-specific classification).Multiple step instances

[0262] FIG. 11 illustrates a flow-chart diagram for an approach in that steps 1413 and 1423 of method 1403 of FIG. 10 are performed in multiple instances.

[0263] As already explained for method 1403, obtaining the leakage-area image 1233 (cf. FIG. 9) in step 1413 and processing that image in step 1423 leads to a classification by degree d(t), also to emission data {E}.

[0264] But for each execution of both steps 1413 and 1423, there are at least two conflicting constraints, in view of classification accuracy, and computation resource spending.

[0265] Looking at the classification accuracy constraint, the computer would provide classifications that does not match the reality in all situations. Ideally operating computers that correctly classify the images for all situations are not available.

[0266] Taking any binary classification as an example, with P for presence (of emissions) and A (for absence of emissions), in multiple executions of method step 1413 and 1423 the computer would output true positives (output is P in correspondence to presence in reality of the oven), false positives (output is P, but in contrast to absence in reality), true negatives (output A, in correspondence to absence in reality), and false negatives (output A, but presence in reality).

[0267] As explained, the classifications (cf. FIG. 10 at step 1423) may trigger a change of the set-point (cf. trigger 1433)

[0268] Consequently, false classifications might lead to an incorrect change of the setpoint. In a worse case scenario, the oven would operate correctly without emissions, but it might start emitting substances because the pressure is set wrong.

[0269] Looking at the computation constraint, images capturing and image processing can be improved to increase the classification accuracy (e.g., to increase the share of "true" over the share of "false"). However, such improvement would require, for example, more sophisticated cameras (such as in terms of pixel numbers), suitablelight conditions at any time, a relatively high amount of training images (that might not be available, at least not initially), and more computing resources (in terms of CPU, memory consumption etc.).

[0270] The solution to overcome the constraints takes different time intervals into account: performing method step 1413 and 1423 (time interval Tl) requires less time than gradually decreasing the pressure (time interval T2). Flowcharts in such figures are not scaled to time, but nevertheless, method 1403 as illustrated in FIG. 10 shows a time disbalance.Q instances

[0271] The execution of steps 1413 and 1423 can be performed in Q instances. In other words, there can be Q detection cycles (with steps 1413 and 1423). FIG. 11 symbolizes such multiple performances with a counter q and a counter check (for example, counter q = 1 to Q). Performing Q instances serially by repeating the step execution is convenient because the overall method executing time Q*T1 would still be less than T2.

[0272] It would also be possible to perform Q instances in parallel (even with different cameras).

[0273] Executing step 1413 and 1423 in Q instances leads to a degree vector D(t) = (dl(t), d2(t), d3(t), ..., dq(t), ..., dQ(t)) with Q degrees dq(t). FIG. 9 illustrates D(t) next to d(t) as being degree 1223. The degree vector can used as emission data {E}.

[0274] In case of serial execution, there would be Q different time-points, but all time points would be within repetition interval Q*T1. The time-points could therefore be summarized to a single "t" standing for the interval between the first performance (q=l) and the last step performance (q=Q).

[0275] As a result, the degree vector D(t) serves as the basis to determine if the trigger (to change the set-point) has to be applied. Having a degree vector D(t) (with Q elements) instead of a single degree d(t) increases the accuracy. Of course, some of Q elements are "false" classifications (or in other granularities, classification that are incorrect otherwise). Occasionally, for some instances, degree might not evenbe available.

[0276] The evaluation (to set the trigger) is performed by applying trigger rule in step 1433'. The trigger rules in step 1433' are pre-defined. Just to give an illustrative example, for the binary classification, the trigger could be set if the majority of the Q degree values is P (present). For example, the Q = 10 vector D(t) = (P, P, P, A, A, P, P, P, A, P) has more P than A, and the trigger would be set.

[0277] The trigger rules in step 1433' can be learned by a tool that applies machine learning (such as by a neural network or the like). Training would be performed accordingly.

[0278] For example, steps 1413 and 1423 should be repeated for Q = 100, so that vector D(t) would have 100 elements dq(t). T1 should be about 3 seconds (for capturing the image, transmitting the image to the computer, processing etc.), and the overall time Q*T1 would be 3*100 seconds (or 5 minutes). The seconds and minutes are given for illustration only.

[0279] As the reaction time (between changing the set-point of the control loop and having the emission stopped) is assumed to be equal or larger than Q*T1, is does not matter if the set-point is modified a couple of minutes earlier or later.

[0280] For example, the computer could apply a rule to generate the trigger if at least J = 50 (of the Q = 100 vector elements) are "P" (emissions are present).

[0281] This example with thresholds based on the J / K share is much simplified. Other rules (or further rules) are possible as well. For example, if J = 20 consecutive degrees from dk(t) to d(q+19)(t) would be "P", the trigger could be generated as well.Trigger rule adaptation

[0282] Using Q instances still has the constraint that the accuracy of individual degrees d(t) that belong to vector D(t) is not improved. But the trigger rule (in step 1433') has an accuracy as well. As the accuracy also depends on visibility circumstances (i.e., the camera being directed to leakage-area 130), the rule can be made dependent on visibility.

[0283] As used herein, the term "visibility" stands for the quality by that leakage-area image 1230 corresponds to the reality at leakage-area 1130. Visibility can be detected and represented by data. For this embodiment, visibility is the visibility by optical cameras.

[0284] The skilled person can apply sensor or the like to collect data that describe visibility circumstances, for examples according to the following:• At night, the optical camera has to rely on artificial light (light from lanterns, flash light during exposure etc.), during the day, the camera takes the images from natural light.• Day-light conditions vary. Precipitations like rain or snowfall darken the scene (between the camera and the oven). The sky could be cloudy or not.• Rain or snow, dust and the like may hit the objective of the optical camera and the image may be corrupted.

[0285] As illustrated, by step 1463, the computer identifies visibility circumstances. As illustrated by line 1473, the computer uses that visibility circumstances to modify the trigger rule.

[0286] The visibility circumstances modify the conditions by that the trigger rule generates the trigger (to change the set-point).

[0287] The computer can aggregate these and other visibility circumstances to a visibility score. For convenience of explanation, the score should be a real number between 0 and 1, from "no visibility at all" to "best conditions". In case of zero, the method execution would fail because the images would be black.

[0288] For simplicity of explanation, the visibility score - although determined by the computer - can be differentiated into "good visibility" and "poor visibility".

[0289] To stay with the above Q = 100 example, for "good visibility", it would be sufficient to generate the trigger if J = 30 (of the Q = 100 vector elements) are "P" (emissions are present), for "poor visibility", the trigger would have to set if J = 70 (again out of 100) are "P".Further aspects of trigger rule adaptation

[0290] The following discussion for rule adaptation does not have to differentiate the cameras, the discussion applies to optical cameras and to thermographic cameras.

[0291] Ovens are being operated during decades, and it can be expected that some ovens are in operation for more than 40 years, or even more than 60 years. Simplified, older ovens suffer from emissions more than younger ovens. The elapsed operation time can optionally be used as a modifier for the rules.Repetition adaptation

[0292] Adapting the rule can be accompanied by adapting the repetition rate Q (i.e., the number of instances, the instance cardinality), as symbolized by line 1483. For example, poor visibility would demand for more repetitions (relatively high Q), and good visibility would allow less repetitions (relatively low Q).Oven-specific classification

[0293] FIG. 12 illustrates leakage-area images (A), (B) ... (E) in symbolic views that show further aspects that ECN 1253 can optionally apply to obtain the classification {E} (i.e., emission data in the form of classes). Visible emissions show certain behavior (visible for optical cameras), and the neural network could by trained to such behavior. Training network 1252 to become ECN 1253 has been explained with FIG. 8 and additional or alternative features could be trained to the network as well, via annotations, or non-supervised. In other words, the models that are explained next could be implemented by network 1252 / 1253.

[0294] It is contemplated to train a semantic segmentation model to classify areas inside the images. For example, the symbolic illustrations show bold dashed lines for emissions (as in (A), (B)). This leads to image areas showing emissions.

[0295] For example, the symbolic illustrations differentiate the walls or other external elements of the oven, given here by thin vertical lines (as in (A), (B), (C). As the walls do not move, they do not move on the images either. With multiple images in the training set, the network learns where the walls are located, even if some imagesshow the walls covered by emissions.

[0296] Semantic segmentation could also differentiate the leakage-area into emission origin and emission distribution areas, not only by semantic segmentation but also by approaches that are explained next.

[0297] Semantic segmentation can identify directions by that visible emissions are "arranged". View (C) shows emissions in a line that is slightly increasing and shows emissions that are somehow turbulent (rather random direction, caused by weather influence, etc.).

[0298] It is contemplated to process multiple images taken from consecutive points in time. As explained already, there is sufficient time (T1 < T2) available (within T2). As the emissions would usually move, and as the wall etc. do not move, moving elements on the image could be identified, by a properly trained network. In the example, view (D) shows emissions moving, with a temporal aspect (during At) and with a spatial aspect (displacement Ax, Az in pixels, corresponding to locations at the oven, cf. the coordinates in FIG. 7). In other words, speed can be detected.

[0299] It is contemplated to process multiple images taken from consecutive point in time not only in view of speed but also in view of speed changes.

[0300] On the way from the origin, emissions loose density. Density could be estimated letting the network learn to see the walls through the "cloud", as in (A). Some wall structures (e.g., the door itself, the door frame, buckstays or beams) would be "strong candidates" for where emissions could be expected.

[0301] Movements (in terms of speed or speed change) are not equal over the image (cf. the density lost or the like), and it is possible to let the network derive images that look like histograms. On the way from the origin, emissions no only loose density but also loose speed. Such phenomena would be identifiable from derived images as well.

[0302] Method to evaluates images are known in the art, as optical flow detection or the like, with discussions in the following:• Dileep K. Appana, Rashedul Islam, Sheraz A. Khan, Jong-Myon Kim: A video-based smoke detection using smoke flow pattern and spatial-temporal energy analyses for alarm systems (Information Sciences, volumes 418-419, December 2017, Pages 91-101; doi.org / 10.1016 / j.ins.2017.08.001).• Jinkyu Ryu and Dongkurl Kwak: A Study on a Complex Flame and Smoke Detection Method Using Computer Vision Detection and Convolutional Neural Network, 2022 (Fire 2022, 5, 108K; mdpi.com / 2571-6255 / 5 / 4 / 108).

[0303] Performing advanced image processing (such as explained here with FIG. 12) is further enabled by the available time. As explained, the computer has a sufficiently large time interval available (smaller than T2) to potentially perform Q instances during Q*T1.

[0304] As ovens are not necessarily located under roofs, precipitations such as snow could appear on the images as well, cf. symbol (E). However, snow falls from the clouds and would move in top-down direction (here in coordinate z), potentially in some wind-related further directions (x). It is contemplated to train the network to detect such precipitation.

[0305] The detection of such or similar events could potentially lead to the following:• to perform the detection of precipitations as step 1463 "identify conditions" with the consequence to modify the rules (as explained with FIG. 11) and / or to change the number of instances Q,• to stop the interaction with controller 1173 (so that the pressure set-point is not changed) to avoid incorrect operation of the oven,• to inform the operator that changing the set-point according to emissions is temporarily disabled.Further details to obtain {E}

[0306] Classifying images to emission degrees is not the same as - for example - recognizing numbers or letters.

[0307] Historical images 1232-m can be collected under different light conditions: with sunshine, at night, on rainy days and so on. As the sun moves, the images are taken at different times of the day.

[0308] As explained, historical images 1232-m are applied (if annotated) to train NC network 1252 (cf. FIG. 8) to enable the NC network to perform method 1403. The same images - if annotated accordingly - can be used to train network 1272.

[0309] In case that oven elements (such as doors) have different external views, for example, if the doors look differently at the extraction side and at the push side of the battery (cf. FIG. 3), the degree classification can be made more accurate by using two sets of historical images (for each side separately).

[0310] Activities such as installing cameras 1140 and computer, as well as running method 1403 do not substantially interfere with the operation of the oven(s). As explained already, the only additional activity is allowing controller 1173 to receive set-point p_set, and that is not required if the NC network is used to obtain {E} only.

[0311] In comparison to inspections by humans, executing method 1403 is possible whenever the one or more ovens are operative. This helps to monitor emissions continuously. The approach may also reduce the number of inspections by humans, so that the health-related risks for the inspectors are reduced.

[0312] Taking further data into account is convenient: some doors will be recognized to suffer from emissions more than other doors, and the doors can be categorized accordingly. The above-mentioned rules - such as the trigger rule - can be adapted in step 1473 by taking door categories into account.

[0313] In the scenario in that the set-point is changed until the oven emits gases, the magnitude of the pressure difference (before and after the change) in relation to the emission degree (detected by steps 1413, 1423 or otherwise by inspection) is a quality indicator for a particular oven. When the same pressure difference is applied, some doors are more likely to leak than others. Such indicators can be taken as a further input for step 1473 as well.Further scenarios with details

[0314] Having described an option to obtain emission data {E} in connection with FIGS. 7-12, the description now returns to discuss the above-mentioned scenarios for coke- ovens.

[0315] From a high-level perspective, network 253• imitates cause-and-effect relations, or• suggests IF-THEN-relations (for activities).

[0316] This is simplified, a particular insertion with {X} does not cause a particular {P} but there is a particular {P} that fits {X} best (IF-THEN).

[0317] Having explained the first scenario, the description now turns to other scenarios.

[0318] It is possible to combine the data types, for example, in the following:• The set combination {E} | {02} is a concatenation for monitored emissions and for monitored oxygen share, and a network can process such a set combination as if it would be a single set.• The set combination {P} | {E} is a concatenation of the value sequence and the emissions {E}. Processing as a single set is possible as well.

[0319] Such and other set combinations can be used during training **2 and during operation **3.

[0320] The description continues with examples, from the view point of operator 193 (cf. FIG. 1). Again, there is the overall goal to reduce emissions. There is a further goal to mitigate risk (such as the risk of undesired operation or failure due to excess oxygen.First scenario with a further aspect

[0321] Operator 193 intends the insertion of coal-cake 113 with particular parameters {X}, and network 253 would provide one or more corresponding sequence {P}. In the above arrow notation for network INPUT -> network OUTPUT that would be {X} -> {P}-

[0322] There could be more than one {P} at the OUTPUT. For example, it might provide{P}a and {P}|3. Operator 193 may have the choice, but {P}a may lead to more emissions {E} than {P}|3, or vice versa. Operator 193 may not have guidance, but information regarding expected emissions would assist operator 193.Second scenario

[0323] Training the network with a set combination {P} | {E}1 (at the INPUT of network 252 to be trained) would allow to have set combination {P} | {E} 13 at the OUTPUT of network 253 during operation. This is no longer the {P}a and {P}P choice, but operator 193 can prefer the operation {P} | {E} with the lowest {E}.

[0324] As a side-note, {E} does not have to be a single value, such as a category, but could also be an indication of a change: "emissions becoming more severe", "emissions becoming more relaxed", etc.

[0325] In terms of the arrow notation, the scenario is {X} -> {P}, {E}. The arrow is not given in parenthesis, because both {P}, {E} would be obtained from the neural network.

[0326] In terms of training 402 and method 403 in FIG. 4, during training, neural network252 has received plurality 501 of historical sequences 501-m of time-variant gasflow parameters {P} 11 as the ground-truth together with historical emission data {E} 11 so that in step data-processing 423, neural network 253 provides preliminary value sequence 503 (i.e., {P}| 3) that is optimized for minimal emissions.

[0327] Or, data-processing 423 received set 303 of input parameters by neural network253 involves the use of a neural network that has also been trained with historical sets of emission data {E}| 1} from reference operations of a reference stampcharging coke-oven as a ground-truth for emissions. As a consequence, data- processing 423 not only provides the preliminary value sequence 503 (i.e., {P}) but also provides an estimation of emission data {E}| 3}.

[0328] Regarding the training, cf. 402 in FIG. 4, the second scenario offers two alternatives: Neural network 253 has been trained either (i) simultaneously with historical datasets (301-m) of input parameter values and historical sets of emission data, or (ii) consecutively with the historical data-sets 501-m that are the sequences of timevariant gas-flow parameters as the ground-truth {P}| 1, and subsequently - during re-training - with historical data-sets 501-m of emission data {E} 11 as a further ground-truth.

[0329] As explained, emissions can be monitored by cameras (cf. FIGS. 7-12). To executemethod 403, the historical data-sets of emission data ({E} 11}) can have been obtained by applying an emission detecting technique that uses a camera (cf. item 1143 in FIG. 9) and that provides emission classification by an emission classifier network (cf. item 1253 in FIG. 9).Third scenario

[0330] Training the neural network with a set combination {P} | {02}l (at the INPUT of network 252 to be trained) would allow to have set combination {P} | {02} | 3 at the OUTPUT of network 253 during operation. This is no longer the {P}a and {P}|3 choice, but operator 193 can prefer the operation {P} | {02} with the optimal 02 share. In that sense, the third scenario is similar to the second scenario.

[0331] In terms of the arrow notation, the scenario is {X} {P}, {02}.

[0332] In term of method 403 in FIG. 4, data-processing 423 the received set of input parameters by neural network 253 involves the use of a neural network that has also been trained with historical sets of oxygen data ({02} 11} from reference operations of (**1) of a reference stamp-charging coke-oven as a ground-truth for oxygen data. Consequently, data-processing 423 not only provides the preliminary value sequence (503, i.e., {P} 13) but also provides an estimation of oxygen data {02} 13.

[0333] As in the second scenario (with the emissions), the neural network can have been trained with historical sets of input parameter values and of oxygen data (i) simultaneously, or by (ii) by re-training.Fourth scenario

[0334] The second and third scenarios can be combined. Training the network with a set combination {P} | {E}| {O2}1 (at the INPUT of network 252 to be trained) would allow to have set combination {P} | {E} | {02} | 3 at the OUTPUT of network 253 during operation.

[0335] In terms of the arrow notation, the scenario is {X} -> {P}, {E}, {02}.Fifth scenario

[0336] In view of the IF-THEN-relations, identifying the value sequence {P} is potentially not always desired. Operator 193 may have the intention to insert the cake according to {X} and control the device 163 according to {P}. Assuming that the network has been trained with {X}, {P} -> {E}, operator 193 would learn an estimated (or predicted) emission {E}. Potentially, the intended {X}, {P} combinations would be replaced by a combination with less critical {E} values.Sixth scenario

[0337] The scenario is an extension to the fifth scenario, with oxygen data {02} as a further information at the output of the network.Further aspects of the computer system

[0338] As explained in detail above, computer system 203 identifies a parameter value sequence for a time-variant gas-flow parameter to be applied to a pressure-control device in a gas-collecting main of an offtake piping system of at least one stampcharging coke-oven that is adapted to process a coal-cake. Computer system 203 (cf. FIG. 1 simplified as computer 203) is adapted for that purpose by comprising modules to perform the computer-implemented method, and computer (system) 203 is therefore the method-executing computer. The computer-implemented method (cf. FIGS. 4-6) can be applied to a plurality of coke-ovens. However, it is not required that method-executing computer 203 is to be located in proximity to the coke-ovens (e.g., to be located in the same plant or hall) and that the computing functionality would have to be implemented in the oven-controllers.

[0339] It may be advantageous to locate computer 203 remotely to the controller(s) of the coke-ovens, for example, in a data-center among other computers (e.g., so-called cloud computers). This remote approach allows offering the method in application- as-a-service or software-as-a-service settings.Generic computer

[0340] FIG. 13 illustrates an example of a generic computer device which may be used withthe techniques described here. FIG. 13 is a diagram that shows an example of a generic computer device 900 and a generic mobile computer device 950, which may be used with the techniques described here. 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 appropriate computers. Generic computer device may 900 correspond to the computer system 200 of FIG. 3. Computing device 950 is intended to represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smart phones, driving assistance systems or board computers of vehicles and other similar computing devices. For example, computing device 950 may be used as a frontend by a user (e.g., an operator of a blast furnace) to interact with the computing device 900. The components shown here, their connections and relationships, and their functions, are meant to be exemplary only, and are not meant to limit implementations of the inventions described and / or claimed in this document.

[0341] Computing device 900 includes a processor 902, memory 904, a storage device 906, a high-speed interface 908 connecting to memory 904 and high-speed expansion ports 910, and a low speed interface 912 connecting to low speed bus 914 and storage device 906. Each of the components 902, 904, 906, 908, 910, and 912, are interconnected using various busses, and may be mounted on a common motherboard or in other manners as appropriate. The processor 902 can process instructions for execution within the computing device 900, including instructions stored in the memory 904 or on the storage device 906 to display graphical information for a GUI on an external input / output device, such as display 916 coupled to high speed interface 908. In other implementations, multiple processors and / or multiple buses may be used, as appropriate, along with multiple memories and types of memory. Also, multiple computing devices 900 may be connected, with each device providing portions of the necessary operations (e.g., as a server bank, a group of blade servers, or a multi-processor system).

[0342] The memory 904 stores information within the computing device 900. In oneimplementation, the memory 904 is a volatile memory unit or units. In another implementation, the memory 904 is a non-volatile memory unit or units. The memory 904 may also be another form of computer-readable medium, such as a magnetic or optical disk.

[0343] The storage device 906 is capable of providing mass storage for the computing device 900. In one implementation, the storage device 906 may be or contain 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 configurations. A computer program product can be tangibly embodied in an information carrier. The computer program product may also contain 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 the memory 904, the storage device 906, or memory on processor 902.

[0344] The high speed controller 908 manages bandwidth-intensive operations for the computing device 900, while the low speed controller 912 manages lower bandwidth-intensive operations. Such allocation of functions is exemplary only. In one implementation, the high-speed controller 908 is coupled to memory 904, display 916 (e.g., through a graphics processor or accelerator), and to high-speed expansion ports 910, which may accept various expansion cards (not shown). In the implementation, low-speed controller 912 is coupled to storage device 906 and low- speed expansion port 914. The low-speed expansion port, which may include various communication ports (e.g., USB, Bluetooth, Ethernet, wireless Ethernet) may be coupled 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, e.g., through a network adapter.

[0345] The computing device 900 may be implemented in a number of different forms, as shown in the figure. 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 arack server system 924. In addition, it may be implemented in a personal computer such as a laptop computer 922. Alternatively, components from computing device 900 may be combined with other components in a mobile device (not shown), such as device 950. Each of such devices may contain one or more of computing device 900, 950, and an entire system may be made up of multiple computing devices 900, 950 communicating with each other.

[0346] Computing device 950 includes a processor 952, memory 964, an input / output device such as a display 954, a communication interface 966, and a transceiver 968, among other components. The device 950 may also be provided with a storage device, such as a microdrive or other device, to provide additional storage. Each of the components 950, 952, 964, 954, 966, and 968, are interconnected using various buses, and several of the components may be mounted on a common motherboard or in other manners as appropriate.

[0347] 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 that include separate and multiple analog and digital processors. The processor may provide, for example, for coordination of the other components of the device 950, such as control of user interfaces, applications run by device 950, and wireless communication by device 950.

[0348] Processor 952 may communicate with a user through control interface 958 and display interface 956 coupled to a display 954. The display 954 may be, for example, a TFT LCD (Thin-Film-Transistor Liquid Crystal Display) or an OLED (Organic Light Emitting Diode) display, or other appropriate display technology. The display interface 956 may comprise appropriate circuitry for driving the display 954 to present graphical and other information to a user. The control interface 958 may receive commands from a user and convert them for submission to the processor 952. In addition, an external interface 962 may be provide in communication with processor 952, so as to enable near area communication of device 950 with other devices. External interface 962 may provide, for example, for wired communication in some implementations, or for wireless communication in other implementations,and multiple interfaces may also be used.

[0349] The memory 964 stores information within the computing device 950. The memory 964 can be implemented as one or more of a computer-readable medium or media, a volatile memory unit or units, or a non-volatile memory unit or units. Expansion memory 984 may also be provided and connected to device 950 through expansion interface 982, which may include, for example, a SIMM (Single In Line Memory Module) card interface. Such expansion memory 984 may provide extra storage space for device 950, or may also store applications or other information for device 950. Specifically, expansion memory 984 may include instructions to carry out or supplement the processes described above, and may include secure information also. Thus, for example, expansion memory 984 may act as a security module for device 950, and may be programmed with instructions that permit secure use of device 950. In addition, secure applications may be provided via the SIMM cards, along with additional information, such as placing the identifying information on the SIMM card in a non-hackable manner.

[0350] 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 in an information carrier. The computer program product contains 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 the memory 964, expansion memory 984, or memory on processor 952 that may be received, for example, over transceiver 968 or external interface 962.

[0351] Device 950 may communicate wirelessly through communication interface 966, which may include digital signal processing circuitry where necessary. Communication interface 966 may provide for communications 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, through radio-frequency transceiver 968. In addition, short-range communication may occur, such as using a Bluetooth, WiFi, or othersuch transceiver (not shown). In addition, GPS (Global Positioning System) receiver module 980 may provide additional navigation- and location-related wireless data to device 950, which may be used as appropriate by applications running on device 950.

[0352] Device 950 may also communicate audibly using audio codec 960, which may receive spoken information from a user and convert it to usable digital information. Audio codec 960 may likewise generate audible sound for a user, such as through a speaker, e.g., in a handset of device 950. Such sound may include sound from voice telephone calls, may include recorded sound (e.g., voice messages, music files, etc.) and may also include sound generated by applications operating on device 950.

[0353] The computing device 950 may be implemented in a number of different forms, as shown in the figure. For example, it may be implemented as a cellular telephone 980. It may also be implemented as part of a smart phone 982, personal digital assistant, or other similar mobile device.

[0354] Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, 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 that are executable and / or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0355] These computer programs (also known as programs, software, software applications or code) include machine instructions for a programmable processor, and can be implemented in a high-level procedural and / or object-oriented programming language, and / or in 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 discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machineinstructions 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.

[0356] To provide for interaction with a user, the systems and techniques described here 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 a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; 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.

[0357] The systems and techniques described here can be implemented in a computing device that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a client computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), and the Internet.

[0358] The computing device can include clients and servers. A client and server 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.

[0359] A number of embodiments have been described. Nevertheless, it will be understood that various modifications may be made without departing from thespirit and scope of the invention.

[0360] In addition, the logic flows depicted in the figures do not require the particular order shown, or sequential order, to achieve desirable results. In addition, 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 following claims.References**1 collecting phase, during that data is collected for use a neural network, for data also 11**2 training phase, during that the collected data serves to train the neural network**3 operation phase, during that the trained neural network operates to control the oven, for data also 13{P} parameter value sequence {P} for a time-variant gas-flow parameter, with {Pl, P2, ...} being a value sequence for time-slots AtAt time-slots for the value sequenceAT time-to-insert, modality sub-parameter for inserting the coal-cake k, K time-series index for the parameter value sequence m, M elements in the training set, i.e., M number of pairs 301-m / 501-m identifiable by index m q, Q instance identifier and instance counter in an approach to obtain emission data from camera images{X0} modality parameter set, with parameters that relates to the operation of the oven and that represents the modality to charge the oven with the cake{XI} property parameter set, with parameters for the coal-cakeZ cardinality that indicates how many different classes of the value sequences the trained network could provide103 stamp-charging coke-oven113 coal-cake to be processed by the oven (i.e., in operation)133 02 sensor163 pressure value of the oven (in operation), as part of the offtake piping183 gas-collecting main GCM in the offtake piping system, up-stream before the value, down-stream after the pressure-control device203 computer (with the network), method-executing computer252 neural network (in training)253 neural network (in operation)30# plurality of input parameters such as XO, XI and others301-m historical parameters for training303 actual set of input parameters, for operation401 collecting historical data, i.e., the training set402 training the network403 method to identify the value sequence, during operation, with parameter processing by the network4x3 method steps50# value sequence {pl, p2, ...} for the time-variant gas-flow parameter:501-m historical value sequence for training, with 301-m / 501-m standing for a training pair with the historical input parameters and the corresponding historical value sequence, for training the network503 the particular sequence to be identified for operation9xx computer in general xxxx 4-digit-references to explain an implementation to obtain emission data {E} by processing camara images

Claims

1. Claims1. Computer-implemented method (403) to identify a parameter value sequence (503, {P} 13, {Pl, P2, P3, ...} 13) for a time-variant gas-flow parameter to be applied to a pressure-control device (163) in a gas-collecting main (183) of an offtake piping system (143, 143', 183) of at least one stamp-charging coke-oven (103) that is adapted to process a coal-cake (113), the method (403) comprising: receiving (413) a set (303) of input parameters, with at least a modality parameter set ({X0} 13) that represents a charging modality by that the coal-cake (113) is to be inserted into the stamp-charging coke-oven (103), and a coal-cake property parameter set ({XI} 13) that represents a property of the coal-cake before insertion; data-processing (423) the received set (303) of input parameters by a neural network (253) that is a classifier and that has been trained (402) in advance with training data (301-m / 501-m), wherein - during training - the input of the neural network (252) has received a plurality (301) of historical data-sets (301-m) of input parameters from reference operations (**1) of a reference stamp-charging coke-oven and wherein the output of the neural network (252) has received a plurality (501) of historical sequences (501-m) of time-variant gas-flow parameters as the ground-truth; wherein data-processing (423) provides a preliminary value sequence (503, {P} 13); and providing (443) the preliminary value sequence (503, {P} 13) to the pressure-control device (163) of the stamp-charging coke-oven (103) during cake insertion.

2. Method (403) according to claim 2, wherein the modality parameter set ({X0} 13) comprises one or more of the following parameters:a first modality parameter (X0.1) that represents a speed increase in an acceleration phase; a second modality parameter (X0.2) that represents the duration of the acceleration phase; a third modality parameter (X0.3) that represents a speed of insertion; a fourth modality parameter (X0.4) that represents a duration of insertion at regime speed; a fifth modality parameter (X0.5) that represents a speed decrease in a deceleration phase; a sixth modality parameter (X0.6) that represents a speed decrease in a deceleration phase; and a seventh modality parameter (X0.7) that represents the duration of the deceleration phase.

3. Method (403) according to any of claims 1 or 2, wherein the coal-cake property parameter set ({XI}) comprises one or more of the following parameters: a first property parameter (XI.1) that represents the share of volatile matters in the coal-cake (113); a second property parameter (XI.2) that represents the humidity in the coal-cake (113); a third property (XI.3) that represents the density of the coal-cake (113); and a fourth property parameter (XI.4) that represents the height of the coal-cake (113).

4. Method (403) according to any of claims 1 to 3, wherein providing (443) the preliminary value sequence (503) is performed conditionally after testing (433) toconfirm that the preliminary value sequence (503) has its start values within tolerances given by a control loop.

5. Method (403) according to any of claims 1 to 4, wherein providing (443) is performed simultaneously with controlling the pressure-control device (163) by a control loop with controlling the pressure-control device (163) such that the oxygen amount of the gas in downstream direction of the gas-collecting main (183) remains below a predefined threshold.

6. Method (403) according to any of claims 1 to 5, wherein providing (443) is performed simultaneously with controlling the pressure-control device (163) by a control loop with controlling the pressure-control device (163) such that the amount of gas emissions ({E}) from the oven (103) remains within a pre-defined range.

7. Method (403) according to any of claims 1 to 6, wherein the time-variant gas-flow parameter for the pressure-control device (163) of the oven (103) is selected from the following: a pressure parameter, a water level parameter, an angular or linear location parameter of a mechanical element, and a parameter to control a plurality of ammonia waterjets.

8. Method (403) according to any of claims 1 to 7, wherein - during training - the neural network (252) has received the plurality (501) of historical sequences (501-m) of timevariant gas-flow parameters as the ground-truth together with historical emission data so that in step data-processing (423), the neural network (253) provides the preliminary value sequence (503) such that it is optimized for minimal emissions.

9. Method (403) according to any of claims 1 to 8, wherein data-processing (423) the received set (303) of input parameters by the neural network (253) involves the use of a neural network (253) that has also been trained with historical sets of emission data ({E} 11}) from reference operations (**1) of a reference stamp-charging coke-oven as a ground-truth for gas emissions, so that data-processing (423) not only provides thepreliminary value sequence (503) but also provides an estimation of emission data ({E}| 3}).

10. Method (403) according to claim 9, wherein the neural network (253) has been trained either(i) simultaneously with historical data-sets (301-m) of input parameters and historical sets of emission data, or(ii) consecutively with the historical data-sets (501-m) that are the sequences of timevariant gas-flow parameters as the ground-truth ({P}| 1), and subsequently - during retraining - with historical data-sets (501-m) of emission data ({E} 11) as further groundtruth.

11. Method (403) according to any of claims 8 and 9, wherein the historical data-sets of emission data ({E}| 1}) have been obtained by applying an emission detecting technique that uses a camera (1143) and that provides emission classification by an emission classifier network (1253).

12. Method (403) according to any of claims 1 to 8, wherein data-processing (423) the received set (303) of input parameters by the neural network (253) involves the use of a neural network (253) that has also been trained with historical sets of oxygen data ({02} 11}) from reference operations of (**1) of a reference stamp-charging coke-oven as a ground-truth for oxygen data, so that data-processing (423) not only provides the preliminary value sequence (503) but also provides an estimation of oxygen data ({02}).

13. Method (403) according to claim 12, wherein the neural network (253) has been trained (i) simultaneously, with historical data-sets of input parameters ({X}| 1) and of oxygen data ({02} 11), or (ii) consecutively with the historical data-sets (501-m) that are the sequences of time-variant gas-flow parameters as the ground-truth ({P}| 1), andsubsequently - during re-training - with historical data-sets (501-m) of oxygen data ({02} 11) as further ground-truth.

14. Method (403) according to any of claims 1 to 13, wherein the step applying (443) the preliminary value sequence (503) to the pressure-control device (163) of the stampcharging coke-oven (103) during cake insertion is followed by: collecting (453) feedback with emission data ({E}) that represents emissions from the stamp-charging coke-oven (103), and / or with oxygen data {02} from a sensor (133) in the gas-collecting main (183), and rewarding (463) the neural network (253) during reinforced learning.

15. Method (403) according to claim 14, wherein collecting (453) feedback with emission data ({E}) applies an emission detecting technique that uses a camera (1143) and that provides emission classification by an emission classifier network (1253).

16. Use of the computer-implemented method (403) according to any of claims 1 to 15, being applied to a plurality of coke-ovens (103, 103'), wherein the method-executing computer (203) is located remotely to the controller of the coke-ovens (103, 103').

17. Computer system (203) to identify a parameter value sequence (503, {P} 13, {Pl, P2, P3, ...} 13) for a time-variant gas-flow parameter to be applied to a pressure-control device (163) in a gas-collecting main (183) of an offtake piping system of at least one stamp-charging coke-oven (103) that is adapted to process a coal-cake (113), the computer system being adapted by comprising modules to perform a computer- implemented method (403) according to any of claims 1 to 16.

18. Computer program product that - when loaded into a memory of a computer and being executed by at least one processor of the computer causes the computer to perform the steps of the method according to any of claims 1 to 16.

19. Computer-implemented method for training a neural network (252) to enable the neural network (252) to identify a parameter value sequence (503, {P} 13) for a timevariant gas-flow parameter to be applied to a pressure-control device (163) in a gascollecting main (183) of an offtake piping system of at least one stamp-charging coke- oven (103) that is adapted to process a coal-cake (113), the method (402) comprising: at the input of the neural network (252), applying a plurality (301) of historical datasets (301-m) of input parameters from reference operations (**1) of a reference stamp-charging coke-oven, the input parameters with at least a modality parameter set ({X0} 11) that represents a charging modality by that coal-cake had been inserted during the reference operations, and a coal-cake property parameter set ({XI} 11) that represents a property of the coal-cake before insertion; at the output of the neural network (252), applying a plurality (501) of historical sequences (501-m, (503, {P} 11)) of time-variant gas-flow parameters as the groundtruth.

Images