Method for generating artificial training data for performance of a machine learning method in training of artificial intelligence for use in a classification method
Patent Information
- Application Number
- EP2023735601
- Authority / Receiving Office
- EP · EP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2022-06-20
- Filing Date
- 2023-06-20
- Publication Date
- 2025-04-23
AI Technical Summary
Current methods for screening luggage using X-ray diffraction devices suffer from high false alarm rates due to imprecise material differentiation, leading to unnecessary manual checks and reduced throughput, as they require extensive and costly real-scanning-based training data for accurate artificial intelligence evaluation.
A generation method for artificial training data using variational simulation of momentum transfer functions, allowing for the creation of a large number of labeled data points without extensive real scans, which can then be used to train artificial intelligence for accurate classification of materials in luggage.
This approach significantly reduces the effort and cost of generating training data, enabling a high level of security with a low false alarm rate by using artificial intelligence to evaluate X-ray diffraction scans with improved accuracy and efficiency.
Smart Images
Figure 1.1
Abstract
Description
[0001] Generation method for artificial training data for carrying out a machine learning process when training an artificial intelligence for use in a classification process
[0002] The present invention relates to a generation method for artificial training data for carrying out a machine learning method of an artificial intelligence for use in a classification method of objects in luggage based on a scan with an X-ray diffraction device, a learning method for training an artificial intelligence and a computer program product for carrying out the generation method and / or the learning method.
[0003] It is well known that in security-critical areas, especially at airports, a high security standard is maintained by x-raying baggage and checking it for dangerous materials and objects. At airports, this applies in particular to so-called checked baggage, which must be completely inspected, i.e., every single piece of baggage, before it is loaded into the cargo hold of an aircraft. Airports typically have transmission devices, such as computer tomography scanners, capable of x-raying baggage. This allows individual objects to be identified with a certain degree of probability, particularly as part of an image analysis of the transmission scans, and to assess their dangerousness, for example, based on their shape.
[0004] In addition, such transmission methods can be used to obtain information about the material properties of objects. In other words, transmission methods can be used, for example, to determine the range in which the average density of a detected object lies, so that at least an indication of a material with a similar density can be assigned to this object. However, this assessment is too imprecise to be able to distinguish dangerous from non-dangerous objects with the desired level of certainty. In other words, objects made of non-dangerous material are also pre-selected as "dangerous" if their density is similar to the density of a dangerous material. This means that a large number of objects trigger an alarm even though they do not contain such a dangerous material.Such an alarm is called a false alarm and is undesirable because it leads to unnecessary effort, especially in the form of manual baggage checks.
[0005] To reduce this false alarm rate, it is also already known to supplement the transmission devices so that pieces of luggage for which an alarm has been issued are subjected to a further, additional test. This additional test could, for example, be an X-ray diffraction device, which does not perform a transmission scan, but can record and evaluate X-rays diffracted by the piece of luggage when irradiated with X-rays. However, this step results in a significant reduction in speed. To evaluate the scan information from an X-ray diffraction device, a so-called momentum transfer function is created, which must be generated for all individual volume components, also known as voxels, and then evaluated.The accuracy of this momentum transfer function depends on the number of photons in the volume components and thus on the speed of the scan with the X-ray diffraction device. The faster the scan is performed, the lower the accuracy of such momentum transfer functions, but the higher the desired throughput of luggage items through such an X-ray diffraction device. Previous solutions therefore have to make do with relatively high speed and correspondingly low quality of the scan information from the X-ray diffraction device in order to perform the subsequent analysis.
[0006] The object of the present invention is therefore to at least partially remedy the disadvantages described above. In particular, the object of the present invention is to combine, in a cost-effective and simple manner, a high degree of reliability in the detection and verification of dangerous alarm objects with the lowest possible false alarm rate when using an X-ray diffraction device. In particular, the aim is to extract the maximum amount of information from the available X-ray diffraction results.
[0007] The above object is achieved by a generation method having the features of claim 1, a learning method having the features of claim 11, and a computer program product having the features of claim 15. Further features and details of the invention emerge from the subclaims, the description, and the drawings. Features and details described in connection with the generation method according to the invention naturally also apply in connection with the learning method according to the invention, and vice versa, so that with regard to the disclosure of the individual aspects of the invention, reference is always made to each other.
[0008] The core concept of the invention is based on using artificial intelligence for the classification of momentum transfer functions obtained during data processing in X-ray diffraction devices. However, such artificial intelligence obtains its functionality and accuracy through a machine learning process, which can also be referred to as a machine training process. Such a machine training process requires a large amount of training data that exhibits a correlation between input information and output information with sufficient accuracy—in this case, between momentum transfer functions at the input and material classes at the output of the artificial intelligence. This training data would typically have to be created laboriously using an X-ray diffraction device for different materials.
[0009] In order to compensate for the lack of accurate training data in a generation method according to the invention and simultaneously avoid the effort of generating the training data through actual scanning, the generation of artificial training data is proposed. This artificial training data is then used in a learning method explained later to implement a machine learning method for training an artificial intelligence, which can then be used in a classification process for objects in luggage based on a scan with an X-ray diffraction device. Such a generation method comprises the following steps:
[0010] Selecting at least one material-specific reference momentum transfer function, - Applying a variation simulation to the selected reference momentum transfer function to generate a plurality of variation momentum transfer functions,
[0011] - Storing the generated plurality of variational momentum transfer functions as artificial training data with the selected reference momentum transfer function as reference information.
[0012] As already indicated, the use of artificial intelligence in a classification process requires a training process for this artificial intelligence. The present invention refers in particular to artificial intelligence in the form of one or more neural networks with weighted nodes. Of course, however, other and / or more complex embodiments can also be understood under the term "artificial intelligence" within the scope of the present invention. The crucial core idea of using artificial intelligence is not to predetermine an algorithmic and thus physically definable relationship, but rather to train this relationship between input parameters and output parameters using a large amount of data in the form of the artificial intelligence's training data.
[0013] While many problems already addressed with artificial intelligence systems involve a large amount of data, such as image recognition, speech recognition, or similar, this is not the case for X-ray diffraction systems, particularly due to their highly sensitive nature in the security sector. Rather, for the use of artificial intelligence, the necessary training data would have to be generated entirely manually and additionally labeled. This means that artificial intelligence has not yet been used in the evaluation of X-ray diffraction scans. Rather, such a use, based on the current state of the art, would require a very large number of test scans, for example, in the range of several thousand scans.This extraordinarily high level of effort has been avoided to date, as the costs and time required are very high compared to the benefits of artificial intelligence. A generation method according to the invention now makes it possible to reduce this considerable effort to such an extent that training data can be provided cost-effectively, which can then be used to train an artificial intelligence in the evaluation of an X-ray diffraction device. This generation method is based on a few measured reference momentum transfer functions and can therefore be implemented using a few real test scans as a basis. An example of one generation method is explained in more detail below.
[0014] For example, an explosive composition can be searched for as a relevant material during later operation when classifying objects. In a generation method according to the invention, a momentum transfer function specific to the material of this explosive composition is then selected as the reference momentum transfer function. This reference momentum transfer function can be taken from a database. However, it is preferred if the reference momentum transfer function is recorded specifically for the respective material before selection, in particular on the same device type or even the identical device of the X-ray diffraction device. In order to achieve very high accuracy for the reference momentum transfer function, the scanning process for generating the reference momentum transfer function can be carried out, in particular, at a significantly reduced speed and thus with increased accuracy.
[0015] However, for the implementation of a generation method according to the invention, it is fundamentally irrelevant whether the reference momentum transfer function still needs to be generated or can already be selected from an existing database. What is crucial is that at least one material-specific reference momentum transfer function is selected. Two or more reference momentum transfer functions can also be selected, all specific to the same material.
[0016] This one reference momentum transfer function already represents a first parameter correlation of the artificial training data, as it represents a correlation between a momentum transfer function and the material specific to this momentum transfer function. The core idea of the invention is based on providing this one reference momentum transfer function with a sampling factor in a further step and, through variation, achieving a large number of different momentum transfer functions that, through the variation based on the reference momentum transfer function, automatically exhibit the same material specificity. In other words, the variation generates a multitude of momentum transfer functions suitable for the same material. To achieve this, a variation simulation is carried out in the second step of the generation process.The starting point for this variational simulation is always the selected reference momentum transfer function, so that preferably the same starting point is always selected for a very large number of variational simulation runs. The variational simulation in the sense of the present invention serves to vary the well-resolved reference momentum transfer function and generates a multitude of variational momentum transfer functions through mixing and sampling. For all generated variational momentum transfer functions, the dominant reference (ground truth) and thus the connection to the material of the reference momentum transfer function is retained.
[0017] To ensure that the correlation to the specific material is maintained even when the momentum transfer function is actually being varied, a corresponding relationship can be used in the variational simulation. This could, for example, be the artificial variational intelligence discussed later. Simple algorithmic relationships, simulation models, statistical distribution functions, linear combinations of spectra, or similar are also conceivable for developing the variational simulation and implementing a sampling functionality.
[0018] As a result of this second step, a multitude of variational momentum transfer functions is obtained, for example, with more than 100,000 or up to 500,000 different variational momentum transfer functions. In other words, all of these variational momentum transfer functions exhibit at least small differences in their detailed design from the reference momentum transfer function. However, the aforementioned differences between all variational momentum transfer functions and the reference momentum transfer function are so small that they do not change the relationship to the specific material of the reference momentum transfer function. A degree of change between the momentum transfer functions, defined, for example, in the form of a change value, is therefore below a limit value beyond which the variational momentum transfer function would no longer be specific to the respective material of the reference momentum transfer function.
[0019] In other words, the variational simulation leads to the availability of a multitude of variational momentum transfer functions, which correlate with, or at least partially even agree with, a multitude of actually determinable momentum transfer functions, without these momentum transfer functions having to be generated by real scans with great effort.
[0020] In the final step of a generation method according to the invention, the multitude of generated variational momentum transfer functions are stored as artificial training data, with this storage being marked or "labeled" in the sense of artificial intelligence. This means that the artificial training data not only contains the variational momentum transfer function, but also, correlated with it, the associated specific material as reference information. In other words, at the conclusion of the generation method according to the invention, a data set, a database, is available which contains a large number of artificial, labeled training data in the form of a large number of variational momentum transfer functions specific to a material and the material specificity as reference information. This was achieved without having to perform a large number of real measurements.
[0021] Of course, in a production method according to the invention, two or more different reference momentum transfer functions, either actually scanned or taken from a database, can also be used for one and the same material. The step of applying a variation simulation is not limited to the application of a single variation simulation. Here, too, it is fundamentally conceivable to use two or more different variation simulations. Last but not least, in anticipation of the later explanation, the production method according to the invention can of course also be carried out for two or more different materials.
[0022] It can be advantageous if, in a generation method according to the invention, the variation simulation has a statistical distribution function to avoid redundant variational momentum transfer functions. This statistical distribution function can, in particular, be separate from a simulation model. Such a distribution function can ensure that momentum transfer functions generated by the variational simulation, which have already been generated from the variational simulation in a previous step, are not transferred multiple times into the training data in a redundant manner. The statistical distribution function can therefore be integrated downstream, but also into the variational simulation itself as an integrated variation parameter. This ensures that redundancy in the training data for the variational momentum transfer functions is avoided or at least reduced.
[0023] Further advantages can be achieved if, in a generation method according to the invention, the variation simulation comprises a simulation model that simulates at least part of the scan with the X-ray diffraction device. Such a simulation model can, for example, be a mathematical and / or physical simulation model that, in computer-implemented form, represents part of the variation simulation. In particular, such a simulation model can be specific to a type or a specific design of an X-ray diffraction device. Such simulation models can also be specific to different materials, provided these would lead to different characteristics during scanning in the X-ray diffraction device.In other words, the simulation model serves as a kind of backward simulation, since it generates a possible starting point for the classification analysis in the form of the multitude of variational momentum transfer functions from a known result, namely the material to be identified and classified as a reference, in the reverse direction.
[0024] It is also advantageous if, in a generation method according to the invention, the variation simulation has at least one variation parameter which is varied during use and in particular represents at least one of the following real factors when performing a scan with an X-ray diffraction device:
[0025] - Mixing with other materials, poor demarcation from other objects and / or segments, - Distortion of the momentum transfer axis,
[0026] - Photon sampling.
[0027] The above list is not exhaustive. It is also possible to use two or more different variation parameters. In particular, the variation parameters relate to physical and / or real-world factors that, when performing the scan with a real X-ray diffraction device, lead to a variation in the measured results and thus to a variation in the generated momentum transfer functions. The variation parameters therefore serve to represent variations that occur in reality in the variation simulation. Poor differentiation from other materials can result, for example, from mixing with other materials.
[0028] Furthermore, it can be advantageous if, in a generation method according to the invention, the steps of applying and storing are repeated for at least one further material-specific reference momentum transfer function. This results in the same generation steps being carried out for an identical specific material with a different initial value, so that greater accuracy and / or a larger number of different, yet nevertheless significant training data can be achieved. Additionally or alternatively, however, another material can also be used, for which one or more reference momentum transfer functions are specifically selected for the generation method. This makes it possible to build a database that contains training data not just for a single material, but for a variety of different materials.It is preferred if the materials for the reference pulse transfer functions contain hazardous materials and / or non-hazardous materials.
[0029] A further advantage can be achieved if, in a generation method according to the invention based on the implementation of different reference impulse transfer functions, the artificial training data is stored in a reference database for later use in a learning process. As already explained in the previous paragraph, a reference database can thus be created which is capable of providing not only a single reproduction of a material, but rather of establishing a correlation with a multitude of different materials. This material database or reference database contains, in particular, at least dangerous alarm materials such as explosives or narcotics. However, harmless materials are preferably also included in order to be able to conduct an open search and / or a negative classification as "non-dangerous" during later classification.For example, it's possible to train a specific classification that's precisely tailored to a particular material. An open classification can also be trained on artificial intelligence in principle, enabling it to actively recognize or exclude different materials.
[0030] Furthermore, it can be advantageous if, in a generation method according to the invention, the variation simulation has a termination criterion with regard to the number of variational momentum transfer functions to be generated. In principle, a higher number of variational momentum transfer functions leads to greater accuracy for the subsequent learning process. However, since limits must be observed with regard to the computational effort during the generation process and, at the same time, it has been determined that, beyond a certain range of variation, no new information in the form of non-redundant variational momentum transfer functions is generated by the variation, such a termination criterion can limit the generation process to a reasonable scope. A particularly simple embodiment of such a termination criterion is a maximum number of, for example, 100,000, 200,000, or 500.000 generated variational momentum transfer functions. However, it is also possible in principle to provide a more complex design of the termination criterion, for example, with regard to the density of the distribution of the variational momentum transfer functions, their different change values between the variational momentum transfer function and the reference momentum transfer function, and the like.
[0031] It is also advantageous if, in a generation method according to the invention, the variation simulation comprises artificial variation intelligence, which is used in generating the plurality of variational momentum transfer functions. While a classical algorithmic relationship can in principle also be selected for the variational simulation, artificial variational intelligence can also be used. It can be based, in particular, on design data or on real measurements on the X-ray diffraction device and can therefore be part of the sampling process, i.e., the variational simulation itself. In this case, a classical training with real training data from the identical X-ray diffraction device can serve as the basis for this artificial variational intelligence.
[0032] Furthermore, it is advantageous if, in a generation method according to the invention, the reference momentum transfer function has been generated using the same X-ray diffraction device for which the artificial intelligence is to be trained for the classification method. In other words, the generation method and the learning method are consistently carried out for a particular and specific type of X-ray diffraction device. The output value in the form of the reference momentum transfer function is correlated with the type of X-ray diffraction device on which the classification method is to be later applied.
[0033] It is further advantageous if, in a generation method according to the invention, part of the variation simulation is manually adjusted. This allows for manual readjustment or adjustment of the variation, so that, for example, in weak areas of variation and the multitude of variation momentum transfer functions, a manual adjustment can be made to supplement and fill these areas. The manual variation can be achieved, for example, by manually adjusting the variation parameters accordingly.
[0034] A further object of the present invention is a learning method for training an artificial intelligence for use in a classification method for classifying objects in luggage based on a scan with an X-ray diffraction device, comprising the following step:
[0035] - Performing machine learning for artificial intelligence using the artificial training data generated by a generation method having the features of the present invention.
[0036] Thus, a learning method according to the invention offers the same advantages as those explained in detail with reference to a generation method according to the invention. It is advantageous if a neural network is used as artificial intelligence, so that a machine learning method for neural networks can be used.
[0037] A learning method according to the invention can be further developed by using the variational momentum transfer functions for at least one material to be detected as artificial training data. This artificial training data is thus capable of detecting a material and / or excluding such a material. It is thus a material-specific neural network with material-specific training.
[0038] Additionally or alternatively, it is advantageous in a given learning method if the variational momentum transfer functions that differ from a material to be detected are used as artificial training data. This leads to the possibility of negative selection, i.e., the exclusion of a material by detecting other materials. Thus, as soon as such artificial intelligence has detected another material, such a material, especially one classified as "non-hazardous," can rule out an alarm situation. This is particularly useful if this second stage of the X-ray diffraction device is to be used for further evaluation and resolution of initial alarms from transmission devices. A combination with active training methods according to the previous paragraph is of course also conceivable.
[0039] It is advantageous if the artificial training data is used without the reference information in a learning method according to the invention. This can be referred to as a free or open learning method, whereby the training data is used without reference information, but with the respective reference as input and / or as a test value.
[0040] Furthermore, the present invention provides a computer program product comprising instructions which, when executed by a computer, cause the computer to perform the generation method according to the present invention and / or the learning method according to the present invention. Thus, a computer program product according to the invention provides the same advantages as those explained in detail with reference to a learning method according to the invention and with reference to a generation method according to the invention.
[0041] Further advantages, features, and details of the invention will become apparent from the following description, which describes embodiments of the invention in detail with reference to the drawings. The features mentioned in the claims and in the description may be essential to the invention individually or in any combination. They show schematically:
[0042] Fig. 1 shows an embodiment of an X-ray diffraction device,
[0043] Fig. 2 shows an embodiment of a method according to the invention,
[0044] Fig. 3 shows a further embodiment of a production method according to the invention,
[0045] Fig. 4 shows a further embodiment of a production method according to the invention,
[0046] Fig. 5 shows a further embodiment of a production method according to the invention,
[0047] Fig. 6 shows an embodiment of a learning method according to the invention,
[0048] Fig. 7 an application situation in the classification of objects.
[0049] Figure 1 schematically shows an X-ray diffraction device 100, in which a detailed analysis of an object O in a piece of luggage G is to be performed. This X-ray diffraction device 100 can be part of a baggage handling system, for example, at an airport, and is in particular, as shown by an example in Figure 7, downstream of a transmission device 200. If, during operation, an object O is detected in a piece of luggage G, which, like 100% of all other pieces of luggage G, passes through a transmission device 200, and is defined as an alarm object because it may contain hazardous material, this piece of luggage G is subjected to further analysis. Only those pieces of luggage G, i.e., pieces of luggage G with objects O that have been predefined as alarm objects, are selected for further testing, here in an X-ray diffraction device 100.This further analysis results in a scan being performed with the X-ray diffraction device 100, meaning that the X-ray diffraction device 100 is capable of determining a momentum transfer function, in particular for each individual volume element of the piece of luggage G and / or the object O. To evaluate whether these momentum transfer functions of the object O are relevant because they correspond to a hazardous material or not, an artificial intelligence K1 is used, which has been trained on such specific materials using an appropriate learning method. This training, and especially the generation of the necessary training data TD, is explained in more detail below.
[0050] Figure 2 shows how a reference momentum transfer function RI, for example, specifically recorded on an X-ray diffraction device 100, can be used as the starting value for the generation process. A variation simulation VS then changes and varies this reference momentum transfer function RI many times, so that a single reference momentum transfer function RI generates a multitude of different variation momentum transfer functions VI, represented here by dots. While all of these variation momentum transfer functions VI differ from one another and also from the reference momentum transfer function RI, the degree of differentiation is so small that they still retain the same relevance and specificity for the associated material, which represents the material-specific starting value for the reference momentum transfer function RI.In other words, all generated variational momentum transfer functions VI are equally specific to the same material, as is the initial value of the reference momentum transfer function RI. Therefore, all generated variational momentum transfer functions VI can be saved as training data TD with the same reference information REF, here the material M1 as an example. Figure 2 shows how a very large number of, for example, several 100,000 different variational momentum transfer functions VI can be generated from a single input parameter in the form of a single reference momentum transfer function RI, which, however, are all consistently specific to the initial material that forms the basis of the reference momentum transfer function RI. Figure 3 is a further development of the embodiment in Figure 2. Here, the variational simulation VS is combined with a statistical distribution function STV, for example in the form of a Gaussian curve.This is intended to ensure that the degree of redundancy of the generated variational momentum transfer function VI is reduced. This statistical distribution function STV can be integrated into the variational simulation VS to generate, output, and / or store only non-redundant variational momentum transfer functions VI in the training data TD.
[0051] Figure 4 also shows a further development of the embodiments of Figures 2 and 3. Here, the variation in the variation simulation VS is provided and / or supplemented by a simulation model SM, which is based on a simulation of a scan run in an X-ray diffraction device 100. Additionally or alternatively, variation parameters VP can be adjusted either automatically and / or manually to ensure an even greater degree of variation while maintaining the same specificity for the material of the reference momentum transfer function RI.
[0052] The embodiment of Figure 5 uses an artificial variational intelligence (VKI) as part of the variational simulation (VS). This can also be based on training data obtained from the X-ray diffraction device 100. Additionally, it is shown here that a different reference momentum transfer function (RI) is now specified for a different, second material, for which the same generation process is performed iteratively, so that now training data with reference information (REF) is generated not only for a first material, here M1, but also for a second material, here M2. The training data (TD) correlate with specific materials, so that all training data (TD) together form a reference database (RD).
[0053] Figure 6 schematically illustrates how the training of the artificial intelligence Kl can be carried out using the training data TD and, in particular, a reference database RD. Existing machine learning methods, so-called deep learning algorithms or similar, can be used, with the decisive advantage of the invention being directed at the provision of the artificially generated and specific training data TD. Figure 7 again illustrates the use of the artificial intelligence Kl, as already explained at the beginning.
[0054] The above explanation of the embodiments describes the present invention exclusively by way of examples. Of course, individual features of the embodiments can be freely combined with one another, provided they are technically feasible, without departing from the scope of the present invention.
[0055] List of reference symbols
[0056] 100 X-ray diffraction device
[0057] 200 transmission device
[0058] 0 object
[0059] G piece of luggage
[0060] RI reference impulse transfer function
[0061] VI Variational momentum transfer function
[0062] VS variation simulation
[0063] VP variation parameters
[0064] VKI artificial variational intelligence
[0065] STV statistical distribution function
[0066] SM simulation model
[0067] TD training data
[0068] RD Reference Database
[0069] REF Reference information
[0070] Small artificial intelligence
Claims
Patent claims 1. Generation method for artificial training data (TD) for carrying out a machine learning procedure in training an artificial intelligence (Kl) for use in a classification procedure of objects (0) in luggage (G) based on a scan with an X-ray diffraction device (100), comprising the following steps: - Selection of at least one material-specific reference impulse transfer function (RI), - Applying a variational simulation (VS) to the selected reference impulse transfer function (RI) to generate a variety of variational impulse transfer functions (VI), - Storing the generated multitude of variation impulse transfer functions (VI) as artificial training data (TD) with the selected reference impulse transfer function (RI) as reference information (REF).
2. Generation method according to claim 1, characterized in that the variational simulation (VS) has a statistical distribution function (STV) for avoiding redundant variational momentum transfer functions (VI).
3. Production method according to one of the preceding claims, characterized in that the variation simulation (VS) comprises a simulation model (SM) which simulates at least part of a scan procedure using the X-ray diffraction device (100).
4. A production method according to one of the preceding claims, characterized in that the variation simulation (VS) has at least one variation parameter (VP) which is varied during application and in particular represents at least one of the following real factors when performing a scan with an X-ray diffraction device (100): - Mixing with other materials - Poor demarcation from other objects and / or segments - Distortion of the momentum transfer axis - Photon sampling generation method according to one of the preceding claims, characterized in that the steps of application and storage are repeated for at least one further material-specific reference momentum transfer function (RI). Generation method according to claim 5, characterized in that, based on the execution for different reference momentum transfer functions (RI), the artificial training data (TD) are stored in a reference database (RD) for later application in a learning process. Generation method according to one of the preceding claims, characterized in that the variation simulation (VS) has a termination criterion regarding the number of variation momentum transfer functions (VI) to be generated.A generation method according to any of the preceding claims, characterized in that the variation simulation (VS) comprises an artificial variation intelligence (AI) which is used in the generation of the plurality of variation momentum transfer functions (VI). A generation method according to any of the preceding claims, characterized in that the reference momentum transfer function (RI) has been generated using the same X-ray diffraction device (100) for which the artificial intelligence (AI) is to be trained for the classification method. A generation method according to any of the preceding claims, characterized in that a part of the variation simulation (VS) is manually adjusted. Learning method for training an artificial intelligence (Kl) for use in a classification procedure for classifying objects (0) in luggage (G) based on a scan with an X-ray diffraction device (100), comprising the following step: - Performing machine learning for artificial intelligence (Kl) using artificial training data (TD) generated by a generation method with the features of any one of claims 1 to 10. A learning method according to claim 11, characterized in that the variational momentum transfer functions (VI) for a material to be detected are used as artificial training data (TD). A learning method according to one of claims 11 or 12, characterized in that the variational momentum transfer functions (VI) that differ from those of a material to be detected are used as artificial training data (TD). A learning method according to any of the preceding claims, characterized in that the artificial training data (TD) are used without the reference information (REF).Computer program product comprising instructions which, when the program is executed by a computer, cause the computer to perform the generation method with the features of one of claims 1 to 10 and / or the learning method with the features of one of claims 11 to 14.