DE102021212110A1 - Process for automatic wear monitoring and anomaly detection of refractory linings in coke ovens - Google Patents

Abstract

The present invention relates to a method for detecting defects in a refractory lining in a chemical reaction chamber, in particular a coke oven, the method having the following steps: a) acquiring at least a first recording of the inside of the refractory lining, b) performing an image analysis at least a first image for detecting defects,c) outputting potential defects.

Description

The invention relates to a method for detecting defects in the refractory lining of a coke oven.

Coke ovens have chambers that have a width of a few 10 cm, for example 40 to 50 cm, and a length of several meters and a height of a few meters. These chambers are operated in a temperature range of up to 1800 °C, for example. The walls are made of refractory bricks. These walls can show a variety of damage over time. In the worst case, holes, cracks or gaps can lead to an undesired gas exchange between the oven chamber and heating wall, which occurs during the process, and this gas could then be released into the environment unintentionally and uncontrolled. It is important to prevent this. In addition, repairs become disproportionately more complex and expensive when the defects become larger. It is therefore expedient to identify defects at an early stage and rectify them before substances are released into the environment or the repair effort becomes disproportionately high.

From the EP 2 113 552 A1 a device for inspecting wall surfaces of coke ovens is known.

From the JP 2016 065 225 A a device for the automatic detection of carbon adhesions in a coke oven is known.

From the post-published DE 10 2021 101 102 discloses an aircraft and a method for inspecting equipment in coking plants to detect sources of error.

Wall damage can only be inspected when the coke oven door is open. For operational reasons, the coke oven cannot be switched off, which means that an inspection takes place when the coke oven is hot. One of the problems with visual inspection is the temperature inside the coke oven chamber. For a person standing in front of an opening, this is already a burden. In addition, there is the burden of escaping chemical substances. At the same time, the tunnel-shaped character makes visual inspection difficult. A qualified inspection usually requires a great deal of experience, which in turn reduces the number of people with the necessary qualifications. If there are also travel restrictions, for example due to a pandemic, a check can be extremely difficult.

The object of the invention is to provide a method for the reliable and early diagnosis of defects in a refractory lining. The occupational safety of the people involved should also be increased and delays in the ongoing operating process should be kept as short as possible by inspection.

This problem is solved by the method with the features specified in claim 1 . Advantageous developments result from the dependent claims, the following description and the drawings.

The method according to the invention serves to detect defects in a refractory lining in a chemical reaction chamber, in particular a coke oven. However, the method can also be applied to other chemical reaction spaces with a refractory lining. Examples include rotary kilns, steam reformers, syngas reformers, reactors, boilers, steam and exhaust gas lines, post-treatment furnaces, melting furnaces, chimneys and chimneys. The method is particularly suitable for a reactor for continuous operation. After emptying and before refilling, a corresponding operation can be carried out at the operating temperature, so that it is not necessary to shut down and cool down the reactor for inspection. The coke oven is particularly suitable for this process. On the one hand, a coke oven is usually made of masonry, which can result in different types of damage than, for example, a furnace vessel with a refractory lining. On the other hand, several coke ovens are usually set up next to each other in a battery, so that the failure and repair of one coke oven can have a direct impact on all the other coke ovens in the battery. Furthermore, the geometry of the coke ovens is such that the individual coke oven chambers are comparatively long, high and narrow, for example, have a length of more than 10 m and a width of around half a meter. At an operating temperature in the range of 1000 °C to 2000 °C, there is a difficulty in inspection due to the geometry and the temperature load, but at the same time there is also a need for regular inspection of the system for the reasons already discussed above.

1. a) Erfassen wenigstens einer ersten Aufnahme der Innenseite der feuerfesten Verkleidung,
2. b) Durchführen einer Bildanalyse wenigstens eines ersten Bildes zur Erkennung von Defekten,
3. c) Ausgeben potentieller Defekte

The procedure has the following steps:

1. a) capturing at least a first shot of the inside of the refractory lining,
2. b) performing an image analysis of at least one first image to identify defects,
3. c) reporting potential defects

The detection in step a) takes place from outside the chemical reaction space. The detection preferably takes place through an opening, in particular through an opening which is used in regular operation, for example to introduce educt, discharge product or to bring in or discharge tools. Using the example of the coke oven, two openings are particularly relevant for the implementation of the method. Firstly, the opening through which the coke is removed from the coke oven (coke side). On the other hand, the opening through which the coke is pressed out of the coke oven (machine side). The openings for supplying the educt coal, which are arranged above, are not suitable because they are not accessible during normal operation due to the coal located there. Alternatively, of course, maintenance hatches or windows can also be used. It is better, however, if openings are used which are used in regular operation anyway, as this disrupts operation less.

The detection in step a) can take place by taking a picture with a camera. However, a plurality of recordings with one camera or a recording of a plurality of recordings with a plurality of cameras can also be used. If a multi-camera system is used, these are preferably fixed to one another in a fixed spatial arrangement, for example using a tripod, which simplifies the superimposition of the recordings to generate an image of the chemical reaction space, in particular of the coke oven.

The detection in step a) can be done by taking a picture with a light field camera. A light field camera, also known as a plenoptic camera, has the advantage that focusing does not take place before the picture is taken, but the various focal planes can be calculated from the recorded picture. As a result, in chemical reaction spaces at great depths, such as in particular coke ovens, rapid and consistently sharp detection can be carried out in a simple manner.

The detection in step a) can be done by recording with a laser scan system. Here, the surface is usually scanned line by line and column by column. Therefore, recording with such a scanning system is comparatively slow compared to a camera that records all information at the same time. On the other hand, the size of the laser point allows more precise scanning and, if necessary, more information can be obtained.

Capturing in step a) can take place by means of a freehand recording with a camera. This has the advantage that it is comparatively simple. But the spatial arrangement is not reproducible. Therefore, in the case of freehand photography, it is advantageous to take more photographs and then stitch them together to obtain an image of the entire refractory lining.

The detection in step a) can take place by taking a picture with a camera which is reproducibly positioned in a predefined position in front of an opening, for example by means of a tripod. This has the advantage that the spatial arrangement of the image and thus the determination of the position of defects is simplified. For this purpose, attachment points for the tripod and positioning marks for the tripod can be arranged at the opening or in the vicinity of the opening.

The capturing in step a) can take place in particular with a focal length between 25 mm and 75 mm based on the 35mm format. Correspondingly shorter focal lengths result for smaller sensors, and correspondingly longer focal lengths for larger sensors. A sensor with 20 to 200 megapixels is preferred as the sensor size in order to have sufficient resolution on the one hand and to limit the amount of data to be processed on the other.

If the chemical reaction space, in particular the coke oven, has two or more openings, recordings are preferably made of both or all of the openings.

For the purposes of the invention, the image includes at least two-dimensional information of at least one intensity distribution. An image recorded using a commercially available camera has three color channels, each of which has a two-dimensional intensity distribution (brightness). In order to expand the information range, more than just three color channels can also be recorded, for example by additional recordings with an IR camera (usually a commercially available camera without an IR blocking filter and a VIS blocking filter instead) or a thermal imaging camera. However, images can also be composed of individual recordings (which also represent images within the meaning of the invention). For example, by merging (stitching) a larger panorama image can be assembled from several exposures. Likewise, for example, a three-dimensional image can be generated from a large number of recordings by means of photogrammetry, such an image also showing, strictly speaking, only a two-dimensional surface, but with the spatial structure in three-dimensional space. The term recording is in the sense of the invention tion is therefore used for an unchanged image (image from the camera, original image, non-processed image).

Thus, the first image in step b) can be the same as the first recording in step a) if the first recording is used unchanged as the first image. Alternatively, the first image can be generated from the first recording, for example by combining it with other recordings. Thus, the first image is equal to the first shot or the first image is generated from the first shot.

• b1) mit einem ersten neuronalen Netzwerk eine Merkmalskarte erstellt wird
• b2) die in der Merkmalskarte potentiell interessanten Bereiche mit einem Gebietsvorschlagsnetzwerk untersucht werden, wobei Bereiche identifiziert werden, welche eine Anomalie enthalten könnten, wobei Ankerboxen erzeugt werden,
• b3) mit einem zweiten neuronalen Netzwerk in den Ankerboxen eine Ausrichtung eines Zielbereichs in jeder Ankerbox durchgeführt wird,
• b4) mit einem multivarianten Clustering-Algorithmus die Schwere jeder Anomalie in jeder Ankerbox analysiert wird und erkannt wird, welche Anomalien Defekte sind.

An image analysis is carried out with the help of a self-learning algorithm. Performing an image analysis has the following sub-steps:

• b1) a feature map is created with a first neural network
• b2) the areas of potential interest in the feature map are examined with an area proposal network, whereby areas are identified which could contain an anomaly, whereby anchor boxes are generated,
• b3) a second neural network in the anchor boxes is used to align a target area in each anchor box,
• b4) using a multivariate clustering algorithm, analyzing the severity of each anomaly in each anchor box and identifying which anomalies are defects.

Instead of the term feature map, the terms feature mapping or feature tableau are also common, but in particular the English term feature map.

To put it very simply, the creation of the feature map in step b1) is a selection of potentially interesting areas, for example rectangular areas within the image, which can contain “something”. No statement is made here as to whether the potentially interesting area actually contains something or what it contains. Only potentially interesting areas are selected, which are subjected to further, more detailed analysis. As a result, areas that obviously contain nothing are the first to be removed from further evaluation, thus reducing the amount of data. At the same time, the size of the potentially interesting areas in the feature map is smaller than the image itself, so that you only have to work with a significantly reduced amount of data in each potentially interesting area. The feature map therefore only contains the potentially interesting areas.

In step b2), the potentially interesting areas of the feature map are now examined. An area proposal network is used for this. Instead of the term area proposal network, regions proposal network, network for the proposal of regions are also common, but in particular also the English term Region Proposal Network.

An anchor box is an area, such as a rectangular area, that contains an anomaly. In this step, all potentially interesting areas of the feature map are sorted out, which ultimately do not contain any anomalies.

Putting it simply, the alignment in the anchor boxes of a target area in each anchor box in step b3), carried out with a second neural network, is the selection of the pixels of the image that belong to the anomaly. This eliminates all pixels that show something else within the anchor box, i.e. in particular undamaged refractory cladding. This achieves a final focussing exclusively on the anomaly.

After the anomalies have been clearly delimited in this way, the severity of each anomaly in each anchor box is analyzed in step b4) using a multivariate clustering algorithm. This also recognizes which anomalies are defects. The position of the anomaly must be taken into account here. An anomaly with, for example, 100 pixels in size, which is very close to the opening and thus to the camera, must of course be assessed differently in terms of severity than the same anomaly that is in the same image but at a position, for example 5 m deep inside the chemical reaction space, in particular the coke oven, and whose real size is correspondingly larger for a size of 100 pixels, for example, which is the same in the image. This would significantly increase the severity of the defect inside.

For example, with a refractory lining in the form of masonry, the texture of the masonry in the image can be used to scale the size. This is particularly possible if the recordings are always made from a reproducible location and there is a first recording without defects, for example after the creation or renovation. In this way, for example, in the event of buildup, the size of the area can be easily estimated from the size of the bricks visible in the "original image".

In a further embodiment of the invention, the detection in step a) and the image analysis in step b) are carried out separately. esp In particular, the image data are transmitted before the analysis, in particular via the Internet or a mobile phone connection, to a server which is designed to carry out the image analysis in step b). Other transmission methods, such as satellite communication or Bluetooth, can also be used, as can transmission via USB stick or CD/DVD. To avoid transport times, digital transmission methods are preferred as far as possible.

In a further embodiment of the invention, the detection in step a) takes place with different focal lengths. For example, the areas close to the capture device (the camera) are recorded with a focal length between 20 mm and 50 mm in relation to the 35mm format. The areas further from the detection device (the camera) are recorded with a longer focal length, for example between 75 mm and 300 mm in relation to the 35mm format. Correspondingly shorter focal lengths result for smaller sensors, and correspondingly longer focal lengths for larger sensors. The different focal lengths result in different angular resolutions, which then again lead to a comparable resolution of the refractory lining, preferably due to the different distance to the sensor.

In a further embodiment of the invention, the first neural network is a convolution-based neural network. The common English-language term for a convolution-based neural network is convolutional neuronal network. Compared to conventional neural networks, the convolution-based neural network has proven to be particularly suitable for creating the feature map due to the high level of accuracy and processing speed.

In a further embodiment of the invention, the first neural network is trained using a priori knowledge. For this purpose, images are used which have been evaluated by a specialist, where defects are marked, labeled and given their severity. The larger this starting data situation is, the more reliably the first neural network can work. This also applies to the second neural network and the multivariate clustering algorithm. Everyone is preferably trained with this a priori knowledge. This training enables the system to reliably identify and classify defects right from the start.

Alternatively, the system can also be integrated into a new or newly renovated system, for example, and the first recordings can thus be classified as free of defects. It is true that the system is then initially unable to identify defects and assess their severity. However, images in which deviations occur are then presented to an expert, for example, who then carries out the assessment. In this way, the system then learns when the first anomalies occur. The disadvantage is that the system is heavily dependent on support at the beginning, but no a priori knowledge is required for this, but regular intervention by the specialist.

In a further embodiment of the invention, the detection in step a) takes place at the operating temperature of the chemical reaction space, in particular of the coke oven. Fluctuations are of course possible. If, for example, a coke oven is opened and the coke is ejected, ambient air naturally reaches the walls and cools them down, in contrast to the time when the coke oven is closed. However, such fluctuations occur during normal operation with each opening and are therefore understood as the operating temperature within the meaning of the invention.

In a further embodiment of the invention, the detection in step a) takes place in a spectral range from 350 nm to 1700 nm, preferably in a spectral range from 350 nm to 1100 nm, particularly preferably in a spectral range from 350 nm to 800 nm. An Si semiconductor can preferably be used for the range from 350 nm to 1100 nm, and an InGaAs semiconductor for the range from 1100 nm to 1700 nm. For example and preferably, the detection can take place by means of a CCD sensor or a CMOS sensor. In this case, the detection does not have to take place completely over the entire spectral range and it does not have to take place in a spectrally resolved manner. It can be captured as a pure black and white image. However, detection is preferably carried out as in commercially available camera systems in the three primary colors with overlapping sensitivity ranges and thus completely over the visible spectrum. However, this is made more difficult by the fact that, for example, at a temperature of around 2000 °C, the maximum of the black ray is between 1200 and 1300 nm. Especially a spectrally non-resolving sensor in the NIR range can have difficulties in capturing details when the dynamic range of the sensor reaches its limits.

In a further embodiment of the invention, the detection in step a) takes place with a multi-camera system. As already explained above, it is advantageous if the multiple cameras are in a fixed spatial relationship to one another. This can be done, for example, by mounting it on a common tripod.

In a further embodiment of the invention, the detection in step a) takes place with a moving image camera, for example a video camera. The use of a motion picture camera has the advantage that due to the large number of pictures, there is a good overlap between two pictures taken one after the other. In this case, the film sequence consists of the recordings. A first image can therefore easily be generated from such a film sequence. Since recordings that can be combined comparatively easily are created in this way, this can also be easily recorded by untrained personnel. Advantageously, a first conversion of the recordings then takes place as soon as possible after the recording, in particular if the first image is then transmitted, for example via the Internet or a mobile network, to a server to carry out step b). A significant data reduction can be achieved in this way.

In a further embodiment of the invention, a large number of images are recorded in step a) and these are combined by means of photogrammetry to form a virtual image of the chemical reaction space, in particular of the coke oven. In addition to the recordings, a model of the chemical reaction space, in particular the coke oven, can also be used for this purpose. As a result, a particularly good spatial assignment of the pixels is generated, which facilitates subsequent localization, a statement of the position of the defects. The virtual image created by photogrammetry is then analyzed in step b).

In a further embodiment of the invention, the detection in step a) takes place through at least one first opening and one second opening of the chemical reaction space, in particular of the coke oven. In the case of the coke oven, one opening is the so-called machine side, on which the coke is pushed out of the oven by a machine. The other opening is opposite the opening from which the coke exits the coke oven. By recording from both sides, the distance between the camera and the defect is limited to a maximum of half the length, especially for the inner area, which makes sense given the tunnel-like geometry of the coke ovens.

In a further embodiment of the invention, each image in which a specific defect cannot be clearly identified in step b4) is presented to a person skilled in the art for a decision. On the one hand, this allows the system to continuously learn. On the other hand, this reduces the need for manual intervention by the specialist to a minimum. This step is inherently slow and limited to the availability of a skilled person. Accordingly, the image classified by the person skilled in the art is preferably used for training the multivariate clustering algorithm. Usually only the trained anomalies can be recognized by the a priori knowledge. A specific effect is therefore presented, for example, if it cannot be clearly assigned to a type of anomaly, for example if the probabilities for two different types are of the same order of magnitude. By resorting to the expert, a clear assignment can then be made. However, if a new type of defect is actually found, it can also be trained on this basis.

In a further embodiment of the invention, the output in step c) takes place in the case of minor defects with a prognosis as to when a serious defect is to be expected. In this way, in particular, a decision can be made as to whether an immediate repair is necessary or whether it is still possible to wait until the next maintenance cycle. This prognosis is all the more reliable the more regularly this method is carried out on the relevant device, so that a development over time is preferably already known, which can be used as a data basis for the prognosis.

In a further embodiment of the invention, the output in step c) includes the type of defect, the severity of the defect and the position of the defect. This information can be provided in tabular form, but preferably also graphically. More preferably, an output is made in such a way that the chemical reaction space, in particular the coke oven, is divided into different segments and all defects for each segment are output cumulatively. This reveals segments in which smaller defects also accumulate, indicating a potentially larger problem overall.

In a further embodiment of the invention, the severity of the defect is output in step c) in three to five degrees of severity. If three degrees of severity are distinguished, these can be, for example, good, moderate and critical. If five degrees of severity are distinguished, these can be, for example, good, light, moderate, critical and fatal. Critical means, for example, that an immediate repair is necessary, fatal that further operation is impossible for safety reasons. However, the classification of the degrees of severity can, for example, also be based on a forecast of the remaining operating time, ie for example >1 year, >1 month, >1 week and <1 week.

In a further embodiment of the invention, the output in step c) takes place cumulatively for all defects within the chemical reaction space, in particular the coke oven, and one Evaluation is determined on the basis of all detected defects for a repair requirement.

In a further embodiment of the invention, the outputting includes defects selected from the list of broken edge, spalling, deposit, breakthrough point, ceramic weld point, joint widening.

In a further embodiment of the invention, a feature map is created in step b1) on the one hand on the basis of simple structures such as color intensity, shading, geometry, corners and edges. On the other hand, more complex features are also taken into account that can be characteristic of the respective anomalies, such as alignment, extent, or recurring patterns.

In a further embodiment of the invention, the method is repeated, wherein the repetition interval is selected from the group at each opening, daily, weekly, monthly, quarterly, annually. On the one hand, a continuous check makes it easier to find changes, on the other hand, a forecast for the remaining operating time can be made more well-founded the shorter the interval is.

• b) weitere Bilder, welche bereits früher von der gleichen Stelle des chemischen Reaktionsraums, insbesondere des Koksofens, erfasst wurden.

In a further embodiment of the invention, the image analysis in step

• b) further images, which were previously captured from the same location of the chemical reaction space, in particular the coke oven.

• 1 Wand
• 2 Wand mit ersten Defekten
• 3 Wand mit ersten Defekten und Ankerboxen
• 4 Wand mit Ankerboxen und analysierten Anomalien
• 5 Wand mit zweiten Defekten
• 6 Wand mit zweiten Defekten und Ankerboxen
• 7 Wand mit Ankerboxen und analysierten Anomalien

The method according to the invention is explained in more detail below with reference to an exemplary embodiment illustrated in the drawings.

• 1 Wall
• 2 Wall with first defects
• 3 Wall with first defects and anchor boxes
• 4 Wall with anchor boxes and analyzed anomalies
• 5 Wall with second defects
• 6 Wall with second defects and anchor boxes
• 7 Wall with anchor boxes and analyzed anomalies

For the sake of simplicity, a piece of masonry is shown in the figures as a section of a refractory lining. Defects are shown very large for clarity. The figures are therefore only to be understood as sketchy and schematic.

In 1 a perfect refractory lining without defects is shown. This picture would emerge, for example, immediately after the construction of a coke oven. This image is an example of a side wall in a coke oven, where the angle would be even steeper towards the back. For the sake of clarity, however, the angle has been selected here in such a way that the principle of the method can be seen better. For simplicity, the ceiling and floor and the opposite wall are also omitted. A real recording would create a tunnel-like impression.

2 shows the same wall with the first defects. In the rear area, a breakthrough is indicated schematically by the omission of four bricks, in the front area attachment to the wall in the form of a hatched circle. In 3 shows the wall with the first defects after step b2). Two anchor boxes show the two regions where defects are located. After step b3) results in 4 picture shown. The second neural network detects the two defects down to the pixel, indicated by the overlying small shaded areas.

5 shows the wall with second defects. The only difference is that these are closer together. Thereby the anchor boxes, as in 6 shown, partially overlapping. Nevertheless, in step b3) the pixel-precise analysis of the defects is carried out just as accurately as in 7 shown.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• EP 2113552 A1 [0003]
• JP 2016065225 A [0004]
• EN 102021101102 [0005]

Claims (19)

Method for detecting defects in a refractory lining in a chemical reaction chamber, in particular a coke oven, the method comprising the following steps: a) capturing at least a first shot of the inside of the refractory lining, b) performing an image analysis of at least one first image to identify defects, c) outputting potential defects, with the detection in step a) taking place from outside the chemical reaction space an image analysis being carried out with the aid of a self-learning algorithm, the image analysis being carried out having the following sub-steps: b1) a feature map is created with a first neural network b2) the areas of potential interest in the feature map are examined with an area proposal network, whereby areas are identified which could contain an anomaly, whereby anchor boxes are generated, b3) a second neural network in the anchor boxes is used to align a target area in each anchor box, b4) using a multivariate clustering algorithm, analyzing the severity of each anomaly in each anchor box and identifying which anomalies are defects. procedure after claim 1 , characterized in that the first neural network is a convolution-based neural network. procedure after claim 2 , characterized in that the first neural network is trained using a priori knowledge. Method according to one of the preceding claims, characterized in that the detection in step a) takes place at the operating temperature of the chemical reaction space, in particular the coke oven. Method according to one of the preceding claims, characterized in that the detection in step a) takes place in a spectral range from 350 nm to 1700 nm, preferably in a spectral range from 350 nm to 1100 nm. Method according to one of the preceding claims, characterized in that the detection in step a) takes place with a multi-camera system. Method according to one of the preceding claims, characterized in that the detection in step a) takes place with a video camera. Method according to one of the preceding claims, characterized in that in step a) a large number of recordings are recorded and these are combined by means of photogrammetry to form a virtual image of the chemical reaction space, in particular of the coke oven, the virtual image created by means of photogrammetry in step b ) is analyzed. Method according to one of the preceding claims, characterized in that the detection in step a) takes place through at least one first opening and one second opening of the chemical reaction chamber, in particular of the coke oven. Method according to one of the preceding claims, characterized in that each image in which a specific defect cannot be clearly identified in step b4) is presented to a specialist for a decision. procedure after claim 10 , characterized in that the image classified by the expert is used to train the multivariate clustering algorithm. Method according to one of the preceding claims, characterized in that the output in step c) in the case of minor defects takes place with a prognosis as to when a major defect is to be expected. Method according to one of the preceding claims, characterized in that the outputting in step c) includes the type of defect, the severity of the defect and the position of the defect. Method according to one of the preceding claims, characterized in that the severity of the defect is output in step c) in three to five degrees of severity. Method according to one of the preceding claims, characterized in that the output in step c) takes place cumulatively for all defects within the chemical reaction space, in particular the coke oven, and an evaluation based on all detected defects for a need for repair is determined. Method according to one of the preceding claims, characterized in that the output includes defects selected from the list of broken edge, spalling, deposit, breakthrough point, ceramic weld point. Method according to one of the preceding claims, characterized in that a feature map is created in step b1) on the basis of color intensity, shading and geometry. Method according to one of the preceding claims, characterized in that the method is repeated, the repetition interval being selected from the group at each opening, daily, weekly, monthly, quarterly, annually. Method according to one of the preceding claims, characterized in that the image analysis in step b) takes into account other images which have already been captured earlier from the same point in the chemical reaction space, in particular the coke oven.

Images