Method for monitoring the operation of a direct reduction plant, associated electronic monitoring device and reduction plant

Abstract

The instant technology concerns, inter alia, a method for monitoring the operation of a direct reduction plant (1) that comprises a direct reduction shaft furnace (10), the method comprising: s1: acquiring an image (Img) of an ensemble of pellets (8) output by the shaft furnace, s2: analysing said image using an algorithm suitable for detecting one or more clusters (9) of pellets and, optionally, one or more pieces of shaft internal crust, and s3: emitting an obstruction-risk signal if one or more clusters or pieces of crust are detected in step s2.

Description

Method for monitoring the operation of a direct reduction plant, associated electronic monitoring device and reduction plant

[0001] The technical field is that of direct reduction of iron, and industrial installations thereof.

[0002] A widespread method for producing steel is to reduce iron oxides in a blast furnace using coke. But this manufacturing route releases significant quantities of CO2, both for producing coke from coal and for producing hot metal in the blast furnace.

[0003] Another manufacturing route is based on so-called “direct reduction methods”. Among them are methods, developed in particular by the companies Midrex Technologies and Energiron, in which iron ores are reduced in a shaft furnace, without coke, using a reducing gas to produce Direct Reduced Iron or DRI (also called indifferently DRI pellets, reduced iron ore, and pellets of reduced iron ore, in this document). DRI then usually undergoes further processing in an electric arc furnace to produce liquid steel.

[0004] In such a method, pellets of iron ore are charged to the top of the shaft furnace and then go down through the shaft furnace, traversing the reducing gas and being reduced to DRI pellets. The pellets are then discharged at an outlet of the furnace shaft located at its bottom. The reducing gas mainly contains CO and H2. The furnace shaft typically comprises an upper, reduction section where most of the iron ore reduction takes place, and a lower section for product homogenization, and possibly for iron carburization and cooling. The lower section typically has a lower part in a conical shape that gradually narrows and ends at the shaft outlet. The direct reduction route allows for substantial reductions of CO2 emissions, compared to blast furnace routes. Yet, obstructions of the furnace shaft regularly happen, requiring to stop the production for days to empty the shaft and unblock its outlet.

[0005] In this context, a method according to claim 1 is provided.

[0006] During the operation of the direct reduction shaft furnace, pellets may clusterize, that is stick to each other and agglomerate. Such clusters can be small (for instance less than 5 pellets) or may be re-broken into pellets easily. In these cases, they are evacuated at the shaft outlet without causing any obstruction. But the pellets clusters may become bigger or more solid, for instance due to a temperature increase in a reduction section of the shaft, and then obstruct the shaft outlet.

[0007] In this method, the material output by the shaft is inspected, by computer vision, to determine if there a risk of obstruction of the shaft based on the presence of one or more clustersin the output material, and, possibly, based on characteristics of the detected cluster(s) (dimension, number, nature....). This allow for detecting that cluster formation in the shaft is significant before the clusters are too big or too solid to be evacuated at its outlet.

[0008] The algorithm for analysing the image of the ensemble of pellets may be suitable also for detecting the presence of one or more pieces of shaft internal crust. In this case, the risk of obstruction is determined based on the presence of: one or more clusters or one or more pieces of crust, in the material output by the shaft furnace. Indeed, shaft obstruction is sometimes caused by pieces of crust fragmented and detached from the shaft wall and during the operation of the shaft furnace. This crust is a crust of agglomerated iron ore and possibly other materials forming on the inner side of the shaft wall. When big, such pieces of fractured crust can block the shaft outlet (and cause damages). Detecting them in the output material allows for detecting that the crust fractures and detaches from the wall, in the shaft, before pieces too big to be evacuated detach from the wall.

[0009] The instant technology also concerns a monitoring method in which the algorithm for analysing the image of the ensemble of pellets is suitable for detecting one or more pieces of crust (by not necessarily for detecting clusters of pellets), the obstruction-risk signal being emitted if such a piece of crust is detected.

[0010] Other parameters (for instance: reducing gas input or output pressure, cluster breaker torque) may be employed, in combination to information related to cluster(s) or piece(s) of crust detection, to determine if there is a risk of obstruction.

[0011] When it is determined that there is risk of obstruction, one or more actions may be taken to mitigate this risk, in particular to reduce clusters formation in the shaft furnace. The direct reduction plant may thus be controlled to adjust its operating conditions so that clusters formation is reduced, for instance by reducing a temperature within in a reduction region of the shaft furnace (reduced compared to a previous, instant value of that temperature, or compared to a normaloperation value of that temperature, or compared to a value entered by an operator). This temperature decrease within the shaft furnace is for instance obtained by reducing the temperature of the reducing gas injected in the shaft, or by adjusting a content of exothermal and / or endothermal compounds in the reducing gas.

[0012] It is however noted that decreasing the temperature within the reduction region usually causes a decrease of the efficiency of the iron ore reduction in the shaft, and thus reduces the process yield. It is thus beneficial to be able, like here, to detect risks of shaft obstruction, as it allows for reducing this temperature only when it is useful for avoiding a shaft obstruction, and to otherwise use a temperature value set to optimize the ore reduction efficiency.

[0013] The method according to the invention may comprise one or several additional features, defined in claims 2 to 19, considered alone or in combination.

[0014] The instant technology also concerns a method for producing reduced iron according to claim 20, an electronic monitoring system according to claim 21 , and a direct reduction plant according to claim 22. The instant technology also concerns a computer program whose execution on a computer (possibly connected to relevant sensors, actuators or controller), makes the computer to execute the method for monitoring presented above. The instant technology also concerns a non-transitory computer-readable medium storing such a computer program.U i : ! b L : > I > U • •

[0015] The instant technology will now be described in more detail and illustrated by examples without introducing limitations, with reference to the appended figures.

[0016] Figure 1 is a schematic representation of a direct reduction plant.

[0017] Figure 2 is a schematic representation of an electronic monitoring system of the plant of figure 1 .

[0018] Figure 3 schematically represents steps of a method for monitoring the plant of figure 1 .

[0019] Figure 4 is an image of an ensemble of pellets output by a shaft furnace of the plant of figure 1 , this image being acquired and analysed for monitoring the operation of the plant.

[0020] Figure 5 is another such image.

[0021] Figure 6 schematically represents results of an analysis of the image of figure 5.

[0022] Figure 7 is a histogram relative to items detected in the image of figure 6.

[0023] Figure 8 represents a dimension D of different clusters detected in the image of figure 5.

[0024] Figure 9 represents another image of an ensemble of pellets output by a shaft furnace of the plant of figure 1 , acquired and analysed for monitoring the operation of the plant.

[0025] Figure 10 schematically represents results of a segmentation of the image of figure 9.

[0026] Figure 1 1 is a histogram of a dimension the items resulting from said segmentation, represented in figure 10.

[0027] As above mentioned, the instant technology concerns, inter alia, a method comprising: inspecting the material output by a direct reduction shaft furnace, by computer vision, to detect one or more clusters of pellets, or one or more pieces of shaft internal crust, and determining if there is a risk of obstruction of the direct reduction shaft furnace based on this material inspection.

[0028] Optionally, this method comprises controlling the operation of the shaft furnace so as to reduce cluster formation in the furnace, when it is determined that there is a risk of obstruction.

[0029] A direct reduction plant according to the instant technology is described first, with reference to figures 1 and 2. A method for monitoring this plant, according to the instant technology, is described then with reference to figures 3 to 1 1 .Direct reduction plant

[0030] Figure 1 schematically represents a direct reduction plant 1 that comprises a direct reduction shaft furnace 10 (also called indifferently the shaft furnace, the furnace or the shaft, in this document) and a gas processing and conditioning unit 20 for feeding the shaft furnace 10 with a reducing gas 21 .

[0031] The direct reduction plant 1 also comprises:- a conveyor 2, here a belt conveyor, for evacuating pellets 8 of reduced ore output by the shaft furnace 10,- an image acquisition device, here a digital camera 3, arranged so that it can acquire images of the pellets 8 evacuated by the conveyor 2,- an electronic monitoring system 5 (figure 2).

[0032] The shaft furnace 10 comprises, from top to bottom:- a charging device 1 1 for charging iron ore that contains oxidized iron, such as hematite or magnetite;- an upper, reduction section 12, where most of the iron ore reduction takes place,- a lower section 13, for product homogenization and possibly for iron carburization and for cooling,- an outlet 14 of the shaft furnace, located at the bottom thereof, where the reduced iron ore is discharged.

[0033] The lower section 13 has a lower part that gradually narrows and ends at the shaft outlet 14. Here, the lower part of the lower section 13 has a conical shape.

[0034] The iron ore is charged in the shaft furnace from the top, in the form of pellets, and then travel through the shaft 10, mostly by gravity. It first traverses the reduction section 12, where the iron ore is reduced by the reducing gas 21 which is injected in the shaft furnace and typically flows in counter-current to the descending iron ore. Then, the iron ore traverses the lower section 13 and it is then discharged, in the form of pellets of reduced iron ore, at the outlet 14.

[0035] As represented in figure 1 , the reducing gas 21 is injected in the furnace shaft at the bottom of the reduction section 12, and, after reacting with the iron ore, it exits the furnace shaft at its top in the form of a top gas 22.

[0036] The main component of the reducing gas 21 is dihydrogen H2 or carbon monoxide CO. In practice, it usually contains both H2 and CO. It is injected at high temperature in the shaft, typically between 750°C and 1200°C.

[0037] In the example of figure 1 , the top gas 22 is processed in a processing unit 27 to produce a process gas 23. The processing unit 27 comprises for instance one or more a scrubbers for water removal, a gas compressor, and a splitter for diverting part of the gas (the diverted part being used as a fuel gas in a reformer 28). The process gas 23 is then mixed with natural gas (or with a methane-rich gas) to produce a methane containing gas 24 input in the reformer 28. The reformer is typically a catalytic reformer where the methane CH4 reacts with carbon dioxide CO2 and with water H2O to produce dihydrogen H2 and carbon monoxide CO. The reformer outputs a reformed gas 26, composed mainly of dihydrogen H2 and carbon monoxide CO and whose temperature is typically above 800°C, or even above 900°C. One or more additional gas supplies may be connected to the reformer 28, for instance to supply water in an adjustable manner, or to supply dihydrogen H2, heated in adjunct tube(s) of the reformer and then added to the reformed gas (to increase its H2 content).

[0038] The piping connecting the reformer 28 to the shaft furnace, for conveying the reformed gas 26 to the shaft furnace, may, like here, comprise an inlet 29 for adding controllable amounts of methane CH4 and / or of oxygen to the reformed gas 26. The pipping in question may also comprise a derivation, to allow the reformed gas, or part of it to flow on demand through an optional gas cooler 31. Once conveyed to the shaft furnace, and with the possible additional O2 / CH4, the reformed gas forms the reducing gas 21 injected in the shaft furnace.

[0039] The camera 3 is positioned so as to visualize the conveyor belt, from above (while being possibly tilted), to acquire images of the material output by the shaft furnace 10. The field of view F of the camera may encompass the whole width over which the material is distributed, on the conveyor belt (to be able to inspect most of the material output by the shaft furnace). The width and / or length of the zone visualized by the camera 3 may be from 10 times to 200 times a pellet diameter (average diameter), or from 15 to 100 times a pellet diameter. It may be from 10 cm to 1 .5 m. The camera resolution allows for acquiring images with at least 5 pixels per pellet diameter, or even with at least 10 pixels per pellet diameter. The camera may be controlled so as to acquire images with a frame high enough to inspect the material output by the shaft furnace with no cutoffs regarding the length of material inspected: for instance, if the field of view F covers a length of 0.75 m while the speed of the conveyor belt is 1 .5 m / s, the camera is controlled so as to acquire, and transfer images with a frame rate of 2 images per second at least.

[0040] The camera 3 is operatively connected to the electronic monitoring system 5.

[0041] The electronic monitoring system 5 comprises at least a processor and a memory.

[0042] It is configured, for instance programmed to implement the method for monitoring the direct reduction plant 1 described further below. This method may comprise controlling the direct reduction plant 1 , based on the camera-based inspection of the material output by the shaft furnace, so as to prevent or at least limit shaft obstructions by clusters of pellets or by pieces of fractured crust.

[0043] The electronic monitoring system 5 comprises one or more data interfaces (like network or communication cards or chips) for receiving and emitting data and / or signals, in particular for receiving the images Img acquired by the camera 3, and possibly for receiving data or signals acquired by other sensors 4 of the direct reduction plant like pressure, flow rate or temperature sensors, communicating with a human-machine interface 6,- sending setpoints values to a low-level controller of the direct reduction plant, or sending control signals to actuators of the direct reduction plant.

[0044] As represented in figure 2, the electronic monitoring system 5 comprises a monitoring module 51 and a control module 52. The monitoring module is configured for analysing the image Img transmitted by the camera, to detect the presence of one or more pellets clusters or of one or more pieces of crust, and optionally to characterize them. In the embodiment described here, the monitoring module is further configured for determining if there is a risk of obstruction of the shaft furnace, based on said image analysis. If such a risk is detected, the control module outputs an obstruction risk signal s, transmitted to:- the Human-Machine Interface 6, for displaying a message to alert an operator of that risk (using a display screen, for instance), and to the control module 52.

[0045] The control module 52 is configured for controlling the direct reduction plant so as to take one or more actions to mitigate the risk of obstruction, when detected. To this end, the control module may be configured for sending an adjusted setpoint value sp to the low-level controller of the direct reduction plant, in order for instance to reduce a temperature, or a flow rate of one of the gases above mentioned. The control module may also be configured for driving directly actuators of the direct reduction plant 1 in order to achieve such a modification of the operating conditions in the shaft furnace.

[0046] Apart from this risk-mitigation function, the control module 52 may be configured for controlling the operation of the direct reduction plant during standard (normal) operation (that iswhen no risk of obstruction is detected), for instance by regulating temperatures, gas flows and / or ore flow.

[0047] The monitoring module 51 and the control module 52 may each take the form of a dedicated group of code instructions; in other words, they make each take the form of a dedicated computer program or subprogram. These two modules may also take the form of two distinct electronic units (for instance two distinct computers or two distinct programmable units), operatively connected to each other.

[0048] The electronic monitoring system 5 may take the form of a standalone computer, or of a system of computers connected to each other. The electronic monitoring system 5 may also be implemented in a distributed manner (somehow “virtually”), using so-called “cloud” resources (computing and storing resources distributed among distinct physical systems in a network).

[0049] The direct reduction plant according to the instant technology could be arranged differently than as above described. For instance, the gas processing and conditioning unit could be configured differently, in order to use mostly H2 for the reduction or the iron ore (and limiting the use of CO as much as possible). Besides, another type of conveyor (such as a bucket conveyor) could be used, instead of the belt conveyor 2 of figure 1 . The digital camera could be located so as to acquire images of ensembles of pellets at a different location than on the conveyor (for instance in a storage zone for storing the reduced pellets). Yet, acquiring images of ensembles of pellets on the conveyor is beneficial; indeed, it allows for detecting clusters early, just after they have been output by the shaft; and the pellets are spread in the conveyor, which is favourable for image-based inspection.Method for monitoring the operation the direct reduction plant

[0050] The method for monitoring the direct reduction plant 1 comprises (figure 3):- s1 : acquiring an image Img of an ensemble of pellets output by the shaft furnace (see figures 4, 5 and 9),- s2: analysing said image using a detection algorithm configured for detecting one or more clusters of pellets 9 and, optionally, one or more pieces of shaft internal crust, and- s3: emitting an obstruction-risk signal s if one or more clusters or pieces of crust are detected in step s2.

[0051] The obstruction-risk signal s is emitted on the condition that one or more clusters or pieces of crust are detected in step s2. Yet, one or more additional conditions may be required to emit the obstruction-risk signal s. The additional condition(s) are for instance relative to features (like a dimension) of the detected clusters or crust pieces; such a condition may be that the numberand / or the dimension of the detected clusters are above a given threshold. In practice, the method comprises a step of determining if there is a risk of obstruction of the direct reduction shaft furnace, based on the results of the analysis of said image Img by the detection algorithm. In this embodiment, step s2 comprises the image analysis step, and the step of determining if there is a risk of obstruction (in which condition the obstruction-risk signal s is emitted).

[0052] The obstruction-risk signal s may be a binary signal, specifying that there is a risk of obstruction, or not. The obstruction-risk signal s may also be a more elaborate signal, providing further information regarding that risk such as a risk value (degree of risk), or a degree of confidence for that risk assessment, or a nature of the risk (blocking by pellets clustering, of by fragmented crust pieces). The obstruction-risk signal s may also take the form of a histogram or of statistical data relative to detected cluster(s) or piece(s) of crust; in this case, the fact that there had been a positive detection of such items is reflected directly by the histogram or the data in question. The obstruction-risk signal s may also convey the analysed image of the pellets on the conveyor, possibly enriched to evidence the detected cluster(s) or piece(s) of crust in that image (for instance highlighted, or coloured), this enriched image being then displayed by the humanmachine interface 6.

[0053] When the obstruction-risk signal s is emitted, it is transmitted: to the human-machine interface 6, which displays information specifying that there is a risk of obstruction of the shaft furnace (during step s5); this may be achieved by specifying that pellets clusters are formed in the shaft furnace, or by displaying an image highlighting the detected cluster(s); other clusters related information may also be displayed, and / or- to the control module 52, which (during step s4): o determines if an action is to be taken to reduce clusters formation and / or crust fragmentation in the shaft furnace 10, o and, in this case, controls one or more elements of the direct reduction plant to achieve this action.

[0054] In the embodiment described here, the obstruction-risk signal s is transmitted both to the human-machine interface 6 and to the control module 52.

[0055] In step s4, features (that is, characteristics) of detected cluster(s), or of detected piece(s) of crust, such as their number and / or dimension, may be taken into account to determine if an action is to be taken and which action, and / or to parametrize that action (for instance so set the value of a temperature decrease to be applied). These features may be taken into account either directly (as such); they may also be taken into non-directly, by determining the action in question based on the obstruction-risk signal s, which itself is determined based on the features in question.

[0056] As represented in figure 3, steps s1 and s2 are repeatedly executed (there are executed several times successively), to continuously monitor the material output by the shaft furnace, and emit the obstruction-risk signal s if appropriate.

[0057] Steps s2 and s4 are now described in more details, for the exemplary embodiment considered here.Step s2: computer vision, image analysis, risk assessment

[0058] In a first implementation of the method, to detect clusters of pellets and possibly pieces of crust, in the image Img of an ensemble of pellets 8 output by the shaft furnace 10, a trained algorithm for object detection (i.e.: object identification) and segmentation is used. This algorithm is based on a convolutional neuronal network (CNN). The object detection and segmentation algorithm is a pre-trained algorithm (pre-trained for object detection and segmentation, using a large, generic purpose images database) fine-tuned using labelled, training images of ensembles of pellets in which clusters of pellets are identified (for instance delineated manually).

[0059] In practice, 10 to 50 such training images of ensembles of pellets (for the fine tuning) enable to obtain a trained object detection and segmentation algorithm efficiently detecting clusters (and, optionally pieces of crust), for instance with an average precision over a test set above 0.6 or even above 0.7. Larger sets of training images may also be used for the fine tuning. The manual labelling of the training images may be done by delineating each cluster (and possibly each piece of crust) in the image by a polygon line. When using the Mask R-CNN algorithm mentioned below, this annotation may be exported as a “sing file” in the COCO JSON format (JavaScript Object Notation, in the format databases of ‘Common Object in Context’ type).

[0060] The object detection and segmentation algorithm may, like here, be of the type of the Mask-R CNN algorithm, which is part of the Dectectron2 library which was developed by the company Meta. More particularly, the object detection and segmentation algorithm may comprise the following steps: a. determining a boundary box for each item detected, b. determining what kind of object is in the boundary box (in other words, determining a class for the item, by regression), and c. determining a mask for that item (segmentation), wherein steps b. and c. are executed in parallel.

[0061] According to an optional feature, the trained algorithm is trained to detect pieces of fragmented crust, in addition to clusters of pellets. To this end, training images of ensembles of pellets including one or more pieces of crust are also employed during the training. In this case, the trained algorithm may be trained also to distinguish clusters of pellets from pieces of crust,that is, to attribute them to different classes. The identification of a big agglomerate as a piece of crust, achieved by the trained algorithm, may be refined (eg: confirmed, or possibly invalidated) based on one or more geometric features of the agglomerate considered. For instance by analysing the border line delineating the mask for that agglomerate to detect angular corners or vertices (which indicate a piece of fragmented crust, while an absence of such a sharp end indicates a cluster of pellets); or based on a shape factor such as the circularity or the compactness of the agglomerate (as clusters of pellets or more round-shaped than pieces of crust).

[0062] According to another optional feature, the trained algorithm is also trained to distinguish fines-cemented from other clusters of pellets (i.e.: to attribute to different classes). Fines-cemented clusters are typically much harder to break and leads to higher risks of obstruction. An imagebased detection of such fines-cemented clusters is possible as their visual appearance differs from the appearance of clusters weakly cemented to each other (for which the pellets of the cluster are more distinguishable from one another than for fines-cemented pellets).

[0063] In step s2, in addition to the detection (and class-identification) of objects and the segmentation, one or more characteristics of detected objects are determined, in this first implementation. More particularly, a dimension of the object is computed, such as a width, a length, a diameter or an area of the region. In this exemplary embodiment, the dimension in question is the apparent diameter D of the region, computed as the square root of 4.S / with S the area of the region (alternatively, this diameter could be computed as the smallest cord for that region, for instance).

[0064] Figure 6 represents results of the computer vision and image-analysis technics above described, applied to the acquired image Img represented in figure 5, the object detection and segmentation algorithm being the Mask R-CNN algorithm fine-tuned as above explained. The Mask R-CNN algorithm is described in more details in the article: “mask R-CNN” by Kaiminq He et al., arXiv:1703.06870v3, 24 Jan 2018, doi: 10.48550 / arXiv.1703.06870.

[0065] As represented in figure 6, a total of nine clusters of pellets were detected in the image Img of figure 5. Each detected cluster is enclosed in a bounding box with a label “cluster” and a confidence index. Most of the clusters detected have a 100% confidence index. Figure 7 summarizes the dimension distribution of the nine detected clusters in the form of a histogram. The number N of clusters, in each of the three dimension-classes considered, is indicated. The equivalent diameter of the single pellet in the image is in the range of 5 to 15 mm, with an average diameter of the single pellets is 10 mm. Seven of the detected clusters are in the size range of 33 to 53 mm in their equivalent diameter D (about the size of equivalent 4 average pellets). One mid-size detected cluster is in the size range of 53 to 75 mm in diameter (the size of equivalent 6 average pellets) and one large detected cluster has a size range of 73 to 93 mm in diameter (the size of equivalent 8 average pellets). Figure 8 represents the values of the effective diameter D for the detected pellet number i, i=1 ..9 (nine pellets detected), for the image Img of figure 5.

[0066] It is noted that a trained algorithm of a type different from the one above described could be used, in alternative implementations of the method. For instance, in a second implementation, the detection of clusters of pellets and pieces of crust is based on a segmentation (not necessarily with an object identification), and then a dimension analysis for the regions resulting from the segmentation, an abnormally big region (among the regions resulting from the segmentation) being identified as a cluster of pellets, or piece of crust). In this case, clusters of pellets are detected because they correspond to abnormally big items. Abnormally big means significantly bigger than a pellet; for instance with a width or with a diameter higher than two times, or even higher than three times a width or a diameter of a pellet. Said width or diameter of a pellet is for instance an average width, or an average diameter of the pellets output by the shaft furnace 10. It may also be a maximum width, or a maximum diameter for these pellets. Anyhow, it is representative of the dimension of the pellets output by the shaft furnace 10.

[0067] To detect the presence of such an abnormally big item,:- the image Img is segmented, and then, for each region resulting from the segmentation of the image, a dimension of the region is computed, such as a width, a length, a diameter or an area of the region.

[0068] A cluster of pellets, or a piece of crust is detected each time one of these regions has a dimension significantly higher than the dimension of a pellet, in practice higher than the dimension of a pellet times a given coefficient. This coefficient may be from 2 to 10. It is for instance equal to 2, to 3 or even equal to 10 or 12. The dimension in question may be the equivalent diameter of the region, above defined, for instance.

[0069] Regarding the segmentation, it can be achieved using a trained convolutional neuronal network (CNN), such as Linet or Resnet (eg: Resnet18), which is a pre-trained one, fine-tuned using labelled, training images of ensembles of pellets in which clusters of pellets are identified (as above explained).

[0070] Alternatively, the segmentation step can also be achieved using a particulates separation technique by direct image processing, such as watershed segmentation. In this case, the segmentation step comprises for instance the following successive operations:- converting the image Img to a gray-scale image;using a binary thresholding method, for instance, Otsu binarization, Adaptive Mean thresholding, or Global thresholding, to separate the foreground (pellets and clusters) from the background; using a morphological transformation, for instance an erosion and dilation method, to separate boundaries of the pellets and clusters from the pellets and clusters themselves;- then, using a marker-based Watershed segmentation (a region-based technique that takes into consideration topographic features such as mountains, valleys, and basins to segment individual items, by simulating flooding in the 3D landscape corresponding to the grey values of the pixels in the image, and identifying barriers separating flooding basins); in particular, the markers for the marker-based Watershed segmentation are marking foreground elements, which are pellets or clusters of pellets, preliminary identified thanks to the operation of boundary separation (by erosion and dilatation).

[0071] Figure 10 shows the results of this watershed segmentation procedure on the image Img of figure 9. In figure 10, the different regions identified, separated one form another, are each attributed a random grey level (so that the result of the segmentation is visible, in figure 10).

[0072] Figure 1 1 summarizes the dimension distribution of the regions in figure 10 (regions resulting from the segmentation of the image Img of figure 9), in the form of a histogram. The number N of clusters, in each of the dimension-classes considered, is indicated. The dimension in question is, again, the equivalent diameter D above-mentioned.

[0073] The equivalent diameter of the single pellet in the image Img of figure 9 is in the range of 5 to 15 mm, with an average diameter of the single pellets is 10 mm. A majority of the particles (regions) resulting from the segmentation are individual single pellets. Figure 1 1 also shows that a number of clusters are present in the image Img. When a particle has an equivalent diameter D higher or above 20 mm (that is 2 times the average diameter of pellets), it is considered to be a cluster of pellets. Some of the detected clusters even have an equivalent diameter D higher than 30 mm (more than 3 times the average dimension of pellet). In figure 1 1 , each class of dimension is indicated as an interval, in mm (for instance, the interval [20, 24] for the pellets with D from 20 to 24 mm).

[0074] In this second implementation of the method, the results of the image analysis comprise:- an indication that one or more cluster(s), or piece(s) of crust have been detected ; this indication might be a binary one ; it may also be the above mentioned histogram (which indicates a positive detection when one of the big dimension classes is populated), or any equivalent data representative of the distribution of dimensions of the items present in the image,- and characteristics of the detected cluster(s) or piece(s) of crust, here the number and dimensions of the detected pellets or pieces of crust.

[0075] Other characteristics of the detected items may also be determined.

[0076] For instance, for each abnormally big item, a nature of the agglomerate, as being either a cluster of pellets or a piece of fragmented crust, may be determined. This determination may be a direct geometric computation, for analysing the border line delineating that region to detect angular corners of vertices (which indicate a piece of fragmented crust, while an absence of such a sharp end indicates a cluster of pellets). The nature of the detected iron-ore agglomerates may also be determined using a dedicated trained classifier, the sub-image (that is, the region) corresponding to the detected iron-ore agglomerate being input in that trained classifier. In the instant method, when the optional feature of determining the nature of the agglomerate (either a cluster of pellets or a piece of crust) is not implemented, any abnormally big item is considered to be a cluster of pellets, by default.

[0077] According to another optional feature, it is determined, for each detected cluster, whether the cluster is fines-cemented or not. It may be achieved using a dedicated trained classifier, the sub-image (that is, the region) corresponding to each detected cluster of pellets being input to that trained classifier.

[0078] Once the acquired image Img has been analysed (should it be according to the first implementation above described, or according to the second or to yet another implementation), the presence of a risk of obstruction, and optionally a quantification of this risk, are evaluated, based on the results of the computer vision and image analysis.

[0079] Here, it is determined that there is a risk of obstruction when the results of the computer vision and image analysis fulfil a predetermined condition.

[0080] In a first example, the predetermined condition is that one or more clusters or pieces of crust are detected in the image Img.

[0081] In a second example, the predetermined condition (more stringent than in the first example above) is that:- a number of clusters or pieces of crust, detected in the image Img, is above a given number No, or that- at least one of the clusters or pieces of crust, detected in the image Img, has a dimension above a given critical dimension Do.

[0082] In a third example, the predetermined condition is that a number of clusters or pieces of crust, detected in the image Img and having a dimension above the given critical dimension Do, is above a given number No’.

[0083] Other examples are possible, for the predetermined condition in question.

[0084] A quantification of the risk, in the form of a risk value, may also be determined based on the results of the computer vision and image analysis. This risk value is for instance increasing when the number and / or the dimension of the detected clusters (or pieces of crust) increases. Besides, when the optional detection of fines-cemented cluster(s) is implemented, the presence of one or more fines-cemented cluster(s) leads to a higher risk value than its absence.

[0085] To determine if there is a risk of obstruction, and optionally to quantify this risk, additional signals may be taken into account (in addition to the results of the computer vision and image analysis), in particular: a cluster breaker torque signal, a pressure signal for the top gas 22 or a pressure signal for a pressure difference between the reducing gas 21 input in the shaft furnace and the top gas 22 output by the shaft furnace. For instance, in a case for which the clusters detected in the image Img are numerous small clusters, an unusually high value of the cluster breaker torque will lead to the conclusion that there is a risk of obstruction while a normal value of that torque will lead to the conclusion that there is no risk of obstruction.Step s4: action taken to mitigate a risk of obstruction

[0086] Step s4 is presented now in the case for which the optional determination of the nature of the detected iron-ore agglomerates (either clusters of pellets, or pieces of fragmented crust) is not implemented. The action or actions taken to reduce the risk of obstruction are then actions for limiting the formation of clusters of pellets in the shaft furnace 10. Pellet sticking, for a given pellet quality and chemistry, is a phenomenon that depends on temperature of the pellet bed in the reduction shaft. Therefore, most of these actions aim at reducing a temperature Tb of the bed of pellets located in a lower part of the reduction section 12. The bed temperature can be modified in two main ways: a) by modifying the temperature of the injected gas and b) by modifying the extent of the endothermic reactions happening in the shaft. These reactions are namely of these types: endothermic iron oxide reduction reaction by H2 or endothermic hydrocarbons cracking or in-situ hydrocarbon reforming with CO2 or H2O. In practice, one or more of the following actions can be taken:- a1 ) reducing a temperature Tg of the reducing gas 21 injected in the shaft furnace 10, by: reducing a flow rate of 02 added to the reducing gas and / or by reducing a reducing gas heater temperature and / or by reducing a reformer temperature (at the reformer output),- a2) increasing an hydrocarbons content in the reducing gas 21 (by increasing a flow rate of hydrocarbons added to the reducing gas through the inlet 29); Typically, injecting more natural gas (injecting more CH4) through the inlet 29,- a3) Increasing a H2 content in the reducing gas 21 , so that there is a higher amount of endothermic iron oxides reduction by hydrogen; this can be achieved by increasing the amount of water that is incorporated in the reducing gas through steam injection for instance,- a4) adding a flow of cold gas to the reducing gas; this cold gas is preferably of a chemistry similar, or even identical to the reducing gas; when available, the gas cooler 31 installed after a heater or after the reformer can be used to cool the hot reformed gas,- a5) reducing the total flow rate of the reducing gas 21 injected in the shaft furnace,- a6) increasing a pellets feeding rate, at the charging device 11 .

[0087] The 02 contained in the reducing gas 21 promotes combustion and exothermal reactions. So, reducing the amount of 02 (action a1 )) allows for reducing the temperature of the reducing gas Tb.

[0088] CH4 (or another hydrocarbon) reacts within the shaft according to an internal reforming reaction which is endothermal (and catalysed by iron). And so, increasing the amount of CH4 in the reducing gas (action a2)) allows for reducing temperature Tb.

[0089] For actions a1 ), a2), and a3), the response time, between triggering the action and observing a decrease of Tb, is typically small, or the order of a few minutes.

[0090] Actions a1 ) to a5), are usually preferred to action a6), as allowing to reduce temperature Tb much faster than action a6).

[0091] Actions a1 ) to a5) may each be executed so as to reach a targeted flow rate for the gas concerned, to reach a targeted gas temperature or to reach a targeted gas composition for the reducing gas. Actions a1 ) to a5) may also be each executed (using close-loop control) so as to obtain a targeted decrease of the temperature Tb (temperature which is typically measured by dedicated temperature sensors placed within the shaft furnace, in the reduction section).

[0092] Anyhow, the targeted decrease of temperature Tb is typically of the order of a few Celsius degrees and may be from 2 to 20°C, or even from 2 to 30°C. Indeed, compared to a nominal value of Tb, a decrease of 3°C for instance reduces substantially the tendency of pellets to stick to each other, while still leading to an acceptable reduction efficiency of oxidized iron.

[0093] This temperature decrease may be determined depending on the risk value (determined in step s2): the higher the risk value, the higher the targeted temperature decrease.

[0094] As above mentioned, when a risk of obstruction is detected, the main objective is to reduce temperature Tb. More precisely, when a risk of obstruction is detected, temperature Tb is reduced compared to a normal operation setpoint for temperature Tb. The normal operation setpoint may be a nominal setpoint, recommended in the absence of abnormality of fault, given the composition and / or size of the pellets. The normal operation setpoint may be determined so as to optimize ametallization ratio of the pellets output by the shaft (while remaining below a maximum temperature limit).

[0095] The normal operation setpoint may be determined depending, inter alia, on:- a composition and / or size of the iron-ore pellets,- a feed rate with which the pellets are fed into the shaft furnace,- a composition of the reducing gas 21 .

[0096] The normal operation setpoint may be entered by an operator using the human-machine interface 6, or may be computed by the electronic monitoring system 5.

[0097] In addition to triggering one or more of actions a1 ) to a6), the following long-term action a7) may be triggered, when there is a risk of obstruction: coating the pellets with an antiagglomerating coating, before they enter the shaft furnace, or, if coating is already applied, change the amount of applied coating (typically in the range of 1 to 4 tons of coating agent per ton of pellets), using a controllable coating device of the direct reduction plant 1 .

[0098] When the optional detection of fines-cemented clusters is implemented, the following complementary action a8) may be triggered (in addition to one or more of actions a1 ) to a6)): controlling a sieving device of the direct reduction plant 1 to enhance pellets sieving (for instance by increasing the sieving time, or by adding a supplementary sieving step).

[0099] Step s4 is now presented in the case for which the optional determination of the nature of the detected iron-ore agglomerates, as being either clusters of pellets or pieces of fragmented crust, is implemented.

[0100] When no piece of crust is detected in image Img, one or more of the actions above described are taken.

[0101] On the contrary, if a piece of crust is detected, the action taken is to reduce the feed rate with which pellets are fed into the shaft furnace, so as to reduce its throughput.

[0102] In the examples above described, the action(s) taken to mitigate the risk of obstruction are determined and triggered automatically by the electronic monitoring system 5. Yet, in other embodiments of the instant method, such actions could be triggered manually, by an operator of the plant. Else, a message suggesting to take such actions could be displayed by the human-machine interface 6, and triggered if validated by the operator.

[0103] Besides, in the example above described, the results of the image analysis are mostly processed in step s2, to determine the obstruction risk signal s. In other embodiments, the obstruction risk signal s may take the form of the raw results of the image analysis, these results being then analysed in step s4 to determine if an action is to be taken, which one, and to what extent. More generally, the different operations of the instant method may be organised in stepsY1 in a different manner than in figure 3. And more than one image, for instance a sequence of successive images could be analysed together to determines if there is a risk of obstruction.

Claims

1 . A method for monitoring the operation of a direct reduction plant (1 ) that comprises a direct reduction shaft furnace (10), the method comprising:- s1 : acquiring an image (Img) of an ensemble of pellets (8) output by the shaft furnace (10),- s2: analyzing said image (Img) using an algorithm suitable for detecting one or more clusters (9) of pellets and, optionally, one or more pieces of shaft internal crust, and- s3: emitting an obstruction-risk signal (s) if one or more clusters (9) or pieces of crust are detected in step s2.

2. A method according to claim 1 wherein, in step s2, the algorithm for detecting one or more clusters (9) of pellets and, optionally, one or more pieces of shaft internal crust comprises a trained convolutional neuronal network, which is a pre-trained convolutional neuronal network fine-tuned using training images that are images of ensembles of iron-ore pellets, at least some of said ensembles of iron-ore pellets each including a cluster of pellets or a piece of shaft internal crust.

3. A method according to claim 1 or 2 wherein, in step s2, the algorithm for detecting one or more clusters (9) of pellets is trained algorithm for both object identification and image segmentation.

4. A method according to claim 3, wherein the algorithm for object identification and image segmentation comprises the following steps:- a) determining a boundary box for each item detected,- b) identifying a type of object to which the item in the boundary box belongs,- c) determining a mask for an area occupied by the item, wherein steps b) and c) are executed in parallel.

5. A method according to claim 1 or 2, wherein, in step s2, a cluster of pellets, or optionally a piece of shaft internal crust, is detected in the image (Img) when an item, having a dimension higher than a dimension of a pellet times a given coefficient, is detected.

6. A method according to anyone of claim 5 wherein:step s2 comprises: o segmenting said image (Img) into distinct regions , o then, for each region resulting from the segmentation of the image, computing a dimension of the region, such as a width, a length, a diameter or an area of the region,- one or more clusters (9) of pellets or, optionally, of one or more pieces of shaft internal crust, being detected when the dimension of one or more of said regions is higher than: a dimension of a pellet times a given coefficient.

7. A method according to claim 6, wherein said segmentation comprises applying a markerbased watershed segmentation and image processing.

8. A method according to claim 6 and claim 2, wherein said segmentation is achieved using the trained convolutional neuronal network.

9. A method according to anyone of claims 5 to 8, wherein step s2 comprises determining if the item, whose dimension is higher than the dimension of a pellet times the given coefficient, is a piece of shaft internal crust or not.

10. A method according to anyone of the preceding claims, wherein step s2 comprises determining one or more characteristics of the detected cluster(s) of pellets among: a dimension, a position, a shape characteristic, a number of clusters in the image, a density of clusters in said image.1 1 . A method according to anyone of the preceding claims, wherein step s2 comprises determining, for the one or more detected clusters, whether the cluster is fines-cemented.

12. A method according to claim 8 or 9, wherein step s2 comprises determining the obstruction-risk signal (s) taking into account the one or more characteristics of the detected cluster(s) and / or the presence of fines-cemented cluster(s).

13. A method according to claim 10, wherein the obstruction-risk signal (s) is determined taking further into account a cluster breaker torque signal and / or a gas pressure signal.

14. A method according to anyone of the preceding claims, wherein the obstruction-risk signal (s) is transmitted to a human-machine interface (6), the method then comprising:- s5: displaying information specifying that there is a risk of obstruction of the shaft furnace (10).

15. A method according to anyone of the preceding claims further comprising:- s4: triggering one or more actions to mitigate a risk of obstruction of the shaft furnace (10), if one or more clusters (9) or pieces of crust are detected in step s2.

16. A method according to the combination of claim 15 and claim 10, wherein the triggering of one of said actions and / or a parametrization of said action is decided depending on the one or more characteristics of the detected cluster(s) of pellets, or depending on a risk value determined based on said one or more characteristics.

17. A method according to claim 15 or 16 wherein the one or more action(s) are one or more of: reducing an 02 injection to a reducing gas (21 ) injected in the shaft furnace (10), increasing a CH4 or other hydrocarbons content in the reducing gas (21 ), increasing an H2 content in the reducing gas (21 ), reducing a temperature of the reducing gas (21 ) using a gas cooler (31 ), reducing a total flow rate of the reducing gas (21 ) injected in the shaft furnace, reducing a temperature setpoint for an inner temperature in a reduction region (12) of the shaft furnace (10).

18. A method according to claim 17, further comprising triggering one or more of the following additional actions:- controlling a coating device of the direct reduction plant (1 ) to coat pellets with an anti-agglomerating coating before they enter the shaft furnace (10),- controlling a sieving device of the direct reduction plant (1 ) to enhance pellets sieving.

19. A method according to anyone of claims 15 to 18 wherein, in case of detection of one or more pieces of crust, the action triggered is reducing a pellets feeding rate.

20. A method for producing reduced iron using a direct reduction plant (1 ), wherein pellets (8) of iron ore are reduced in a direct reduction shaft furnace (10) using a reducing gas (21 ), wherein the operation of the direct reduction plant (1 ) is monitored using the method according to anyone of claims 1 to 19.21 . An electronic monitoring system (5) comprising at least one processor and one memory, configured for executing the method according to anyone of claims 1 to 19.

22. A direct reduction plant (1 ) comprising a direct reduction shaft furnace (10), a gas processing and conditioning unit (20) for feeding the shaft furnace (10) with a reducing gas (21 ), an image acquisition device (3) arranged for acquiring an image of an ensemble of pellets output by the shaft furnace, and the electronic monitoring system (5) of claim 21 , the electronic monitoring system (5) being operatively connected to the image acquisition device (3), to a human-machine interface (6), and to one or more actuators of the direct reduction plant (1 ) or to a low-level controller of the direct reduction plant (1 ).

23. A computer program, comprising instructions whose execution on a computer (5) make the computer to execute the method according to anyone of claims 1 to 19.

Images