Method for training a neural network for determining the geometry of defects on the basis of non-destructive measurement methods

Abstract

The present invention relates to a method for training a neural network for determining a geometry of one or more real defects of a metallic, in particular magnetisable, object, in particular a pipe or a tank, examined by means of at least one first non-destructive measurement method, wherein the neural network is trained in a training EDP unit with at least one training data set which comprises associated input and output data sets, wherein an output data set comprises geometry data representing at least one section of the object having at least one defect, and an input data set comprises at least one measurement result assigned to the geometry data, wherein at least for a part of the training data set, a simulation measurement result, obtained by way of a simulation of the at least one first non-destructive measurement method, is used as a measurement result, said simulation measurement result being enriched with real measurement data obtained during performance of the non-destructive measurement method at a defect-free region of the object or of a reference object.

Description

[0001] Method for training a neural network to determine the geometry of defects based on non-destructive measurement techniques

[0002] The present invention relates to a method for training a neural network to determine the geometry of defects based on at least one non-destructive measurement method. The invention further relates to a method for determining the geometry of a defect using the neural network trained in this way, and to a computer program product comprising instructions for carrying out this method.

[0003] Non-destructive testing methods can be used to examine objects for defects. The results of these tests are evaluated to determine the type and / or size of the defects. Based on this, it can then be determined how much stress the object can withstand, or whether and to what extent repairs or replacement are necessary. To prevent unnecessary measures, it is crucial to describe the defects and their geometry as precisely as possible.

[0004] Defects within the meaning of the present invention are, in particular, areas with metal loss due to corrosion, cracks, or other weakening of the object's wall. A pipeline is an example of an object to be examined in this way. Pipelines are regularly inspected using non-destructive testing methods. Typically, so-called "smart pigs" are used as devices for carrying out such non-destructive testing. The aim of these investigations is to determine safe operating conditions for the pipeline. This includes, in particular, a safe operating pressure. Safe operating conditions depend on the pipeline's load-bearing capacity. This, in turn, depends on the condition of any welds and on the number and size of defects.

[0005] The load-bearing capacity of a pipeline can be determined based on the maximum pressure at which the pipeline is expected to rupture, the so-called "burst pressure." The operating pressures within the pipeline must maintain a sufficient safety margin below this maximum pressure. Accurate prediction of the maximum pressure makes it possible to operate the pipeline at the highest possible, yet still safe, operating pressure. Especially when transporting gaseous media, the transport capacity of a pipeline depends heavily on the operating pressure within the pipeline.

[0006] Accurate determination of the maximum pressure requires precise measurement of the dimensions of defects. It is known to perform magnetic flux leakage (MFL) measurements, where the data is evaluated by specially trained personnel to determine the size of (corrosion) defects in the object. The defects identified in the data are parameterized as rectangular boxes and evaluated. The assumptions necessary for this evaluation of the measurement results, also known as "sizing," are proprietary. Furthermore, the interpretation of the measurement results is strongly influenced by the experience of the evaluators. The most widely used industry standard, AP 1163, describes the disadvantages of this simplified approach. It is well documented that the quality of this approach depends heavily on the expertise of the person performing the evaluation.The evaluations are always interpretations of the data obtained through a measurement run using the respective non-destructive measurement method, influenced by subjective factors.

[0007] From EP 3 722 800 A1, it is known that the geometries of defects can be determined by inverting at least parts of reference data sets obtained by non-destructive testing methods using a neural network trained for this task. The defect geometries obtained in this way are not necessarily sufficiently accurate, so that after inversion, a verification of an accuracy measure and, depending on the accuracy measure, a further iterative adjustment of the defect geometry may be necessary until a satisfactory accuracy measure is achieved. The object of the present invention is to provide an improved method for training such a neural network. Furthermore, it is an object of the invention to provide a method for determining the geometry of a defect using such a trained neural network.Furthermore, the object of the invention is to provide a computer program product containing instructions for carrying out the method for determining the geometry of a defect using a neural network trained in this way.

[0008] The problem is solved by a method according to claim 1. The problem is further solved by a method according to claim 17 and a computer program product according to claim 22. Advantageous embodiments of the invention can be found in the dependent claims relating to these claims and in the following description.

[0009] The inventive method for training a neural network suitable for determining the geometry of one or more real defects of a metallic object examined by means of at least one non-destructive measurement method is carried out using a training computer unit and at least one training data set. The metallic object examined is, in particular, a magnetizable object, especially a pipe or a tank. The training computer unit can be a commercially available computer unit comprising at least one processing unit such as a CPU and at least one memory connected to the processing unit. The computer unit can further comprise one or more communication interfaces for communication with other electronic devices. Furthermore, the computer unit can include or be connected to input and output devices.Ideally, the training computer unit should have multiple graphics processing units (GPUs). Such a training computer unit makes training the neural network particularly easy.

[0010] The training dataset comprises related input and output datasets. An output dataset includes geometric data representing at least one section of the object containing at least one defect. An input dataset includes at least one measurement result associated with the geometric data. The method according to the invention is characterized in that, for at least a part of the training dataset, a simulation measurement result obtained by simulating at least one non-destructive measurement method is used as the measurement result. This simulation measurement result is enriched with real measurement data obtained by performing the non-destructive measurement method on a defect-free area of ​​the object or a reference object.

[0011] Measurement data assigned to a defect-free area of ​​the object can be identified much more easily in the actual measurement data compared to determining the size of defects. The effort required to train people to mark defect-free areas in the actual measurement data is significantly less than the effort required to train people to determine the size of defects. The actual measurement data obtained in a defect-free area of ​​the object can therefore be made available relatively easily.

[0012] The use of such a training dataset allows for a specification and thus an exact knowledge of the geometry data representing a section of the object and, in particular, of the geometry of one or more defects represented by the geometry data.

[0013] By simulating at least one initial non-destructive testing (NDT) method, a simulation measurement result is calculated that would be expected when performing at least one initial NDT method on an object with precisely the geometry represented by the specified geometric data, including defects. Using such simulation measurement results simplifies the creation of the training dataset, as it eliminates the need to locate real objects exhibiting various defects or defect geometries to represent the entire range of defects and defect geometries of interest. No geometric measurements need to be performed on these defects to accurately measure them. In particular, no measurements using at least one NDT method need to be carried out on the corresponding objects with the corresponding defects.If geometric data of a section of an object, including a defect, obtained through geometric measurements are available, these can of course be used to obtain the corresponding simulation measurement result by simulating at least one first non-destructive measurement method.

[0014] By using appropriate simulation measurement results, the method according to the invention makes it possible to easily provide training data for a wide range of possible geometries of defects.

[0015] Simulation measurement results can be provided with consistent quality. In particular, the quality of the

[0016] Simulation measurement results are not dependent on the quality of a real measurement of a real defect or a section of an object containing the defect.

[0017] Many different defect geometries can be used to calculate the simulation measurement results. These geometries can represent typical, real-world defects. The geometry data representing the defects reproduces the geometry of the respective defect as accurately as possible within the spatial resolution used to generate the simulation measurement results. A neural network trained with simulation measurement results based on such defect geometries can output a defect geometry when evaluating real measurement data that closely approximates the actual defect geometry. The defect geometry then does not need to be approximated by "boxing," the process of determining the minimum dimensions of a predefined shape, typically a cuboid, that still completely contains the defect.

[0018] In particular, the geometry data has a spatial resolution that is at least one order of magnitude, preferably at least two orders of magnitude, above the spatial resolution of geometry data determined by means of boxing.

[0019] A more precise knowledge of the geometry of defects identified using at least one non-destructive measurement method makes it possible to determine the load limit of an examined object, in particular a pipeline or a section of pipeline, more accurately.

[0020] The load limits determined on the basis of such more precise data on the geometry of defects are regularly higher than the load limits determined by "boxing" defects.

[0021] The simulation results obtained by simulating at least one non-destructive measurement method do not include effects such as noise, which inevitably occur in a real measurement and lead to deviations between real measurement data and the simulation results. Likewise, the simulation does not account for deviations between a device used to perform at least one non-destructive measurement method and its representation in a simulation environment. Noise and deviations of a real device used to perform at least one non-destructive measurement method compared to the device modeled for a simulation are typically not considered in simulations.

[0022] In the method according to the invention, the simulation measurement results are enriched with real measurement data obtained during the execution of the at least one first non-destructive measurement method on a defect-free area of ​​the object or a reference object. The real measurement data measured in a defect-free area of ​​the object or a reference object using the at least one first non-destructive measurement method contain effects such as noise and systematic distortions that result from the deviations of the real device used to perform the at least one first non-destructive measurement method from the ideal device modeled for the simulation.

[0023] Because the actual measurements are performed in a defect-free area of ​​the object or a reference object, the real measurement data contains no effects attributable to defects. The object or the reference object does not need to be completely defect-free; the presence of defect-free areas is sufficient. After performing at least one initial non-destructive measurement procedure on the object or the reference object, defect-free areas can be identified in the obtained real measurement data and used to enrich the simulation measurement results.

[0024] Using a training dataset obtained in this way, the neural network is trained to "see through" both noise and systematic distortions caused by deviations of the devices used to perform at least one initial non-destructive measurement procedure from the model of the devices used in the simulation. A neural network trained in this manner is particularly well-suited for determining the geometry of one or more real defects based on measurement data obtained during at least one initial non-destructive measurement procedure on an object.

[0025] The actual measurement data can be obtained using different devices and / or on different reference objects.

[0026] Enriching the simulation results with real-world measurement data also makes it easy to provide a particularly large training dataset. For this purpose, the same simulation result can be enriched with different real-world measurement data, which, for example, relate to different defect areas of the object or the reference object.By enriching the simulation measurement results with aspects of a real measurement, a large training dataset can be easily provided, which on the one hand allows a large degree of control over the actual geometries of the defects within the training dataset as well as the associated expected response of these geometries when performing at least one first non-destructive measurement procedure on the corresponding geometry, the simulation measurement result, and on the other hand makes at least some aspects of real measurements available for training the neural network.

[0027] The inventive method for training a neural network thus makes it possible in a simple way to train a neural network which, when evaluating measurement data obtained from carrying out at least one first non-destructive measurement method on an object to be examined, can reproduce the geometry of one or more real defects with particularly small deviations.

[0028] Ideally, the deviations are so small that subsequent iterative adjustment of the defect geometry or the calculation of a prediction dataset by simulating the first non-destructive measurement procedure based on the defect geometries determined by the neural network, as well as a comparison of the prediction dataset with the measurement result obtained during the first non-destructive measurement procedure to determine an accuracy measure, can be omitted. The determination of one or more real defects using a neural network trained with the method according to the invention can then be carried out particularly quickly and / or particularly easily and cost-effectively.

[0029] Preferably, the first non-destructive measurement method is an MFL measurement method, in particular an axially oriented MFL measurement method or a circumferentially oriented MFL measurement method, or an eddy current measurement method. Using one of these measurement methods as the first non-destructive measurement method makes the procedure particularly easy to perform. It is then easy to determine, after performing a measurement on the object or a reference object, whether this raw data originates from a measurement on a defect-free area of ​​the object and can be used to enrich the simulation measurement results with real measurement data. Prior evaluation of the raw measurement data is unnecessary.

[0030] The raw measurement data selected in this way can be used for enrichment without further analysis. In particular, prior analysis of the raw data is unnecessary in order to obtain a comprehensive noise-free database or a database from which noise can be extracted.

[0031] Preferably, the raw measurement data from the execution of the first non-destructive measurement procedure on a defect-free area of ​​the object or a reference object, optionally after normalization and / or other mathematical operations in which the relative sizes of the raw measurement data to each other are preserved, are used as real measurement data to enrich simulation measurement results.

[0032] This allows systematic distortions to be modeled realistically in a simple way. A neural network trained in this way can determine the geometry of a defect particularly well from real measurement data.

[0033] Preferably, the respective simulation measurement result is enriched with real measurement data by superimposing it. This represents a particularly simple way to enrich a simulation measurement result with real measurement data. Specifically, the superimposition of the simulation measurement result with real measurement data is achieved by adding the simulation measurement result and the real measurement data.

[0034] Preferably, those parts of the simulation measurement result that correspond to a region of the geometry data containing or representing a defect are superimposed with real measurement data. By using a training dataset in which precisely these parts of the simulation measurement result are superimposed with real measurement data, the neural network is trained, during training with such a dataset, to ignore noise and systematic deviations between the simulation of the at least one first non-destructive measurement method and / or the model of the device for carrying out the at least one first non-destructive measurement method in the simulation when evaluating and determining the geometry of a defect.

[0035] Furthermore, and particularly preferably, the entire simulation measurement result is superimposed with real measurement data. This makes superimposing the simulation measurement result with real measurement data, and thus creating the training dataset, particularly easy.

[0036] Preferably, the simulation result and the real measurement data are normalized before being superimposed. This normalization ensures that the simulation result does not dominate the real measurement data in the training dataset, or vice versa. This way, the simulation result, superimposed with real measurement data, can be achieved to represent the measurement data expected when performing at least one initial real measurement procedure on the object with a real device particularly well.

[0037] If the non-destructive testing (NDT) method is an MFL (magnetic field-free) method, the simulation measurement result and the actual measurement data can be normalized, particularly with respect to the respective background magnetization. The background magnetization is easy to determine. Normalizing the simulation measurement result and the actual measurement data to the background magnetization is relatively straightforward. Preferably, for determining the defect-free areas in which the NDT is performed to obtain actual measurement data for enriching the simulation results, areas around welds are excluded as inherently containing defects. Therefore, no measurement data from these areas is used to enrich the simulation results, regardless of whether defects are actually present in these areas.

[0038] Similarly, areas around fixtures and other known deviations from the predominant shape of the object's surface, especially a pipeline, can be excluded as being defective by definition.

[0039] In particular, the areas that are defined as having defects can extend 20 cm in front of and behind a weld seam or 20 cm around internal components, as well as other known deviations from the predominant shape of the object's surface, especially a pipeline.

[0040] In the area surrounding welds and / or internal components, changes can occur in the distance between a device used to perform at least one initial non-destructive measurement procedure and the object, particularly a pipeline. These changes occur regardless of whether the weld or the area surrounding it has defects. The change in the distance between the device used to perform at least one initial non-destructive measurement procedure and the object regularly influences the actual measurement data in a systematic manner.

[0041] Treating such areas as inherently defective for the purpose of determining defect-free regions in which the non-destructive measurement procedure is carried out, and excluding them from the actual measurement data in a defect-free region, makes it easy to obtain real measurement data that includes statistical effects such as noise as well as design-related effects of the real measuring device, but is largely free from effects attributable to larger and locally limited deviations in the geometry of the object and their influence on the execution of a real measurement procedure.

[0042] The procedure is then particularly easy to perform. A neural network trained in this way enables a particularly reliable determination of the geometry of defects.

[0043] In particular, the actual measurement data are obtained by performing at least one first non-destructive measurement procedure on a smooth, and especially weld-seam-free, wall section of the object or a reference object. A smooth wall section of the object is one that has no geometric defects. Furthermore, the wall section has no elements that influence the positioning of a device for performing at least one first non-destructive measurement procedure on the wall section of the object. On a smooth wall section, the device can be positioned relative to the wall section as intended.Especially with a device that ideally rests against a section of the object's wall for measuring purposes, welds or internal components of the object can otherwise cause the device to not rest against the wall section, but rather the wall section and the device to form an angle to each other. This would introduce a further, systematic distortion into the actual measurement data. A wall section that only exhibits surface roughness due to manufacturing is considered smooth in this context.

[0044] If it is desired that the training dataset also be supplemented with such systematic distortions, then real measurement data, in which a device for performing at least the first non-destructive measurement procedure is not ideally aligned with the object, can of course be used to create the training dataset or a further training dataset. However, to train the neural network on the fundamental deviations, it is advantageous if, at least for a portion of the training dataset, the real measurement data used for superimposing the simulation measurement results does not contain such additional distortion. In this way, a particularly reliable neural network can be trained using this method.

[0045] The training dataset can also include input and output datasets representing geometric data for a defect-free section of the object, as well as the measurement results associated with this geometric data. The measurement results associated with such defect-free geometries of the object can be simulation-generated measurements. These can also be enriched with real-world measurement data obtained during non-destructive testing of a defect-free area of ​​the object or a reference object. Alternatively, if the geometry of the defect-free areas of the object or reference object is known, the real-world measurement data can be used directly as measurement results. Using such a training dataset, the neural network is also trained to recognize defect-free areas of the object.A neural network trained in this way can better distinguish between defects and defect-free areas and is better suited to determine the geometry of defects.

[0046] Preferably, the neural network is a convolutional neural network. Such a neural network can be trained particularly easily with the training dataset to determine the geometry of one or more real defects of an object examined using at least one initial non-destructive measurement method.

[0047] Preferably, the neural network comprises an input layer for receiving the input data sets of at least one training data set, and an output layer for outputting data on the geometry of the object or sections of the object. The output of such data occurs regardless of whether a defect is present or not. The input layer allows the receipt of a 2D or pseudo-3D representation of the measurement results. Here, each point in a coordinate system is assigned a measurement value from the measurement method, corresponding to the specified resolution. The output of the output layer can be in the same coordinate system and also comprise a 2D or pseudo-3D representation of the geometry of the object or a section of the object. Each point in a coordinate system within the specified resolution is assigned a value describing the geometry of the object, in particular the geometry of a defect.This can involve depth information representing the depth of a corrosion site or crack at the corresponding point in the coordinate system. Spatially resolved input data sets from the execution of the non-destructive testing procedure can thus be translated into spatially resolved representations of potential defects in the object. Such a neural network is particularly well-suited for determining the geometry of one or more defects in a metallic object examined using at least one initial non-destructive testing method.

[0048] Preferably, the neural network is trained by providing it with the input data from at least one training dataset. Based on the input data, the neural network generates predictive output datasets, each comprising at least one predictive geometry. These predictive output datasets are then compared to the output datasets of the at least one training dataset. Based on the discrepancies between the predictive output datasets and the output datasets of the at least one training dataset, parameters of the neural network are adjusted. In the context of neural networks, these parameters are also referred to as weights. The direction and extent of the adjustment of the parameters or weights of the neural network are determined by at least one training algorithm.During the training of the neural network, the neural network itself, and thus the functionality of the training computer unit, is changed.

[0049] These steps are preferably repeated until a specific error measure is reached. The number of steps required depends, among other things, on how much the parameters or weights of the neural network are changed in each individual step. The error measure is predefined. In particular, the error measure determines what deviation between the predicted output data sets and the actual output data sets is acceptable. A threshold value can be used as the error measure, which must be either undercut or exceeded.

[0050] Preferably, at least one training dataset is generated, at least in part, from a database containing geometry data and the associated simulation measurement results from the simulation of at least one initial non-destructive measurement method. This allows for particularly quick and easy creation of a training dataset. The simulation measurement results do not need to be recalculated repeatedly but can be at least partially retrieved from an existing database. This makes the process simpler and more cost-effective.

[0051] Preferably, the input datasets also include at least one piece of information about the type of defect, and the neural network is configured to output information about the type of defect when defects are identified. Using such a training dataset, the neural network is further trained to output the type of defect when a defect is identified. The type of defect could be, for example, corrosion, a crack, an inclusion, delamination, or similar. A neural network trained in this way can output additional information about any defects.

[0052] Preferably, the at least one training dataset further comprises input and output data sets obtained by measuring various possible internal components of the object using at least one first non-destructive testing method. Internal components are understood here to be elements that are located in areas that are locally limited compared to the object's other dimensions. If the object under investigation is a pipeline, it typically has various measuring devices for monitoring the pipeline. For example, the pressure and / or temperature in the pipeline can be monitored. Furthermore, pipelines can include branches, valves, motors for actuating the valves, and similar internal components. In addition, pipelines can be provided with, for example, cathodic corrosion protection. This, too, can locally influence certain non-destructive testing methods.The input data sets comprise at least one measurement result obtained from the measurement of the respective component using at least one initial non-destructive measurement method, and the output data sets comprise at least the relevant geometric data of the respective component. Training a neural network with a training data set supplemented with corresponding data on components makes it possible to provide a neural network that can not only determine the presence of defects and their geometry, but also recognize the presence of a component and output corresponding geometric data relevant to the respective component.

[0053] Since it can be difficult to calculate a simulation measurement result for typical installations and / or to provide the geometry data of the installation relevant for at least the first non-destructive measurement method, real measurement data is used for this purpose. However, simulation measurement results of corresponding installations can also be used.

[0054] Measurements of various possible internal components of the object are preferably performed on components without geometric defects. The real-world measurement data obtained from these components are used to enrich at least some of the simulation measurement results. In particular, measurement results obtained by superimposing simulation measurement results with real-world measurement data from components without geometric defects can be used to train the neural network to recognize the geometry of defects with sufficiently good accuracy, even when these defects occur in the vicinity of corresponding internal components. The associated geometric data is obtained by combining the relevant geometry of the respective internal component with the geometry of the corresponding defect.A neural network trained in this way is particularly useful because the geometry of defects in the area of ​​installations can otherwise often only be determined with difficulty and / or with great uncertainty due to possible changes caused by the respective installation in the data obtained by the respective non-destructive measurement methods.

[0055] Preferably, the input datasets include at least information about the type of component present, and the neural network is configured to output information about the type of component when components are identified. With a suitable training dataset comprising input datasets appropriately supplemented with information about the type of component, and with a neural network that outputs information about the type of component when components are identified, a neural network can be trained that can output information about any components present in addition to the geometry of defects.

[0056] Preferably, the measurement results of the at least one training dataset are each assigned to a geometrically larger section of the object than the geometry data corresponding to the respective measurement result. In particular, the geometry data are assigned to a section of the object that lies centrally within the section of the object to which the measurement results are assigned. When measuring an object with the at least one non-destructive measurement method, defects in a section of the object can be detected by a device for carrying out the at least one method according to the invention without the device being placed directly in the area of ​​the defect. Particularly in magnetic flux leakage measurements, a defect can cause changes in the area of ​​the magnetic flux leakage lines that extend into a region outside the defect itself and can be detected there by a corresponding device.Because the measurement results in at least one training dataset are assigned to a geometrically larger section of the object than the geometry data corresponding to the respective measurement result, this effect can be taken into account when training the neural network.

[0057] To use such a trained neural network, measurement data obtained during at least one non-destructive measurement procedure on the object must be divided into corresponding sections and provided to the neural network as input data. The neural network then outputs geometric data belonging to a correspondingly smaller geometric section of the object under investigation. Because the input data represents a larger section of the object than the respective output data, the quality of the defect geometry estimation is further improved with such a trained neural network. The individual sections of the data obtained from the measurement overlap at their edges in such a way that the resulting geometric data representing a smaller section of the object under investigation are at least seamlessly adjacent to one another or also overlap.

[0058] Preferably, the method for training the neural network comprises a further step in which, after completion of training with the at least one training dataset, the neural network undergoes further training. A fine-tuning training dataset is used for this further training. This fine-tuning training dataset is created by performing at least one initial non-destructive measurement procedure on at least one part of the object and using the measurement results obtained in this way, as well as the corresponding geometric data obtained from a geometric measurement, particularly of the defective sections of the object, for the creation of the input and output datasets.Such a fine-tuning training dataset can be used when at least one section of the object under investigation, for example, a pipeline, is measured using a separate geometric measurement method, such as a laser scan, thus obtaining the object's actual geometry. The neural network, trained on the basis of at least one training dataset enriched with simulation measurement results from real-world measurements, can then be further trained to deliver improved results when evaluating a specific object measured using at least one non-destructive measurement method. The fine-tuning training dataset contains information about both the device used to perform the non-destructive measurement method and the specific object itself.

[0059] Preferably, the input data sets comprise at least one further measurement result from another non-destructive testing method, corresponding to the respective geometry. The neural network is configured to receive the corresponding number of measurement results from different non-destructive testing methods, evaluate them jointly, and output a geometry of one or more defects. Each input data set then receives at least one further measurement result obtained from another non-destructive testing method or a simulation of another non-destructive testing method. Using such a training data set, the neural network is trained to evaluate further information based on another non-destructive testing method to determine the geometry of a defect.In particular, when the measurement methods used are linearly independent, the neural network now has additional information available to determine the geometry of a defect. Combining the results of several non-destructive testing methods with the neural network allows for better prediction of defect geometry. This is especially true when the different non-destructive testing methods are sensitive to different types and / or orientations of defects. Different non-destructive testing methods are often defect-specific, so using at least two different methods can lead to synergistic effects.While, for example, MFL measurement data often detect corrosion-specific defects and EMAT measurement methods are more commonly used for crack detection, a combination of these two non-destructive measurement methods can determine a wider range of potential defects. Furthermore, the quality of the prediction of defect geometry is improved by the neural network trained in this way.

[0060] Particularly preferably, the first non-destructive testing method is an axially oriented MFL, a circumferentially oriented MFL, or an eddy current testing method, and the at least one further non-destructive testing method is another testing method from this group of methods or an EMAT or ultrasonic testing method. When used in pipelines, for example, an axially oriented testing method can detect defects with a circumferential extent particularly well, while defects whose extent extends axially along the pipeline are hardly detected by such a method. Conversely, a circumferentially oriented MFL method detects axially oriented defects particularly well, while it is significantly less able to detect circumferentially oriented defects. If the output data sets of the at least one

[0061] Since the training dataset includes measurement results for both axially oriented and circumferentially oriented MFL measurement methods, the neural network can be trained to evaluate both sets of results simultaneously, thus achieving a better determination of defect geometry. The quality of a correspondingly trained neural network for determining the geometry of one or more defects in a metallic object examined using at least one first non-destructive measurement method and another non-destructive measurement method is further improved.

[0062] The invention is further solved by a method for determining one or more actual investigated defects using a neural network trained according to one of the preceding claims in an evaluation unit. The training unit can be used as the evaluation unit. Typically, the requirements for the training unit to train a neural network within a reasonable timeframe are higher than the requirements for an evaluation unit to use the trained neural network to determine defect geometries. For economic reasons, it is therefore advantageous to use a different unit as the evaluation unit than the training unit, which can then be used for other purposes.The neural network is fed at least one reference dataset of the object, generated based on at least one initial non-destructive measurement method. The geometry is then preserved by the neural network inverting at least parts of this reference dataset. This inversion of the reference dataset parts by the trained neural network is performed quickly and reliably. This is particularly important when processing very large reference datasets resulting from the inspection of correspondingly large objects.

[0063] Preferably, the method comprises the step of generating the reference data set, in which at least one device for carrying out the at least one non-destructive measurement method measures or inspects the object and thus acquires the reference data set. Alternatively, a reference data set obtained by a previously performed non-destructive measurement method can also be used.

[0064] A major advantage of using such a trained neural network is that the reference dataset of the measured object does not need to be obtained completely from a single measurement. The reference dataset can also be obtained by combining two or more sequentially performed non-destructive testing methods. This is particularly relevant in the context of pipelines being measured or inspected using an inline inspection pig. With previously known methods for determining the geometry of one or more actual defects under investigation, a complete inspection run with such an inline inspection pig is necessary. If parts of such an inline inspection pig fail, preventing certain sections of the pipeline from being inspected, the run must be repeated until a complete run is achieved.When using the trained neural network, it may be sufficient to repeat the run if, during the second run, a potential failure of parts of the inline inspection pig prevents other sections of the pipeline from being inspected. The neural network can then be provided with a reference dataset from these two incomplete measurement runs for evaluation. The method is particularly easy to implement.

[0065] Particularly preferably, the inner wall of an object designed as a pipeline is measured, wherein the device is an inline inspection pig. With an object designed as a pipeline and a device designed as an inline inspection pig, a measurement run can cover several hundred kilometers of the pipeline. Correspondingly large amounts of data are generated during such a measurement run, in which the inner wall of the object designed as a pipeline is measured. For the evaluation of such a large amount of data, a particularly fast method for inverting the reference data set is especially advantageous. The neural network trained with the method according to the invention represents such a method, with which the reference data set can be inverted particularly quickly and thus cost-effectively, and the geometry of defects can be determined.

[0066] Preferably, the object is subjected to pressure at least during operation and is in particular designed as an oil, gas or water pipeline, and the method includes as a further process step a determination of a load-bearing limit of the object based on the geometries of the defects determined from the at least one reference data set.

[0067] Because the neural network can determine the geometry of defects with exceptional reliability, the subsequent determination of the object's load-bearing capacity can also be performed with exceptional reliability. With a more reliable determination of the load-bearing capacity, the safety margins typically applied when considering the load-bearing capacity can be reduced. The reserves of the object under investigation can be better utilized.

[0068] The method preferably includes a step in which the operating conditions of the object are adjusted based on the determined load-bearing limit. This can be achieved by setting operating parameters with a sufficient safety margin from the determined load-bearing limit during the object's operation. The more precisely the load-bearing limit is determined—in this case, by improving the geometry of defects during the inversion of parts of at least one reference dataset using the neural network—the smaller the required safety margin can be, and the object's reserves can be better utilized. This is particularly relevant when the object is an oil or water pipeline, or especially a gas pipeline. The speed at which a fluid such as oil or water can be conveyed through a pipeline depends on the pressure gradient within the pipeline.Higher pressure gradients result in higher flow velocities of liquids. The pipeline's transport capacity is increased. In gas pipelines, this effect is further compounded by the fact that the density of the gas transported in the pipeline increases at higher pressure. Therefore, at higher pressure and the same pressure gradient, more mass of a gas can be transported in a gas pipeline than at lower pressure.

[0069] The load-bearing capacity limit is typically determined by the most severe defect. With a better estimation of the defect geometry and an improved determination of the load-bearing capacity limit, a more informed decision can be made regarding which parts of an object, particularly a pipeline, require immediate repair and which parts will only need to be repaired, reinforced, or replaced during later, possibly regularly scheduled, maintenance work.

[0070] Furthermore, the invention comprises a computer program product including instructions for carrying out a method for determining the geometry of one or more real investigated defects using a neural network trained according to the method according to the invention.

[0071] Further advantages and units of the invention can be found in the following description of the figures. Figure 1 schematically shows the creation of a training data set for training a neural network;

[0072] Fig. 2 shows the creation of a training dataset with additional

[0073] Information about the defect-free area of ​​the object;

[0074] Fig. 3 shows the creation of a training dataset with additional

[0075] Information on the building's fixtures and fittings;

[0076] Fig. 4 shows the training of a neural network;

[0077] Fig. 5 shows a device for carrying out a non-destructive measurement method;

[0078] Fig. 6 shows the creation of a fine-tuning training dataset;

[0079] Fig. 7 shows the use of a trained neural network for

[0080] Determination of the geometry of one or more defects.

[0081] Fig. 8 Geometry data showing a defect

[0082] represent a section of the object; Fig. 9 represents the geometry data corresponding to Fig. 8.

[0083] Simulation measurement result;

[0084] Fig. 10 and Fig. 11 show real measurement data obtained during the non-destructive measurement procedure on a defect-free area of ​​a reference object;

[0085] Fig. 12 shows a measurement result obtained by combining the simulation measurement result according to Fig. 9 and a section of the real measurement data according to Fig. 11;

[0086] Fig. 13 shows a measurement result obtained by combining the simulation measurement result of Fig. 9 with a part of the real measurement data according to Fig. 10;

[0087] Fig. 14 shows at least one non-destructive method.

[0088] Measurement method generates part of a reference data set of an object;

[0089] Fig. 15 shows a result obtained by inverting the reference data set. Fig.

[0090] 7 Geometry obtained by the neural network of the

[0091] The object includes the geometry of the defects; Fig. 1 shows a flowchart for creating a training dataset 4. The training dataset 4 comprises at least one output dataset 8 and one input dataset 6. The output dataset 6 comprises geometry data 10 representing an object 28, with at least some of the geometry data 10 including defects. Based on the geometry data 10, simulation measurement results 14 are generated by simulating at least one first non-destructive measurement method. Real measurement data 16 are obtained from defect-free areas of an object 28 or a reference object using the at least one non-destructive measurement method. The simulation measurement results 14 are enriched with the real measurement data 16 and used as measurement results 12. These form an input dataset 6 of the training dataset 4.Using measurement results 12, which were obtained from simulation measurement results 14 enriched with real measurement data 16, as input data set 6, makes it possible to easily create a training data set 4 with which a neural network 2 can be trained particularly well to determine the geometry of defects after carrying out at least one first non-destructive measurement procedure on the object 28.

[0092] Fig. 2 shows how a training dataset 4 created according to Fig. 1 can be supplemented with information on defect-free areas of the object 28. For this purpose, the real measurement data 16 from a defect-free area of ​​the object 28, along with the associated geometry data 11 of the corresponding defect-free area of ​​the object 28, are used as input datasets 6 and output datasets 8. These form a further part of the training dataset 4. The use of a training dataset 4 that also contains information about defect-free areas of the object 28 makes it possible to train a neural network 2 to better distinguish between defects and defect-free areas of the object 28 when determining the geometry of defects, and thus to better determine the geometry of defects.

[0093] Fig. 3 shows how a training dataset 4 created according to Fig. 1 can be supplemented with information about any internal components of object 28. For this purpose, defect-free internal components of object 28 are measured using at least one first non-destructive measurement method. The real measurement data 17 of the defect-free internal components obtained in this way can, together with the associated relevant geometry data 11 of the defect-free internal components, form output datasets 8 and input datasets 6 of a training dataset 4. With such a training dataset 4, the neural network 2 can be trained to recognize internal components and their associated geometry.

[0094] However, it is also possible to superimpose the real measurement data 17 of defect-free components with simulation measurement results 14 belonging to geometry data 10 with defects and to use them as measurement results 12 in the input data set 6. The geometry data of the output data set 8 belonging to these measurement results 12 can then be obtained by combining the geometry data 10 with defects and the geometry data 11 of defect-free components in a manner corresponding to the enrichment process. In this way, it is possible to obtain a training data set 4 with which a neural network 2 can be trained to accurately determine the geometry of defects even when defects occur in the vicinity of components.

[0095] A neural network 2 is trained using such a training dataset 4, as shown in Fig. 4. The input datasets 6 of the training dataset 4 are passed to the neural network 2 via an input layer 2'. The neural network 2 determines predictive output datasets 18, which are output via an output layer 2". The predictive output datasets 18 and the output datasets 8 are compared in step 22, and in step 24 a training algorithm adjusts parameters or weights of the neural network 2. These steps are repeated until an acceptable error level is achieved when comparing the predictive output datasets 18 and the output datasets 8.

[0096] Fig. 5 shows a device 26 for carrying out the at least one first non-destructive measurement method in the form of an inline inspection pig, which is arranged in an object 28 designed as a pipeline for carrying out the at least one first non-destructive measurement method.

[0097] Based on the measurement procedure thus performed, a fine-tuning training dataset 4' can be created. This is shown in Fig. 6. For this purpose, a geometric measurement must be carried out on at least one section of the object 28 to determine the actual geometry of the object 28. This can be done by a laser scan. The geometric data obtained in this way are used as fine-tuning geometry data 10' in the fine-tuning output dataset 8'. The data belonging to the section of the object 28 examined by the geometric measurement, obtained with the at least one first non-destructive measurement method, are used as fine-tuning measurement results 12' in the fine-tuning input dataset 6'. With a fine-tuning training dataset 4' created in this way, a neural network 2 trained according to Fig. 4 can be further trained according to the procedure shown in Fig. 4, but using the fine-tuning training dataset 4'.This trains the pre-trained neural network 2 on the specific object 28 under investigation and the specific device 26 used to perform at least the first non-destructive measurement procedure. A neural network 2 trained in this way for the specific application can determine the geometry of defects in the specific object 28 under investigation particularly well.

[0098] Based on the neural network 2 trained using a method according to Fig. 4 or further trained according to Fig. 6, the method shown in Fig. 7 can be used to determine the geometry of one or more real defects under investigation. A device 26 for carrying out at least one first non-destructive measurement method, in this embodiment an inline inspection pig, is used to examine the object 28 under investigation, in this embodiment a pipeline. At least one reference data set 20 is generated from this examination. The reference data set 20 is provided to the neural network 2. The neural network 2 then outputs the geometry of one or more defects of the object 28 under investigation. Based on the geometries of defects thus determined, the load-bearing capacity limit of the pipeline is determined in step 30 of the method according to Fig. 7.Based on the load-bearing capacity limit determined in this way, the operating conditions of the pipeline can be set in step 32.

[0099] In the illustrated embodiment, two reference data sets 20 are passed to the neural network 2. One reference data set 20 is obtained using a first non-destructive measurement method, and the second reference data set 20 is obtained using a second non-destructive measurement method that differs from the first. The neural network 2 must be configured and trained to accept the two different reference data sets 20. The neural network can then perform a parallel evaluation of both reference data sets 20.Because the various non-destructive measurement methods are typically more sensitive to certain types and / or orientations of defects than to others, a neural network 2 trained to evaluate different reference data sets 20 can typically determine the geometry of defects more accurately than a neural network that only accepts one of the reference data sets 20 for evaluation.

[0100] Fig. 8 shows a graphical representation of geometric data 10 representing a section of object 28 with at least one defect, which forms part of an input dataset 8 of a training dataset 4 for training a neural network 2. The section of object 28 represented by the geometric data has several defects. In Fig. 8, these are indicated by lines of equal height that contain depth information. The defects shown are corrosion defects. The geometric data 10 can either be generated synthetically or obtained by measuring an object with corresponding defects using a geometric measurement technique such as a laser scan.

[0101] Fig. 9 shows a simulation measurement result 14 obtained by simulating at least one first non-destructive measurement method. The simulation measurement result 14 in Fig. 9 represents the same area of ​​an object 28 as the geometry data 10 of Fig. 8. It can be seen that the simulation measurement result 14 shows deflections in the area of ​​the defects. The individual lines of the simulation measurement result 14 represent the response signal that would be registered, under the ideal conditions assumed for the simulation, by a sensor of a device 26 for carrying out the at least one first non-destructive measurement method, which would be moved along this line past an object 28 having the geometry represented in Fig. 8. In the areas that are far enough away from the defects, these lines are straight under the ideal conditions assumed for the simulation.In the area of ​​defects, the measurement signal of the respective sensor changes. Depending on the type of non-destructive testing method, a sensor in such a device can register a change even before it reaches the defect or after it has already passed it. The illustrated embodiment shows a simulation measurement result of a magnetic flux leakage measurement (MFL measurement).

[0102] Alternatively, for such a measurement a simulation measurement result 14 can be used, which represents a larger section of the object 28, or only a section of the geometry data 10 shown in Fig. 1 can be used for the creation of the training data set 4.

[0103] Figures 10 and 11 show real measurement data 16 obtained by performing at least one first non-destructive measurement method on a defect-free area of ​​an object or a reference object. The object is a pipeline that was measured circumferentially and axially. The signals recorded by the individual sensors of the device 28 used for this purpose run essentially horizontally in the defect-free area. However, it can be seen that the lines are not smooth but exhibit slight deflections. These are caused by deviations of the conditions prevailing during a real measurement from the idealized conditions assumed for the simulation. For example, the geometry of the pipelines may exhibit manufacturing-related deviations from an ideal shape, or surface roughness may lead to fluctuations in the signal.The speed at which the device 28 is moved past the object 26 can vary. All these variations can contribute to the deviations shown in Fig. 10, the noise.

[0104] Fig. 11 shows alternative real measurement data 10 to Fig. 10 in a similarly defect-free area of ​​object 28. It can be seen that there is an offset of the lines at approximately the axial coordinate 400. At the axial coordinate 700, there is a strong spike in all signals, which then drop off. This indicates that the device 26 was passed at this point by a part of object 28 extending around the entire circumference of the pipeline, which affected all sensors of the device 26 equally. The pipeline has a circumferential weld seam at this point. A white line is visible in the area of ​​the circumferential coordinate at approximately 100. This is due to the failure of the sensor of the device 26 assigned to the corresponding area.

[0105] Figures 10 and 11 thus illustrate, using noise and weld seams as examples, that real measurement data 16 can be expected to contain effects that are not reflected in the simulation measurement results 14. Therefore, measurement results 12 obtained by enriching the simulation measurement results 14 with real measurement data 16 are used to carry out the procedure.

[0106] Figures 12 and 13 show corresponding measurement results 12. The measurement result 12 according to Figure 12 was obtained by enriching the simulation measurement result 14 according to Figure 9 with a section of the real measurement data 16 according to Figure 11. Figure 13 shows a measurement result 12 that was obtained by enriching the simulation measurement result 14 according to Figure 9 with parts of the real measurement data 16 according to Figure 10.

[0107] The enrichment of the simulation measurement results 14 with the real measurement data 16 is carried out for the measurement results 12 shown in Fig. 12 and Fig. 113 by adding the simulation measurement result 14 and parts of the real measurement data 16. Prior to this, the simulation measurement result 14 and the real measurement data 16 were normalized. Since the illustrated embodiment involves MFL measurement data, the normalization was performed with respect to the respective background magnetization.

[0108] The training dataset 4 for training a neural network 2 comprises input and output datasets 6, 8, where geometric data 10, as shown in Fig. 8, are each assigned to measurement results 12, as shown in Figs. 12 and 14. By superimposing the same simulation measurement result 14 with different real-world measurement data 16, a large training dataset 4 can be easily generated. The same geometric data 10 can be combined with different measurement results 12, which are based on the same simulation measurement result 14 enriched with different real-world measurement data 16 or different sections of the real-world measurement data 16.The use of such a training data set 4 makes it easy to train a neural network 2 which can easily perform an inversion of an object with defects obtained by measuring a defect with the reference data set 20 obtained by at least one first non-destructive measurement method and can thus determine the geometries of defects of the object 28 under investigation with particular reliability.

[0109] Fig. 14 shows a reference data set 20 of an object generated based on at least one first non-destructive measurement method. In the example shown in Fig. 14, this data consists of magnetic flux leakage data. However, one or more other non-destructive measurement methods can also be used to obtain one or more reference data sets 20 of the object. The reference data set 20 is provided to the previously trained neural network 2. The neural network 2 then outputs the geometry of the object, including defects, as shown in Fig. 15. The defects are again represented in Fig. 15 as lines of equal height.

[0110] For evaluation, the reference dataset 20 can be divided into individual sections whose size corresponds to the size of the measurement results 12 used in training the neural network 2. The neural network 2 then outputs a geometry for each of these sections, the size of which corresponds to the geometry data 10 used in the training dataset 4. The geometry shown in Fig. 15 can then be assembled from these individual sections of the geometries 10.

Claims

Patent claims 1. Method for training a neural network (2) for determining the geometry of one or more real defects of a metallic, in particular magnetizable, object (28), in particular a pipe or a tank, examined by means of at least one first non-destructive measurement method, wherein the neural network (2) is trained in a training computer unit with at least one training data set (4) comprising related input and output data sets (6, 8), wherein an output data set (8) comprises at least one section of the object (28) having at least one defect, geometry data (10) representing, and an input data set (6) comprises at least one measurement result (12) associated with the geometry data (10), characterized in that at least for a part of the training data set (2) the measurement result (12) is a simulation measurement result (14) obtained by a simulation of the at least one first non-destructive measurement method.which is enriched with real measurement data (16) obtained during the execution of the non-destructive measurement procedure on a defect-free area of ​​the object (28) or a reference object.

2. Method according to claim 1, characterized in that the first non-destructive measuring method is an MFL measuring method, in particular an axially oriented MFL or a circumferentially oriented oriented MFL measurement method, or an eddy current measurement method.

3. Method according to one of the preceding claims, characterized in that the simulation measurement result (14) is enriched by real measurement data (16) by superimposing the simulation measurement result (14) with real measurement data (16).

4. Method according to claim 3, characterized in that the parts of the simulation measurement result (14) which are assigned to an area of ​​the geometry data (10) which has a defect are superimposed with real measurement data (16).

5. Method according to one of claims 3 or 4, characterized in that the simulation measurement result (14) and the real measurement data (16) are normalized before superimposition.

6. Method according to one of the preceding claims, characterized in that when enriching the simulation measurement results (14) with real measurement data (16), areas around welds and / or components of the object or a reference object are excluded as having defects by definition when carrying out the non-destructive measurement method on a defect-free area of ​​the object (28) or a reference object.

7. Method according to one of the preceding claims, characterized in that the neural network (2) is trained by passing the input data sets (6) of the at least one training data set (4) to the neural network (2), on the basis of which the neural network (2) generates prediction output data sets (18) which each comprise at least one prediction geometry, wherein a comparison of the prediction output data sets (18) and the output data sets (8) of the at least one training data set (2) is carried out and parameters of the neural network (2) are adjusted based on the deviations between the prediction output data sets (18) and the output data sets (8).

8. Method according to claim 7, characterized in that these steps are repeated until an error measure is reached.

9. Method according to one of the preceding claims, characterized in that the at least one training data set (2) is formed at least partially on the basis of data from a database containing geometry data (10) and the associated simulation measurement results (14).

10. Method according to one of the preceding claims, characterized in that the output data sets (8) also include at least one piece of information about the type of defect and the neural network (2) is configured to output information about the type of defects when defects are identified.

11. Method according to one of the preceding claims, characterized in that the at least one training data set (4) further comprises input and output data sets (6, 8) obtained by measurements of various possible incorporations of the object (28) using the at least one first non-destructive measurement method, wherein the input data sets (6) comprise at least one measurement result obtained during the measurement of the respective incorporation using the at least one first non-destructive measurement method and the output data comprise at least the relevant geometry data (11) of the respective incorporation.

12. Method according to one of the preceding claims, characterized in that at least for a part of the training data set (2) a simulation measurement result (14) which is enriched with real measurement data (16) obtained by measuring at least one installation of the object with the at least one first non-destructive measurement method is used as measurement result (12).

13. Method according to claim 11 or 12, characterized in that the output data sets (8) include at least one piece of information about the type of the respective installation and the neural network (2) is configured to output information about the type of installations when installations are identified.

14. Method according to one of the preceding claims, characterized in that the measurement results (12) are each assigned to a geometrically larger section of the object (28) than the geometry data (10) corresponding to the respective measurement result (12).

15. Method according to one of the preceding claims, characterized in that, based on the execution of at least one first non-destructive measurement method on at least one part of the object (28) and the fine-tuning measurement results (12') obtained, as well as the fine-tuning geometry data (10') corresponding to these fine-tuning measurement results (12') obtained by a geometry measurement, in particular of the sections of the object (28) having defects, a fine-tuning training data set (4') is created and the trained neural network (2) is further trained with the fine-tuning training data set (4').

16. Method according to one of the preceding claims, characterized in that the output data sets (8) include at least one further measurement result (12) of a further measurement system assigned to the respective geometry. non-destructive measurement methods and the neural network (2) is set up to receive the corresponding number of measurement results (12) from non-destructive measurement methods, evaluate them jointly and output a geometry of one or more defects.

17. Method according to claim 16, characterized in that the first non-destructive measuring method is an axially oriented MFL, circumferentially oriented MFL or an eddy current measuring method and the at least one further non-destructive measuring method is another measuring method from this group of measuring methods or an EMAT or ultrasonic measuring method.

18. Method for determining the geometry of one or more real, investigated defects using a neural network (2) trained according to one of claims 1 to 16 in an evaluation computer unit, wherein at least one reference data set (20) of the object (28) generated on the basis of at least one first non-destructive measurement method is passed to the neural network (2), wherein the geometry is obtained by inversion of at least parts of the reference data set (20) by the neural network (2).

19. The method of claim 18, further comprising the step of generating the reference data set (20) by at least one device (26) for The object (28) is measured or inspected and the reference data set (20) is recorded, at least by first carrying out a non-destructive measurement procedure.

20. Method according to claim 19, characterized in that the inner wall of an object (28) designed as a pipeline is measured and the device (26) is an inline inspection pig.

21. Method according to one of claims 18 to 20, characterized in that the object is an object (28) that is subjected to pressure at least during operation and is designed in particular as an oil, gas or water pipeline, and a load-bearing limit of the object (28) is determined on the basis of the geometry of the defects.

22. Method according to claim 21, characterized in that the operating conditions of the object (28) are adjusted on the basis of the determined load-bearing limit.

23. Computer program product comprising instructions for carrying out a method according to any one of claims 17 to 21.

Images