EP3639199A1

EP3639199A1 - Method for rating a state of a three-dimensional test object, and corresponding rating system

Info

Publication number: EP3639199A1
Application number: EP19756319.0A
Authority: EP
Inventors: Stefan SUWELACK; Markus Stoll; Steffen SLAVETINSKY
Original assignee: Renumics GmbH
Current assignee: Renumics GmbH
Priority date: 2018-06-29
Filing date: 2019-06-28
Publication date: 2020-04-22
Also published as: US20210124982A1; DE102019209553A1; WO2020001711A1; US11954923B2

Abstract

A method for rating a state of a three-dimensional test object by taking into consideration a prescribed assessment task and by using a rating system comprising a neural network, wherein a training data record is provided and/or used that comprises multiple state data points from one or more three-dimensional training objects, wherein the training data record comprises a known state rating in regard to the prescribed assessment task for each of the state data points, wherein the rating system is adapted to the prescribed assessment task by parameterising the neural network of the rating system by using the training data record in a training process, and wherein an execution process involves the adapted rating system being used to calculate a state rating for a prescribed state data point of the test object. In addition, a corresponding rating system and a computer program product are disclosed.

Description

METHOD FOR EVALUATING A CONDITION

A THREE-DIMENSIONAL TEST OBJECT AND APPROPRIATE EVALUATION SYSTEM

The invention relates to a method for evaluating a state of a three-dimensional test object.

Furthermore, the invention relates to an evaluation system for evaluating a state of a three-dimensional test object.

Finally, the invention relates to a corresponding computer program product.

It is known from practice that in many areas of the product life cycle, such as in development or maintenance, the precise assessment of the physical condition of a mechanical or technical system is an important criterion for engineering decisions. These decisions include, for example, the design of safety margins or maintenance intervals. Measurements or simulations create a digital image of the mechanical or technical system (digital twin) on which the system state is given as a discretized scalar or vector field, for example in the form of a temperature, pressure or voltage distribution.

In practice, the physical states of a mechanical system are largely assessed manually by experts. The assessment is based on the personal experience of the expert. During the assessment, a tensor field is typically represented by suitable computer-based visualization programs. In this regard, reference is made, for example, to the non-patent literature “Kratz, A., Schoeneich, M., Zobel, V, Burgeth, B., Scheuer mann, G., Hotz, /. and Stommel, M. (2014, March): Tensor visualization driven mechanical component design. In Visualization Symposium (PacificVis), 2014 IEEE Pacific (pp. 145-152). IEEE ”. In addition to the tensor field, additional information such as time stamps or simple numerical data, for example frequencies, is often used for the assessment. In the However, manual assessment is particularly disadvantageous in that it is very time-consuming and heavily dependent on the personal and subjective experience of the expert. The manual assessment is also susceptible to human volatile errors and is difficult to reproduce. Fig. 1 illustrates the manual assessment by an expert, which in practice is a typical procedure for the assessment of mechanical systems.

In order to counter the described disadvantages of manual assessment, it is desirable to automate the assessment process. For this, it is necessary to map the application-specific human expert knowledge on the computer. To this end, research has been carried out in the field of knowledge-based engineering (KBE) since the 1990s. For example, refer to the non-patent literature “Verta gen, W J., Bermell-Garcia, P., van Dijk, R. E. and Curran, R .: A criticai review of Knowledge-Based Engineering: An Identification ofresearch chaiienges. Advanced Engineering informatics 26 (2012), 5-15'Ne riesen. The basic idea is to describe the corresponding decision paths using rules and ontologies. In order to be able to apply these methods to sub-symbolic data such as tensor fields, application-specific features such as statistical variables within regions or components must first be extracted. 2 illustrates an automatic assessment of the system status of a mechanical system, whereby in some cases the definition of application-specific features enables rule-based automation.

Although such rule-based automation solutions can be very valuable in individual cases, they also have considerable disadvantages: The formulation of application-specific rules and features is extremely complex and the knowledge mapped in this way can usually only be transferred poorly or not at all to other applications become. Furthermore, the accuracy lags far behind that of human experts.

In comparison to the rule-based KBE, methods are also known from practice which carry out an automatic assessment of the system status on the basis of the definition of features and machine learning processes. However, this is problematic table that in applications on system states which are given as tensor fields, application-specific features for dimension reduction still have to be extracted first. 3 illustrates the automatic assessment of the system status of a mechanical system using simple machine learning methods, the machine learning methods requiring the definition of application-specific features for assessing system statuses. Practical methods use complex methods such as spectral analyzes, diffusion maps or stochastic methods to extract the application-specific features. For example, refer to the non-patent literature "Garcke, J., Iza-Teran, R., (2017): Mach ine Learning Approaches for Data from Car Crashes and Numerical Car Crash S / mu / at / ons, NAFEMS 2017" and "Martin , /., & Bestie, D. (2018): Automated eigen mode Classification for airfoiis in the presence of fixation uncertainties. Engineering Applications of Artificial Intelligence, 67, 187-196 ”. However, the main disadvantages of these approaches are the high effort involved in identifying suitable application-specific features (feature engineering) and the significantly lower level of accuracy compared to that of the human expert.

The interpretation and classification of the respective system status of a mechanical or technical system is currently usually the responsibility of human experts who make a decision based on their experience based on a visualization of the system status. This approach is subjective and prone to errors. Previous approaches known in practice for automating such classifications are based on rule-based approaches. Predefined characteristics are extracted (for example maximum value, mean value, histograms), which are then classified using a simple method (for example threshold value method). However, these approaches are highly problem-specific, difficult to maintain, and do not achieve the accuracy of human experts.

The present invention is therefore based on the object of designing and developing a method for evaluating a state of a three-dimensional test object of the type mentioned at the outset such that an improved Status assessment, especially with regard to efficiency and / or accuracy, is possible. Furthermore, a corresponding evaluation system and a corresponding computer program product are to be specified.

According to the invention, the above object is achieved by the features of claim 1. Thereafter, a method for evaluating a state of a three-dimensional test object is given, taking into account a predefined assessment task and using an evaluation system comprising a neural network, wherein a training data set is provided and / or used, which contains several state data points from one or more comprises three-dimensional training objects, with the training data record relating to the predefined assessment task comprising a known status assessment for each of the status data points, the neural network of the assessment system using the training data record being parameterized in a training process in order to adapt the assessment system to the predefined assessment task - is standardized, and a status assessment is calculated in an execution process with the adapted assessment system for a given status data point of the test object.

The above object is further achieved by the features of claim 17. According to this, an evaluation system for evaluating a state of a three-dimensional test object is specified taking into account a predetermined evaluation task, the evaluation system comprising a, preferably deep, neural network, the evaluation system being designed in such a way that a training data set can be used which contains several state data points from one - comprises one or more three-dimensional training objects, the training data record comprising a known status evaluation for each of the status data points in relation to the predefined assessment task, the assessment system being further designed such that the neural network is adapted to the given assessment task can be parameterized in a training process using the training data set, and that after the adaptation to the predefined assessment task in an execution process for a predefined status data point of the test object, a Z. inventory assessment is predictable. Finally, the above object is achieved by the features of claim 18 relating to a computer program product.

In the manner according to the invention, it was first recognized that it is of considerable advantage if, as part of an automatic assessment of the states of a test object, a definition or identification of application-specific features and rules, which usually takes place in advance, can be largely avoided. According to the invention, it has also been recognized that by using a neural network, in particular by using a deep neural network, automation can be brought about which, after carrying out a training process, enables a state of the test object to be assessed fully automatically or is assessable. As soon as the neural network has been adapted or trained for a specific predefined assessment task, the neural network can recognize and evaluate the state of predefined input data, namely the state data point of the test object, on the basis of previously learned states in the training process on the basis of the training data set , For this purpose, the condition evaluation for the condition data point of the test object is calculated by a suitably trained neural network.

The present invention therefore proposes a method in which a state of the three-dimensional test object is assessed taking into account a predetermined assessment task and using an assessment system which comprises a neural network. According to the invention, a training data record is initially provided and / or used, which comprises several state data points from one or more three-dimensional training objects. The training data record includes a known status evaluation for each of the status data points in relation to the specified assessment task. To adapt the evaluation system to the specified assessment task, the neural network of the evaluation system is then parameterized using the training data set in a training process. With the neural network adapted to the predefined assessment task, the adapted or trained assessment system can then calculate an assessment of the condition for a given status data point of the test object in an execution process. Consequently, the method according to the invention enables improved condition assessment, in particular with regard to efficiency and / or accuracy.

In contrast to approaches from the prior art, which are highly problem-specific, difficult to maintain and do not achieve the accuracy of human experts, the invention and / or advantageous refinements of the invention enable a completely data-driven teaching-in of the evaluation system, so that the Evaluation system can evaluate or assess the state of a three-dimensional test object. The basis for this is a training data record which provides a problem-specific set of states with a known state evaluation. The information can then be used to teach the assessment system to the specified assessment task, whereby a generic assessment algorithm of the assessment system, which is preferably based on a deep neural network, is trained with the training data of the training data set. In contrast to conventional approaches known from the state of the art, no features selected a priori are required for this.

At this point it should be pointed out that the terms “three-dimensional test object” and “three-dimensional training object” - in particular within the scope of the claims and preferably within the description - can be understood in the broadest sense as three-dimensional, preferably objective, objects, which, depending on the construction of the object and / or assume various states due to the properties inherent in the object. In the context of the invention and / or in the context of advantageous embodiments of the invention, the test object or the training objects can be provided, simulated and / or used in the form of a digital twin or in the form of a digital virtual model. The digital twin can be generated by measurements and / or by simulations, namely by generating a digital image of the three-dimensional object, the state being given as a discretized tensor field on the digital image. Furthermore, the three-dimensional test object or the three-dimensional training objects can be designed as digital virtual CAD models on which the state is given as a discretized tensor field. Furthermore, it should be noted that the expression “assessment task” - in particular within the scope of the claims and preferably within the scope of the description - is to be understood in the broadest sense, namely as possible assessments and / or assessments of states that relate to a three-dimensional Test object or in relation to a three-dimensional training object can be made.

With regard to the expression “condition assessment”, it should be pointed out - in particular within the scope of the claims and preferably within the description — that a condition assessment can be understood to mean the result of the assessment task for a given condition data point.

With regard to the expression “discretization”, it should be noted that in physical reality, taking into account a macroscopic view, there may be continuous distributions of measured values. Since only a finite number of values can be stored in the computer / computer when creating a digital twin or a digital virtual model, a discrete subset of the potentially infinite number of measured values can therefore advantageously be formed (discretization).

With regard to a “deep neural network”, it should be pointed out - in particular in the context of the claims and preferably in the context of the description - that a deep neural network - compared to a conventional or simple neural network - is distinguished by a larger number distinguished on layers. A deep neural network can advantageously comprise at least six layers.

In order to be able to train a deep neural network in a suitable manner, special mathematical formulations or calculation rules may be required. A convolutional network, that is to say a “convolutional neural network”, can therefore advantageously be provided as the deep neural network. With regard to further details on convolutional neural networks and the functional principle of deep neural networks, reference is made to the non-patent literature “LeCun, Y, Bengio, Y, & Hinton, G. (2015): Deep learning. nature, 521 (7553), £ 3 ”. The evaluation system can advantageously be implemented as a software system on a powerful computer. In a further advantageous manner, the computer could have highly parallel computing units in the form of powerful graphics units. The evaluation system can be modular. A data module can enable the administration and / or use of the training data. The neural network could be parameterized via a training module and the parameterized neural network could finally be used for automatic assessment via an evaluation module. Furthermore, different interface modules can enable easy access to the evaluation system, for example via web-based protocols.

The three-dimensional test object can advantageously comprise a component, a mechanical system, an electromechanical system and / or an electrochemical system. A constructed technical system can be regarded as a system here, which can consist of several parts and components. For example machines, plants, motor vehicles, etc. and / or parts thereof. This means that the states of a wide variety of technical systems can be evaluated efficiently.

In a further advantageous manner, the training objects of the training data set can be selected problem-specifically in such a way that the predetermined assessment task can be applied to the training objects and to the test object. The training objects and the test object are thus related to one another in such a way that the specified assessment task is defined both for the training objects and for the test object. As a result, the neural network of the evaluation system can be effectively trained for the assessment task with the correspondingly selected training data set or with the status data points and the associated known status assessments. A trained / adapted neural network enables the calculation of a condition assessment by the assessment system based on a predetermined condition data point of the test object. In an advantageous embodiment, the state data points of the training objects and the state data point of the test object can each comprise spatial state data. The spatial status data can advantageously be data which describe and define the status of the test object or of the training object. This means that the spatial status data represent the status of the three-dimensional object to be assessed. The spatial status data can be represented in the form of a spatial status field, the status data being distributed in three-dimensional space according to a predefined pattern or according to a predefined arrangement. Accordingly, the spatial state data with scanning points can span a three-dimensional scanning space - possibly around the test object or around the training object, the scanning points in the scanning space being corresponding to the three-dimensional object (ie test object and / or training object) a predefined arrangement are distributed. The spatial state data indicate the sample values prevailing at the sample points - for example physical measured values and / or sample values calculated on the basis of simulation values - in relation to the three-dimensional object. Efficient processing of the spatial status data can thus be implemented by the neural network of the evaluation system, and completely data-driven teaching of the evaluation system can be implemented.

The spatial state data comprised by a state data point of a training object can advantageously define a state of the training object. Furthermore, the spatial state data included in the state data point of the test object can define a state of the test object. In this way, a state of the three-dimensional object (ie test object and / or training object) can be defined with available or provided spatial state data, which must be evaluated or determined taking into account the specified assessment task. In order to provide the training data record, the “known” condition assessment belonging to the spatial condition data of a condition data point can be determined manually by an experienced expert. Furthermore, the one to be determined in advance Known condition assessment also takes place by evaluating historical data. A training data record can thus be provided with which the neural network of the assessment system can be taught in in an efficient manner.

In an advantageous embodiment, the spatial status data of a status data point can be represented as a discretized scalar, vector and / or tensor field. A tensor can represent a sample value - such as a measured value and / or a sample value calculated on the basis of simulation values - at a sample point. The tensor field can then include a predeterminable number of spatially distributed samples represented as a tensor with reference to the respective sampling points. A tensor can be seen as the generalization of scalar values (tensor 0th level), vectors (tensor 1st level) or matrices (tensor 2nd level). A particularly advantageous representation of the spatial status data can thus be realized, which enables the spatial status data to be processed efficiently by the neural network of the evaluation system.

In an advantageous embodiment, the spatial status data of a status data point can be converted into the form of a uniform grid. The grid can advantageously be a numerical grid, in particular a Cartesian grid. This enables efficient processing of spatial status data by the neural network of the evaluation system.

In an advantageous embodiment, the spatial status data of a status data point can be represented as a tensor of level N + 3, where N is the order of the status data per grid point. With this form of representation, the spatial state data in the form of a uniform Cartesian grid can be analyzed particularly efficiently by means of a neural network, preferably a deep neural network. In other words, N here is the order of the discretized tensor field of the input data for the neural network of the evaluation system. For a scalar field like a temperature field wise N = 1. For a voltage field, N = 2 applies. If the status data is also resolved in time, the dimension increases by an additional 1. In a particularly advantageous manner, the spatial status data - for example measurement and / or simulation data - of a status data point is transferred to this representation, so that efficient processing is made possible by the neural network of the evaluation system.

In an advantageous embodiment, the spatial status data of a status data point can represent a physical quantity. The spatial status data of the status data point can thus define a physical status of the test object or of a training object. The spatial status data of a status data point of the test object can then be analyzed by the evaluation system, so that the status evaluation is calculated / ascertained on the basis of the underlying spatial status data for the status data point.

In an advantageous embodiment, the spatial status data of a status data point can include a temperature field, a deformation field, a speed field, a voltage field, a pressure field, a displacement field and / or an electromagnetic field. The respective state field is advantageously present in a discretized form. A wide variety of properties or corresponding states of a three-dimensional test object to be examined can thus be efficiently evaluated using the spatial state data of a state data point and taking into account the specified assessment task.

In an advantageous embodiment, the spatial status data of a status data point can represent a digitization status of the respective three-dimensional object. The spatial state data of a state data point can preferably include the local resolution of a numerical grid and / or a quality feature of a numerical grid. For example, the equilibrium of triangles could be considered as a quality feature of the numerical grid. A digitization state of a digitized three-dimensional test object can thus be evaluated in an efficient manner. The state data points of the training objects and the state data point of the test object can advantageously include at least one additional parameter in addition to the spatial state data. In this way, further parameters which - possibly - have an influence on the state of the three-dimensional object can be taken into account when determining or calculating the state evaluation for a state data point. This provides more precise results.

In an advantageous embodiment, the training process can include a parameter optimization step, with a status evaluation being calculated by the evaluation system for each state data point of the training data set in the parameter optimization step. The known status evaluation of the training data record belonging to the respective status data point can be compared with the calculated status evaluation, so that network parameters of the neural network are adapted as a function of the comparison. The network parameters are the weights of the neural network of the evaluation system, which are required for the calculations of the neural network. By changing or adapting the network parameters or the weights of the neural network, the generic evaluation algorithm of the evaluation system can be adapted and trained for the specified evaluation task. This enables an efficient calculation of the condition evaluation for a condition data point of the test object.

The parameter optimization step can advantageously be carried out several times, possibly with different training data sets. A more precise evaluation of a state of a test object can thus be realized.

In an advantageous embodiment, application-specific network parameters of the neural network of the evaluation system can be determined in the training process by back propagation, the training data set and initial network parameters (ie initial weights) of the neural network being used as input data for back propagation. The initial network parameters / weights can be understood here as network parameters that are used before a first trai- initial process. Thus, an efficient calculation of the condition evaluation for a condition data point of the test object can be made possible

The neural network of the evaluation system can advantageously be configured in such a way that the spatial state data and the at least one additional parameter are processed separately from one another in a plurality of folding layers before the respectively calculated intermediate results are brought together again and by further layers of the neural Network are processed. A particularly efficient processing of the input data is thus implemented by the evaluation system, whereby an efficient and precise calculation of the state evaluation for a state data point is made possible.

Advantageous embodiments of the invention can evaluate or assess the condition of a three-dimensional test object - such as a mechanical system - on the basis of spatial condition data, in particular on the basis of spatial tensor fields. Advantageous embodiments of the invention can be implemented and categorized according to the type and origin of the underlying status data:

- Spatial status data based on physical measurement data of a three-dimensional test object: Using various measurement methods such as ultrasound, depth cameras, multi-camera systems and / or sensor arrays, spatial status data of mechanical components and / or technical systems can be recorded as test objects. For example, temperature and / or deformation fields are conceivable as a basis for spatial state data.

- Spatial state data based on numerical simulation data of a test object: Physical states of a test object can be simulated using numerical simulation methods and made available in high resolution. With this approach, a wide variety of states can to be examined. For example, here, among other things, voltage, displacement and / or pressure fields are conceivable as the basis for spatial status data.

- Spatial status data based on the digitization of a three-dimensional test object: When digitizing a test object (for example a mechanical component or system) and / or with physical measurement data in relation to the test object, errors can occur under certain circumstances. The assessment of the digitization status is therefore essential in order to estimate these errors. For example, the local resolution and / or a quality feature of the numerical grid are conceivable as the basis for spatial state data.

Advantageous embodiments of the invention can have at least one of the following advantages:

- Improving performance: In particular during product development, the assessment or assessment of the physical status of test objects, such as mechanical systems, primarily serves to increase the performance of the test object. For example, the reduction in weight, the reduction of noise emissions, the increase in energy efficiency, the reduction in air resistance and / or the increase in stability in the event of accidents can be taken into account and improved.

- Improvement of the lifespan: During the product development, the expected lifespan can be estimated above all with simulation studies, for example for the distribution of power or heat. During the use of a test object or a mechanical system, additional measurement results can be used for this purpose.

- Increased security: An analysis of the test object or the technical system, for example using ultrasound or depth cameras, enables the generation of spatial status data, for example in the form of Tensor fields. The automatic assessment and evaluation of this status data enables the operational safety of the test object or the system to be monitored automatically. An automatic assessment / evaluation of digitization states of the test object is also helpful for monitoring the simulation process during product development.

There are now various possibilities for advantageously designing and developing the teaching of the present invention. For this purpose, reference is made on the one hand to the claims subordinate to claim 1 and on the other hand to the following explanation of preferred exemplary embodiments of the invention with reference to the drawing. In connection with the explanation of the preferred exemplary embodiments of the invention with reference to the drawing, generally preferred configurations and developments of the teaching are also explained. Show in the drawing

1 is a schematic view of an illustration of a manual assessment of a mechanical system by an expert,

2 shows a schematic view of an automatic assessment of the state of a mechanical system, where rule-based automation is possible by defining application-specific features,

3 shows in a schematic view an illustration of an automatic assessment of the state of a mechanical system using simple machine learning methods, the machine learning methods requiring the definition of application-specific features for assessing system states,

4 shows a schematic view of a method for evaluating a state of a three-dimensional test object according to an exemplary embodiment of the invention, the state of the test object being automatically assessed, 5 shows in a schematic view an illustration of a training process according to an embodiment of the invention, the neural network being adapted by training for a predetermined assessment task or for a specific application,

6 shows a schematic view of a training process for adapting a neural network of an evaluation system according to an exemplary embodiment of the invention,

7 shows a schematic view of a network architecture of a deep neural network for a rating system according to an exemplary embodiment of the invention, the deep neural network being designed as a convolutional neural network,

8 shows a schematic view of a method for evaluating a state of a three-dimensional test object according to an exemplary embodiment of the invention, the state of the test object being automatically assessed,

9 is a schematic view of an illustration of a temperature field as spatial state data of an accumulator pack for a method according to an embodiment of the invention,

10 is a schematic view of an illustration of a voltage field of an angle calculated by numerical simulation for a method according to an embodiment of the invention,

11 shows a schematic view of the angle according to FIG

10, but without the polygonal grid and with the sampling points shown on the angular geometry, and

12 shows a schematic view of an illustration of a method for evaluating a state of a mechanical system as three-dimensional test object according to an embodiment of the invention.

1 shows in a schematic view an illustration of a manual assessment by a human expert as a typical procedure for evaluating a mechanical system. The system state is specified by a tensor field, which is typically visualized to the expert by a suitable computer-based visualization program. Additional information about the mechanical system to be analyzed is also included, so that the expert ultimately has to assess or assess the system status based on the information and data visualized using his subjective experience.

FIG. 2 shows in a schematic view an illustration of an automatic assessment of the system status of a mechanical system, wherein - in contrast to the procedure according to FIG. 1 - rule-based automation is possible by the definition of application-specific features.

FIG. 3 shows in a schematic view an illustration of an automatic assessment of the system status of a mechanical system, simple machine learning methods being used in addition to the procedure according to FIG. 2. The machine learning processes require the definition of application-specific features to assess the system status of the mechanical system.

4 shows in a schematic view an illustration of a method for evaluating a state of a three-dimensional test object according to an exemplary embodiment of the invention, the state of the test object being assessed automatically. The exemplary embodiment illustrated in FIG. 4 is a method based on a deep neural network, which can be used for the automatic assessment of states of technical systems and / or components. System states of the test object can be automatically be judged. The condition to be assessed of the three-dimensional test object to be examined is specified by spatial condition data in the form of a tensor field. Furthermore, one or more additional parameters can be provided as additional information. The tensor field and any additional parameters that may be provided are analyzed by the trained or adapted deep neural network, so that the status of the test object is evaluated automatically.

The method according to the exemplary embodiment shown in FIG. 4 can include the following:

- Deployment phase: generation or compilation of an application-specific training data record;

- Training process: Training of the neural network to determine the application-specific network parameters (weights of the neural network);

- Execution process: automatic recognition and assignment of the previously learned states (estimation) by calculating the state evaluation for a predefined state data point;

In an initial phase, an application-specific training data set with several status data points is created. Each state data point (sample, training example) of the training data record comprises spatial state data which represent a training example. Furthermore, each status data point can include one or more additional parameters. Finally, the training data record includes a known status evaluation for each of the status data points.

In the context of a training example, the spatial status data provide the basis for describing or defining a status of a three-dimensional training object. The additional parameters are optional additional information that can also influence the state of the training object. The spatial status data, if appropriate, together with one or more State parameters thus form the basis for the state of the respective state data point. The well-known condition assessment is based on the spatial condition data and - if available - the optional additional parameters. This known status evaluation of the training examples or the training data set can be determined manually by experienced human experts. Ideally, training examples including such manual (known) status assessments for a problem or for an application are already available, since appropriate assessments have already been carried out manually in the past.

5 shows in a schematic view an illustration of a training process according to an exemplary embodiment of the invention, the neural network being adapted by means of training for a predetermined assessment task or for a specific application with regard to a three-dimensional test object. After the training process, the neural network can automatically assess or assess the system states of the test object.

In the training phase, the network parameters, i.e. the weights of a neural network are adjusted on the basis of a training data set that includes application-specific training data. The neural network is thus adapted or parameterized by the training for a specific application, taking into account a specified assessment task.

If necessary, the training process can be carried out with different training data, based on the trained weights of a previous training process, so that subsequent training can be carried out.

As soon as the neural network has been adapted or trained for the specified assessment task, it can determine the status of a test object in an execution process on the basis of specified or provided input data - spatial status data, if necessary with one or more additional parameters - and on the basis beforehand in the training phase Recognize and evaluate learned states. The result of this condition assessment can be, for example, a numerical value (regression) or the assignment to a predefined class (classification). 6 shows in a schematic view an illustration of a training process for adapting a neural network of an evaluation system according to an exemplary embodiment of the invention.

The evaluation system comprises a component for converting the spatial state data, which belongs to the input data, into a data format based on a uniform grid, which can be processed efficiently and easily with neural networks.

As a further component, the evaluation system comprises a modular network architecture based on a convolutional neural network (ConvNet), which allows an efficient status assessment and optionally allows the processing of one or more additional parameters as additional information. Furthermore, the modular network architecture of the evaluation system enables the neural network to be retrained. With regard to further details on convolutional neural networks and the functional principle of deep neural networks, reference is made to the non-patent literature “LeCun, Y, Bengio, Y, & Hinton, G. (2015): Deep tearning. nature, directed.

The evaluation system of the embodiment shown in FIG. 6 is designed in such a way that a training process and an execution process can be carried out, so that it is possible to evaluate the state of the test object using the neural network architecture of the evaluation system. In the execution process, the condition assessment for a condition data point of the test object is calculated by the assessment system trained for the specified assessment task.

6 schematically shows the structure of the training process. Training examples, which are provided in the form of the training data record, serve as input. Depending on the application, a large number of training examples may typically be required. For example, this could be several hundred thousand training examples. The following entries are used for each training example in the course of the training process: - Spatial condition data

- One or more additional parameters (optional)

- Initial network parameters for the neural network

- Known condition assessment for the raining example

The output or the result of the training process are the updated, application-specific network parameters, so that the neural network or the evaluation system is adapted to any test object, taking into account the specified assessment task.

As part of the data processing, the spatial state data are converted into a form that can be processed effectively and effectively for neural networks. For this purpose, for example, the conversion of the spatial state data into a tensor field on a Cartesian grid lends itself. Before the first training, ie before the training process, the network parameters (weights) of the neural network can be initialized on the basis of an initialization function.

The calculations or “estimates” of the neural network are then compared with given expert assessments, namely the known status assessments, and an error calculation is carried out. A parameter optimization is then carried out. In a suitable manner, this parameter optimization step can be carried out until a termination condition is fulfilled. As a termination condition, for example, falling below a predetermined error limit or reaching a predetermined — possibly maximum — number of optimization steps is conceivable.

As part of the data preparation, the spatial state data are processed in such a way that they are represented in the form of a uniform, preferably a Cartesian, grid. The spatial state data can thus be effectively analyzed using deep neural networks.

This form of representation can be interpreted as a tensor of level N + 3, where N is the order of the tensor field of the input data (for a scalar field like a Temperature field then applies, for example, N = 1, for a voltage field N = 2, if the data is also temporally resolved, the dimension increases by 1). In an advantageous application, the input data, for example measurement or simulation data, are first to be converted into this representation.

Spatial state data, which are obtained from measurement results, are usually represented on the basis of point clouds. In order to convert such a tensor field into a uniform lattice, the sampling points located in a lattice cell can, for example, be summarized by averaging the respectively assigned tensors. Other possibilities known to those skilled in the art from practice are also conceivable.

Spatial state data obtained from numerical simulations are usually represented by tensor fields based on unstructured, polygonal networks. In order to convert such a tensor field into a uniform grid, the polygonal network can, for example, be scanned at the centers of the grid cells of the uniform grid. Other possibilities known to the specialist from practice are also conceivable.

Especially with large mechanical systems as test objects, the presentation of spatial status data in the form of uniform grids requires a lot of storage space. Measures can therefore advantageously be taken to reduce the memory consumption. It may be advisable to examine only a small part of the test object separately, for example by dividing the grid into different regions. Furthermore, the use of adaptive data structures for storing the spatial state data on the uniform grid is conceivable. Various data structures known from the practice of computer graphics are conceivable for this, such as the use of octrees. An octree is a rooted tree, the nodes of which have either eight direct successors or no successors at all. Three-dimensional data sets can thus be subdivided hierarchically in a suitable manner. 7 shows a schematic view of a network architecture of a deep neural network for a rating system according to an exemplary embodiment of the invention, the deep neural network being designed as a convolutional neural network.

The input data for the deep neural network are, on the one hand, the spatial status data of the test object and, on the other hand, one or more additional parameters of the test object, such as weight, frequency information, material types, etc. The spatial status data lie as a tensor field on a Cartesian Grid in front.

For the implementation of deep neural networks, various architectures are conceivable that are suitable for the assessment of the input data. FIG. 7 shows a possible embodiment of a rating system based on a Convolutional Neural Network (ConvNet). First, the additional parameters and the spatial status data are processed separately from each other in several folding layers - each with an optional subsequent normalization, activation and pooling layer - before the results are merged and processed by further layers. The spatial state data is already available here in a processed form, namely as a tensor field on a Cartesian grid. Furthermore, the exact number of folding layers and the special configuration (normalization, activation and pooling layers) can vary depending on the application.

After the convolutional layers of the neural network shown in FIG. 7, one or more completely connected neuron layers can follow, in order to fix the output of the neural network to a specific dimensionality, as is often desired, for example, in the case of a classification. In the case of various applications, final processing can then be carried out using special activation / normalization functions (for example voting). The sum of all network parameters (weights) for each layer determine the behavior of the neural network. These network parameters can be determined in a training process, which is given as an example with the embodiment shown in FIG. 6. In the embodiment according to FIG. 7, a folding layer is implemented as a combination of folding step, activation step, normalization step and pooling step. These steps can be carried out several times in succession in different configurations and sequences within the neural network. Furthermore, several foldings can be carried out, including normalization and activation, in order to then carry out a pooling step.

The name “folding layers” comes from the mathematical operation “folding” of the same name, which is the basis of the folding layer. A convolution (/ * g) of two functions / and g with R ⁿ (C results as follows:

In applications of convolutional networks as a neural network, multidimensional discrete data is usually used. Accordingly, the discrete fold can be used. An input /, for example a temperature field, is then folded in the folding layer on a three-dimensional grid with a folding kernel K in order to produce the output or the result of the folding. The parameters optimized during the training correspond to the parameters of the convolution kernel K. For three-dimensional input data, this results in the discrete convolution using a formula as follows:

S (i,;, / c) here denotes the result of the folding of the input data / instead of i, j, k with a folding kernel K of size pi, h, r.

In order to accelerate the training and / or to carry it out more robustly, the inputs and outputs of different layers within the neural network can be normalized. This enables faster and more robust training. This can be done in the normalization step based on the expected value and the variance of the training data. Activation layers can be used to produce a nonlinear output from the linear output of a fold. Here, for example, the sigmoid function is useful:

Furthermore, only partially differentiable functions such as the Rectified Linear Unit 3 (ReLU3) and 6, such as leakyReLU or random functions such as dropout layer (set randomly selected values of the output to 0) can be used.

The following applies to the ReLU activation function: y = max (0, x)

The following applies to the LeakyReLU activation function:

G x if x> 0

y = 1

ί * x otherwise

Pooling layers are used to summarize the output of previous operations. Here, a summary of a defined neighborhood is generated according to a certain procedure. Max pooling can be used as a strategy for pooling, in which a right-angled neighborhood is combined by choosing its maximum value.

The network parameters of the deep neural network according to the embodiment shown in FIG. 7 can be determined using a back propagation method. Various optimization methods known from practice are conceivable. For example, the stochastic gradient descent or a modified version such as the Adam method can be used. In this case, the entire data set is advantageously not used for optimization, but instead repeats randomly selected subsets of the training data set for the calculation of the gradient. With the help of this gradient, which is calculated on a target function, the necessary adjustment of the parameters is determined. In machine learning, an error function can serve as the objective function, which quantifies the deviation between the prediction of the neural network and the expert opinion, namely the known condition assessment.

Different methods can be used to calculate the error of a prediction, depending on the desired function and error tolerance of the estimator. In the case of a classification, the error can be determined, for example, as the mean square error between input and output or using the cross entropy.

The cross entropy between the annotation y 'and the prediction y is calculated as follows:

With regard to an initialization of the network parameters (weights) of the neural network, these can be initialized with random values before the first training. This can be done in a suitable manner, for example with the "Xavier" initialization function, in which the weights are selected at random from the interval defined as follows:

With this interval, random sampling can be carried out as part of the "Xavier" initialization, the number of input neurons and out corresponding to the number of output neurons. In this case, however, a lot of training data is required before the neural network achieves a good assessment accuracy. Transfer learning offers another possibility: Here, the neural network is initially trained for a related task for which a large amount of training data is available. Then, in further training steps, the neural network is adapted to the specific target application (post-training). Various embodiments are conceivable for this procedure. One possibility may be that the parameters of the completely connected neuron layers are randomly reinitialized before the post-training, but the other parameters are taken from the previous training. In this case, it may make sense to leave the network parameters (weights) of the folding layers constant.

8 shows a schematic view of an illustration of a method for evaluating a state of a three-dimensional test object according to an exemplary embodiment of the invention, the state of the test object being automatically assessed. After the determination of the application-specific network parameters has been completed as part of the training, the evaluation system can be used as part of the execution process for automatic evaluation or assessment of the state of the test object.

9 shows a schematic view of a sketched illustration of a temperature field of an accumulator pack 1 for a method according to an exemplary embodiment of the invention. 9 is intended to illustrate the result of a thermal 3D imaging, wherein the battery pack 1 was photographed from different directions during a charging process with a depth camera and a thermal imaging camera, and a three-dimensional point cloud 2 was subsequently generated by image fusion. 9 shows temperatures in the range between 15 ° C. and 45 ° C. according to the temperature division 3 on the vacuum pack 1. There is then a temperature for each point on the visible surface of the battery pack 1. A spatial state field of the temperature on the surface of the battery pack 1 can thus be created from various 2D thermal images and depth information, in order then to be supplied as spatial state data to an evaluation system according to an embodiment of the invention. The evaluation system can evaluate the state of the battery pack 1 based on the discretized temperature field. Defects in the battery pack 1 can not only reduce the performance of a system driven by the battery pack 1, but also represent a dangerous source of explosion.

So far, human experts have evaluated or assessed such data. Using an evaluation system or performing a method according to an embodiment of the invention, it is now possible to automatically predict or calculate the probability of error and also the type of possible sources of error.

For this purpose, several hundred status data points from training objects are first collected and made available as part of a training data set. Within the scope of the exemplary embodiment relating to the battery pack 1, each state data point of the training data record comprises the following:

- Spatial state data: point cloud with temperature information per sampling point in the scanning space - obtained from sensor data, and

- Additional parameters: age of the pack, current charging current

Furthermore, the training data record comprises a known status evaluation for each of the status data points. The known condition assessment is determined in advance by an experienced expert, whereby the following assessments or condition assessments are distinguished:

- Accumulator pack free of defects

- Accumulator pack with errors, no security risk

- Accumulator pack with errors, security risk

In order to convert the temperature field defined on a point cloud into a uniform grid, the resolution of the grid is first defined (for example 256x256x256 cells). Depending on the size of the battery pack, the size of the grid cells is then calculated by dividing the dimensions by the number of cells in each spatial direction. Then for each cell all Points P_i sought that are in the cell. The temperature value for each grid cell then results from the averaging of the temperature values at points P_i.

Good results can be achieved with various neural network configurations. A particularly advantageous configuration of the neural network is as follows:

- Folding layers for spatial state data: The folding layers use normalization and activation by means of leakyReLu, max-pooling is also carried out every two to three layers. A total of six to nine folding layers.

- Folding layers for the additional parameters: A folding layer including normalization and activation using leakyReLu.

- Additional folding layers: two to three folding layers including normalization and activation using leakyReLu.

- Finally, two completely connected neuron layers and post-processing using Softmax normalization follow.

Initially, the network parameters (weights) of the neural network are initialized using "Xavier" initialization. In the case of existing data, the described post-training strategy is followed.

As described, a three-dimensional point cloud with temperature information per sampling point serves as input for the evaluation system. The age of the battery pack 1 and the current charging current also serve as input for the evaluation system. The three-dimensional point cloud with the temperature information per sampling point forms the spatial state data in the form of a discretized temperature field. The age and the current charging current represent the additional parameters. The evaluation system can then automatically judge the state of the battery pack 1 based on the spatial state data, namely the discretized temperature field, and on the basis of the additional parameters and can distinguish the following cases:

- Accumulator pack free of defects

- Accumulator pack with errors, no security risk

- Accumulator pack with errors, security risk

10 shows in a schematic view an illustration of a voltage field of an angle 4 calculated by numerical simulation for a method according to an embodiment of the invention. The voltage field of the angle 4 is shown as a scalar field (von Mises voltage) on a polygonal grid 5 according to the voltage division 6, the polygonal grid 5 in FIG. 10 being designed as a tetrahedron grid.

An angle should be built as light as possible for a mechanical system without compromising reliability and service life. For this purpose, the voltages (force per surface) are determined in the angle with the aid of a numerical simulation. The result is then available as a tensor field on a polygonal grid. For each node in the grid there is a specific stress tensor (3x3 matrix). For a manual assessment, this tension tensor can be converted into a scalar value (von Mises tension) and visualized for the “assessing” engineer, for example as outlined in FIG. 10.

Based on this visualization, a human expert is currently assessing whether the angle is designed to be stable enough based on his experience. With a method or with an evaluation system according to an embodiment of the invention, it is possible to automatically classify certain patterns in the voltage field of the angle and thus to evaluate the state of the component.

Several hundred state data points are collected to create a training data set, each state data point comprising the following: - Spatial state data: discretized voltage field on a polygonal grid with voltage information per sampling point - obtained from numerical simulation, and

- Additional parameters: Material type used

- Angle correctly simulated, voltage field perfect, good design

- Angle correctly simulated, tension field problematic, construction must be changed

- Angle incorrectly simulated, simulation must be repeated

FIG. 11 shows a schematic view of the angle 4 according to FIG. 10, but without the polygonal grid and with the sampling points 7 shown on the angle geometry of the angle 4.

So that the spatial status data of a status data point can be analyzed by the neural network of the evaluation system, the spatial status data are subjected to data preparation. For this purpose, a scanning is carried out in order to transfer the voltage field defined on the tetrahedron lattice shown in FIG. 10 to a uniform lattice. First, an even grid is defined. Then, for each center point P of a grid cell, it is identified in which tetrahedron T of the original grid P is located. The stress tensor at P is then calculated by linearly interpolating the stress sensors at the four points in the tetrahedron T. Fig. 1 1 illustrates the position of the sampling points on the geometry of the angle.

A particularly advantageous configuration of the neural network of the evaluation system can be carried out as in the network configuration in the exemplary embodiment relating to the battery pack, but no folding layers are used to process additional parameters. Initially, the network parameters (weights) of the neural network are initialized using "Xavier" initialization. In the case of existing data, the described post-training strategy is followed.

As described, the input to the evaluation system is a spatial field data on a voltage field on a polygonal grid, which was calculated using a numerical simulation. The material type of the angle used also serves as additional information for the evaluation system.

The evaluation system can then automatically judge the state of the angle based on the spatial state data, namely the discretized voltage field and on the basis of the additional parameter, and can distinguish the following cases:

- Angle correctly simulated, voltage field perfect, good design

- Angle incorrectly simulated, simulation must be repeated

The voltage state of an angle can thus be automatically assessed as a test object.

12 shows in a schematic view an illustration of a method for evaluating a state of a mechanical system as a three-dimensional test object according to an exemplary embodiment of the invention. To optimize the noise comfort and / or the longevity of mechanical systems - such as a body of an automobile or an aircraft fuselage / wing - the vibration or oscillation behavior of the system, ie the test object, must be analyzed and analyzed optimize. For this purpose, essential vibration forms, such as torsion, are preferably identified in measurement and / or simulation data. The three-dimensional test object or the mechanical system can be evaluated or examined with regard to its vibration behavior with the exemplary embodiment according to FIG. 12. The exemplary embodiment according to FIG. 12 is characterized in that there is a point cloud with associated displacement vectors for a three-dimensional test object, namely a complex mechanical system such as, for example, a motor vehicle body or an aircraft fuselage or an aircraft wing. A three-dimensional displacement field is thus provided as the state data point. This state data point represents a spatial waveform as the state of the test object. Furthermore, it is conceivable that this status data point, together with the assigned excitation frequency, which can optionally be assigned, represents a spatial waveform as the status of the test object.

The spatial state field of the displacement vectors, i.e. the three-dimensional displacement field can be obtained by a dense measuring field on sensors or by simulation. The evaluation system can then evaluate the state of the system with respect to the vibration behavior based on a discretized displacement field. Furthermore, it is conceivable that the evaluation system evaluates the state of the system in relation to the vibration behavior based on a discretized displacement field and the associated excitation frequency.

The state of vibration of the test object or mechanical system in relation to different excitation frequencies is decisive for the noise comfort and the durability of the mechanical system. The evaluation of the individual vibration forms is essential for this, i.e. the assignment of the state data point to physically relevant basic patterns such as torsion. In this way, it can be avoided that these relevant basic vibration forms resonate with known external excitation frequencies and that the mechanical system to be investigated is emitted to a large extent or is destroyed.

So far, evaluations or evaluations of such data have been performed manually by human experts. Using an evaluation system or carrying out a method according to an embodiment of the invention, it is now possible to identify critical excitation frequencies for essential fundamental waveforms automatically. For this purpose, several hundred status data points from training objects are first collected and made available as part of a training data set. In the context of the exemplary embodiment relating to the vibration behavior of the mechanical system, each state data point of the training data set comprises the following:

- Spatial state data: point cloud with displacement vector per sampling point in the scanning space - preferably obtained from sensor or simulation data, and

- Additional parameters (here optional): frequency of the excitation signal

Furthermore, the training data record comprises a known status evaluation for each of the status data points. The known condition assessment is determined in advance by an experienced human expert, the assessment of the vibration forms and condition assessments depending on the precise type of mechanical system. For an automobile body or an aircraft fuselage / wing, for example, one or more of the following assessments or status assessments can be distinguished or taken into account:

- torsion

- Vertical bend

- transverse bend

Further assessments or status assessments are conceivable.

In order to convert a displacement field defined on a point cloud into a uniform grid, the following procedure can be used, for example. First, the resolution of the grid is determined (for example 256x256x256 cells). The size of the grid cells is then calculated depending on the size of the mechanical system by dividing the dimensions by the number of cells in each spatial direction. All points P_i which are in the cell are then sought for each cell. The three-dimensional displacement vector for each The grid cell then results from the averaging of the displacement vectors at the points P_i.

- Folding layers for spatial state data: The folding layers use normalization and activation by means of leakyReLu, a max pooling is also carried out every two layers. A total of six layers of folds.

- Further folding layers: two folding layers including normalization and activation using leakyReLu.

Initially, the network parameters (weights) of the neural network can be initialized using "Xavier" initialization. The described post-training strategy can be followed for existing data.

As described, a three-dimensional point cloud with displacement information per sampling point serves as input for the evaluation system. Furthermore, the frequency of the excitation can serve as an optional input for the evaluation system. The three-dimensional point cloud with the displacement information per sampling point forms the spatial state data in the form of a discretized displacement field. The excitation frequency represents a possible additional parameter. The evaluation system can then automatically judge the state of the mechanical system based on the spatial state data, namely the discretized displacement field and possibly based on the optional additional parameter, by identifying a basic physical waveform in the state data point. In the example of an automobile body or an aircraft fuselage or an aircraft wing, the following cases can be distinguished or taken into account:

- torsion and / or

- vertical bend and / or

- transverse bend

Other cases are conceivable.

Taking into account the exemplary embodiment according to FIG. 12, a method can be configured as follows:

The method is used to evaluate a state of a three-dimensional test object, taking into account a predetermined assessment task and using an evaluation system that comprises a neural network. A training data record is provided and / or used, which comprises several status data points from one or more three-dimensional training objects. The training data record includes a known status evaluation for each of the status data points in relation to the specified assessment task. To adapt the assessment system to the specified assessment task, the neural network of the assessment system is parameterized in a training process using the training data set. In an execution process, a condition assessment is calculated or ascertained with the adapted assessment system for a given status data point of the test object.

In the context of this configuration, the test object is a mechanical system, for example a vehicle body, an aircraft fuselage, an aircraft wing or at least a component thereof. It can be used as a predefined assessment task to determine at which frequency a predefinable (or a certain physically relevant) waveform occurs. The condition assessment of the test object can thus be a vibration form (vibration pattern). A state data point - in the context of this embodiment - is a three-dimensional displacement field.

With regard to further advantageous refinements of the method according to the invention or of the evaluation system according to the invention, reference is made to the general part of the description and to the appended claims in order to avoid repetition.

Finally, it should be expressly pointed out that the above-described exemplary embodiments of the method according to the invention or of the evaluation system according to the invention only serve to discuss the claimed teaching, but do not restrict it to the exemplary embodiments.

Claims

Expectations

1. Method for evaluating a state of a three-dimensional test object, taking into account a predetermined assessment task and using an evaluation system that comprises a neural network,

wherein a training data record is provided and / or used, which comprises several state data points from one or more three-dimensional training objects,

wherein the training data record includes a known condition assessment for each of the condition data points in relation to the specified assessment task,

whereby the neural network of the evaluation system is parameterized in a training process using the training data set in order to adapt the evaluation system to the specified evaluation task, and

whereby a status assessment is calculated in an execution process with the adapted assessment system for a given status data point of the test object.

2. The method according to claim 1, characterized in that the test object comprises a component, a mechanical system, an electromechanical system and / or an electrochemical system.

3. The method according to claim 1 or 2, characterized in that the training objects of the training data set are selected problem-specifically in such a way that the predetermined assessment task can be applied both to the training objects and to the test object.

4. The method according to any one of claims 1 to 3, characterized in that the state data points of the training objects and the state data point of the test object each comprise spatial state data.

5. The method according to claim 4, characterized in that the spatial state data comprised by a state data point of a training object define a state of the training object, and the spatial state data comprised by the state data point of the test object define a state of the test object.

6. The method according to claim 4 or 5, characterized in that the spatial state data are presented as a discretized scalar, vector and / or tensor field.

7. The method according to any one of claims 4 to 6, characterized in that the spatial state data are converted into the form of a uniform grid.

8. The method according to any one of claims 4 to 7, characterized in that the spatial status data are represented as a tensor of level N + 3, where N is the order of the status data per grid point.

9. The method according to any one of claims 4 to 8, characterized in that the spatial state data represent a physical quantity.

10. The method according to any one of claims 4 to 9, characterized in that the spatial state data comprise a temperature field, a deformation field, a speed field, a voltage field, a pressure field, a displacement field and / or an electromagnetic field.

1 1. The method according to any one of claims 4 to 10, characterized in that the spatial state data represent a digitization state, preferably wherein the spatial state data comprise the local resolution of a numerical grid and / or a quality feature of a numerical grid.

12. The method according to any one of claims 1 to 1 1, characterized in that the state data points of the training objects and the state data point of the test object include at least one additional parameter in addition to spatial state data.

13. The method according to any one of claims 1 to 12, characterized in that the training process comprises a parameter optimization step, wherein in the parameter optimization step a status evaluation is calculated by the evaluation system for each state data point of the training data set, the known one belonging to the respective state data point The condition assessment of the training data record is compared with the calculated condition assessment, and network parameters of the neural network are adapted as a function of the comparison.

14. The method according to any one of claims 1 to 13, characterized in that the parameter optimization step, possibly with different training data records, is carried out several times.

15. The method according to any one of claims 1 to 14, characterized in that application-specific network parameters of the neural network of the evaluation system are determined in the training process by back propagation, the training data set and initial network parameters of the neural network being used as input data for the back propagation.

16. The method according to any one of claims 1 to 15, characterized in that the neural network of the evaluation system is configured in such a way that the spatial status data and the at least one additional parameter are processed separately from one another in a plurality of folding layers before the respectively calculated intermediate results be brought together again and processed by further layers of the neural network.

17. Evaluation system for evaluating a state of a three-dimensional test object, taking into account a predetermined evaluation task, in particular for carrying out a method according to one of claims 1 to 16, wherein the evaluation system comprises a, preferably deep, neural network, the evaluation system being designed in this way, that a training data record can be used which comprises several status data points from one or more three-dimensional training objects, the training data record in Includes a known condition assessment for each of the condition data points in relation to the specified assessment task,

wherein the evaluation system is further configured such that the neural network can be parameterized in a training process using the training data set in order to adapt to the predefined assessment task, and that after the adaptation to the predefined assessment task in an execution process for a predefined status data point of the test object a condition assessment can be calculated.

18. Computer program product with program code which is stored on a machine-readable carrier and which provides and / or executes a method for evaluating a state of a three-dimensional test object according to one of claims 1 to 16.