Real-time control of industrial ovens based on the degree of waste generated by computer recognition.

Abstract

The computer (200) obtains a pressure setpoint (p_set) for a programmable controller (170) associated with the oven (100) and controlling the gas pressure (p(t)) inside the oven. The controller (170) receives pressure data (p(t)) from a pressure sensor and interacts with a pressure valve. The computer (200) obtains a leak area image from a camera (140) installed outside the oven (100) showing an area (130) on the outer surface of the oven (100) where gas emissions may be present. The computer (200) processes the image with a pre-trained network and classifies the degree of emissions (d(t)). Applying predefined rules and based on the classified degree (d(t)), the computer (200) changes the setpoint (p_set) for the programmable controller (170).

Description

Technical Field

[0001] The present disclosure generally relates to industrial production processes and equipment for performing such processes. More specifically, the present disclosure relates to a computer system, method, and computer program product that assist in environmental monitoring of technical equipment such as ovens through emissions leak diagnosis.

Background Art

[0002] In the industry, reactors such as ovens perform industrial production processes. Put simply, the process can be a continuous and uninterrupted process or a periodic process. The cycle usually starts when the oven takes in materials and ends when the product is output. The oven exposes the materials to a relatively high temperature, and for solid materials, the process also involves substances that are gases. Some of the gases are useful by-products, and the gases are usually recovered by means such as gas recovery pipes.

[0003] The process is described and controlled by parameters (or process variables), among which there are the temperatures of materials and gases, as well as gas pressure, quantity (materials, gases, auxiliary materials), the characteristics of chemical reactions occurring in the materials, and many other parameters.

[0004] To keep the parameters within a pre-defined acceptable range, the oven has sensors, actuators, controllers, and other equipment for implementing a control loop. Such a control loop usually applies set values such as a pressure set-point, a temperature set-point, etc. The implementation of the control loop is well-known in the art. For example, the controller receives pressure data from at least one pressure sensor and interacts with a pressure valve to adjust the pressure to the set-point pressure.

[0005] Parameters may change during process execution; for example, gas pressure may decrease or increase. If such changes are necessary (to support the process), the settings are typically modified over time.

[0006] The gas pressure inside an oven is related to the air pressure outside the oven. In the case of "overpressure," as used here, the absolute pressure (of the process gas) inside the oven is higher than the absolute pressure (of the air) outside the oven. In the case of "underpressure," the absolute pressure inside is lower than the absolute pressure outside.

[0007] Ovens have openings such as doors, holes, lids, connection systems, intake systems, and removal systems. Since openings are usually closed during the process, it is convenient to use the term "closure" instead.

[0008] The primary function of such a closure is to fill the oven with material and to remove the (processed) material from the oven. When properly closed, the closure prevents the movement of material through it. Opportunities to open the closure are usually process-related; simply put, the closure is opened at the beginning and end of the process cycle, and otherwise, as the term suggests, the closure is closed.

[0009] The closures have an additional function of preventing gas flow. To put it simply, in the case of overpressure, process gases (i.e., gas emissions) must not be released from the oven, and in the case of low pressure, air must not enter the oven. The closures are usually equipped with seals to prevent gas emissions, but such seals are exposed to the gas temperature. The ability to stop the gas decreases over time, for example, due to frequent opening and closing, which also degrades the seal condition over time. On average, such a cycle may occur once a day. Leaks in the closures occur from time to time, and leaks have undesirable side effects such as gas emissions.

[0010] There are several aspects of gas emissions that require further attention: the gas may contain materials that can accumulate in or near enclosed areas. These deposits (including tar) are difficult to remove and can degrade seals. Depending on the chemical composition of the process gas, it may be toxic or flammable. Therefore, air pollution is a further concern due to potential safety issues for (human) operators and environmental impacts.

[0011] A human operator periodically searches for discharge by, for example, visually inspecting the oven and its closure. The operator determines the intensity (concentration) of the discharge, its duration, and other phenomena.

[0012] However, there are various limitations: operators take special safety precautions (wearing protective clothing, masks, etc.) when approaching the oven. Because operators are not always present, it is difficult to determine the discharge period. Evaluations by human operators lack the accuracy of sensor measurements, and evaluations may vary depending on the operator.

[0013] Furthermore, the constraints are not limited to the operator but also apply to the equipment that provides the control loop. Changes to tolerances, sensor and actuator structures, and even the design of the closures, could ultimately lead to process failure or damage to equipment, including the oven itself.

[0014] US2014 / 0002639A1 discloses an approach to autonomously detect the diffusion of chemical substances into the air by analyzing images from detection cameras.

[0015] DE102021101102A1 discloses the use of unmanned aerial vehicles (UAVs) to monitor coke ovens. The UAVs can measure temperature with IR cameras, measure gas concentrations with chemical sensors, and take pictures of the ovens with onboard cameras.

[0016] US10,059,884B2 discloses a coke oven, points out that the pressure inside the oven changes over time, and also mentions undesirable emissions. [Overview of the Initiative]

[0017] With the overall goal of reducing or eliminating gas emissions from ovens into the environment due to leaks, the computer interacts with one or more ovens with essentially a single setting (one point): the computer causes the oven controllers to update the pressure setpoint. This setpoint of pressure inside the oven is updated to a specific value at which emissions are expected to be at a desirable level that satisfies two conditions: (i) avoiding or minimizing emissions, and (ii) continuing to efficiently carry out processes inside the oven.

[0018] The controller implementation loop, which includes data from the pressure sensor, is modified only by controlling the pressure setpoint (of the controller). The method-performing computer can be considered a further (or second) control loop for controlling the emission degree. The computer performs the method (computer implementation) to obtain the pressure setpoint of the controller that controls the gas pressure inside the oven.

[0019] The area on the external surface of the oven where gas emissions are visible is called the leak region. A computer acquires leak region images from a camera located outside the oven. The computer processes the leak region images using a pre-trained network and classifies the degree of emissions in the leak region images. There are several classes of classification. In one embodiment, the network simply recognizes the presence or absence of emissions (two classifications: present and absent), while in another embodiment, the network classifies the emissions at a higher granularity.

[0020] This network is trained on training data that includes historical images taken as reference material. Artificial annotations are associated with the historical images.

[0021] The computer applies predefined rules to change settings based on the degree to which the waste is classified.

[0022] To improve the accuracy of waste detection, image-related activities (i.e., image acquisition and processing) can be performed in multiple instances (processing units).

[0023] A computer implementation method for obtaining pressure setpoints for a programmable controller associated with an oven and controlling the gas pressure inside the oven is disclosed. The programmable controller receives pressure data from at least one pressure sensor and interacts with a pressure valve.

[0024] The computer acquires leak region images from a camera installed outside the oven (i.e., outside the oven), showing areas on the outer surface of the oven where gas emissions may be present (i.e., where they might appear). Hereafter, this area will be referred to as the leak region.

[0025] The computer processes leak region images using a pre-trained network to classify the degree of leakage. The network is pre-trained with training data that includes historical images taken as reference and artificial degree annotations on the historical images. Applying predefined rules and based on the classified degree of leakage, the computer modifies settings for a programmable controller.

[0026] Optionally, the computer performs a step of acquiring a leak area image from an optical camera so that the leak area image indicates an area where visible gas emissions may be present.

[0027] Optionally, the computer performs the step of obtaining a leakage area image from the thermography camera to indicate an area where gas emissions that can be recognized by the temperature gradient towards the background of the image may exist.

[0028] Optionally, in a processing step, the computer uses a pre-trained network to classify the degree of emissions into two binary categories: a first degree of no emissions and a second degree of emissions.

[0029] Optionally, in a processing step, the computer uses a pre-trained network to classify the second degree into a plurality of subclasses. The second degree includes subclasses such as: present as low emissions, present as medium emissions, present as high emissions.

[0030] Optionally, the computer performs the step of obtaining a leakage area image and, within a time interval shorter than the time interval required for the controller to interact with the pressure valve to actually change and stabilize the gas pressure to a set value inside the oven, processes the leakage area image in a plurality of instances (processing units).

[0031] Optionally, by performing the steps in a plurality of instances, the computer obtains a plurality of leakage area images, processes the plurality of leakage area images individually, and classifies the degree individually for each instance, thereby obtaining a plurality of degrees represented by a degree vector.

[0032] Optionally, when applying pre-defined rules, the computer evaluates the degree vector according to the degree distribution within the degree vector.

[0033] Optionally, the computer evaluates the degree vector according to the distribution by either: specifying the share between the binary degrees of no emissions and emissions, and specifying the rate of change between the degrees.

[0034] Optionally, when used with an optical camera, the computer further identifies the situation by having the camera acquire multiple leak region images. Such visible conditions are selected based on one of the following: the quality of light (natural or artificial), the presence or absence of precipitation in the scene between the camera and the oven, or the presence or absence of dust on the camera's objective lens. When applying predefined rules to evaluate the degree vector, the computer adapts the predefined rules according to the situation.

[0035] Optionally, the computer indirectly identifies the situation by evaluating the following data representing the camera environment: light intensity in leak area images, light characteristics by separating daylight from artificial light, light characteristics by separating sunlight from moonlight, quality and quantity of precipitation reaching the oven, and detection of meteorological precipitation.

[0036] Optionally, the computer identifies the situation by processing leak area images. For example, the computer can use an additionally trained network for this purpose. This additional network identifies rain, snow, or other items between the oven and the camera. Simply put, the presence of such items indicates a worsening of the situation.

[0037] Optionally, the computer processes leak area images using a pre-trained network trained on either (1) training data containing historical images taken as reference from an oven, or (2) training data containing historical images taken as reference from a physically different oven.

[0038] Optionally, the computer acquires images of leak areas in multiple ovens of the battery. The camera is mounted on a vehicle whose primary purpose is to transport materials between the multiple ovens. The vehicle moves with the camera.

[0039] A computer system is disclosed for acquiring pressure setpoints for a programmable controller associated with an oven and for controlling the gas pressure inside the oven. The programmable controller receives pressure data from at least one pressure sensor and interacts with a pressure valve. The computer system is adapted to include a module for performing a computer implementation method.

[0040] This disclosure also relates to the use of a computer implementation method for obtaining a pressure setpoint for controlling the gas pressure inside an oven selected from coke ovens, furnaces, iron-making industry devices, a piece of equipment for steel-making, cement reactors, concrete reactors, and chemical reactors.

[0041] A computer program (or computer program product) is also disclosed. The computer program causes a computer to perform steps of a computer implementation method, which are stored in the computer's memory and executed by at least one of the computer's processors.

[0042] Furthermore, a computer implementation method is disclosed for training a network (on oven leak region images, by processing leak region images showing the area of ​​the outer surface of the oven from which gas emissions can be released, as previously described) in order to classify the degree of emissions. The computer connects historical images to the input of the network and artificial annotations to the historical images at the output of the network. The artificial annotations include emission classes as ground truth (reference values ​​or baseline data). Connecting historical images includes changing the oven pressure setpoint by a specific pressure difference until the oven emits gas visible to the camera. Images showing the oven and the gas visible to the camera are captured as historical images. Connecting artificial annotations (to the network output) includes using observations of specific emission degrees as annotations (or, optionally, using specific pressure differences as annotations).

[0043] From an overall perspective, this disclosure also relates to a computer implementation that can be considered a combination of a first method for training a network and a second method for using the trained network. Either method can be performed by first and second computing functions implemented by one or more physical computers.

[0044] The second method of execution is used in conjunction with the oven to obtain the pressure setpoint of a programmable controller that controls the gas pressure inside the oven. The programmable controller receives pressure data from at least one pressure sensor and interacts with a pressure valve. The computer obtains leak area images from a camera installed outside the oven, which represent areas on the outer surface of the oven where gas emissions may exist, as follows:

[0045] The first implementation method uses a method to train a network that classifies the degree of emissions in oven leak region images by processing leak region images that show areas on the outer surface of the oven from which gas may be released (i.e., leak regions). The first computing function connects historical images to the input of the network and connects artificial annotations to the historical images to the output of the network, where the artificial annotations include emission classes as ground truth. The first implementation method yields a trained network that classifies the degree of emissions in leak region images (i.e., a pre-trained network).

[0046] The second computing function processes leak area images using this trained network to classify the degree of waste. Applying predefined rules, the second computing function modifies the programmable controller settings according to the classified degree of waste. Optional features have already been outlined in this summary. [Brief explanation of the drawing]

[0047] Embodiments of the present invention will be described in detail with reference to the accompanying drawings: [Figure 1] This outlines the network phases and image processing activities applied to an oven in operation. [Figure 2] This shows an overview (overall picture) of the loop control applied to the oven. [Figure 3] A perspective view of the oven battery is shown, illustrating multiple setpoint control loops. [Figure 4] A single oven, its closure, and visible discharge in the leak area are shown, along with an introduction to the degree of discharge. [Figure 5] This shows the training phase in which the neural network is trained. [Figure 6] This shows two control loops applied to a single oven, equipped with a controller, camera, and pre-trained network. [Figure 7] This diagram shows a computer implementation method for obtaining pressure setpoints for a programmable controller related to an oven. [Figure 8] Figure 7 shows a flowchart for a technique where several steps of the method are performed in multiple instances. [Figure 9] To illustrate further techniques for improving classification accuracy, images from various situations are presented. [Figure 10] This indicates a general-purpose computer to be used to perform one of the methods. [Modes for carrying out the invention]

[0048] Figure 1 shows an overview of phases **1, **2, and **3 related to the image processing activity. The network applies image processing to one or more ovens 100 in operation.

[0049] From a general perspective, this description focuses on computer-implemented image processing using a network. The network applies machine learning techniques, including training. Images may (or may not) show gas emissions 120 from one or more ovens 100. With the overall goal of reducing (and, if possible, eliminating) gas emissions (from ovens into the environment due to leaks), the description describes a method for detecting gas emissions at present (i.e., while the ovens are operating).

[0050] The intention is to identify countermeasures such as changing the pressure inside the oven.

[0051] To explain the network, the explanation will divide it into phases, and Figure 1 shows: • During the preparation phase**1, data is collected for use in one or more networks. • The data collected during that time plays a role in the training phase (i.e., becoming a trained network)**2. • In the meantime, one or more trained networks play a role in controlling the operation of the oven, as outlined in the current operation phase**3.

[0052] The oven may be "on" during all phases of operation.

[0053] Measures can be taken after image processing via the network. Immediate measures may include adapting the current oven operation (such as changing the pressure in real time).

[0054] In the current operation phase**3, a computer-implemented method is executed to obtain (or change) the pressure setpoint "p_set" of the programmable controller that controls the gas pressure inside the oven. This is described in detail as method 403 in Figures 7-8. This method is executed when one or more ovens are in a processing state where leakage is not expected, ideally (e.g., when the closure is closed). The computer is expected to execute this method repeatedly to detect immediately if any discharge occurs (see Figures 7-8).

[0055] The time interval T1 represents the time required for the computer to execute a step sequence to acquire data that describes the waste (for example, to classify the waste). The step sequence includes acquiring (one or more images) and classifying (one or more) images.

[0056] The time interval T2 represents the time required for the programmable controller (which operates on other components of the oven) to take action until the discharge stops. The duration of T2 depends primarily on the processes inside the oven, and the time interval T2 is • The time it takes to operate the pressure valve (i.e., to open and close the valve), • Allow time to change the pressure (to the new pressure setting) and for the internal pressure of the oven to stabilize (at the new pressure setting). Includes.

[0057] T2 is for a single oven. If multiple ovens are operating in parallel, assume that T2 is the same for all ovens.

[0058] Time interval T2 follows time interval T1 with a relatively short interval—a so-called trigger interval—the computer then evaluates the emission classification by applying predefined rules and finally triggers (activates) the controller to change the pressure setpoint (optionally via an auxiliary controller). Here, the duration of the trigger interval is negligible.

[0059] Note that T1 is shorter than T2. ​​The effect of this temporal imbalance can be applied to a favorable degree depending on the method. Simply put, the time interval T1 (see Figure 7) can be estimated in seconds, and the time interval T2 can be estimated in minutes.

[0060] In the description, the current point in time is indicated by "t". Unless otherwise specified, t refers to the time when the method execution begins (see method 403 with start line 499 in Figure 7). Since the discharge is not expected to start or stop within T1, the notation t applies to the entire time interval T1 (i.e., until the end of the method execution).

[0061] The collection of historical data (regarding emissions) and the assignment of annotations to it are performed in preparation phase**1. In training phase**2, one or more networks need to be trained (from network 252 to network 253, Figure 5). Training phase**2 can begin as soon as a sufficient amount of training data has been collected (and annotated).

[0062] The current operation phase**3 includes the computer's runtime, which executes methods such as method 403, as described in Figures 7 and 8.

[0063] During the current operating phase 3, one or more ovens will change their discharge behavior, which can be measured over longer intervals such as T2.

[0064] Figure 1 shows the oven's operation over time with a bold horizontal line. The line is not scaled, but there are some gaps that symbolize that the oven is not always under pressure. Therefore, since discharge is expected only during "pressure ON time," data collection for use in the network (phase**1) is largely limited to "pressure ON." Furthermore, the network works to control the oven's operation (phase**3) only during "pressure ON." Network training (phase**2) is independent of the oven's operation and can be performed during ON / OFF.

[0065] Data regarding the on / off status is always available, so this explanation will not go into further detail, but this method is assumed to be performed only when the pressure is on.

[0066] Data representing emission intensity (concentration) is given as a degree (or class). While the degree is not the same as an exact measurement, the accuracy (and granularity) is sufficient to identify countermeasures for emissions, such as changing pressure settings.

[0067] The explanation concludes with a detailed description of the camera's implementation, image preprocessing, and other related topics.

[0068] control loop Figure 2 shows a very simplified overview of the first and second control loops applied to oven 100. Both loops interact to reduce gas emissions from oven 100 into the environment due to leakage.

[0069] As shown on the right side of the figure, the pressure controller 170 implements a first control loop for controlling the pressure p(t) inside the oven 100. The pressure controller 170 receives pressure data p(t) from at least one pressure sensor and interacts with a pressure valve (see Figure 6 for details). The pressure controller 170 opens and closes the valve according to the difference Δp between the actual value p(t) and the set value p_set. As used herein, the term “pressure valve” includes any equipment that can directly or indirectly change the pressure inside the oven.

[0070] When the temperature inside oven 100 changes, the pressure p(t) also changes, but the pressure controller 170 maintains the pressure at the set value p_set.

[0071] As shown on the left side of Figure 2, the computer 200 implements a second control loop (in the functions of the discharge classifier and auxiliary controller).

[0072] The second control loop adapts the setpoint (p_set) from the first control loop, but adapts it using different input:discharge rates. Computer 200 performs a (computer implementation) method to obtain the pressure setpoint p_set of the controller 170 that controls the gas pressure inside the oven 100 (see the first control loop, method in Figure 7).

[0073] The second control loop has a set value for the desired minimum discharge level (d_set, e.g., no discharge or low discharge). From a broad perspective, computer 200 implements several functions: (i) Computer 200 functions as an emission degree sensor to obtain d(t). This function is implemented in T1) by a module called an "emission classifier". (ii) Computer 200 also functions as a comparator that generates triggers during negligibly short trigger intervals, according to the deviation Δd from d_set of degree d(t). (iii) When a trigger is generated, computer 200 also functions as a set-point generator, processing further parameter data (from the oven) and generating a new pressure setpoint (or checking the old pressure setpoint). In other words, the set-point generator updates p_set by applying rule logic. Other parameters include, for example, temperature, and information about the process inside the oven indicating whether pressure changes are permissible for the particular process. Rule logic can simply automate the operator's response when they notice discharge (to reduce the pressure by a certain amount).

[0074] In more detail, the computer 200 implements an emission level sensor by using a neural network (or equivalent machine learning tool) to process images from one or more leak areas 130 (of the oven 100) captured by one or more cameras 140. The explanation symbolizes one or more cameras by arrows pointing to the leak areas; see, for example, Figures 3 and 6.

[0075] Other functions, (ii) trigger generator and (iii) setpoint generator, can be implemented in the auxiliary controller.

[0076] The two control loops "contact" at only one point: the second control loop provides p_set (for example, through updates). Feedback from oven 100 to computer 200 is transmitted via images, transcending the network and rule logic (as an emission classifier).

[0077] Skilled operators have implemented the (first) control loop for oven pressure for centuries (or even longer). Theoretically, it doesn't even require the application of sophisticated electronics; the first control loop can be implemented solely through mechanics.

[0078] The situation (picture) is completely different for the second control loop, particularly its degree sensing function (i): This requires taking images from a specific area of ​​the oven (such as leak area 130), preprocessing the images (e.g., receiving artificial annotations, Phase**1), training the network (Phase**2), and running the trained network (Phase**3).

[0079] It is helpful to explain the second loop by referring to a coke oven, a prominent example of an oven. Based on the description herein, experts can apply the teachings to other ovens (or general reactors).

[0080] Camera and visibility The image broadly represents the characteristics of the leakage region 130 using data. The data for each pixel corresponds to the individual positional properties of the leakage region 130, and its description illustrates several examples in Figures 4 and 9.

[0081] Since cameras, especially digital cameras for taking images, are well known in the art, skilled personnel can position the camera in the leakage area 130, select an appropriate optical system (i.e., lens system), select a sensor with appropriate technology (e.g., charge-coupled CCD, complementary metal-oxide-semiconductor CMOS, or other technology), make the camera mechanically robust to withstand industrial environments (i.e., vibration, dust, moisture, etc.), implement the transmission of image data from the camera to a computer, etc. (e.g., via a network for industrial use), and so on.

[0082] To illustrate this technique, a skilled operator might use a so-called industrial camera equipped with dustproof features (i.e., a cover) or some dust removal features (similar to a window wiper). The camera would be remotely operated (i.e., manual exposure would not be required). Such a camera could also be used as a surveillance camera, and potentially existing surveillance cameras could also function as camera 140.

[0083] While it is not necessary to describe the details of such cameras in this document, the following points should be noted, considering that human involvement is involved.

[0084] Computer displays are ubiquitous in industrial environments, but in most situations (i.e., not during Phase 3), there is no need to show images to human users (such as oven operators).

[0085] In other words, the drawings herein symbolically represent images showing objects in the leakage area, including doors with and without discharge; however, in actual implementations, the images are processed by computer 200.

[0086] As a result, camera 140 can be implemented to operate in either of two wavelength ranges using a single camera:

[0087] (i) The camera 140 operating in the first radiation range is an optical camera that captures images in visible light (i.e., approximately 380 to 750 nanometers). Typically, the light from the camera is reflected light. Typically, the sensor of an optical camera is optimized to provide an image for human viewing, and the sensor is implemented as an RGB sensor or as a sensor that applies other color schemes.

[0088] (ii) The camera 140 operating in the second radiation range is a thermographic camera that takes images using infrared radiation. The infrared radiation is emitted from the leak region 130. The oven door is distinguishable from the environment because it is relatively hot when in operation. The temperature difference that can be captured by the thermographic camera allows the exhaust to be distinguished from the door. The IR spectrum is relatively broad, and the usable wavelength range is 8 to 14 μm. Typically, the sensors of thermographic cameras are optimized to provide data representing the temperature for each pixel.

[0089] In images captured with a thermal imaging camera, gas emissions can be identified by the temperature gradient toward the background of the image. While measuring the temperature gradient is not necessary, the gas emissions are expected to be at a higher temperature than the oven surface (displayed as the background). The temperature difference is estimated to be in the range of 50 to 400 Kelvin.

[0090] As used here, the emissions are understood to be “visible to a camera” in both radiation ranges. The emissions are “visible” to both the human eye and an optical camera (i.e., the first range), and the emissions are “thermally visible” to a thermographic camera (invisible to the human eye).

[0091] To elaborate, the images are linked to human annotation during preparation phase**1. The annotations become the ground truth for training in phase**2. In phase**1, images captured by optical cameras are displayed to experts on a computer screen, while images captured by thermographic cameras are displayed in a adapted manner. For example, individual temperature values ​​can be displayed in different colors in a so-called heat map.

[0092] The assignment of colors to temperature needs to allow experts to distinguish the emissions from the rest of the image. In other words, if experts cannot see the emissions in the (heatmap-adapted) image, they cannot annotate the degree of emissions, etc.

[0093] The hybrid approach addresses the constraint that images to be annotated must be captured in two versions using two cameras. By positioning the thermal imaging camera to align with the optical camera, both cameras will effectively view the same leak area at the same time.

[0094] The optical camera version (using an optical camera) is shown to experts (Phase**1), and the training data for Phase**2 includes annotations combined with images from the thermographic camera version.

[0095] hybrid method Image processing from the thermographic camera in the second control loop (based on emission classification) is expected to be less accurate than image processing from the optical camera.

[0096] However, a thermal imaging camera can compensate for this potential drawback if it serves as a backup. Ovens constantly emit heat day and night, and a thermal imaging camera can always provide images (showing leak areas). In contrast, an optical camera only provides images at night (or under other "poor" conditions) if the oven is illuminated by artificial light (energy consumption is not negligible). (We will not discuss the exception of nighttime oven fires here, as such fires are not emissions regulated by pressure settings.)

[0097] Therefore, the operation of the second control loop (in Phase 3) depends on the light conditions (i.e., the visible conditions outlined below, Figure 8), • Using only optical cameras, • Select the most reliable degree of data from both parallel cameras. This is categorized as: • Thermographic camera only (backup scenario when the optical camera cannot provide images).

[0098] Multiple ovens powered by batteries, such as a coke oven, are a notable example. Figure 3 shows perspective views of ovens 100-1, 100-n, and 100-N connected to battery 100-BATT, as well as the setpoint control loops specific to each oven. The ovens can be arranged in parallel with the battery (oven 100-n "touches" oven 100-(n+1), with the ovens positioned adjacent to each other) or in other ways. Several control devices (controller computer, user interface, etc.) can be applied to the battery (typically to control each oven individually).

[0099] An oven can be a coke oven. Taking a coke oven as an example, it is also useful in the following case: a coke oven is usually installed with such a battery, and a single battery 100-BATT can have N ovens 100-n. Typical values ​​range from N=30 for one battery to N=75 for another.

[0100] Although there is an oven-specific controller 170-n (see Figure 2) and an oven-specific setting value p_set_n, the computer 200 can be implemented as a single computer that multiplexes its operation to provide the oven-specific setting value. Figure 3 symbolizes the computer 200 receiving an image from the camera (arrow symbol) and providing the setting values ​​individually as vectors (p_set_1, ..., p_set_n, ..., p_set_N) with values ​​for ovens 100-1, 100-n, and 100-N, respectively. Multiplexing (i.e., sequential execution) is merely an example; an expert can process multiple loops with multiple setting values ​​using, for example, parallel processing (or a hybrid method including parallel and sequential processing).

[0101] Figure 3 also introduces coordinate conventions, but these are merely examples. Each oven 100-n has a width X (usually equal for all N ovens), and the ovens 100-n are arranged vertically. Each oven 100-n has a length Y and a height Z, and considering the material, Y is divided into a "push-in side" and a "removal side". X, Y, and Z can be assumed to be substantially equal for all n=1 through N.

[0102] The camera symbols indicate areas where gas emissions may occur ("leakage areas"): Camera 140-ex may capture images from the XZ plane on the "extraction" side of the battery (shown here on the left), Camera 140-top may capture images from the XY plane on the top of the battery, and Camera 140-push may capture images from the XZ plane on the "intake" side of the battery.

[0103] For batteries, ovens with doors typically look similar, and pre-processed images can be applied to match the image. For example, numbers may be painted on the surface of the oven, but these painted numbers do not affect the output. Since the numbers are visible in the image, they can be removed from the image during pre-processing.

[0104] Alternatively, for example, if the captured image shows standardized elements of the oven (such as door hinges), the image can be pre-processed to match the hinge placement.

[0105] Leakage regions do not need to perfectly correspond to the XZ or XY planes; rather, all ovens within the battery may have one or more areas (leakage regions) where leakage is expected, for example, in the aforementioned closures. Leakage can occur where the heating wall and the material chamber come into contact with each other.

[0106] Leakage areas can be divided by the area (index n) belonging to each individual oven, and further divided by the components belonging to each oven. For example, in the case of an oven with doors on both sides, there are leakage areas on the removal side and the push-in side.

[0107] The dashed lines represent the camera trajectory, which is (optionally) implemented by a movable camera: • The retrieval track (the line on the left), • The upper surface track, and • Push-side trajectory This indicates.

[0108] Simplification: To cover a wide area, a single camera may move along multiple paths. For example, camera 140-top may start monitoring the area of ​​oven 100-1 on the push-in side, move to oven 100-N, and then return to oven 100-1 on the take-out side from oven 100-N for monitoring.

[0109] The camera does not need to move along a linear trajectory; it can be installed in a fixed position, and its movement (tilt, pan, etc.) can be directed towards a specific leak area.

[0110] Images captured by the camera are associated with metadata indicating the time the image was taken, the specific oven (referenced by index n, etc.), and the location of the leak (referencing the removal side, top surface, and push-in side). It is also possible to use finer-grained coordinates, as shown in Figure 4 for example.

[0111] Coke ovens have been known in the field for decades, and the process in a coke oven is a distillation process. Coke is the main product, but gas is a by-product (untreated coke oven gas). Numerous schematic diagrams are included in the patent literature, and even in IPC class: C10B Decomposition and carbonization of carbonaceous materials for the production of gas, coke, tar or similar.

[0112] For example, Figure 1 of EP1065254B1 (hereinafter referred to as reference '254') shows a perspective view of such a battery having a coke oven. Each oven consists of a carbonization chamber and a combustion chamber and has closures / openings, such as a filling hole (or lid) for filling the material from the top, and doors (for pushing in coke) located at both ends of the carbonization chamber. This reference also shows a vehicle (or "vehicle") and defines the longitudinal direction (where Figure 3 is repeatedly shown).

[0113] The paper by Ghosh et al. is also summarized in Figure 1 (NKGhosh and L. PARTHASARATHY, "Air Pollution Control in Coke Ovens," Clean Technology for the Metallurgical Industry (EWM-2002), January 24-25, 2002, National Institute of Metallurgy (CSIR), Jamshedpur).

[0114] Although terminology may vary from reference to reference, gas leaks can occur in closures (or openings) such as doors (XZ plane, push-in and push-out (door)), filling holes (filling / filling hole / lid, upper XY plane (filling hole)), and risers (upper, see Ghosh et al.).

[0115] During the distillation process, the gas pressure decreases (from a relatively high value to a relatively low value). The relative gas pressure (i.e., overpressure) is set to a predefined setpoint p_set. For example, the setpoint can be given as a relative pressure to atmospheric pressure, and such a setpoint is between 50 Pa and 200 Pa. In other units, this is approximately 5 to 20 millimeters of water column. The gas pressure is adjusted (in combination of the first and second loops) to minimize gas emissions from the door at the start of the distillation process and minimize air intake from the door at the end of the distillation process. In other words, the setpoint changes depending on the process.

[0116] Substances produced by gas emissions can (upon cooling) turn into the tar deposits described above. These deposits are difficult to remove from the surface and shorten the lifespan of the door.

[0117] control loop Having discussed some of the basic concepts of an oven (for example, a coke oven), the explanation will now detail the control loop for a single oven 100.

[0118] This explanation distinguishes the components active in phases **1, **2, and **3** described above. In other words, throughout this explanation, the notations **1 / **2 / **3 represent similar but different components in these phases. Notation **0** (as in Figures 2-4) does not require phase distinction. This rule does not mean that all components are required in all phases: for example, network 253 (Figure 6) is only required during operation **3.

[0119] In this explanation, the phases are divided from the perspective of the neural network by computer 200 (see Figure 2) providing the emission level d(t). From that perspective, training takes place in phase**2. Preparation phase**1 is the phase in which images and observations are collected over time to obtain the training set for the network (see Figure 5, historical data). Training phase**2 is the phase in which the network is being trained, and operation phase**3 is the phase in which the trained network is used as part of the second control loop.

[0120] Figure 4 shows the oven 100, the closure 110, and the visible discharge 120 (i.e., the discharge visible to the camera), and illustrates the discharge rate d(t). In the figure, coordinates Z and X are repeated. Coordinate X can be conveniently further divided into "left," "center," and "right," but the simplified scale is simply for explanatory purposes.

[0121] On the left side of Figure 4, an oven 100 (for example, a coke oven belonging to a battery, see Figure 3) with a closure 110 is shown. In this example, the closure 110 is a door on the XZ plane (i.e., the take-out or push-in side of the oven battery). However, regardless of where the closure 110 is located in the oven 100, it should be kept closed during operation.

[0122] However, there are visible emissions 120, such as gas emissions from oven 100, which are visible to the camera. In this diagram, emissions are symbolized by bold dashed lines. Emissions may occur in (or beyond: originate from) the leak region 130. In this example, the leak region 130 is shown to include the door and the upper part of oven 100.

[0123] As already mentioned, the term "leakage area" refers to the area on the outer surface of the oven where gas emissions visible to a camera may exist. The following further subdivisions of the leakage area are useful:

[0124] (i) The outer surface of the oven where waste may leak, such as around the closure, can be labeled as an "emission source" area.

[0125] (ii) The outer surface of the oven where the discharge is still visible after it has leaked can be labeled as the “discharge distribution” area.

[0126] For the sake of explanation, we can simplify and assume that the "emission source" and "emission distribution" regions are separate. Computer-vision techniques can be used to distinguish between them (see Figure 9 for consideration, if properly trained), and the inherent characteristic of emissions moving away from emission sources (emissions are fugitive) can be used to determine whether or not emissions are present.

[0127] Since leaks are an exception that should be prevented, leak areas should be referred to as "potential leak areas." However, for the sake of simplicity, the explanation will only use the label "leak area."

[0128] In this example, the discharge 120 should originate in the right-hand portion (letter R in X-coordinate) of the leak area 130 near the door. (The diagram is simplified without further distinction between the source and distribution area.) In simplification, the discharge 120 is assumed to be visible to the camera at the location where it originated.

[0129] Figure 4 shows a leak area image 230 in which the door is (at least partially) shown in its central portion and the discharge 230 is shown. Since the discharge is in the image, the display changes from 120 to 230. The figure is simplified and here an image captured under ideal conditions is shown, where virtually all of the discharge is visible in the image captured by the camera. The description will discuss the accuracy (under non-ideal conditions) in Figure 8 (and partially Figure 9) below.

[0130] Figure 4 shows images 232-1, 232-2, 232-m, and 232-M together on the right. The images were taken in preparation phase**1 (as historical data) and have already been annotated in phase**1. Annotations 222-1, 222-2, 222-m, and 222-M are artificial annotations, and here they are labeled with the notation R, L, M (by X coordinate) and "no emission" or "emission present". The # symbol further categorizes "present" as "present as low emission#", "present as medium emission##", and "present as high emission###".

[0131] Annotating images with two main categories (present, absent) and subcategories (low, medium, high for present) is simply for illustrative purposes. The number of categories corresponds to the number of categories that the operating network (253 in Figure 6) can distinguish.

[0132] The collection of leak region images 232-m with annotations 222-m serves as training data (see references marked **2, see Figure 5 for details).

[0133] Inspection procedures are typically standardized. For example, the U.S. Environmental Protection Agency (EPA) has defined an air pollution test method for determining visible emissions (VE) from coke ovens ("Method 303 – By-product coke oven batteries"). For Europe, the German Federal Environment Agency provides a useful reference in the form of "Information sheet on best available technologies in steel production under Industrial Emissions Directive 2010 / 75 / EU of March 2012 [Industrial Emissions Directive 2010 / 75 / EU of the European Parliament and Council of 24 November 2010]". This document also refers to Method 303. Ghosh et al. discuss further criteria.

[0134] Furthermore, the results of such inspections can be used as annotations. While the annotating expert may not need to view the images themselves, the images to be annotated must be captured by the camera during the inspection. Skilled individuals can apply data binding techniques to associate specific annotations with specific images. For example, an expert inspects a specific oven, enters a specific emission level (and other data) into a database, and confirms that the camera captures an image. However, the expert does not need to view this image again.

[0135] Furthermore, image 232-m and annotation 222-m do not need to be created for specific ovens in which the network operates during Phase 3.

[0136] Inside the battery (see 100-BATT in Figure 3), the ovens appear almost identical. Each oven may have an individual label (such as a specific number), and sometimes the name of the battery operator (or manufacturer) may be printed on the surface. The style of such labels or nameplates may vary from battery to battery.

[0137] Such relatively small differences in the appearance of the oven (or battery) do not affect the emissions. Annotating experts (in Phase **1) ignore such differences.

[0138] To avoid overfitting during subsequent training (Phase 2), image preprocessing can be used to match individual labels (etc.) to their positions. For example, if the number is visible in the image, it can be removed.

[0139] training Figure 5 shows training phase 2**2 by neural network 252 being trained to become neural network 253. The annotations function as ground truth.

[0140] Verifying a network after it has been trained requires the expertise (know-how) of a skilled professional. Therefore, details regarding verification will be omitted for simplicity.

[0141] Acquisition of further training data Whenever an ejecta is detected during operation (i.e., during Phase 3), it is classified (by the operator) and can be linked to an image taken during operation. Such annotated images 232-m / 222-m can be used for follow-up training (i.e., a new instance of Phase 2). Over time, the accuracy of ejecta detection improves.

[0142] For a particular battery, it may not yet be possible to obtain enough training data (see Figure 5, images 232-m / 222-m) for the network to classify its degree. This may involve using the network that was initially trained (based on historical data from other batteries) and continuously retraining the network (if historical data from the oven on which network 253 is being applied is available).

[0143] The emissions (used in the training data) may only occur occasionally. This is expected in ovens with relatively new batteries and / or relatively new doors. A new door seal will prevent emissions better than an old door seal.

[0144] From an operational (Phase 3) perspective, this is beneficial. If there is insufficient historical training data (Phase 1 and 2), the network will not be optimally deployed (through training). In other words, due to a lack of battery-specific training data, the network will not be able to improve the accuracy of that particular battery.

[0145] However, there is a simple method to enhance the training data. In the case of a battery 100-BATT with individual pressure control (each oven 100-n, see Figure 3), the pressure can be changed—at least temporarily—until an ejection (the smallest ejection visible to the camera) appears. The ejection will be displayed in the image. Annotation does not need to rely on the image: experts already know that ejection will occur, so they do not need to look at the image. The pressure difference that produces the ejection sample created for this purpose is known. The data describing this pressure difference serves as the degree of ejection. Note that even with the same pressure difference, individual ovens may show different degrees of ejection.

[0146] Two control loops Figure 6 shows the two control loops (from Figure 2) applied to a single oven 103, which includes controllers 173 and 263, a camera 143, and a pre-trained network 253. The layout in Figure 5 is similar to the layouts in Figures 2 and 3, with the setpoint control loop on the left and the pressure control loop on the right. The figure shows an oven in operation, so reference **3 is used.

[0147] The pre-trained network 253 and auxiliary controller 263 can be implemented by the computer 200 (see Figures 2-3). Network 253 implements the discharge classifier (see function (i) above), and controller 263 implements the trigger generator and setpoint generator (functions (ii) and (iii)).

[0148] The oven 103 is illustrated with a door 113 (an example of a closed section) in a side view facing the YZ plane. The oven 103 is significantly simplified by showing only one door. (In Figure 3, this corresponds to doors on both sides.)

[0149] The oven 103 has a pressure sensor 153 that provides a numerical value representing the internal pressure p(t) at any given time t, and a programmable controller 173 (or "programmable logic controller PLC") receives a pressure setpoint p_set. The programmable controller 173 instructs a pressure valve 163 to increase or decrease the pressure inside the oven 103. The configuration comprising the sensor 153, the programmable controller 173, and the valve 163 is known in the art.

[0150] In this embodiment, the controller 173 does not always receive the set value p_set from the controller 263, but only when an ejection is detected (i.e., when a trigger is generated).

[0151] A system with adjustable pressure corresponds to valve 163. An expert can select the optimal system. For example, a system that independently controls the pressure in each oven of a coke battery is commercially available from Paul Wurth SA (32, rue d'Alsace, L-1122 Luxembourg, LUXEMBOURG) under the trademark SOPRECO®. In simplified terms, such a system is part of the discharge pipe, and the discharge orifice is modified to control the gas flow rate through the pipe. Further details are described, for example, in EP2160449B1.

[0152] Gas emissions 123 are released from the oven 103 through leaks (such as from a closed door), and are expected to originate primarily in the leak area 133 on the outer surface of the oven 103. A small portion of the emissions may be released from other areas, but this is irrelevant to the classification of the degree. The emissions 123 are visible to a camera (see the "Camera and Visibility" section).

[0153] Camera 143 captures image 233 from region 133, so most of the waste is visible in the image (at least the data represents the waste). The figure shows one camera, but other cameras can be placed in other leak regions as described in Figure 3 (see cameras 140-ex, 140-top, and 140-push).

[0154] Image 233 arrives at the pre-trained network 253 virtually simultaneously with its capture (i.e., at time t). The signal propagation delay from camera 143 can be ignored. Of course, image 233 arrives without annotation. The pre-trained network 253 provides a degree d(t) for the categories or classes described above (see Figures 3 and 4). Ideally, the accuracy should be the same as that determined by a human expert. (See Figure 8 for an example of achieving higher accuracy using a vector D(t) instead of d(t).)

[0155] Simply put, the auxiliary controller 263 processes d(t) and (if necessary) provides a new pressure setpoint. Instead of receiving d(t), the controller 263 can receive a trigger (see Figure 7). Optionally, the controller 263 may also be implemented by a pre-trained neural network.

[0156] Figure 6 is simplified by not illustrating metadata, and because the method can be used with multiple ovens (such as batteries), experts can use data binding or other techniques to ensure that p_set is applied to the oven from which the image was taken. In other words, data binding ensures that p_set_n is secured and that the image is associated with index n of a particular oven.

[0157] As already explained, the programmable controller 173 (see Figure 6) receives pressure data p(t) from at least one pressure sensor 153 and interacts with the pressure valve 163. This maintains the pressure at the pressure setpoint p_set. The controller 173 operates when the oven 103 is pressurized (see ON in Figure 1).

[0158] Figure 7 shows a flowchart of a computer-implemented method 403 for obtaining the pressure setpoint p_set of a programmable controller 173 associated with the oven 103 and controlling the gas pressure p inside the oven 103.

[0159] The figure shows Method 403 within a dotted rectangle. The double lines indicate that the execution of Method 403 is repeated periodically (while on, see Figure 1).

[0160] Method 403 is shown starting at line 499. Each method execution begins at a point in time labeled “point t” in Figure 1. Method 403 is shown with two alternative ends. On the right, Method 403 ends with step 443, which modifies the setpoint p_set of the programmable controller 173. Often, a new setpoint is chosen so that the pressure decreases gradually, but having a new setpoint may include leaving p_set unchanged.

[0161] On the left side, if the detected emission level is below a predefined threshold (or other identifier), method 403 is repeated, and therefore step 443 is unnecessary. In other words, method 403 terminates with a follow-up activity (step 443) if a trigger is generated, or terminates if no trigger is generated.

[0162] As mentioned above, once method 403 is completed, it is repeated again (starting from point 499). From a computational efficiency standpoint, there may be a waiting period (from the end of 433 / 443 to the start of 499). This waiting period may be related to T2.

[0163] Figure 7 shows method 403 for a single oven, but method 403 can be performed in successive iterations (in parallel or in combination thereof, see Figure 3, which outlines computer 200) for N ovens in a battery (see battery 100-BATT). Method 403 can be multiplexed for multiple ovens in one battery (see Figure 3), and the execution time T1 of the step sequence 413 / 423 for obtaining classification (function (i)) is still sufficiently short.

[0164] A camera 143 installed outside oven 103 acquires a leak region image 233 (step 413). Image 233 shows regions 130 and 133 on the outer surface of ovens 100 and 103 where gas emissions 120 and 123 visible to the camera may be present. As described, regions 130 and 133 are leak regions.

[0165] The computer is equipped with a neural network 253 as an emission classifier, which processes the leak region image 233 using the pre-trained network 253 (step 423) to classify the degree 223 (or d(t), see Figure 6) of the emissions 123 on the leak region image 233.

[0166] As explained in Figure 5, network 253 is trained with training data 232-m and 222-m, which include historical images 232-m taken as reference, and annotations 222-m indicating the degree of artificiality paired with the historical images 232-m.

[0167] Step sequences 413 / 423 are executed during T1 (see Figure 1).

[0168] By applying the predefined rules in step 433, the computer changes the setting value (p_set) for the programmable controller 173 according to the classification degree 223(d(t)) of the visible waste 123 (step 443). As shown in Figure 6, this setting change can be implemented outside the network, within the controller 263.

[0169] In this example, the (new) setting is configured so that the pressure p(t) of oven 103 gradually decreases.

[0170] Predefined rules Generating triggers does not require setting adjustments to occur in all situations. Experienced users can implement the first control loop (see the right side of Figure 2) with additional dimensions. The rules applied in step 433 may include processing additional data. For convenience, the following will be given as an example: Depending on the process conditions inside the oven, changes in pressure can be permitted or blocked. When the process is finished—when the door is opened in a timely manner—it is undesirable to reset the pressure. • There may be predefined limits on the number of times the setpoint can be changed (and the resulting actions involving the pressure valve). Such limits can be implemented to prevent the valve from being subjected to excessive load or operation. The rules can be set to match optimization goals, such as keeping energy consumption low.

[0171] Experienced users may apply the rule in step 433 without further explanation here. However, it should be noted that the accuracy of the rule input data—the degree from step 423 and the correspondence with any alternative measurements (such as visual inspection)—will affect the result of the rule.

[0172] Becoming more accurate Since Method 403 has been described in the context of the control loops of one or more ovens in Figure 7, this description will focus on optional techniques for more precisely changing the pressure setpoint. These optional techniques require obtaining the degree with greater precision (in step 423).

[0173] This method is illustrated in Figure 8 (multiple step instances, rule fitting) and Figure 9 (oven-specific classification).

[0174] Multiple step instances Figure 8 shows a flowchart of a technique in which steps 413 and 423 of method 403 in Figure 7 are performed in multiple instances.

[0175] As already explained in Method 403, a leak region image 233 (see Figure 6) is obtained in step 413, and by processing that image in step 423, a classification by degree d(t) is obtained.

[0176] However, for each execution of both steps 413 and 423, • Classification accuracy, and • Consumption of computing resources, From this perspective, there are at least two competing constraints.

[0177] Considering the limitations of classification accuracy, computers provide classifications that do not always match reality. An ideally functioning computer that correctly classifies images in all situations is not available.

[0178] Taking an arbitrary binary classification as an example, where P is the presence (of the waste) and A is the absence (of the waste), in multiple executions of method steps 413 and 423, the computer will: • True positives (output P corresponds to actual presence in the oven) • False positives (output is P, contrasting with actual absence) • True negatives (corresponding to actual absence in output A), and • False negatives (output A, but actually exist) Outputs.

[0179] As explained, classification (see left side of Figure 7 in step 423) can trigger a change in the setting value (see trigger on the right side of Figure 7 and step 433).

[0180] As a result, false classifications can lead to incorrect changes in settings. In the worst case, the oven may function correctly without any emissions, but start emitting material because the pressure setting is incorrect.

[0181] Looking at the computational constraints, we can improve image acquisition and image processing to increase classification accuracy (for example, increasing the share of "true" values ​​compared to the share of "false" values). However, such improvements would include, for example, • A more sophisticated camera (regarding pixel count, etc.) • Always under appropriate lighting conditions, • A relatively large amount of training images (which may not be available at least initially), and • More computing resources (CPU, memory consumption, etc.) It is required.

[0182] A solution to overcome the constraints is to take different time intervals into account: the time required to perform method steps 413 and 423 (time interval T1) is shorter than the time required to gradually reduce the pressure (time interval T2). Although the flowchart does not have a time scale in its figure, it nevertheless shows that method 403 has a time imbalance as shown in Figure 7.

[0183] K (individual) instances Steps 413 and 423 can be performed with K instances. In other words, a K detection cycle is possible (steps 413 and 423). Figure 8 illustrates such multiple executions with a counter K and counter checks (e.g., counter k=1 to K). It is convenient to sequentially execute K instances by repeating step executions because the total execution time K*T1 of the method is still shorter than T2.

[0184] Furthermore, it is possible to run K instances in parallel (even if they are different cameras).

[0185] Performing steps 413 and 423 on the K instance yields a degree vector D(t) = (d1(t), d2(t), d3(t), ..., dk(t), ..., dK(t)), where dk(t) is the degree of K. Figure 6 shows that the D(t) next to d(t) has a degree of 223.

[0186] In sequential execution, there are different time points K, but all time points are within the iteration interval K*T1. Therefore, the time points can be summarized as a single "t" representing the interval between the first execution (k=1) and the last step execution (k=K).

[0187] As a result, the degree vector D(t) serves as the basis for determining whether a trigger needs to be applied (to change the setting value) (see Figure 7). Having a degree vector D(t) (containing K elements) instead of a single degree d(t) improves accuracy. Of course, some of the K elements are "false" classifications (or other granularities, classifications that would otherwise be incorrect). In some cases, a degree may not even be available.

[0188] The evaluation (for setting a trigger) is performed by applying the trigger rule in step 433'. Step 433' in Figure 8 is an extended version of step 433 (Figure 7), but the result (whether or not a trigger is generated) is the same.

[0189] The trigger rules for step 433' are predefined. For example, in binary classification, a trigger can be set if the majority of the K degree values ​​are P (present). For instance, in vector D(t) = (P,P,P,AA,P,P,P,A,P) with K=10, P is greater than A, and the trigger is set. (Whether to change the pressure setting is determined in step 433, Figure 7.)

[0190] The trigger rule in step 433' can be learned by a machine learning tool (such as a neural network). Training is then performed accordingly.

[0191] For example, steps 413 and 423 need to be repeated for K=100, so the vector D(t) will have 100 elements dk(t). T1 is approximately 3 seconds (image acquisition, sending the image to the computer, processing, etc.), so the total time K*T1 is *100 seconds (or 5 minutes). The seconds and minutes are for illustrative purposes only.

[0192] Since the reaction time (between a change in the control loop setting and the cessation of discharge) is assumed to be K*T1 or greater, it does not matter if the setting is changed a few minutes earlier or later.

[0193] For example, a computer can apply a rule to generate a trigger if at least J=50 (out of 100 vector elements K) is "P" (meaning waste exists).

[0194] This example, with a threshold based on the J / K share, is a significant simplification. Other rules (or further rules) are possible. For example, a trigger could also be generated when the degree of continuity J=20 from dk(t) to d(k+19)(t) is "P".

[0195] Trigger rule adaptation A constraint exists that using K instances does not improve the precision of individual degrees d(t) belonging to the vector D(t). However, trigger rules (step 433') also have precision. Since precision also depends on the visible conditions (i.e., the camera is pointed towards the leak region 130), rules can be created based on visibility.

[0196] The term "visibility," as used herein, refers to the quality by which the leak region image 230 corresponds to reality in the leak region 130. Visibility can be detected and represented in data. In this embodiment, visibility is visibility by an optical camera.

[0197] Skilled individuals can apply sensors and other equipment to collect data describing the visible situation, for example, according to the following: At night, optical cameras must rely on artificial light (such as light from a lantern or a flash during exposure), but during the day, cameras capture images using natural light. • Sunlight conditions vary. Rain or snow can darken the scene (between the camera and the oven). The sky may be cloudy or not. Rain, snow, dust, etc., can hit the objective lens of an optical camera and damage the image.

[0198] As shown in the diagram, step 463 allows the computer to identify a visible condition. As shown by line 473, the computer uses that visible condition to modify the trigger rule.

[0199] The conditions for trigger rules, which generate triggers (to change the setting value), are modified based on visible circumstances.

[0200] The computer compiles these and other visible conditions into a visibility score. For ease of explanation, the score should be a real number between 0 and 1, ranging from "not visible at all" to "best condition." If the score is zero, the image will be black, and the method will fail.

[0201] To simplify the explanation, the visibility score—which is determined by a computer—can be categorized into "good visibility" and "poor visibility."

[0202] Focusing on the above example of K=100, in the case of "good visibility," if J=30 (out of the K=100 vector elements) is "P" (discharge exists), it is sufficient to generate a trigger. In the case of "poor visibility," if J=70 (again out of 100) is "P," a trigger must be set.

[0203] Further aspects of trigger rule compliance In the following discussion on rule compliance, there is no need to distinguish between cameras, and the discussion will cover optical cameras and thermal imaging cameras.

[0204] Ovens are assumed to have been in operation for decades, with some having been in operation for over 40 years, and even over 60 years. Simply put, older ovens produce more waste than younger ones. Optionally, elapsed operating time can be used as a rule modifier.

[0205] In relation to Figure 7, the explanation already mentioned that further data can be processed in rule 443 (for changing the setting value). Similarly, further data can be used in trigger rules. For example, the probability of emissions being produced may differ depending on the process phase. An oven starting a new process is more likely to produce "smoke" than an oven that is about to finish its process (i.e., when coke is almost ready).

[0206] Repeated adaptation This rule can be met by matching the iteration rate K (number of instances, instance cardinality), which is represented by line 483. For example, poor visibility increases the number of iterations (relatively high K), while good visibility decreases the number of iterations (relatively low K).

[0207] Oven-specific classifications Figure 9 shows leak region images (A), (B)...(E) in a symbolic view, illustrating further aspects to which the neural network 253 can be selectively applied to obtain classification. Visible ejecta exhibit specific behaviors (visible with an optical camera), and the neural network can be trained to such behaviors. Training network 252 to become network 253 is described in Figure 5, and additional or alternative features can also be trained by annotation or unsupervised. In other words, the models described below can be implemented by networks 252 / 253.

[0208] It is assumed that a semantic segmentation model will be trained to classify regions within an image. For example, in symbolic illustrations (e.g., (A), (B)), a thick dotted line indicates the waste product. This connects to the image area representing the waste product. ·For example, the symbolic illustration divides walls and other external elements of the oven with thin vertical lines (e.g., (A), (B), (C)). Since the wall does not move, it does not move in the image either. With multiple images in the training set, the network learns where the wall is even if some images are covered with emissions.

[0209] With semantic segmentation, it is also possible to divide the leakage area into the emission source and emission distribution areas, which is possible not only with semantic segmentation but also with the method described next.

[0210] With semantic segmentation, the direction can be specified by "arranging" the visible emissions. Display (C) shows the emissions with a single line that slightly increases (showing an upward trend: slightly increasing) and also shows emissions in some form of turbulence (more random directions caused by the influence of weather, etc.).

[0211] It is assumed that a plurality of images taken at consecutive time points are processed. As already explained, there is available time (T1 < T2) within (T2). Since the emissions usually move and walls etc. do not move, a properly trained network can identify the moving elements in the image. In the example, display (D) shows the moving emissions in terms of the temporal aspect (during Δt) and the spatial aspect (displacements Δx, Δz corresponding to the position of the oven in pixel units, refer to the coordinates in Figure 4). In other words, the speed is detectable.

[0212] It is assumed that a plurality of images taken at consecutive time points are processed considering not only the speed but also the speed change.

[0213] The density of the emissions decreases while moving away from the emission source. If the network is made to learn to see the wall through the "cloudiness" as in (A), the density can be evaluated. Some wall structures (e.g., the door itself, door frame, soffit beam, or beam) are "strong candidates" for the location where emissions are expected.

[0214] Movement (in terms of velocity and velocity changes) is not equal on the image (see density loss, etc.), and the network can derive images similar to histograms. As the material moves from the source, not only the density but also the velocity of the emissions decreases. Such phenomena can also be identified from derived images.

[0215] Methods for evaluating images, such as optical flow detection, are discussed below. • Dileep K. Appana, Rashedul Islam, Sheraz A. Khan, Jong-Myon Kim: Video-based smoke detection using smoke flow patterns and spatiotemporal energy analysis for alarm systems (Information Sciences, volume 418-419, December 2017, pp. 91-101; doi.org / 10.1016 / j.ins.2017.08.001). • Jinkyu Ryu and Dongkuri Kwak: A study on complex fire and smoke detection methods using computer vision detection and convolutional neural networks, 2022 (Fire 2022, 5, 108K; mdpi.com / 2571-6255 / 5 / 4 / 108).

[0216] Advanced image processing (as illustrated in Figure 9) becomes even more feasible with available time, and as explained, the computer has the potential to execute K instances during K*T1 if the time interval is long enough (shorter than T2).

[0217] The implementation of advanced image processing can be considered a further embodiment of performing step 433 or 433' (see Figures 7-8).

[0218] Since ovens are not always located under a roof, precipitation such as snow may appear in the images (see symbol (E)). However, snow falls from clouds and moves from top to bottom (coordinate z here), and may also move in another direction (x) related to the wind. The network is intended to be trained to detect such precipitation.

[0219] The detection of such or similar events has the potential to lead to the following: • Precipitation detection is performed as step 463 "Identifying Conditions," and the rules are modified (explained in Figure 8) and / or the number of instances K is changed based on the results. • To prevent the oven from malfunctioning, stop the interaction with controller 173 (so that the pressure setting does not change). • Inform the operator that the setting value will be temporarily disabled depending on the type of waste being discharged.

[0220] Further details The explanation then explores details that are generally applicable.

[0221] Figure 3 shows a battery equipped with multiple ovens and a track for moving cameras. Reference '254' mentioned above shows a more detailed battery, including a coal-filling vehicle that moves over the coke oven (see the upper track in Figure 3) and guide vehicles (or general vehicles) positioned on the removal and pushing sides (see the tracks on both sides). The vehicles are primarily responsible for loading and unloading materials into the ovens, but they can also serve a secondary role in carrying cameras (140-ex, 140-top-140-push, etc.).

[0222] Using existing cars (or vehicles) can be advantageous because cars are controlled to reach a specific oven. In other words, existing battery-powered navigation systems can be synergistically reused.

[0223] Classifying images by their degree of output is not the same as recognizing numbers or letters, for example.

[0224] Historical images 232-m can be collected under various lighting conditions, including daylight, nighttime, and rainy weather. As the sun moves, images are taken at different times of the day.

[0225] As mentioned above, the historical image 232-m (if annotated) is applied to train network 252 (see Figure 5), enabling the network to perform method 403. The same image—and therefore properly annotated—can be used to train network 272.

[0226] If oven elements (such as doors) look different from the outside, for example, if the door on the battery removal side and the door on the push-in side look different (see Figure 3), using two sets of historical images (for each side) can allow for more accurate classification of the degree of difference.

[0227] Activities such as installing camera 140 and computer 200, and operating method 403, do not substantially interfere with the operation of the oven. As already explained, the only additional activity is to enable controller 173 to receive the set value p_set.

[0228] Experts are familiar with the automation layer. It is also possible to install Computer 200 as a so-called Level 2 automation module.

[0229] Compared to human inspection, Method 403 can be performed at any time when one or more ovens are operating (e.g., while ON in Figure 1). This allows for continuous monitoring of emissions. This technique can also reduce the number of human inspections and mitigate health risks to inspectors.

[0230] It is useful to take more data into consideration: if a door is recognized as being more susceptible to the effects of discharge than other doors, it can be classified accordingly. The rules mentioned above—such as trigger rules 433 / 433'—can be adapted to step 473 by taking the door category into account.

[0231] In scenarios where the setpoint is changed until gas is released from the oven, the magnitude of the pressure difference (before and after the change) relative to the degree of release (detected in steps 413, 423, or during inspection) becomes a quality indicator for that particular oven. Even with the same pressure difference acting, some doors may leak more easily than others. These indicators can also be used as further inputs in step 473.

[0232] General-purpose computer Figure 10 shows an example of a general-purpose computer device that can be used with the technology described herein. Figure 10 shows an example of a general-purpose computer device 900 and a general-purpose mobile computer device 950, which can be used with the technology described herein. The arithmetic unit 900 is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers. The general-purpose computer device 900 may correspond to the computer system 200 in Figure 3. The arithmetic unit 950 is intended to represent various forms of mobile devices, such as personal digital assistants, mobile phones, smartphones, driver assistance systems, or vehicle board computers, and other similar arithmetic units. For example, the arithmetic unit 950 may be used as a front-end for a user (e.g., a blast furnace operator) to interact with the arithmetic unit 900. The components shown herein, their connections and relationships, and their functions are illustrative only and are not intended to limit the practice of the inventions described and / or claimed in this document.

[0233] The arithmetic unit 900 comprises a processor 902, memory 904, storage device 906, a high-speed interface 908 connecting memory 904 to a high-speed expansion port 910, and a low-speed interface 912 connecting a low-speed bus 914 to storage device 906. Each component 902, 904, 906, 908, 910, and 912 is interconnected using various buses and can be mounted on a common motherboard or in other ways as needed. The processor 902 can process instructions to be executed within the arithmetic unit 900, including instructions stored in memory 904 or storage device 906 for displaying graphical information for a GUI on an external input / output device such as a display 916 coupled to the high-speed interface 908. In other implementations, multiple processors and / or multiple buses may be used, along with multiple memories and multiple types of memory as needed. Also, multiple arithmetic units 900 are connected, and each device provides a portion of the required operation (e.g., as a server bank, a group of blade servers, or a multiprocessor system).

[0234] Memory 904 stores information within the arithmetic unit 900. In one implementation, memory 904 is a volatile memory unit or unit(s). In another implementation, memory 904 is a non-volatile memory unit or unit(s). Memory 904 may also be in another form that is computer-readable, such as a magnetic disk or an optical disk.

[0235] The storage device 906 can provide large-capacity storage to the arithmetic unit 900. In one implementation, the storage device 906 is or may include a computer-readable medium, including a set of devices such as a floppy disk device, a hard disk device, an optical disk device, or a tape device, flash memory or other similar solid-state memory device, or a device in a storage area network or other configuration. A computer program product can be specifically embodied in an information carrier. A computer program product may include instructions that perform one or more of the above-described methods at runtime. The information carrier is a computer or machine-readable medium, such as memory 904, storage device 906, or memory on the processor 902.

[0236] The high-speed controller 908 manages the bandwidth-intensive operation of the arithmetic unit 900, while the low-speed controller 912 manages the low-bandwidth-intensive operation. Such function assignments are illustrative only. In one implementation, the high-speed controller 908 is coupled to memory 904, a display 916 (e.g., via a graphics processor or accelerator), and a high-speed expansion port 910 that can accept various expansion cards (not shown). In this implementation, the low-speed controller 912 is coupled to the storage device 906 and the low-speed expansion port 914. The low-speed expansion port may include various communication ports (e.g., USB, Bluetooth®, Ethernet, Wireless Ethernet) and may be coupled to one or more input / output devices (devices) via a keyboard, pointing device, scanner, or network device such as a switch or router, e.g., a network adapter.

[0237] The arithmetic unit 900 can be implemented in several different forms, as shown in the figure. For example, it can be implemented as a standard server 920, or multiple times within a group of such servers, or as part of a rack server system 924. It can also be implemented in a personal computer such as a laptop computer 922. Alternatively, components from the arithmetic unit 900 can be combined with other components in a mobile device (not shown), such as device 950. Each such device may contain one or more arithmetic units 900, 950, and the entire system may consist of multiple arithmetic units 900, 950 communicating with each other.

[0238] The arithmetic unit 950 includes a processor 952, input / output devices such as memory 964 and display 954, a communication interface 966, a transceiver 968, and other components. Device 950 may also include storage devices such as a microdrive or other devices to provide additional storage. The respective components 950, 952, 964, 954, 966, and 968 are interconnected using various buses, and some of the components may be mounted on a common motherboard or in other suitable ways as needed.

[0239] The processor 952 can execute instructions in the arithmetic unit 950, including instructions stored in memory 964. The processor may be implemented as a chipset of a chip including multiple separate analog and digital processors. The processor may provide, for example, control of other components of device 950, such as user interface control, applications run by device 950, and coordination of wireless communication by device 950.

[0240] The processor 952 can communicate with the user via a control interface 958 and a display interface 956 coupled to the 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 suitable display technology. The display interface 956 may include suitable circuitry for driving the display 954 to present graphical and other information to the user. The control interface 958 may receive commands from the user and translate them for transmission to the processor 952. Furthermore, an external interface 962 may be provided to enable short-range communication between device 950 and other devices in communication with the processor 952. The external interface 962 may provide, for example, wired communication in one implementation or wireless communication in another, and multiple interfaces may also be used.

[0241] Memory 964 stores information within the arithmetic unit 950. Memory 964 can be implemented as one or more of a computer-readable medium or media, a volatile memory unit or unit(s), or a non-volatile memory unit or unit(s). Extended memory 984 is provided to and can be connected to device 950 via an extended interface 982, and may include, for example, a SIMM (Single In Line Memory Module) card interface. Such extended memory 984 may provide additional storage space for device 950, or may also store applications or other information for device 950. Specifically, extended memory 984 may include instructions for executing or supplementing the processes described above, and may also include secure information. Therefore, for example, extended memory 984 may function as a security module for device 950 and may be programmed with instructions that enable the secure use of device 950. Furthermore, a secure application may be provided via the SIMM card, along with additional information such as placing identification information on the SIMM card in a hack-proof manner.

[0242] The memory may include, for example, flash memory and / or NVRAM memory, as described below. In one implementation, a computer program product is concretely embodied in an information carrier. The computer program product includes instructions that perform one or more of the above-described methods at runtime. The information carrier is a computer or machine-readable medium, such as memory 964, extended memory 984, or memory on processor 952, which can be received, for example, via transceiver 968 or external interface 962.

[0243] Device 950 can communicate wirelessly via a communication interface 966 and may include digital signal processing circuitry as needed. The communication interface 966 can provide communication under various modes or protocols, such as GSM voice calls, SMS, EMS, or MMS messaging, CDMA, TDMA, PDC, WCDMA®, CDMA® 2000, or GPRS. Such communication may occur, for example, via a radio frequency transceiver 968. In addition, short-range communication may occur, such as using Bluetooth®, WiFi, or other such transceivers (not shown). Furthermore, a GPS (Global Positioning System) receiving module 980 may provide device 950 with additional navigation and location-related radio data, which may be used as appropriate by applications running on device 950.

[0244] Device 950 can also communicate audibly using the audio codec 960, which can receive voice information from a user and convert it into usable digital information. The audio codec 960 can similarly generate audible sound for the user, for example, through a speaker in the handset of device 950. Such sound may include sounds from a voice phone, recorded sounds (e.g., voice messages, music files, etc.), and sounds generated by applications running on device 950.

[0245] The arithmetic unit 950 can be implemented in several different forms, as shown in the figure. For example, it can be implemented as a mobile phone 980. It can also be implemented as part of a smartphone 982, a personal digital assistant, or other similar mobile device.

[0246] The various implementations of the systems and technologies described herein can be realized in digital electronic circuits, integrated circuits, specially designed ASICs (application-specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs executable and / or interpretable on a programmable system, which includes at least one programmable processor, a storage system, at least one input device, and at least one output device coupled together for receiving and transmitting data and instructions, for special or general purposes.

[0247] These computer programs (also called programs, software, software applications, or code) contain machine instructions for a programmable processor and can be implemented in high-level procedural and / or object-oriented programming languages ​​and / or assembly / machine language. The terms “machine-readable medium” and “computer-readable medium” as used herein refer to any computer program product, apparatus and / or device (e.g., magnetic disks, optical disks, memory, programmable logic devices (PLDs)) used to provide machine instructions and / or data to a programmable processor, and include machine-readable mediums that receive machine instructions as machine-readable signals. The term “machine-readable signal” refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0248] To provide user interaction, the systems and technologies described herein can be implemented on a computer having a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) and a keyboard and pointing device (e.g., a mouse or trackball) to which the user can provide input to the computer. User interaction can also be provided using other types of devices. For example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback). User input can also be received in any form, such as acoustic input, voice input, or tactile input.

[0249] The systems and technologies described herein may be implemented in a computing device that includes backend components (e.g., data servers), middleware components (e.g., application servers), or frontend components (e.g., a graphical user interface, or a client computer with a web browser that allows a user to interact with an implementation of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system may be interconnected by digital data communication (such as a communication network) in any form or medium. Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the internet.

[0250] A computing unit can include a client and a server. Clients and servers are typically geographically separated and usually interact via a communication network. The client-server relationship arises from computer programs running on each computer that have a client-server relationship with one another.

[0251] Several embodiments are described. However, it will be understood that various modifications can be made without departing from the characteristics and scope of the present invention.

[0252] In addition, the logic flow shown in the figure does not require a specific order or sequence to obtain the desired result. Furthermore, other steps can be provided or removed, and other components can be added to or removed from the described system. Therefore, other embodiments are within the scope of the following claims.

Claims

1. A computer implementation method (403) for obtaining a pressure setpoint (p_set) for a programmable controller (173) associated with an oven (103) and controlling the gas pressure inside the oven (103), that is, a programmable controller (173) that receives pressure data (p(t)) from at least one pressure sensor (153) and interacts with a pressure valve (163), A camera (143) installed outside the oven (100, 103) acquires a leak region image (233) showing the area on the outer surface of the oven (100, 103) (130, 133), which is hereafter referred to as the leak region (130, 133), and in which gas emissions (120, 123) may be present (413); To classify the degree (223, d(t)) of the discharge (123) in the leak region image (233), the leak region image (233) is processed by a pre-trained network (253), where the network (253) is trained with training data (232-m, 222-m), which includes a historical image (232-m) taken as a reference and an artificial degree annotation (222-m) paired with the historical image (232-m); Apply predefined rules (433, 433') and change the setting value (p_set) for the programmable controller (173) according to the degree of classification (223, d(t)) of the waste (123) (443). The above method, including the above.

2. The method according to claim 1, wherein the step of acquiring a leak region image (233) (413) is performed by acquiring a leak region image from an optical camera (143), and the leak region image (233) shows a region (130, 133) in which visible gas emissions (120, 123) may be present.

3. The method according to claim 1, wherein the step of acquiring a leak region image (233) (413) is performed by acquiring a leak region image from a thermographic camera (143), and the leak region image (233) indicates a region (130, 133) in which gas emissions (120, 123) that can be recognized by a temperature gradient toward the background of the leak region image (233) may exist.

4. The method (403) according to any one of claims 1 to 3, wherein in the processing step (423), a pre-trained network (253) classifies the degree of discharge (223) into two binary categories: a first degree of no discharge and a second degree of discharge.

5. The method (403) according to claim 4, wherein in processing step (423), a pre-trained network (253) classifies the second degree into a plurality of subclasses.

6. The method (403) of claim 5, wherein in processing step (423), a pre-trained network (253) classifies the second degree into one of the following subclasses: present as low emission, present as medium emission, and present as high emission.

7. The method (403) according to any one of claims 1 to 6, wherein the steps of acquiring (413) a leak region image (233) and processing (423) the leak region image (233) are performed in multiple instances (K) within a time interval (K*T1) shorter than the time interval (T2) required for the controller (173) to actually change and stabilize the gas pressure (p(t)) to a set value inside the oven (103) through interaction with the pressure valve (163).

8. The method (403) according to claim 7, wherein the step is performed in multiple instances (K) to obtain multiple leak region images (233), and a processing (423) step is performed individually for each instance of the multiple leak region images (233) to obtain multiple degrees represented by a degree vector (D(t)).

9. The method according to claim 8 (403), which includes applying a predefined rule (433, 433') to evaluate the degree vector (D(t)) according to the degree distribution within the degree vector (D(t)).

10. The method according to claim 9 (403), wherein evaluating the degree vector (D(t)) according to the distribution includes either identifying the share of emission present and non-emission between two degree values, or identifying the change rate between the degree values.

11. The camera (143) further includes identifying (413) a situation in which it has acquired (413) multiple leak region images (233), which will be referred to below as a visible situation. The situation is as follows: - The quality of light, whether it is natural light or artificial light. - Whether or not it rains in the scene between the camera and the oven. - Whether or not there is dust on the camera's objective lens. Select from any of the following, and The application of a predefined rule (433') for evaluating the degree vector (D(t)) is performed by adapting the predefined rule (433') according to the visible circumstances. A method used in an optical camera according to any one of claims 1 to 10 (403).

12. The visible conditions are as follows: • Light intensity in the leakage region, - Distinguishing daylight from artificial light, the characteristics of light (light properties) - Characteristics of light by distinguishing sunlight from moonlight. • The quality and quantity of precipitation that reaches the oven, - Detection of meteorological precipitation phenomena, The method according to claim 11 (403), which is indirectly identified (463) by evaluating data representing the camera environment, selected from (463).

13. The visible conditions are identified by processing the leak region image (463), the method according to claim 11 (403).

14. The processing (423) of leak region images (233) by a pre-trained network (253) is performed by a network trained on either: (1) training data (232-m, 222-m) including historical images (232-m) taken as reference from an oven (101), or (2) training data (232-m, 222-m) including historical images (232-m) taken as reference from a physically different oven, according to any one of claims 1 to 13 (403).

15. The method (403) according to any one of claims 1 to 14, wherein the acquisition (413) of leak region images (233) is performed for a plurality of ovens (100-n) of a battery (100-BATT), and a camera (143) is mounted on a vehicle whose primary purpose is to transport material between the plurality of ovens, and the vehicle moves together with the camera.

16. A computer system (200) for obtaining a pressure setpoint (p_set) for a programmable controller (173) associated with an oven (103) and for controlling the gas pressure (p) inside the oven (103), The programmable controller (173) receives pressure data (p(t)) from at least one pressure sensor (153) and interacts with the pressure valve (163). A computer system adapted to include a module for performing a computer implementation method (403) as described in any one of claims 1 to 15.

17. Use of a computer implementation method (403) according to any one of claims 1 to 15 for obtaining a pressure set value (p_set) for controlling the gas pressure (p) inside an oven (103) selected from a coke oven, a furnace, iron and steel equipment, part of a steelmaking facility, a cement reactor, a concrete reactor, and a chemical reactor.

18. A computer program product that, when stored in the memory of a computer and executed by at least one processor of the computer, causes a computer to perform a step of any of the methods of claims 1 to 15.

19. A computer implementation method for classifying the degree of emission (d(t), D(t)) of the exhaust (122) in the leak region image (232) of the oven (102) by training a network (252) to process a leak region image (232) showing areas (130, 133) on the outer surface of the oven (100, 102) from which gas exhaust (120, 122) can be released from the oven (100), wherein the method classifies the degree of emission (d(t), D(t)) of the exhaust (122) in the leak region image (232) of the oven (102), Connecting a historical image (232-m) obtained by changing the oven's pressure setting by a specific pressure difference until the oven emits gas visible to the camera to the input of the network (252); capturing an image showing the oven and gas visible to the camera as a historical image (232-m); and By using observations of specific emission levels, artificial annotations (222-m) including emission classes as ground truth are connected to historical images (232-m) at the network output. A method that includes this.

Images