METHOD FOR DETERMINING THE GEOMETRY OF A DEFECT AND FOR DETERMINING A LOAD-BEARING LIMIT

Description

[0001] The present invention relates to a method for determining the geometry of a defect and a method for determining a load limit of an object that is subjected to backloading at least during operation.

[0002] One of the key tasks of pipeline inspections, especially with so-called intelligent pigs, is to predict safe operating conditions based on the condition of the pipeline. Pipeline operators are particularly interested in the condition of any welds and the number and size of defects. Defects include, for example, areas with metal loss due to corrosion, cracks, or other weakening of the wall of an object intended, in particular, for the storage or transport of liquid or gaseous media. These include pipes, pipelines, or tanks.

[0003] Knowing the maximum pressure applicable to a pipeline ("maximum burst pressure") at which the pipeline will be destroyed is relevant for the operating pressures that can be set within the pipeline. Accordingly, the accurate prediction of this pressure is important. Currently, to calculate this limit, a flaw is only approximated in terms of its length, width, and depth and is thus considered a box. However, this conservative approach is disadvantageous, especially for metal losses due to corrosion on the outside or inside of a pipeline, as the simplified geometric figures necessarily overestimate the current structure of the flaw. This leads to an underestimation of the object's maximum burst pressure and thus to an underestimation of the permissible operating pressures. An object that can be operated at higher pressures, such as a pipeline or a gas tank, can be operated significantly more economically.

[0004] The state of the art for measuring the corrosion of an object is to have specially trained people evaluate scans of magnetic flux leakage (MFL) data based on magnetic flux leakage (MFL) measurements to determine the size of the (corrosion) defects. The signals displayed in the scans are parameterized in boxes and evaluated. The assumptions required for this evaluation of the measurement results, also known as sizing, are proprietary. On the other hand, the interpretation of the measurement results is heavily influenced by the experience of the evaluators. Ultimately, the quality of the predictions can only be ensured by on-site investigations of the pipeline. This, in turn, entails high costs for the operators. The most widely used industry standard, API 1163, describes the adverse effects of this simplified approach.It is well documented that the quality of this approach depends heavily on the knowledge of the observers. The approaches used in practice are always an interpretation of the data obtained from an MFL measurement run, influenced by subjective factors.

[0005] In addition, approaches are known from scientific, i.e. theoretical, considerations in which the most accurate simulation of the measured signals is achieved using forward models through successive variation and iterative methods of an assumed defect geometry. Neural networks, for example, are used here. Theoretically, these approaches can produce solutions in the sense of a resulting defect geometry, but these solutions are not necessarily realistic. This is particularly true for complex data sets, which lead to unexpected, exotic, and incorrect solutions under certain inverse problems. While in precisely defined and delimited test scenarios such scientific models produce a solution to the described problem in the form of defect geometries, this has not yet been successful for real measurement data, which exhibit a multitude of disturbing influences.

[0006] US 2016 / 0187523 A1 describes eddy current sensors whose data are used in a numerical inversion to determine the thickness of nested pipeline pipes.

[0007] US 6,316,937 B1 discloses the use of different proportionality equations in which the stress values ​​measured on the basis of MFL measurements are plotted against the respective defects depending on the material.

[0008] US 2016 / 0178580 A1 discloses the use of a predefined equation to determine the thickness, width and depth of a defect found in MFL measurements.

[0009] US 2017 / 0176629 A1 discloses another method for evaluating eddy current measurements in nested borehole pipes.

[0010] It is therefore an object of the present invention to show a robust way to accurately reconstruct a real defect from its MFL measurement data set and to calculate the load limit of a corrosion-affected object as accurately as possible.

[0011] The object is achieved by a method according to claim 1 and, with regard to the determination of a load limit, by a method according to claim 15. Advantageous embodiments of the invention can be found in the subclaims which refer back to these claims and in the following description.

[0012] According to the invention, the determination of the geometry of one or more defects is carried out by means of a plurality of competing expert routines, each with at least one separate search strategy or at least one separate algorithm, in particular in parallel on a computer unit.

[0013] Such a method according to the invention for determining the geometry of one or more real, examined defects of a magnetizable object, in particular a pipe or a tank, with a reference data set of the object generated on the basis of one or more MFL measurements, comprises an at least partial representation of the object by means of an EDP unit, in particular on or by an at least three-dimensional object grid, and further comprises a determination, in particular a generation, of an initial defect geometry as the starting defect geometry, in particular on the object grid or on an at least two-dimensional defect grid, a determination, in particular a generation, of a first MFL prediction data set as an initial prediction data set on the basis of the initial defect geometry, in particular by simulating an MFL measurement or assigning an MFL data set, and an iterative adaptation of the initial defect geometry to the geometry of the real defect or defects by means of the IT unit, wherein this adaptation is carried out by means of several expert routines running in competition with one another and preferably in parallel to one another, wherein in respective expert routines a respective expert defect geometry is generated by means of at least one separate algorithm or a separate search strategy and on the basis of the initial defect geometry.

[0014] The expert routine has its own algorithm if at least one of the algorithms available to the expert routine for adapting the defect geometry differs at least partially from the algorithms of another expert routine. Stochastic processes can preferably be used to differentiate the algorithms of different expert routines. Each expert routine has at least one algorithm for adapting the defect geometry; preferably, at least one expert routine has several algorithms available. Likewise, the selection of an algorithm based on stochastic processes can be made or specified within an expert routine.

[0015] According to the invention, it is further provided that a respective expert prediction data set is determined on the basis of the respective expert defect geometry, in particular by simulating an MFL measurement or assigning an MFL data set, wherein the expert defect geometry underlying the respective expert prediction data set is then made available to at least several and in particular all of the expert routines as a new initial defect geometry for further adaptation to the geometry of the real defect(s) if the expert prediction data set is more similar to the reference data set than the initial prediction data set. Subsequently, i.e. for the next comparisons of the respective expert defect geometries with the new initial defect geometry, the expert prediction data set belonging to the new initial defect geometry is used as the new initial prediction data set.A measure of similarity can be calculated using a fitness function. The iterative adjustment using the expert routines continues until a stopping criterion is met. The expert prediction data set and the first MFL prediction data set both show MFL fields corresponding to the assumed defect geometry. These can be calculated or simulated.

[0016] The determination of the expert prediction data set can be carried out within the workflow of the expert routine and / or via a program module controlled separately by a monitoring routine.

[0017] The determination of the expert prediction data set based on the respective expert defect geometry is still carried out by simulating an MFL measurement, which is described below, especially when sufficiently large databases with calculated or measured MFL data for the respective defect geometries are not yet available. Alternatively, the expert defect geometry can also be provided with an MFL data set from a sufficiently extensive database. A combined approach is also possible, in which a database is first searched for existing MFL data and only then, if the search is negative, is a simulation performed. Overall, this can lead to a rapid determination of the respective expert prediction data set.

[0018] The method according to the invention is carried out completely and in particular automatically on a computer unit, which may optionally consist of several computers. The associated computer program can be a single program or it can be a program package comprising a plurality of program modules which, for example, due to resource constraints, run in a distributed manner on different computer systems or subunits and can be stored there on respective computer media. A computer has in particular the typical means of a data processing unit such as one or more processors, at least temporary memory (RAM), data communication means, display and / or input units. While the selection of the reference data set can preferably be user-controlled, the determination of the defect geometry during the iterations takes place automatically.Preferably, before the actual iteration, program parameters can be specified for selecting the algorithms available to the expert routines, determining an initial defect geometry, determining the first prediction data set, and / or an expert prediction data set, each of which displays MFL fields. For example, it can be specified whether the initial prediction data set is to be determined via a simulation of an MFL measurement based on a grid representing the object with the defect or whether it is to be loaded from a database via regression. In particular, in the case of simulating an MFL field, the parameters necessary for comparison with the reference data set, such as the direction of magnetization, strength of magnetization, distance of the sensor from the object surface, and / or speed of the measuring device, can be specified.

[0019] By having expert routines competing for IT resources, each using their own search strategies to find their own solutions to determine the geometry of the real defect, the problem inherent in state-of-the-art theoretical approaches of finding isolated solutions is avoided. Compared to manual evaluation of the data sets, the solution found is not only significantly better but also easier to understand and document. This avoids singular solutions, where an algorithm of any kind cannot make any progress and yet the defect geometry is not realistically represented.

[0020] To simulate the flux leakage data associated with a geometry, a representation of the object on or through a three-dimensional grid is usually necessary. The representation is at least partially represented on this grid, in the sense that at least the part of the object with the defect(s) and preferably adjacent areas are represented by or on the object grid. Alternatively, the flux leakage data can also be determined via a database query, for example, using a regression function.

[0021] Although the method according to the invention can be carried out simultaneously for determining one or more defects within a reference data set, for the sake of simplicity, reference will usually only be made to one defect in the following.

[0022] With a clever choice of object representation, especially through or on the at least three-dimensional object grid, the defect geometry can be assigned as a value to the grid elements or grid points for the MFL simulation. Depending on the geometry of the respective grid, interpolations or grid adjustments may be necessary for this.

[0023] In particular, the defect geometries, which form the basis for determining the associated (MFL or expert) prediction data sets, are defined by defect depths, which represent, for example, the depth of corrosion on an object surface, on the grid nodes of a two-dimensional defect grid. The two-dimensional representation of the defects allows the expert routines to work significantly faster than if the adjustment of the defect geometry is performed on a three-dimensional grid.

[0024] To simulate the MFL measurement of a new defect geometry, the defect grid, particularly a two-dimensional one, is preferably interpolated to the grid points of the object grid, whereby the surface of the object to be represented is adapted to the defect depths of the defect geometry. The simulation is then calculated on the object grid, particularly a three-dimensional one. Alternatively, the flux leakage simulation can also be performed on a two-dimensional grid or using a regression model based on a database of MFL data sets derived from finite element method simulations and / or MFL measurements.

[0025] Based on the initial defect geometry, which is obtained or specified, for example, via a look-up table, a database comparison, or a particularly one-time execution of an expert routine, a first MFL prediction data set is determined as the initial prediction data set, in particular by simulating an MFL measurement. The simulation of the MFL measurement is carried out, for example, using a finite element model as a forward calculation. In the simulation of the leakage flux measurement, the necessary parameters are specified according to the actual measurement. This applies in particular to the magnetization direction, the magnetic field strength, and / or the distance of the sensors above the surface of the object. Based on the initial defect geometry, an initial prediction data set is then obtained as a simulated leakage flux measurement.This data set could already be compared with the object's reference data set, but this usually does not lead to meaningful solutions at the beginning of the iteration.

[0026] A separate routine for defining an initial defect geometry as the starting defect geometry is not absolutely necessary, but it does reduce the computing time required in subsequent program runs. Alternatively, the initial defect geometry can already be the result of a run through an expert routine. The initial defect geometry can also be specified in alternative ways, for example, by a completely flat, so-to-speak defect-free geometry.

[0027] The initial defect geometry is used as the starting defect geometry in the iterative approximation process of the competing expert routines. The expert routines themselves, for example, are independent of each other as separate program modules without direct interaction with each other and can be provided with resources, especially computing time, depending on a monitoring routine or a main module.

[0028] Based on the expert defect geometry developed in a respective expert module, an expert prediction data set is determined for each of these geometries, in particular by simulating an MFL measurement. Thus, for each expert defect geometry, which is available in particular as a 2D data set of depth values ​​of the defect, an expert prediction data set is generated as a simulated MFL measurement. The simulation of the MFL fields based on the respective expert defect geometries is carried out according to the previously described calculation of the initial prediction data set. In particular, the calculations based on the respective expert defect geometries are performed in parallel. This can be accompanied by the creation of a database in which the leakage flux fields associated with each defect are stored, with the aim of saving computing time later and for other similar data.

[0029] Before generating the expert prediction dataset, it may be advantageous to adapt, in particular partially refine, the underlying mesh, in particular the defect mesh, and if necessary also the object mesh, for calculating the expert defect geometry. Mesh morphing techniques can be used for this purpose. In these techniques, the object or defect mesh is refined by shifting and / or splitting grid points, particularly in areas of strong gradients, in order to enable more precise determination of the geometry or, subsequently, more accurate simulation. In other areas with weaker gradients, the mesh can be coarsened to save computing time. In this way, the mesh used is automatically adapted for optimal evaluation of the defect geometry. At the same time, this significantly reduces the number of unknowns, which in turn saves computing time.

[0030] If the comparison between the reference data set and the expert prediction data set of an expert routine shows that it is closer to the reference data set than the previous initial prediction data set, the corresponding expert defect geometry is made available as the initial defect geometry for the subsequent iterations and for the corresponding expert routine. Based on this solution, the subsequent expert routines can then start from this geometry in a subsequent iteration step, unless, for example, they have found a better solution during their own ongoing defect geometry determination, which is then made available to further or all expert routines.

[0031] Among the competing expert routines, preference is given to those that, as described below, are more successful in approximating the real measurement data than other competing expert routines, based on the available resources of the computing unit. The computing unit's resources include, in particular, CPU or GPU time and / or memory allocation or prioritization.

[0032] Advantageously, the expert routines (on the IT unit) compete with each other in such a way that the distribution of the IT unit's resources, particularly in the form of computing time, to a respective expert routine depends on a success rate, which particularly takes into account the number of initial defect geometries calculated by the expert routine and made available to one or more other expert routines, and / or depends on a reduction of a fitness function, which particularly takes into account the number of expert prediction data sets generated for the reduction. The competition between the expert routines arises in particular from the fact that the program section designed as a monitoring routine allocates more resources, particularly in the form of computing time, preferably CPU or GPU time, to the respective expert routines if they are more successful than other expert routines.An expert routine is successful when it has found a defect geometry that is more suitable for the reference data set, for example a simulated stray field measurement, which is made available to the other expert routines.

[0033] This can result, for example, in individual, particularly successful expert routines receiving more than 50% of the total available computing time, which significantly reduces the overall duration of the method according to the invention. At the same time, the program can specify that none or individual expert routines do not fall below a certain percentage of computing time in order to avoid the problem of singular and exotic defect geometries or results from the individual routines. This way, in the event that a previously successful expert routine only finds a local and not a global solution, a way out of the deadlock situation otherwise encountered in the state of the art can be found.

[0034] The adjustment using the expert routines continues until a stop criterion is met. This could be, for example, a residual difference between the measured and simulated data. It can also be an external stop criterion, for example, based on the available computing time or a particular, predefined number of iterations or a particular, predefined or predetermined computing time or a computing time determined from the available computing time.

[0035] It has been found that the accuracy of flaw detection is qualitatively improved by the method according to the invention. A resulting calculation of the maximum load capacity shows that, for example, pipelines can be operated significantly longer. The accuracy of flaw detection is significantly increased. Maximum operating pressures resulting from the simulated flaw geometry using the method described above and below can be set up to 50% higher, which significantly reduces maintenance and upkeep costs for the pipeline operator and its operators. For the first time, an adequate determination of the ASME B31G-2012 level 2 approach ("river bottom profile") for the "remaining strength algorithm" can now also be realized for MFL data sets.

[0036] Preferably, the resources of the IT unit are distributed, particularly in the form of CPU time, to a respective expert routine depending on the number of initial defect geometries provided by this expert routine for all expert routines. This can, for example, be a number of slots for calculating the expert prediction data sets in the form of simulated MFL data sets, the number of processor cores processing the computing task in parallel, or the like. Furthermore, within the framework of the computer program product implementing the method according to the invention, it can be provided that this adapts to the resources available in the IT units in the form of processor cores, memory space, memory architecture, graphics cards, etc. By prioritizing particularly preferred expert routines and their algorithms, the detection of the actual defect geometry becomes significantly faster.

[0037] In order to further minimize the problem of singular, local solutions, it is particularly intended to use an additional reference data set that is linearly independent of the first reference data set to determine the geometry of the defect(s). The data sets are linearly independent if they were generated by MFL measurements with magnetizations of the object that are angled to one another. The magnetizations are angled to one another if the respective mean induced magnetic field strengths in the examined area are neither parallel nor congruent. In particular, the angle is between 40° and 140°, preferably between 80° and 100°, and particularly preferably 90°. Based on the initial defect geometry, a further initial prediction data set is determined, in particular by means of a linear independence, i.e.In particular, a further MFL simulation is generated that takes the different magnetizations into account, and an expert defect geometry is only used as the initial defect geometry if the associated expert prediction data sets determined for both independent magnetizations are more similar to the respective reference data sets than the initial prediction data sets determined for the two magnetizations and / or a fitness function that takes both expert prediction data sets into account is improved. By processing the two linearly independent data sets in parallel or concurrently and using an initial defect geometry whose simulated measurement data must be better overall in terms of similarity or a fitness function, the risk of singularities is further reduced. At the same time, the quality of the initial prediction data sets available for all expert routines is improved.The number of iterations can thus be further reduced.

[0038] In particular, the first reference data set is generated using an MFL measurement with axial magnetization, and the second reference data set is generated using an MFL measurement with magnetizations in the circumferential direction of the pipe. The magnetizations of the pipe or an object are perpendicular to each other, so that maximum information content can be obtained from the magnetic flux leakage measurements, which is fully available through the simultaneous consideration of the corresponding reference data sets and the simulated expert prediction data sets during the calculation. Taking the above into account, the process steps described below are analogous when using two reference data sets generated from linearly independent magnetizations.

[0039] By using initial and / or expert prediction data sets based on a forward model to simulate the MFL measurements, particularly using a finite element model, the MFL measurement simulations are performed quickly. The simulation of the leakage flux measurements based on the expert defect geometries can be implemented using a dedicated program module, which is controlled and / or monitored by a monitoring routine and called separately by the individual expert routines. It can also involve multiple modules distributed across individual computer units and made available to a respective expert routine.

[0040] Advantageously, the initial defect geometry is generated using a look-up table, one of the expert routines and / or a machine learning algorithm, which, as described above, improves the overall computing time, especially if a grid adjustment is already performed.

[0041] In particular, the refinement of the object and / or defect grid can take place in the areas where the depth of the simulated defect(s) exceeds a threshold. This threshold can be specified so that only gradients above a certain value lead to a change in the grid. During such a refinement, the total number of gradients of a new expert defect geometry can be taken into account in order to achieve a balance between the adaptation of the respective grid, in particular the object grid, and the subsequent computational operations.

[0042] Refining the grid with the aim of reducing computation time can be performed either on the basis of an initial reference dataset or before calculating the expert prediction dataset. This can also be done using a separate program module or individual submodules of the respective expert routines.

[0043] In particular, the refinement of the object and / or defect grid by grid point shifting and / or splitting particularly advantageously reduces the required CPU time by significantly reducing the number of independent variables that must be used in the forward algorithm to simulate the MFL measurement.

[0044] A grid point shift can also be used to adjust object or defect grids.

[0045] Preferably, a fitness function is used as a measure of the similarity of the expert prediction and reference data sets in order to compare the simulated and measured data sets based on standard routines and accordingly quickly, ie while saving computing time.

[0046] In particular, the initial defect geometry is stored in the form of a two-dimensional data set or a pointer pointing to it in a memory area of ​​the computer unit accessible to all expert routines. This memory area is, in turn, controlled by a monitoring routine, so that individual expert routines can also be prioritized in this regard.

[0047] Instead of using the initial defect geometry, which is stored, for example, in a central memory area accessible to all expert routines, at the beginning of each new iteration, at least one expert routine can adapt its own expert defect geometry at the beginning of a new iteration without adopting the initial defect geometry. For this purpose, an expert routine can have a functional specification in which, for example, an opposing strategy is specifically selected depending on the search strategies used in other expert routines. In such a case, the expert routines can indirectly influence each other. Such an approach can be particularly advantageous if it is determined that a previously always successful routine prefers an unrealistic solution. This can be detected, for example, based on inadmissible values ​​regarding the depth of a defect.If an expert routine that does not adopt the initial defect geometry does not provide improved solutions, it is automatically down-prioritized so that it is given increasingly less computing time.

[0048] The stopping or convergence criterion can preferably be assumed to be a substantial change in the initial defect geometry, or more generally, in the geometry of the defect and / or object grid, that fails to occur after several iterations. The solution found up to that point is then the best. The stopping criterion is preferably chosen such that the observed variations in the flux leakage simulation, which lead to the refinement of the object or defect grid, are substantially lower, e.g., by a factor of 2, than the variations resulting from the individual measurement scatter, which are individually specified, for example, based on so-called "Essential Variables" of the API 1163 standard. This ensures that the accuracy of the final model lies within the range of the accuracy specified by the measurement itself.Accordingly, a comparison of the variation of the expert prediction data set with the measurement variation of the actual data set is preferably used as the stop criterion. The stop criterion that causes a program stop and, in particular, an output or storage of the initial defect geometry calculated up to that point can preferably be specified by pre-configurable program parameters.

[0049] In particular, an expert routine has several algorithms available for adapting the expert defect geometry. These can be approaches from the field of machine learning, stochastic optimization, empirical and / or numerical model functions. In particular, the expert routines can also utilize empirical values ​​from evaluators. Preferably, the different algorithms in an expert routine are either randomly generated or selected using a selection function. This creates a sufficiently diverse approach with which all solutions can be considered in a targeted manner and under competitive conditions.

[0050] The object posed at the outset is also achieved by a method for determining a load-bearing capacity of an object which is subjected to pressure at least during operation and is designed in particular as an oil, gas or water pipeline, wherein in the method a data set describing one or more defects is used as an input data set in a calculation of the load-bearing capacity limit, designed in particular as forward modeling, wherein the input data set is initially generated according to a method described above or below for determining the geometry of a defect. The advantageous representation of the defect geometry, in particular as a non-parameterized true three-dimensional geometry oras a two-dimensional surface with respective depth values, makes simplifications previously considered necessary in the industry superfluous, so that for this reason too an increase in the accuracy of the defect determination as a whole is guaranteed in a way that was previously unattainable.

[0051] Whereas previously accuracy was limited to specifying the point of maximum depth of the defect, the entire profile is now determined with high accuracy. Typically, the accuracy of the maximum depth is reduced to the achievable level depending on the measurement accuracy, i.e. approximately ± 5% of the wall thickness compared to approximately ± 10% of the wall thickness for sizing according to the state of the art described above. However, the prediction of the load limit depending on the geometry of the defect achieves increases in accuracy, for example from previously ± 50% to now ± 5%, especially for critical cases. The advantage of the invention thus lies in particular in an adequate representation of the defect geometry achieved for the first time, which precisely makes this increase possible.

[0052] Further advantages and details of the invention can be found in the following description of the figures. Schematically shown: Fig. 1 a defect determination according to the state of the art, Fig. 2 a schematic representation of a method according to the invention, Fig. 3 a more detailed explanation of part of the Fig. 2 , Fig. 4 Comparison of a result of a method according to the invention with measurement data, Fig. 5 schematic representation of a grid refinement as part of a method according to the invention, Fig. 6 Result of a method according to the invention.

[0053] Individual features of the exemplary embodiments described below can, in combination with the features of the independent claims, also lead to further developments according to the invention.

[0054] In the state of the art, the evaluation of MFL data of a pipe is carried out according to Fig. 1 by means of the definition of boxes, which is particularly based on experience. The boxes shown in the figure have respective dimensions in length, width and depth. The x and y axes are shown in meter units ([m]). A review of the actual defect geometry on which this evaluation is based using a laser scan, i.e. using a direct measurement, has shown that the maximum burst pressure of 4744.69 kPa, which can be determined based on the defect geometry assumed by the MFL data evaluation, is only 55.2% of the maximum burst pressure calculated based on the actual geometry. Due to the state of the art, the operating pressure for safe operation of the pipeline, which is 3621.29 kPa based on the experience-based evaluation, is significantly below a possible safe operating pressure.

[0055] In the method according to the invention, according to one embodiment, the surface of a pipe is represented by a 2D mesh surface. The defect geometry can be described as a vector of depth values ​​D lying on a defect grid 5 ( Fig. 5 ). This defect geometry is compared with the initial defect geometry based on the result of a fitness function F(D) that takes into account the MFL fields associated with the respective geometries. It is assumed that the lower the value of a fitness function, the closer the assumed expert defect geometry is to the real geometry: F D = ∑ M H cal D − H m + R D

[0056] Here, M is the number of data records to be processed simultaneously (real MFL data records), H cal the result of a simulation of the MFL measurement, H m are the measured data from the MFL measurement (reference dataset), and R (D) is a regularization term that can be set as follows: R D = α ∇ D , where α is a scaling term.

[0057] The process sequence according to the invention is described at least in sections below according to Fig. 2 described, whereby a majority of the parallel and competing expert routines 11 are described only with one block 14.

[0058] For example, multiple runs of the same MFL pipeline pig can be combined as input data sets according to Box 2. Both data sets 1 can be filtered beforehand for better merging and aligned with each other (process step 3), for example, to reduce any artifacts or background noise. In addition, a further data set 4, generated based on a linearly independent, additional magnetization and also filtered for alignment to identical grid structures, can be used, so that, according to process step 6, two aligned reference data sets are available, created based on measurement runs with linearly independent magnetizations.

[0059] Data sets that are precisely matched to one another can be treated jointly, whereby the method according to the invention realizes the simultaneous treatment of the data sets by using a fitness function that takes the merged data sets into account.

[0060] In step 7, a first of the reference data sets available in step 6 is selected for further processing. For this purpose, in step 8, an initial defect geometry is assumed as the starting defect geometry, in particular, generated in this case, which is based, for example, on a normalized measurement signal S(x,y) / (max S). For example, the defect geometry can be derived from a threshold function that takes into account the amplitude at grid points where the signal is greater than a certain threshold value / (e.g., 0.2): G x y = 0 , if S x y max S < l S x y max S , sonst .

[0061] The above approximation leads to a number of N-defect depth values ​​at the respective grid points: D i = i wt / N * G , with wt as the thickness of the pipe wall. For such a defect geometry, the fitness function is calculated, and the profile with the lowest function value is used as the input solution: D init = arg min F D i

[0062] This input solution is then provided as the initial defect geometry for the individual expert modules. Initially, the number of parameter values ​​(elements of the vector D) describing the defect geometry can be kept as small as possible to reduce computing time. This is achieved primarily through dynamic grid adaptation. Since the number of depth values ​​corresponds to the number of nodes in the defect grid (5), the number of nodes is also the number of defect parameters. Starting with a comparatively coarse grid, this is successively refined in relevant areas.

[0063] For example, for a given node spacing of 14 mm, a corresponding grid cell size of 14 mm x 14 mm and defect limit values ​​of 30%, 50% and 80% of the wall thickness, the Fig. 5 The refinement shown can be achieved in the relevant grid region, with those cells that exceed the above depth values ​​being successively subdivided. The grid deformation then correlates with the assumed defect geometry, i.e., in regions of large gradients, there is a larger number of grid points.

[0064] According to the method according to the invention, the workflow of a group of expert routines 11 that compete with each other is simulated on the EDP unit. For this purpose, the program can have various modules that can independently and, in particular, not synchronized with each other, input data into specific areas of the EDP unit for further processing there. This is done, in particular, under the supervision of a monitoring routine 9 (Fig. 3a). A plurality of expert routines 11 thus maintains a number of computing slots 13 depending on the above-defined success, e.g., the number of initial defect geometries written into a common memory area 12, in order to generate expert defect geometries and / or to be able to perform associated MFL simulations or, in the case of an independent MFL simulation module, to have them performed. This corresponds to block 14 according to Fig. 2 , which exemplifies several expert routines 11 (Fig. 3a). Starting from the individual calculation slots 13, according to the present exemplary embodiment, the MFL simulations of the individual expert defect geometries are also carried out in the simulation modules 16 for the purpose of creating the expert prediction data sets under the supervision of the monitoring routine 9. The more slots 13 are available for an expert routine, the greater the proportion of IT resources for this expert routine. Preferably, the number of program modules for carrying out MFL simulations is equal to the number of slots. The monitoring routine 9 monitors the number of iterations and the resulting changes in the initial defect geometry and further monitors whether an associated stop criterion has been met. The result is then output according to block 17.

[0065] The number of computing slots 13 available for an expert routine 11 and the simulation routines subsequently made available can vary such that a first expert routine can, for example, utilize up to 50% of the total computing time available for the computing slots and simulation routines.

[0066] As shown, the initial defect geometries are stored in memory area 12. This can be a memory area accessible to the expert routines 11. Log files from the expert routines 11 and monitoring routine 9, as well as instructions for the expert routines 11, which are then implemented independently, can also be stored there. For example, this could be an interrupt command that is issued when the stop criterion is met.

[0067] Preferably, the expert routines 11 are independent program modules that generate new expert defect geometries and input them into the simulation routines 16. Furthermore, the fitness function presented at the beginning can be generated in the expert routines 11 based on the expert prediction data set and compared with the initial prediction data set stored in area 12. If the expert prediction data set is more similar overall to the reference data set or, in the case of linearly independent measurement data sets, to the two reference data sets than the data set stored in area 12, this expert prediction data set is then used as the new initial prediction data set.

[0068] For example, in expert routines 11, a new defect geometry is randomly generated. Machine learning algorithms or empirical rules can be used for this purpose. However, to further improve the convergence of the solutions, the implementation of at least two basic expert routines is advantageously provided, as described below.

[0069] These search strategies, which are preferably always implemented in a method according to the invention, are based on an assumed probability distribution p(x,y) of grid points whose depth value results in a maximum reduction of the fitness function. The probability function is used to identify N grid points (xn,yn). At each of the considered points, the depth function, which describes, for example, the depth of corrosion at the grid location, is changed by ΔD, with the sign of the change being randomly generated. The number of selected points N can also be chosen randomly: D new x y = D x n y n ∓ Δ D , für ausgewählte Punkte D x y , sonst

[0070] By selecting the probability function p (x,y), different expert strategies can be realized, for example: p x y = D x y D x y

[0071] This algorithm implements a variation of the defect depths, favoring the grid points with the greatest depth. Another strategy could be as follows: p x y = H the best x y − H m x y H the best x y − H m x y

[0072] Such an algorithm varies the defect geometry at positions where the simulated measurement signal for the best known solution has the greatest difference to the measured signal.

[0073] Based on this, different expert routines or their algorithms can be constructed by varying the number of grid points to be considered and the ΔD. The following six expert routines can be used as examples: 1. p x y = D x y D x y , N = 1 and Δ D = 1 % Wanddicke 2. p x y = D x y D x y , N = 2 and Δ D = 5 % Wanddicke 3. p x y = D x y D x y , N = 3 and Δ D = 5 % Wanddicke 4. p x y = H the best x y − H m x y H the best x y − H m x y , N = 1 and Δ D = 1 % Wanddicke 5. p x y = H the best x y − H m x y H the best x y − H m x y , N = 2 and Δ D = 5 % Wanddicke 6. p x y = H the best x y − H m x y H the best x y − H m x y , N = 3 and Δ D = 5 % Wanddicke

[0074] The monitoring routine 9 shown in Fig. 3a has two functions in particular: Firstly, it checks whether the stop criterion has been met, and secondly, it allocates the resources of the IT unit between the individual experts based on their success. One measure of success is P = Δ F N , where ΔF is the reduction of the fitness function F by the result of the respective expert routine and N is the number of simulations required for this. An evaluation of the n expert routines can be described as R n = P n ∑ P i . The number of computation slots Ns for an expert routine in one iteration is then N S = int R n N all , where N all is the number of all available slots.

[0075] In simulation routines 16, an MFL measurement for an expert defect geometry is simulated. An expert routine can iterate until it finds a solution whose expert prediction data set is better than the initial prediction data set stored in area 12. If this is the case, expert routine 11 can process another linear independent data set or attempt to achieve further better solutions based on the already improved solution.

[0076] If several data sets from different iterations, which cannot be aligned consistently, have been calculated by the expert routines, the resulting geometries can also be automatically superimposed within the scope of the implementation of the method according to the invention, whereby the maximum depth is taken as a conservative estimate at the individual grid points: D x y = max N D n x y for n = 1... N, where N is the number of data sets that must be processed consecutively. Based on a depth profile resulting from such a superposition of defect geometries, an MFL signal can be simulated. The resulting error can result from the errors of the respective data sets in the individual calculations: E = H cal D − H m

[0077] To demonstrate the efficiency of the proposed method, a variety of test scenarios were conducted, with the following Fig. 4 the data of two MFL inspection runs, which were carried out with linearly independent magnetization, are used.

[0078] The Fig. 4 No. 21 shows an image of a real MFL measurement with magnetization running in the axial direction, while Figure 22 results from a measurement taken in the circumferential direction. The calculated result of the defect geometry, which was achieved using the previously described method according to the invention, is indexed with 23. As in Figure 24, the contour lines evenly divide the area between 0 and 60% metal loss depth. Figure 24 shows the actually scanned and thus directly measured outer surface of the pipe section corresponding to Figures 21 and 22. The result is a very high degree of agreement between the laser scan measurement and the solution achieved using the method according to the invention. This is significantly better than the solution based on the evaluation known from the prior art.It can be assumed that the deviations in the result according to the method according to the invention and that of the laser scan measurement are predominantly due to technical tolerances.

[0079] Based on the conventional approach with established state-of-the-art and Fig. 1 The result of the determination of the defect geometry shown above results in the mentioned maximum burst pressure of 4744.69 kPa. Based on the method according to the invention, the Fig. 1 underlying MFL dataset which is contained in the Fig. 6 shown defect geometry (contour lines at 2 mm depth) and based on this, a maximum burst pressure of 8543.46 kPa. In this case, this comes within 99.4% of the maximum burst pressure determined based on the actual defect geometry determined by laser scanning. Accordingly, a pipeline examined using the method according to the invention can be operated with a safe operating pressure of 6520.53 kPa. This results in a safe operating pressure of 3621.29 kPa compared to the safe operating pressure of 3621.29 kPa based on the evaluation according to the state of the art ( Fig. 1) offers significant advantages for pipeline operators. The method according to the invention allows the condition of a pipe and thus the pressure required for safe operation of the pipeline to be determined much more realistically, while operational safety is still ensured. Using the method according to the invention, with expert routines competing for IT resources, such a result can be made available to pipeline operators more quickly, or at least within the same evaluation time as with the prior art.

Claims

1. Method for determining the geometry of one or more real, examined defects of a magnetizable object, in particular a pipe or a tank, with a reference data set of the object generated on the basis of one or more MFL measurements, comprising an at least partial representation of the object via an EDP unit in particular on or via an at least three-dimensional object grid, and comprising a determination of an initial defect geometry as the initial defect geometry in particular on the object grid or an at least two-dimensional defect grid (15), a determination of a first MFL prediction data set as an initial prediction data set based on the initial defect geometry by simulating an MFL measurement or assigning an MFL data set, and iterative adaptation of the initial defect geometry to the geometry of the real defect(s) via the EDP unit and via several expert routines (11) running in competition and preferably in parallel with each other in such a way that the expert routines compete with each other for resources of the EDP unit, whereby a respective expert defect geometry is generated in respective expert routines (11) via at least one own algorithm and on the basis of the initial defect geometry, a respective expert prediction data set is determined on the basis of the respective expert defect geometry in particular by simulating an MFL measurement or assigning an MFL data set, and the expert defect geometry on which the respective expert prediction data set is based is then made available to at least several in particular all of the expert routines (11) as a new initial defect geometry for further adaptation to the geometry of the real defect(s), if the expert prediction data set is more similar to the reference data set than the initial prediction data set, and then the expert prediction data set belonging to the new initial defect geometry is used as the new initial prediction data set, whereby the iterative adjustment is carried out using the expert routines until a stop criterion is met.

2. Method according to claim 1, characterized in that the expert routines (11) run in competition with one another in such a way that the resources of the EDP unit are distributed to a respective expert routine in particular in the form of computing time, preferably CPU and / or GPU time, as a function of a success rate, in which in particular the number of initial defect geometries calculated by this expert routine and made available for one or more other expert routines (11) is taken into account, and / or as a function of a reduction of a fitness function, in which in particular the number of expert prediction data sets generated for the reduction is taken into account.

3. The method according to one of the preceding claims, characterized in that a fitness function is used as a measure of the similarity of the expert prediction and reference data sets.

4. Method according to one of the preceding claims, characterized in that, in order to determine the geometry of the defect(s), a further reference data set which is linearly independent of the first reference data set with respect to the magnetization is additionally used, and a further initial defect geometry is determined on the basis of the initial defect geometry in particular via a further MFL simulation, which takes into account the linear independence, and an expert defect geometry is only then used as the initial defect geometry, if the associated expert defect geometries determined for both independent magnetizations are more similar to the respective reference geometries than the initial defect geometries determined for the two magnetizations and / or a fitness function taking into account both expert defect geometries is improved.

5. Method according to claim 4, characterized in that the first reference data set was generated via an MFL measurement with axial magnetization and the second reference data set was generated via an MFL measurement with magnetization in the circumferential direction of the tube.

6. Method according to one of the preceding claims, characterized in that initial and / or expert prediction data sets are generated on the basis of a forward model for simulating the MFL measurement and in particular via a finite element model.

7. Method according to one of the preceding claims, characterized in that the initial defect geometry is generated via a look-up table, by one of the expert routines (11) and / or by a machine learning algorithm.

8. Method according to one of the preceding claims, characterized in that the defect grid (5) is refined in regions in which the depth of the simulated defect (s) exceeds a threshold value.

9. Method according to one of the preceding claims, characterized in that the object and / or defect grid (15) is refined before the calculation of a respective expert prediction data set.

10. Method according to one of the preceding differences, characterized in that the initial defect geometry or a pointer referring thereto is stored in a memory area (12) of the EDP unit accessible to all expert routines (11).

11. Method according to one of the preceding claims, characterized in that a change in the initial defect geometry and / or the geometry of the object and / or defect grid (15) and / or the initial prediction data set and / or at least one expert prediction data set which is absent after a plurality of iterations is assumed as a stop criterion.

12. Method according to claim 11, characterized in that a comparison of the variation of the expert prediction data set with the measurement scatter of the real data set is used as a stop criterion.

13. Method according to one of the preceding claims, characterized in that an expert routine (11) is assigned a plurality of algorithms for adapting the expert defect geometry comprising machine learning, stochastic optimization, empirical and / or numerical model functions.

14. Method according to claim 13, characterized in that an algorithm is randomly generated or selected and / or changed by a selection function in an expert routine (11).

15. Method for determining a load-bearing capacity limit of an object which is subjected to pressure at least during operation and is designed in particular as an oil, gas or water pipeline, in which a data set describing one or more defects is used as an input data set in a calculation of the load-bearing capacity limit, characterized in that the input data set is first determined in accordance with a method according to one of the preceding claims.