METHOD FOR ADJUSTING THE FIBER-MAKING EQUIPMENT

Abstract

The invention proposes a method for adjusting at least one fiber-forming parameter of a fiber-forming device used in the manufacture of mineral wool. The method includes: generating shape information of the spinning machine of the fiber-forming device by creating images of the markers on the spinning machine and mapping, using image analysis, the positions of the imaged markers with their corresponding physical positions; quantifying the physical melt parameters relative to the spinning machine by analyzing, on the basis of shape information and using a computer vision system, the imaged spinning machine with the melt portion to be fiberized falling on it; and calculating, on the basis of the melt information, the fiber-forming parameters and their target values, the fiber-forming parameters depending on the melt information, and, if the fiber-forming parameters deviate from the target value, adjusting the fiber-forming device to reduce the deviation from the target value.

Description

[0001] METHOD OF ADJUSTING FIBERISING APPARATUS

[0002] FIELD OF THE INVENTION

[0003] The present invention involves the measurement of one or more melt parameters in a fiberising apparatus used for the manufacture of mineral wool, and the modification of one or more fiberising parameters by adjusting the apparatus. Typically, a melt parameter measured is the location of impact of the melt on the spinner, and the apparatus is adjusted to change this location.

[0004] BACKGROUND

[0005] Mineral wool can be manufactured using a variety of different techniques. Typically, a suitable material, such as various types of rock, is heated in a furnace until molten. This produces mineral melt, which is then centrifugally fiberised with a fiberising apparatus. A fiberising apparatus comprises a series of spinners rotating in an opposite direction to the previous spinner in the series and arranged such that melt poured on to the top spinner is thrown in sequence onto the surface of each subsequent spinner, or thrown off the sequence of spinners. The resulting fibres are collected as wool.

[0006] The performance of the manufacturing process will vary with many factors, for example, the angle of melt impact formed, with respect to an axis of rotation of the spinner, between an upright radial axis of the spinner and the location of impact of the melt. If the angle is too small, there will be an increased amount of melt splash as the melt impacts the subsequent wheels too high up. If the angle is too large, the melt does not make sufficient contact with the subsequent wheels and the manufacture process is less efficient.

[0007] Commonly, an operator will monitor the manufacturing process and attempt to minimize waste and maximise efficiency. There are many adjustments which can be made to the fiberising apparatus, such as the position of the trough from which the melt pours onto the spinners. Attempting to make these adjustments by hand to optimize the process requires careful monitoring. Further, determining that there is a need for adjustment is challenging. This is because it is difficult to monitor the process visually due to the scale of the fiberising apparatus.

[0008] It is known to use cameras to capture video of the melt stream inside the apparatus and allow an operator to view this remotely. However, this still requires skilled assessment of the images to determine if adjustment is necessary, and further requires the operator to take manual action to make the necessary adjustments. As such, the efficient operation of the apparatus cannot be guaranteed.

[0009] There is therefore a need to gain better control over the manufacture process while reducing the amount of operator input required. This needs to be achieved while maintaining the quality of the output and a safe working environment around the apparatus.

[0010] SUMMARY OF INVENTION

[0011] According to a first aspect, there is provided a method for adjusting at least one fiberising parameter of a fiberising apparatus used in manufacture of mineral wool, the method comprising: generating shape information of a spinner of the fiberising apparatus by imaging markers on the fiberising apparatus and mapping, using image analysis, positions of the imaged markers to their corresponding physical location; quantifying physical melt parameters relative to the spinner by analysing, based on the shape information and using a computer vision system, an imaged spinner with melt to be fiberised falling thereon; and calculating, based on the melt parameters, a fiberising parameter and a target value thereof, the fiberising parameter being dependent on the melt parameters, and, if the fiberising parameter deviates from the target value, adjusting the fiberising apparatus to reduce the deviation from the target value.

[0012] The method allows for better control over the manufacture process while reducing the operator input required. Calibration of the computer vision via generation of shape information allows for calculation of three-dimensional (3D) quantities, such as the location of impact of the melt on the spinner, using only a single imaging device. The calculation of melt parameters via computer vision enables constant monitoring of the process without the need for a correspondingly large number of operators or a significant amount of operator time. Additionally, the calculation of target values of fiberising parameters and the subsequent adjustment of the fiberising apparatus further reduces the need for operator input, whilst increasing the control over the fiberising process versus manual adjustment by operators.

[0013] In some examples, the markers are located on the fiberising apparatus, for instance, on a support structure of the spinner. In other words, in these examples, the markers are not located on the spinning wheel, such as only on parts of the fiberising apparatus other than the spinning wheel. In these examples, the location of the markers is determined by computer vision, from which the location of the fiberising apparatus in the field of view of the imaging device. In some examples, the shape information of the spinner is generated with reference to the location of the markers and to preset data about the spinner or spinning wheel. The preset data may, in some examples, be a 3D model or 3D mapping of the spinner or spinning wheel. In other cases, the preset data may be the locations of markers previously captured. In such examples, the shape information may be generated using the location of the fiberising apparatus in the field of view of the imaging device, as determined from the location of the markers, to map the preset data into the field of view of the imaging device.

[0014] The preset data may be generated in some examples by an Al model which uses radiance fields (“NERF”s) or Gaussian splatting to reconstruct an accurate 3D model of the spinner and / or spinning wheel. Alternatively, or in addition, the preset data may comprise LIDAR, RADAR or other ranging or rangefinding measurements of the spinner and / or spinning wheel.

[0015] In other examples, the markers are located on the spinner or on the spinning wheel. These examples are discussed further below, though, it should be understood that all the examples below are also intended to be considered in combination with the previous example, where applicable.

[0016] The imaging device used to observe the melt flow may be calibrated via generation of shape information of the spinner. The shape information typically comprises locations of particular points on the spinner, locations of edges of the spinner, and / or perspective transformations from two-dimensional (2D) points in the field of view of the camera to 3D points in space. Typically, the markers on the spinner comprise a calibration pattern on the surface of the spinner. The calibration pattern may be an integral part of the spinner itself or placed on the surface. Regardless of whether integral or placed on the spinner surface, the calibration pattern may denote at least one (or the) physical edges of the spinner or an upright radial axis of the spinner. During the generation of the shape information, these edges and / or the upright radial axis can be detected, allowing for calculation of distances, such as distances from the spinners edge. Calculating these distances provides a mechanism for ensuring the melt flow stays between the edges of the spinner. There may be present additional markers, or geometric patterns such that the shape, size and location of the spinner can be assessed and accounted for in the computer vision algorithms.

[0017] A calibration pattern may be a regular pattern, such as a chequerboard pattern, or a pattern formed of an array or grid of repeating elements, icons or images. Alternatively, the pattern may be an irregular pattern, which does not repeat, such as a gradient, an irregular array of elements, which may be the same, or may be different, or a more complex image. A calibration pattern may comprise repeating and non-repeating elements, for example, respectively formed of any of the above elements in combination.

[0018] The calibration pattern may be placed on the surface of the spinner as a physical coating, covering, wrap or layer thereupon. Alternatively, the calibration pattern may be formed by projection (of light) onto the spinner. In this case, the calibration pattern may be formed on the surface of the spinner by a projector or other lighting techniques. In one example, the calibration pattern is formed on the surface of the spinner using a laser projector. The calibration pattern in this case could be a laser projected grid of structured light, though, of course, any other calibration pattern discussed above could equally be projected onto the surface of the spinner. The laser projector may be located next to the camera, or in a different location. The shape information may be generated from a single image of the spinner, but, more commonly, may be generated from multiple images of the spinner. The shape information may be generated by any implementation on a computer, including by classical algorithms, by machine learning, or by Al algorithms. In some examples, an Al model which uses radiance fields (“NERF”s) or Gaussian splatting may be used to reconstruct an accurate 3D model of the spinner and / or spinning wheel.

[0019] Many melt parameters can be calculated using the computer vision, such as the stability of the melt flow or the presence of splashes. Typically, at least one of the physical melt parameters comprises a location of impact of the melt on the spinner is calculated. Monitoring the location of impact of the melt is beneficial as the quality of the wool and the amount of waste produced vary strongly with the location of impact.

[0020] Additionally or alternatively, locations of the axial edges of the melt flow, with respect to the axis of rotation of the spinner, may be calculated, such as by the at least one of the physical melt parameters comprising the axial edges of the melt flow with respect to the axis of rotation of the spinner. Measuring the axial edge locations permits detection of the melt flow becoming misaligned from the spinner and flowing off or over the edges of the spinner instead of directly onto the surface. Detecting the misalignment is beneficial as a misaligned melt flow is undesirable, causing wastage.

[0021] Using the measured melt parameters, target values for fiberising parameters of the fiberising apparatus are calculated. These target values are typically calculated to reduce waste and increase the quality of the output. One of the at least one fiberising parameter which may be monitored, and which may have a target value, is a distance between an (i.e. at least one) axial edge of the melt flow and an axial edge of the spinner. In this case, the target value could be set to keep the melt flow a prescribed distance from the edge of the spinner. A further example is the location of a melt trough, which is the trough from which the melt flows onto the spinner. Thus, one of the at least one fiberising parameters may comprise a location of a trough from which the melt flows onto the spinner. This target value can, be calculated as a function of the location of impact of the melt, with the objective of keeping the location of impact optimally positioned to minimize waste. Accordingly, the target value of the at least one fiberising parameter may further be a location of impact of the melt on the spinner. Calculating these target value allows for more accurate adjustment of the apparatus as compared to assessment by eye.

[0022] An optimal location of impact of the melt can be found by consideration of the angle formed, with respect to an axis of rotation of the spinner, between an upright radial axis of the spinner and the location of impact of the melt. Typically, when one of the at least one fiberising parameters comprises a location of a trough from which the melt flows onto the spinner and the markers on the spinner comprise a calibration pattern on a surface of the spinner denoting at least one physical edge of the surface, when the target value of the at least one fiberising parameter is a location of impact of the melt on the spinner, of the at least one fiberising parameter is the location of impact where, with respect to an axis of rotation of the spinner, an angle formed between an upright radial axis of the spinner and the location of impact of the melt is in the range of 15 to 35 degrees (°), the angle being in the same direction as the direction of rotation of the spinner. Calculating an angle instead of an absolute position has the advantage of being independent of the size of the spinner.

[0023] Other fiberising parameters which could be calculated include the energy efficiency of the fiberising apparatus, which could have a target value of an efficiency of at least some predetermined amount, or the level of surface damage of the spinners themselves, where the target could be set such to ensure the spinner has even wear across the whole surface.

[0024] The calculation of the fiberising parameters and the corresponding target values can be based directly on the measured melt parameters, but typically will be based instead on an average of these melt parameters over several measurements. Thus, the at least one fiberising parameter may be calculated from a moving average of the physical melt parameters. As the movement of the melt flow can be highly varied, taking averages of melt parameters like its position or width increases the reliability of the fiberising process and leads to fewer unnecessary or erroneous adjustments of the apparatus.

[0025] If the fiberising parameter deviates from the corresponding calculated target value, the fiberising apparatus can be adjusted to reduce the discrepancy. In the case of the position of the melt trough, the target position could be calculated to keep the location of melt impact in an optimal position. The adjustment of the fiberising apparatus would then comprise moving the melt trough appropriately, commonly by means of a hydraulic system or stepper motors. Accordingly, adjusting the fiberising apparatus may include adjusting the location of the trough with a hydraulic system or stepper motor. Moving the melt trough requires less mechanical complexity than moving the spinners. Moving the melt trough by automatic means reduces the amount of operator input required as well as permitting a much higher frequency of adjustment, compensating for changes in the melt flow faster and allowing for a higher efficiency manufacturing process.

[0026] In the case where the imaging of the melt flow is imaging performed by a camera, the camera will typically be in a fixed location such that it has a constant field of view containing the spinner and the melt. The shape information used to calibrate the computer vision system is dependent on the location of the camera, so it is desirable to know if the camera has moved, in order to recalculate or invalidate the shape information. Detecting movement of the camera can be done by including a calibration target in the field of view of the camera, which could be attached to the fiberising apparatus, and identifying, using image analysis, the location of the calibration target in the field of view. As such, the imaging may be imaging by a camera, the camera being in a fixed location, a field of view of the camera including a calibration target, and further comprising identifying, using image analysis, the location of the calibration target in the field of view and detecting, based on the image analysis, movement in a position of the camera. If the location of the calibration target changes, it can be deduced that the camera has moved. Typically, a human operator will be alerted in the case of camera misalignment. This increases the reliability of the manufacture process, as well as increasing the safety of the human operators around the apparatus. In some examples, the adjustment of the fiberising apparatus may be carried out in automated fashion, which may be by mechanical, electrical, electro-mechanical, hydraulic or pneumatic means. Alternatively, in some cases, the adjustment of the fiberising apparatus may be carried out by a human operator. Therefore, in order to aid the operator, in some examples, the method may further comprise outputting the calculated fiberising parameter and the target value, the fiberising apparatus being adjustable, based on the output, to reduce the deviation from the target value.

[0027] The outputting may be outputting directly to a screen, display, or any (arbitrary) display device. Alternatively, the output may be outputting by other means, including by transmission over a network, physical readouts such as dials, or by light or sound. Typically, the outputting is outputting via a screen, where the screen may also display the camera output with annotations representing the fiberising parameter and the target value.

[0028] In order to increase the accuracy of the shape information, in some examples, the generating shape information comprises imaging second markers on the fiberising apparatus and mapping, using image analysis, positions of the imaged second markers to their corresponding physical location. This can increase the accuracy of the shape information by assisting in locating the spinner in space.

[0029] In some examples, melt flows from a trough onto the spinner. In such examples, it is common that (at least a portion of) the melt will cool when flowing from the trough, forming “slag”, i.e. cooled melt, which remains attached to the trough. This can interfere with the performance of the fiberising apparatus, limiting or interfering with the flow of melt. Accordingly, in some examples, one fiberising parameter (potentially of a plurality of fiberising parameters) comprises a quantity of slag formed on a trough from which the melt flows onto the spinner.

[0030] Further, in some examples, the step of adjusting the fiberising apparatus comprises removal of slag from the trough from which the melt flows onto the spinner. Slag may be removed using mechanical, electro-mechanical, hydraulic, or pneumatic means, or may be removed manually. In some examples, slag is removed using a mechanical arm, by passing the arm through the flow of melt such that the arm impacts the slag. In this way, the arm is used to dislodge slag from the trough, thereby removing it.

[0031] According to a second aspect, there is provided a system for adjusting at least one fiberising parameter of a fiberising apparatus used in manufacture of mineral wool, the system comprising an imager arranged to image the fiberising apparatus; calibration markers on the fiberising apparatus; and a processor. The processor is configured to generate shape information of a spinner of the fiberising apparatus by imaging markers on the fiberising apparatus and mapping, using image analysis, positions of the imaged markers to their corresponding physical location; quantify physical melt parameters relative to the spinner by analysing, based on the shape information and using a computer vision system, an imaged spinner with melt to be fiberised falling thereon; and calculate, based on the melt parameters, a fiberising parameter and a target value thereof, the fiberising parameter being dependent on the melt parameters, and, if the fiberising parameter deviates from the target value, adjust the fiberising apparatus to reduce the deviation from the target value.

[0032] BRIEF DESCRIPTION OF DRAWINGS

[0033] An example process and an example apparatus are described in detail herein with reference to the accompanying figures, in which:

[0034] Figure 1 shows a schematic of an example apparatus;

[0035] Figures 2a, 2b and 2c show a schematic of the elements required for calibration; Figure 3 shows a flow diagram of the method;

[0036] Figure 4 shows an example of the training data used to train the computer vision algorithms, for the determining of melt parameters;

[0037] Figure 5 shows a further example of the training data used to train the computer vision systems, for the generation of shape information;

[0038] Figure 6 shows an example of detection of the width of the melt flow by the computer vision system; Figure 7 shows four examples of training data used to train the computer vision system to detect the physical bounds of a melt flow impacting a spinning wheel;

[0039] Figure 8 shows six further examples of annotated training data, showing a particular spinning wheel, which can be used to train the computer vision system to detect the location and physical bounds of the melt;

[0040] Figure 9 shows six further examples of annotated training data from the same camera and spinning wheel as Figure 8;

[0041] Figure 10 shows six further examples of annotated training data, showing a particular spinning wheel, which can be used to train the computer vision system to detect the location and physical bounds of the melt;

[0042] Figure 11 shows six further examples of annotated training data from the same camera and spinning wheel as Figure 10;

[0043] Figure 12 shows six further examples of annotated training data, showing a particular spinning wheel, which can be used to train the computer vision system to detect the location and physical bounds of the melt flow; and

[0044] Figure 13 shows six further examples of annotated training data from the same camera and spinning wheel as Figure 12.

[0045] DETAILED DESCRIPTION

[0046] An example system comprising a fiberising apparatus and a system for adjusting at least one fiberising parameter of a fiberising apparatus is generally illustrated at 1 in Figure 1. The system has two spinners 10, 11 (also referred to as a “first spinner” and “second spinner” respectively) onto which melt 30 is being poured from a melt trough 20.

[0047] The melt is produced in a furnace and, in some arrangements, passes through a series of troughs before reaching the melt trough, progressively cooling. The spinners rotate, typically, at a high speed. In various examples, this is at a rotation rate approximately 7000 RPM (revolutions per minute). This fiberises the melt, producing wool. The fall of the melt 30, the first spinner 10, and a calibration target 51 are all inside a field of view 41 of a camera 40. The camera is able to be used to capture images as a video or (stills) picture feed. The images captured by the camera are transmitted to the processor 50. In the example shown in Figure 1 , the transmission of the images is carried out using a wired connection. In other examples, the transmission may include or may be carried out fully by a wireless connection.

[0048] The system is able to make measurements of one or more melt parameters, via computer vision analysis with a processor 50 of images captured by the camera 40, determine fiberising parameters, calculate target values of these fiberising parameters, and subsequently adjust the fiberising apparatus if the fiberising parameters differ from the target values.

[0049] In various examples, adjustment of the fiberising apparatus involves changing the location of the melt trough 20. Adjusting the location of the trough changes the location the melt 30 impacts on the spinner 10. The location of impact of the melt is important to the optimization of the manufacturing process. This is because positioning the location of the impact of the melt optimally reduces the amount of waste and the amount of wear on the apparatus. Therefore, in some examples, one of the fiberising parameters is taken as the angle of melt impact 25 with respect to the upright (or vertical) of the axis of rotation of the spinner 13, and the target value is set at an optimal value, in the range of 15 to 35 degrees.

[0050] To allow the melt trough 20 to move, in some examples, the melt trough is connected to a movement mechanism 23 which allow it to move parallel to the axis of rotation of the spinner 10 and parallel to the axis defined by the direction of rotation at the top of the spinner, whilst maintaining a constant vertical distance from the spinner.

[0051] In an example, the movement mechanism 23 may comprise a set of rails. In an example, the set of rails may comprise a rail permitting movement of the melt trough parallel to the axis of rotation of the spinner. In an example, the set of rails may comprise a rail allowing movement of the melt trough parallel to the axis defined by the direction of rotation at the top of the spinner. In a different example, the trough is also free to move along the vertical axis separating it from the spinner.

[0052] The movement mechanism 23 is connected to a controller 21 , which moves the trough in response to a signal from the processor 50. In Figure 1 , the signal is be transmitted using a wired connection. In other examples, the signal transmission may include or may be carried out fully by a wireless connection.

[0053] The controller is able to move the trough by means of an actuator, such as a linear actuator, or hydraulic system in some examples. In other example, the movement of the trough is achieved by use of a stepper motor, or by other means.

[0054] In some examples, to detect if the camera 40 has remained in the same location and orientation since the shape information was generated, the location of a target 51 in the field of view 41 of the camera is recorded. In various examples, the location of this target is detected using the computer vision every time a measurement is taken or every time the camera is operated. If the location of the target has changed, it is deduced that the location or orientation of the camera has changed and therefore the generated shape information is no longer valid. If the shape information is no longer valid, the measurements made by the system cannot be taken as accurate. In several examples, this means the adjustments of the fiberising apparatus are disabled, and the operator is informed.

[0055] Before any measurements of melt parameters can be made, the system 1 must be calibrated via generation of shape information. Figure 2a, 2b and 2c depict at least part of the view of the camera 40 in the field of view 41 as a series of composite images showing information relevant to the shape information generation.

[0056] In Figure 2a, the spinner 10 is depicted with melt 30 flowing from the melt trough 20 as is seen by the camera 40. In Figure 2c, the same field of view is shown but without the melt pouring onto the spinner. Instead, the spinner is shown with a calibration pattern 60 applied to the exterior surface. In the example shown in Figure 2c, the pattern is a chequerboard pattern in black and white. In other examples, the pattern could be integral to the spinner instead of being applied separately. Additionally or alternatively, in other examples the pattern is also able to be a pattern other than a chequerboard pattern, for example, a test pattern or resolution pattern.

[0057] This pattern is detected by the computer vision algorithm running on the processor 50 and used to generate the shape information describing the mapping between 2D pixels in the camera’s field of view and the points in 3D space. In some examples, an applied chequerboard pattern is used. The corners of the squares of the chequerboard pattern are detected using a corner detection algorithm, such as a Moravec, Harris or SUSAN corner detection algorithm. In other examples, binary-based methods or contour-based methods of corner detection could be used. In still other examples, a machine learning algorithm or neural network could be trained on annotated data of chequerboard patterns and used to detect the corners of the chequerboard pattern. In other examples, the calibration pattern may not comprise a chequerboard pattern and other properties of the pattern, such as edges or specific features, may be detected.

[0058] After detection of the location of the corners of the chequerboard pattern, the location of each corner in the 2D field of view of the camera can be mapped onto the known position of the corner in 3D space. In one example, this mapping comprises the calculation of a rotation matrix and a translation mapping the coordinate system of the camera with the origin at the camera’s optical centre onto a 3D co-ordinate system with an origin at the centre of the spinner. Additionally or as an alternative, this mapping comprises calculation of a skew coefficient compensating for non-perpendicular image axes of the camera. Additionally or as an alternative, lens distortion of the camera is compensated for dependent on the focal length of the lens. This distortion may, in some examples, be radial or tangential.

[0059] In Figure 2b a composite image of the spinner 10 with calibration pattern 60 applied and the spinner with falling melt 30 is shown. The computer vision system running on the processor 50 is used to detect the corners of the chequerboard pattern, and from the corners the edges 55 of the chequerboard pattern are located.

[0060] In various examples, the width 53 of the falling melt is also determined. The location of the falling melt 30 in the image captured by the camera 40 is determined by means of machine learning, explained in detail later. The location of the falling melt in the image is transformed using the shape information into 3D space. The difference between the location of the two edges of the melt impacting on the spinner 10 gives the width based on the width of the physical spinner (i.e. the distance from front to back) being known. In some examples, the width of the melt may be determined at the location of impact of the melt.

[0061] In addition to a chequerboard calibration pattern, in the example shown in Figure 2c, the applied calibration pattern 60 also comprises an edge marker 61 and a top marker 63. The edge marker is a line running around the circumference of the spinner and denoting the location of the outer edge of the spinner. In another example, the system 1 could be configured to detect the marker along the inner edge of the spinner instead.

[0062] The edge marker 61 is detected by the computer vision algorithm running on the processor 50. The detection is used to calculate distances from the edge of the spinner 10.

[0063] The top marker 63 is a line along the axis of rotation of the spinner, denoting the very top of the spinner. The top marker is applied by hand to the first spinner 10 or as part of the manufacture process of the spinner. The top marker is either applied as part of the calibration pattern 60, as a supplement to the calibration pattern or independent from the calibration pattern.

[0064] During setup for the calibration, the top marker 63 is positioned at the top (i.e. the highest point or surface) of the first spinner 10. In some examples, this is achieved via matching against or aligning with a marking on the fiberising apparatus (not shown), though in other examples it could be simply aligned by eye (i.e. without use of a marking or marker). The top marker is detected by the computer vision algorithm and used to help define the extent of the spinner.

[0065] The accurate positioning of the markers is important to the performance of the system. In one example, sufficiently good performance is obtained with placement of the edge marker 61 to a precision of ± 2 millimetres (mm), and alignment of the top marker with a radial precision of ± 2°.

[0066] The calibration data is obtained using different exposure settings on the camera to settings used when the camera is imaging flow of the melt. This is because the melt, due to its temperature, is significantly brighter than the spinner under ambient conditions, necessitating higher sensitivity during the calibration measurements.

[0067] In Figure 3, an example schematic 3 is shown, detailing the steps of one example process for adjusting a fiberising parameter. The process begins with a calibration step 70, where the shape information is generated by imaging the spinner 10 with calibration markers 60 and processing this image with the computer vision system running on the processor 50, such as set out above in relation to Figure 2a to 2c.

[0068] Once the system is calibrated, the remaining steps shown in Figure 3 are repeated in sequence until the process is terminated. These steps comprise an image capture step 71 , optional camera misalignment step 72, melt parameter step 73, fiberising parameter step 75 and an adjustment step 77.

[0069] The image capture step 71 comprises capturing an image with the camera 40 and transmitting the image to the processor 50. The image can then be analysed using the machine learning algorithms in the computer vision system, explained in detail later.

[0070] The camera misalignment step 72 comprises measuring the location of the calibration target 51 in the field of view 41 of the camera 40 using the computer vision system running on the processor 50. When located, it is identified whether and ensuring that the location of the calibration target in the field of view is within a predetermined threshold of the location measured at the time of the calibration step 70.

[0071] Using a small threshold value allows for some minor movement in the camera location due to vibration of the apparatus 1 without tolerating a major deviation. Therefore, it is confirmed that the camera 40 has not moved since the generation of the shape information.

[0072] If the camera 40 is identified as having moved, the process is terminated in some examples. This is because the shape information can no longer be relied on to transform between the 2D image and the 3D quantities being measured, making the measurements inaccurate. Such inaccuracies could lead to erroneous adjustments of the system, possibly reducing efficiency or endangering the apparatus or operators.

[0073] In the melt parameter step 73, the physical parameters of the melt are determined via the computer vision system running on the processor 50, trained with suitable training data. In this example, the location of impact of the melt 30 on the spinner 10 is calculated, as well as the width of the melt and the distance of the melt from the edges of the spinner along the axis of rotation.

[0074] In some examples, the measurement of the location of impact is accurate enough to determine the angle of melt impact 25 to ± 2°. Here, the width of the melt is measured at the location of impact of the melt on the spinner. The angle of melt impact is determined relative to the centre of the measured width line 53. This has the disadvantage that the view of the camera may be obscured by the presence of slag in the flow. As such, in other examples, additionally or as an alternative, the width of the melt may be measured below the location of impact, i.e. with increased angle with respect to the angle of impact 25. Measurements made below the location of impact are less likely to be obscured by slag as the slag would have to be larger. However, this can be inaccurate due to geometric distortion as the melt becomes detached from the surface of the spinner as it is thrown to the subsequent spinner 11 . The shape information is unable to account for this detachment so the measurement of the width can be incorrect. In the fiberising parameter step 75, fiberising parameters are calculated as a function of the melt parameters found in the melt parameter step 73 on the processor 50. In this example, one of the fiberising parameters is the angle 25 of melt impact, with respect to an axis of rotation of the spinner, formed between an upright radial axis 13 of the spinner 10 and the location of impact of the melt, the angle being in the same direction as the direction of rotation of the spinner. A second fiberising parameter is the distance between the melt and the radial edges of the spinner, the location of the edges being known from the shape information generated from the calibration marker 61 .

[0075] Each of these fiberising parameters has an associated target value, which is then calculated on the processor 50. In other examples, there may be fiberising parameters without target values. The target value may be known in advance or may require calculation from external factors, depending on what the desired result is.

[0076] In some examples, the target values of fiberising parameters may be calculated to maximise energy efficiency, or to minimise the wear to the apparatus over a long period of operation. In this example, the target values of the fiberising parameters are set to minimise waste and position the melt flow optimally. Therefore, the target value of the angle of melt impact 25 is determined to minimise the amount of waste produced, approximately between 15 to 35 degrees.

[0077] In another example, the target values of the fiberising parameters may be calculated to ensure that the melt does not flow over the front or of the spinner 10 back (i.e. edges with respect to the axis of rotation). The target distance of the melt from the edge of the spinner is set to ensure sufficient distance between the melt and the edge of the spinner.

[0078] In other examples, the target values may be set by the operator, such that the system maintains user-specified values of fiberising parameters automatically. This means the fiberising parameters of the fiberising apparatus can be controlled manually, but with less input from the operator required versus making manual adjustments to ensure fiberising parameters stay constant.

[0079] In the adjustment step 77, the fiberising apparatus is adjusted if necessary. In some examples, to determine whether to adjust the apparatus, the processor 50 compares the difference between the fiberising parameters and the target values of the fiberising parameters to a threshold value.

[0080] In various examples, the target values are compared to moving averages of the fiberising parameters over a small number of measurements. Using a moving average reduces the impact of incorrect measurements, as well as compensating for the random movement of the melt flow. This leads to a delay in several examples, such as to an approximately 10 second delay, on making adjustments to the fiberising apparatus. This delay in regulation can mean that the system is less efficient as a misaligned melt flow will be tolerated for a small amount of time before adjustments are made.

[0081] Different fiberising parameters may have different threshold values. Different threshold values help the process account for the fiberising apparatus being more or less sensitive to a particular fiberising parameter.

[0082] If the difference between a particular fiberising parameter and the respective target value is higher than a threshold, the apparatus is adjusted. In this example, the adjustment required is the movement of the melt trough 20. The processor sends a signal to the controller 21 , which uses mechanical means to move the melt trough along the axes 23 and reposition the melt flow.

[0083] In this example, the melt trough 20 is moved using hydraulics. The system is fitted, in various examples, with inductive sensors to ensure smooth operation and prevent any deviation beyond the hydraulic system’s optimal range.

[0084] In some examples, the process is configured to make small changes incrementally to gauge the impact of the changes on the system and avoid overcompensating and missing the target value. However, in other examples, e.g. in the case of an electronic parameter like a voltage, the apparatus may be adjustable immediately to a desired value.

[0085] Making small changes and repeatedly reassessing the state of the fiberising apparatus means that the relative values of adjustments to the fiberising apparatus are all that is required, as opposed to needing to know absolute values.

[0086] In some examples, the computer vision system uses machine learning algorithms to identify melt parameters from the images captured by the camera 40. These melt parameters include the location of the impact of the melt on the spinner 10 and the location of the edges of the falling melt.

[0087] In Figure 4a, 4b and 4c, examples of the training data used to train the machine learning algorithms in the computer vision system are shown. The machine learning algorithms are of an image classifier type, capable of identifying elements in an image with a given certainty (a function of how closely the element matches the elements in the training data, as a percentage, for example). These algorithms typically require a large corpus of annotated training data, where pairs of images and the desired output from the algorithm for that image are given, and statistical mappings between the training data images and the desired outputs are made. Once this training process is complete, the machine learning algorithms are trained and no longer need access to the corpus of training data.

[0088] In Figure 4a, an image captured by the camera 40 is shown, with melt 30 falling from the melt trough 20 onto the first spinner 10. Shown also as a line (in red in the drawings as originally filed) is an annotation 81 , marking the location of impact of the melt onto the spinner across the full width of the melt. A large series of images like this, with annotations (i.e. the desired output from the machine learning component of the computer vision system), form a corpus of training data for detecting the location of impact of the melt. In Figure 4b, a second example of this type is shown, with a slightly more complicated geometry, as a bright splash 86 is excluded from the annotated region. Training data including anomalies such as splashes and other material helps make the algorithm resilient to their presence in the images. As a further example of this, Figure 4c shows a case with presence of substantial slag 83 (an example of other material that may be considered an anomaly), where the slag is occluding the camera’s view of the melt flow. In this case, the annotated region 81 is split into several parts, and the edges of the melt flow are more difficult to detect.

[0089] In Figure 5, training data for the generation of shape information is shown. In the upper panel, the spinning wheel 10 is shown with a calibration pattern 60 on the surface. In this example, a chequerboard calibration pattern has been applied to the spinning wheel. It provides an edge marker 61 , shown in red in the drawings as originally field, and a top marker 63, shown in red in the drawings as originally filed. In this example, as shown in the lower panel, the computer vision system detected the location of the top marker 63. This is shown by the green line 57 in the drawings as originally filed. The computer vision system further detects the location of the corners of the chequerboard pattern, shown by the red dots 54 in the drawings as originally filed. This allows the calculation of the shape information as expounded above.

[0090] In Figure 6, the detection of the location of the impact of the melt 30 on the spinner 10 is shown, in the presence of slag in the melt flow. The shaded region 87 shown in orange shows the region the computer vision system determines to be falling melt yet to impact the spinner. The shaded region 89 (shown in blue in the drawings as originally filed) shows the region the computer vision system determines to be melt which has impacted the spinner. The computer vision system can determine this in part due to the shape information used to calibrate the system, the key points of which are shown as the lines 55 (shown in orange in the drawings as originally filed). The computer vision system has made an estimate of the width 53 of the melt flow, which does not span the full width of the melt due to the presence of slag disrupting the measurement.

[0091] Figure 7 shows four further examples of annotated training data which can be used to train the computer vision system to detect the location and physical bounds of the melt flow. In these examples, the images are annotated by marking two regions on the raw image. The falling melt region 87 marks the region where the melt has not yet impacted on the spinning wheel. The spun melt region 89 marks the region where the melt has impacted the spinning wheel. Many such images are required to train the computer vision system to recognise these regions in an arbitrary input. By using training data from a large number of different cameras and spinning wheels, the ability to classify spinning and wheels the computer vision system was not trained on is developed.

[0092] Figure 8 shows six further examples of annotated training data which can be used to train the computer vision system to detect the location and physical bounds of the melt flow In these examples, the images are annotated by marking regions on the raw image. The falling melt region 87 marks the region where the melt has not yet impacted on the spinning wheel. The spun melt region 89 marks the region where the melt has impacted the spinning wheel. Six further examples from the same spinning wheel and camera can be seen in Figure 9.

[0093] Figure 10 shows six further examples of annotated training data which can be used to train the computer vision system to detect the location and physical bounds of the melt flow. Six further examples from the same spinning wheel and camera can be seen in Figure 11 .

[0094] Figure 12 shows six further examples of annotated training data which can be used to train the computer vision system to detect the location and physical bounds of the melt flow. Six further examples from the same spinning wheel and camera can be seen in Figure 13.

[0095] Each of the sets of annotated training data shown in Figure 7 to Figure 12 is a subset of a larger set of annotated training data. The original images of the set of annotated training data were collected and annotated to identify regions of falling melt, and melt that has reached the spinner. These were then provided to the computer vision system and used to train the computer vision system.

[0096] Overall, the annotated training data set includes around 2000 images. As can be seen from Figures 7 to 12, a number of different scenarios are incorporated into the annotated training data. These include “clean” images, such as the top row and bottom left images shown in Figure 7. Other images in the annotated training data include splashes, such as in the bottom right image shown in Figure 7. Some images include “shot”, which is typically individual drops of melt that have separated from the rest of the melt. As shown in Figure 4 and Figure 6, slag is shown in some images. This forms on the gutter.

[0097] A number of different relative camera and spinner positions were used to collect the images for the annotated training data set. Across the whole annotated training data set, the annotated training data used around ten different relative camera and spinner positions. In various examples the annotated training data set includes a different number of images and / or a different number of cameraspinner relative positions.

Claims

CLAIMS1 . A method for adjusting at least one fiberising parameter of a fiberising apparatus used in manufacture of mineral wool, the method comprising: generating shape information of a spinner of the fiberising apparatus by imaging markers on the fiberising apparatus and mapping, using image analysis, positions of the imaged markers to their corresponding physical location; quantifying physical melt parameters relative to the spinner by analysing, based on the shape information and using a computer vision system, an imaged spinner with melt to be fiberised falling thereon; and calculating, based on the melt parameters, a fiberising parameter and a target value thereof, the fiberising parameter being dependent on the melt parameters, and, if the fiberising parameter deviates from the target value, adjusting the fiberising apparatus to reduce the deviation from the target value.

2. The method according to claim 1 , wherein the markers comprise markers on the spinner.

3. The method according to claim 1 or claim 2, wherein at least one of the physical melt parameters comprises a location of impact of the melt on the spinner.

4. The method according to any one of the preceding claims, wherein at least one of the physical melt parameters comprises the axial edges of the melt flow with respect to the axis of rotation of the spinner.

5. The method according to any one of the preceding claims, wherein one of the at least one fiberising parameters comprises a location of a trough from which the melt flows onto the spinner.

6. The method according to claim 5, wherein, adjusting the fiberising apparatus includes adjusting the location of the trough with a hydraulic system or stepper motor.

7. The method according to any one of the preceding claims, wherein the markers on the spinner comprise a calibration pattern on a surface of the spinner denoting at least one physical edge of the surface.

8. The method according to claim 5 and claim 7, wherein the target value of the at least one fiberising parameter is a location of impact of the melt on the spinner.

9. The method according to claim 8, wherein the target value of the at least one fiberising parameter is the location of impact where, with respect to an axis of rotation of the spinner, an angle formed between an upright radial axis of the spinner and the location of impact of the melt is in the range of 15 to 35degrees, the angle being in the same direction as the direction of rotation of the spinner.

10. The method according to claim 4 and claim 7, wherein the target value of the at least one fiberising parameter is a distance between at least one axial edges of the melt flow and at least one axial edges of the spinner.11 . The method according to any one of the preceding claims, wherein one of the at least one fiberising parameters comprises an energy efficiency of the fiberizing apparatus.1 . The method according to any one of the preceding claims, wherein the imaging is imaging by a camera, the camera being in a fixed location, a field of view of the camera including a calibration target, and further comprising identifying, using image analysis, the location of the calibration target in the field of view and detecting, based on the image analysis, movement in a position of the camera.

13. The method according to any one of the preceding claims, wherein the at least one fiberising parameter is calculated from a moving average of the physical melt parameters.

14. The method according to any one of the preceding claims, further comprising:outputting the calculated fiberising parameter and the target value, the fiberising apparatus being adjustable, based on the output, to reduce the deviation from the target value.

15. The method according to any one of the preceding claims, wherein the markers are formed by projection onto the spinner.

16. The method according to any one of the preceding claims, wherein the at least one fiberising parameter comprises a quantity of slag formed on a trough from which the melt flows onto the spinner.

17. The method according to claim 16, wherein the step of adjusting the fiberising apparatus comprises removal of slag from the trough from which the melt flows onto the spinner.

18. A system for adjusting at least one fiberising parameter of a fiberising apparatus used in manufacture of mineral wool, the system comprising: an imager arranged to image the fiberising apparatus; calibration markers on the fiberising apparatus; and a processor configured to: generate shape information of a spinner of the fiberising apparatus by imaging markers on the fiberising apparatus and mapping, using image analysis, positions of the imaged markers to their corresponding physical location; quantify physical melt parameters relative to the spinner by analysing, based on the shape information and using a computer vision system, an imaged spinner with melt to be fiberised falling thereon; and calculate, based on the melt parameters, a fiberising parameter and a target value thereof, the fiberising parameter being dependent on the melt parameters, and, if the fiberising parameter deviates from the target value, adjust the fiberising apparatus to reduce the deviation from the target value.

Images