DESIGN AND MANUFACTURE OF A FLOW MACHINE SHOVEL

Description

[0001] The invention relates to a method for designing a turbomachine blade. Furthermore, the invention relates to a method for manufacturing a turbomachine blade. The invention also relates to a computer unit and a machine-readable storage medium.

[0002] Designing a turbomachine blade is a complex process. This is due, among other things, to the fact that, on the one hand, a multiple of blade parameters are typically required to characterize a turbomachine blade, and the blade to be designed must satisfy various boundary conditions (especially from different disciplines such as fluid dynamics, strength of materials, thermal engineering, etc.), and on the other hand, there is often only limited explicit knowledge about the usually complex relationships / dependencies between the required boundary conditions and the blade parameters that enable these boundary conditions to be met.

[0003] Artificial neural networks have proven their worth in numerous technical fields by independently learning (using training data) complex relationships and dependencies between different parameters. In particular, a (trained) artificial neural network can be used to predict the blade parameters of a turbomachine blade, enabling compliance with desired boundary conditions.

[0004] Methods for designing turbomachinery blades, especially turbine blades, are disclosed, for example, in the conference report "A Neural Network Aero Design System for Advanced Turbo-Engines" by Jose M. Sanz (available at: https: / / ntrs.nasa.gov / search.jsp?R=19990019840) and in the article "The Development of Highly Loaded Turbine Rotating Blades by Using 3D Optimization Design Method of Turbomachinery Blades Based on Artificial Neural Network & Genetic Algorithm" by Fan-zhen Zhou et al. (published in Chinese Journal of Aeronautics Vol. 16 No. 4).

[0005] One object of the invention is to improve the design of a turbomachine blade based on the use of neural networks and, in particular, the manufacture of a turbomachine blade based on such a design process.

[0006] This problem is solved according to the invention by a method according to claim 1, by a method according to claim 8, by a computing unit according to claim 9 and by a machine-readable storage medium according to claim 10.

[0007] Advantageous embodiments of the invention are the subject of the further patent claims and the following description.

[0008] In the inventive method for designing a turbomachine blade, predetermined input parameters are transmitted to a neural network system. The neural network system determines and outputs blade parameters based on the transmitted input parameters. According to the invention, the neural network system comprises several separate neural networks, each with an output layer, each of which determines one or more of the blade parameters and outputs them via its output layer. The separate neural networks of the neural network system include a first neural network and a second neural network.Furthermore, according to the invention, the blade parameter(s) determined by the first neural network and output via its output layer is / are different from the blade parameter(s) determined by the second neural network and output via its output layer.

[0009] In other words, the invention provides for the determination of the blade parameters to be calculated by separate neural networks (i.e., not all blade parameters to be calculated by the same neural network). This means that each neural network within the neural network system has fewer (different) blade parameters to calculate compared to the case where all blade parameters are calculated by the same neural network, and this can lead to an improvement in the prediction accuracy of the blade parameters calculated by the neural network system, or a reduction in the (prediction) errors of the blade parameters calculated by the neural network system.

[0010] The "prediction accuracy" of a blade parameter determined by the neural network system refers to how close the value of the blade parameter determined by the neural network system is to the optimal blade parameter value with which the required input parameters can best be realized. The closer the value of the blade parameter determined by the neural network system is to the optimal blade parameter value, the higher the prediction accuracy of this blade parameter (and the lower its (prediction) error).

[0011] Furthermore, distributing blade parameter determination across separate neural networks allows physically related blade parameters, such as different thickness parameters, to be determined using the same neural network, while other related blade parameters, such as different curvature parameters, are determined using a different neural network within the neural network system. This means that the distribution of blade parameter determination across different neural networks can be problem-specific, taking physical relationships into account. In such a case, distributing blade parameter determination across separate neural networks has a particularly beneficial effect on the prediction accuracy of the blade parameters.

[0012] The neural networks of the neural network system are artificial neural networks. For the sake of readability, the term "artificial" is omitted in this description and in the patent claims.

[0013] The turbomachine blade to be designed can be, for example, a turbine blade or compressor blade, especially for a gas turbine plant.

[0014] As mentioned previously, the neural network system, or rather its neural networks, outputs blade parameters. That is, the blade parameters represent output parameters of the neural network system, or rather its neural networks.

[0015] Blade parameters within the meaning of the present invention are those parameters which characterize the turbomachine blade itself and / or its arrangement / alignment in a blade ring.

[0016] Conveniently, each of the neural networks in the neural network system has an input layer in addition to its output layer.

[0017] The input parameters transmitted to the neural network system preferably relate to desired boundary conditions for the turbomachine blade, to desired boundary conditions for a blade ring comprising several such turbomachine blades, and / or to desired boundary conditions for a fluid flow (to be realized using the blade or blade ring).

[0018] Furthermore, at least some of the blade parameters can refer to a blade section or to different blade sections between a hub-side end of the turbomachine blade and a hub-remote end of the turbomachine blade.

[0019] As mentioned previously, the neural network system comprises a first neural network and a second neural network. Additionally, the neural network system can include at least one further neural network that determines one or more blade parameters and outputs them via its output layer. These parameters differ from the blade parameters determined by the first and second neural networks.

[0020] The respective neural network of the neural network system is trained using training data before the blade parameters are determined.

[0021] The training data is preferably a pseudo- or true subset of a baseline training dataset. The baseline training dataset may, for example, include data obtained from previous turbomachine blade designs and / or previous blade design optimizations. In other words, the baseline training dataset may include data from previous design and / or optimization projects. This data may, in particular, be based on measurements and / or simulations.

[0022] It is advantageous to check whether the basic training dataset contains redundant parameters. Any redundant parameters found can then be filtered out. This allows the number of input parameters required to determine the blade parameters to be kept to a minimum and the topology of the neural network system to be kept simple. Furthermore, the training times for the individual neural networks can be kept short.

[0023] Advantageously, the training data used to train the neural networks corresponds to the basic training dataset or, if the basic training dataset contains redundant parameters, to the basic training dataset cleaned of the redundant parameters.

[0024] Checking whether the base training dataset contains redundant parameters can be done, for example, using statistical methods, particularly Pearson correlation coefficients. Before deciding whether to exclude a parameter from the base training dataset from the training data, it is preferable to check whether this parameter is redundant not only from a statistical perspective but also from a physical one. This prevents a parameter from being excluded from the base training dataset simply because a random statistical correlation has been found between this parameter and another parameter from the base training dataset.

[0025] In accordance with the present invention, a parameter from the basic training dataset can be redundant, in particular, if a correlation, such as a linear relationship, exists between this parameter and another parameter from the basic training dataset. Methods such as "input merging" or "adaptive feature selection by using mutual information" can be used, for example, to find correlations between parameters from the basic training dataset.

[0026] The blade parameters determined by the neural network system include geometric blade parameters.

[0027] At least some of the geometric blade parameters can, for example, characterize a blade section (i.e., a profile of the turbomachine blade). Furthermore, at least some of the geometric blade parameters can relate to different blade sections between the hub-side end of the turbomachine blade and the end furthest from the hub.

[0028] The geometric blade parameters can include one or more curvature parameters or one or more thickness parameters.

[0029] In the invention, the geometric blade parameters comprise one or more curvature parameters as well as one or more thickness parameters. Furthermore, the curvature parameter(s) are determined by another neural network within the neural network system, and the thickness parameter(s) are output via its output layer.

[0030] The aforementioned thickness parameter(s) may include, among other things, a profile thickness at half of the rear blade section, a leading edge radius (also called nose radius), a maximum profile thickness, a thickness offset and / or a trailing edge angle, or at least one of these parameters.

[0031] For the purposes of the present invention, the "rear blade section" means that part of a blade section / blade profile that extends from the position (on the chord line) where the maximum profile thickness is located to the trailing edge of the turbomachine blade. The "half" of the rear blade section refers to half the distance from the position of the maximum profile thickness to the trailing edge of the turbomachine blade (along the chord line).

[0032] The curvature parameter(s) in question may include, among other things, a maximum profile curvature, a curvature recession, an angle between the chord line and the skeletal line at the trailing edge and / or a step angle, or at least one of these parameters.

[0033] In the invention, the input parameters transmitted to the neural network system are two-dimensional parameters. These are parameters that are each defined / definable by a pair of values ​​(i.e., by two related values). The input parameters can, for example, represent a Mach number distribution, in particular a suction-side Mach number distribution and / or a pressure-side Mach number distribution. Alternatively or additionally, the input parameters can represent at least one pressure distribution, in particular a suction-side pressure distribution and / or a pressure-side pressure distribution.

[0034] In this context, the Mach number distribution refers to a Mach number as a function of the normalized chord of the turbomachine blade, or the shape of this function. Similarly, the pressure distribution refers to a pressure as a function of the normalized chord of the turbomachine blade, or the shape of this function.

[0035] Furthermore, it is advantageous if, after training the neural networks of the neural network system and before determining the blade parameters, one or more existing neuronal connections are removed from at least one of the neural networks using a pruning method, such as "optimal brain damage," "early brain damage," or "inverse kurtosis." This prevents overfitting. Additionally, the generalizability of the neural network system can be increased. This means that the quality of outputs for which the neural networks were not trained can be improved, and the remaining optimization space for any subsequent optimizations can be reduced.

[0036] Furthermore, the blade parameters determined by the neural network system can include a blade efficiency of the turbomachine blade. Alternatively or additionally, the blade parameters determined by the neural network system can include a logarithmic decrement of the turbomachine blade. The logarithmic decrement of the turbomachine blade is a measure of the damping behavior of the turbomachine blade.

[0037] It is particularly advantageous if the blade parameters determined by the neural network system include, on the one hand, geometric blade parameters and, on the other hand, at least one additional blade parameter, such as the blade efficiency and / or the logarithmic decrement of the turbomachine blade. This makes it possible, in particular, to combine design processes from different disciplines, such as fluid dynamics, structural engineering, and thermal engineering, into a single, unified design process, instead of carrying them out sequentially as is currently the norm. Such a multidisciplinary design process enables shorter design times. Furthermore, such a multidisciplinary design process can significantly narrow down the optimization space for any subsequent optimizations by the individual disciplines.

[0038] The neural network system advantageously includes a third neural network in addition to the first and second neural networks. For example, the first neural network might determine geometric blade parameters and output them via its output layer, the second neural network might determine and output a blade efficiency, and the third neural network might determine and output a logarithmic decrement of the turbomachine blade. In such a case, the neural network system provides multidisciplinary parameters as blade parameters.

[0039] The invention enables a so-called inverse design of the turbomachine blade. This means that, based on predefined input parameters, blade parameters, in particular geometric blade parameters, can be determined and output, with which these input parameters can be implemented.

[0040] The flow parameter(s) in question can be, for example, an approach Mach number, an outflow Mach number, an approach angle, a deflection angle and / or a blade flange speed, or at least include one of these parameters.

[0041] The geometric quantity(s) in question can be, for example, a trailing edge radius and / or a division ratio, or at least include one of these parameters.

[0042] The division ratio in this case is to be understood as the ratio of the slope of a bucket ring (i.e. the distance between two adjacent buckets of the bucket ring along its circumferential line) to a profile depth of the respective bucket.

[0043] Furthermore, the aforementioned thermodynamic quantity(s) can be, for example, a suction-side temperature and / or a pressure-side temperature, or at least include one of these parameters.

[0044] It is advantageous if at least one of the neural networks, in addition to its output layer, has an input layer and a plurality of intermediate layers (also called "hidden layers"). That is, advantageously, at least one of the neural networks is a deep neural network. The individual layers of such a multi-layered neural network are conveniently linked hierarchically. Deep neural networks are particularly well-suited for determining difficult-to-predict (shovel) parameters.

[0045] Preferably, a parameter is determined by at least one of the layers of such a network and transmitted to one of the other layers of this neural network.

[0046] In this way, for example, knowledge gained or learned in determining an easier-to-predict parameter can be used to determine a more difficult-to-predict parameter.

[0047] Furthermore, it can be provided that at least one of the neural networks comprises an ensemble of several neural subnetworks. The neural subnetworks of the ensemble can, for example, differ from one another with respect to their network structure. Preferably, the same input parameters are transmitted to each neural subnetwork of this ensemble.

[0048] Advantageously, each neural subnetwork of this ensemble determines a blade parameter value for the same blade parameter. Furthermore, it is advantageous to average the blade parameter values ​​determined by the ensemble's neural subnetworks, either weighted or unweighted. The resulting averaged blade parameter value is then output as the blade parameter by the neural network to which the ensemble belongs, via its output layer. This averaging process minimizes the prediction error of the output blade parameter, as it averages out the prediction errors of the individual subnetworks.

[0049] The neural network to which the ensemble belongs can, in particular, comprise several such ensembles. These can be hierarchically linked. At least one of the ensembles can, for example, determine a parameter that is then passed as an input parameter to one of the other ensembles in this neural network. In this way, knowledge gained or learned in determining an easier-to-predict parameter can be used to determine a more difficult-to-predict parameter.

[0050] The parameter that is transmitted from one of the ensembles to one of the other ensembles of this neural network is preferably a scoop parameter that is output by the said neural network to which the ensembles belong via its output layer.

[0051] Furthermore, at least one of the subnetworks of the respective ensemble can be designed as a so-called "deep neural network".

[0052] As mentioned at the beginning, the invention relates to a method for manufacturing a turbomachine blade.

[0053] The inventive method for manufacturing a turbomachine blade comprises the following steps: Designing a turbomachine blade according to the inventive method for designing a turbomachine blade, wherein blade parameters are determined, optionally modifying one or more of the blade parameters taking into account at least one optimization criterion, and manufacturing the turbomachine blade according to the determined, optionally modified, blade parameters.

[0054] The second step of the manufacturing process is optional. This means that the turbomachine blade can be manufactured according to the blade parameters determined using the neural network system, without modifying the determined blade parameters taking into account at least one optimization criterion.

[0055] In the production of the turbomachine blade, an additive manufacturing process, such as selective laser melting, selective laser sintering or the like, can be used in particular.

[0056] The invention also relates, as mentioned at the outset, to a computer unit and a machine-readable storage medium.

[0057] The embodiments, design details and advantages described above in connection with the inventive method for designing a turbomachine blade can also refer to the inventive computer unit and the inventive storage medium.

[0058] The computer unit according to the invention is configured to carry out the inventive method for designing a turbomachine blade. The computer unit comprises a neural network system configured to receive predetermined input parameters and to determine and output blade parameters based on the received input parameters. According to the invention, the neural network system of the computer unit comprises several separate neural networks, each with an output layer, each of which is configured to determine one or more of the blade parameters and output them via its output layer. The separate neural networks include a first neural network and a second neural network. The first neural network is configured to determine one or more of the blade parameters and output them via its output layer.Furthermore, the second neural network is set up to determine one or more of the blade parameters and output via its output layer which one(s) differs from the blade parameter(s) determined and output by the first neural network.

[0059] The computer unit can be set up to carry out the specified process steps by equipping the computer unit with a suitable computer program that enables the computer unit to carry out the process steps.

[0060] The machine-readable storage medium according to the invention has a program code by which a computing unit is caused to execute the method according to the invention for designing a turbomachine blade when the program code is executed by the computing unit.

[0061] The machine-readable storage medium can be a component of the computer unit according to the invention or be connected to the computer unit according to the invention.

[0062] The description of preferred embodiments of the invention given so far contains numerous features, some of which are summarized in the individual dependent claims. However, these features can also be considered individually and combined in meaningful ways. In particular, these features can be combined individually and in any suitable combination with the methods, the computer unit, and the storage medium according to the invention. Furthermore, method features can also be considered properties of the corresponding device unit.

[0063] The properties, features, and advantages of the invention described above, as well as the manner in which they are achieved, will become clearer and more readily understandable in connection with the following description of the exemplary embodiments of the invention, which are explained in more detail in relation to the figures. The exemplary embodiments serve to illustrate the invention and do not limit it to the combinations of features specified therein, including functional features. Furthermore, suitable features of each exemplary embodiment can also be considered in isolation, removed from one exemplary embodiment, incorporated into another exemplary embodiment to supplement it, and combined with any of the claims.

[0064] They show: FIG 1 an embodiment of a first computer unit according to the invention for designing a turbomachine blade; FIG 2 a profile of a turbomachine blade; FIG 3 a schematic representation of a neural network in which different layers of neurons of the neural network are shown; FIG 4 a diagram in which a pressure-side and a suction-side Mach number distribution are shown; FIG 5 a schematic representation of a neural network comprising an ensemble of several subnetworks; FIG 6 an embodiment of a second computer unit according to the invention for designing a turbomachine blade.

[0065] FIG 1 schematically shows a first computer unit 2 for designing a turbomachine blade.

[0066] This computing unit 2 comprises a neural network system 4, which is configured to receive specified input parameters 6 and to determine and output several blade parameters 8, which characterize the turbomachine blade to be designed and its intended arrangement / orientation in a blade ring, based on the received input parameters 6.

[0067] The neural network system 4 comprises several separate neural networks 10, each of which has an output layer (see FIG 3 ). In the embodiment according to FIG 1 The neural network system 4 comprises a first neural network 12, a second neural network 14, a third neural network 16 and a fourth neural network 18.

[0068] The input parameters 6, which the neural network system 4 uses to determine and output the blade parameters 8, are transmitted to each of the four neural networks 12-18. In the present embodiment, the input parameters 6n comprise two-dimensional parameters that define a suction-side Mach number distribution. But SS represents a blade cut, and m two-dimensional parameters that define a pressure-side Mach number distribution But PS represent the aforementioned shovel cut (cf. FIG 4 ), where n and m are natural numbers greater than 1. In other words, the input parameters 6 represent n points of a suction-side Mach number distribution. But SS and m points of a pressure-side Mach number distribution But PS represents. Each of the input parameters 6 includes as its first value a normalized profile depth. x i and, as a second value, the corresponding Mach number on the suction side But SS ( X i) or print-side Mach number ButPS ( X i).

[0069] The determination of the blade parameters 8 is divided among the four neural networks 12-18, with each of the four neural networks 12-18 determining one or more of the blade parameters 8 and outputting them via its output layer, and with the blade parameter(s) determined by the respective neural network 12-18 and outputting them via its output layer being / are different from the blade parameters determined and output by the other neural networks 12-18.

[0070] The blade parameters determined by the first and second neural networks 12, 14 are geometric blade parameters 20. In the present embodiment, the first neural network 12 determines curvature parameters 22 of a blade profile (here, by way of example, a maximum profile camber F, a camber offset M, an angle γ 1 between the chord line and the skeletal line at the blade trailing edge (cf. FIG 2 ) and a step angle λ) are determined and output via its output layer. The second neural network 14, on the other hand, determines the thickness parameters 24 of this blade profile (here, an exemplary profile thickness). y 1 / 2 at half of the rear shovel section, a leading edge radius R 0 , a maximum profile thickness D, a thickness offset N and a trailing edge angle w 1) determined and output via its output layer.

[0071] Furthermore, the third neural network 16 determines a logarithmic decrement Ξ of the turbomachine blade as a blade parameter and outputs it via its output layer. The fourth neural network 18 determines a blade efficiency η as a blade parameter and outputs it via its output layer.

[0072] In principle, the neural network system 4 can have a different number of neural networks 10, 12-18. Furthermore, the determination of the blade parameters 8 can, in principle, be distributed differently among the individual neural networks 10, 12-18.

[0073] Before the blade parameters 8 are determined using the neural network system 4, each of the four neural networks 12-18 is trained using training data based on a baseline training dataset containing data from previous design and / or optimization projects. Before training the neural networks 12-18, it is first checked whether the baseline training dataset contains redundant parameters. If the baseline training dataset contains redundant parameters, the data from the baseline training dataset, cleaned of the redundant parameters, serves as the training data. Otherwise, all data from the baseline training dataset is used as training data.

[0074] After training the neural networks 12-18 and before determining the blade parameters 8, one or more existing neuronal connections are, if necessary, pruned in one or more of the neural networks 12-18 using a pruning procedure (see FIG 3) removed.

[0075] The design of a turbomachine blade, as described above, is followed by the manufacturing of the turbomachine blade.

[0076] Before the production of the turbomachine blade, at least one of the blade parameters 8 determined and output by the neural network system 4 can be modified, if necessary, taking into account at least one optimization criterion.

[0077] The production of the turbomachine blade is carried out according to the blade parameters determined by the neural network system 4, possibly modified, for example using an additive manufacturing process.

[0078] FIG 2 shows an exemplary profile of a turbomachine blade 26.

[0079] The turbomachine blade 26 comprises a leading edge 28 and a trailing edge 30. Furthermore, the turbomachine blade 26 has a top surface 32 and a bottom surface 34.

[0080] In FIG 2 Both a chord line 36 and a skeleton line 38 of the turbomachine blade 26 are shown. In addition, in FIG 2 Some of the previously mentioned bucket parameters 8 have been identified.

[0081] Among the in FIG 2 The identified blade parameters 8 include, among others, the profile depth t of the turbomachine blade 26 (i.e., the length of the chord line 36), the leading edge radius R 0 , the maximum profile thickness D, the maximum profile camber F (i.e. the greatest deviation of the skeletal line 38 from the chord line 36), the trailing edge angle w1 (i.e., the angle between the upper surface of the blade 32 and the lower surface of the blade 34 at the trailing edge 30) and the angle γ 1 between the chord line 36 and the skeletal line 38 at the trailing edge 30. In addition, the following are included in FIG 2 The blade parameters 8 identified are the camber offset M (i.e., the distance of the position of the maximum profile camber F to the leading edge 28 along the chord line 36), the thickness offset N (i.e., the distance of the position of the maximum profile thickness D to the leading edge 28 along the chord line 36), and the profile thickness y 1 / 2 at half of the rear blade section.

[0082] FIG 3 shows a schematic representation illustrating the basic structure of a neural network 10.

[0083] The in FIG 3The depicted neural network 10 comprises a multitude of artificial neurons 40, which form several layers. In this case, the neurons 40 form an input layer 42 for receiving input parameters, an output layer 44, as well as a first intermediate layer 46 and a second intermediate layer 48. One or more output parameters, which are determined by the neural network 10 based on the input parameters, are output via the output layer 44 of the neural network 10.

[0084] In the neural network 10 from FIG 3 The input layer 42 and the second intermediate layer 48 comprise three neurons 40, while the first intermediate layer 46 comprises five neurons 40 and the output layer 44 comprises one neuron 40, the number of neurons 40 in each layer 42-48 being chosen only as an example.

[0085] The individual layers 42-48 of the neural network 10 are interconnected by neuronal connections 50, each connecting two neurons 40. In this case, each neuron 40 of the input layer 42 is connected to each neuron 40 of the first intermediate layer 46. Furthermore, the neurons 40 of the first intermediate layer 46 are each connected to each neuron 40 of the second intermediate layer 48.

[0086] Each neuron 40 of the second intermediate layer 48 is in turn connected to the neuron 40 of the output layer 44.

[0087] The four previously mentioned neural networks 12-18 of neural network system 4 from FIG 1 Each of the neural networks 12-18 can have such a structure or a similar structure. In particular, each neural network 12-18 of the neural network system 4 can have a different number of intermediate layers 46, 48, a different number of neurons 40, and a different number of neuronal connections 50.

[0088] FIG 4 shows an axis diagram whose ordinate axis represents a Mach number. But represents and whose abscissa axis represents the normalized profile depth x of a turbomachine blade, where the value 0 of the normalized profile depth x corresponds to the leading edge position of the turbomachine blade and the value 1 of the profile depth x corresponds to the trailing edge position of the turbomachine blade.

[0089] The diagram shows an exemplary suction-side Mach number distribution. But SS around a shovel cut as well as an exemplary pressure-side Mach number distribution But PS: This shovel cut is shown.

[0090] Furthermore, in FIG 4 some points of the suction-side Mach number distribution But SS as well as some points of the print-side Mach number distribution But PS drawn, which, for example, serve as input parameters to the neural network system 4. FIG 1can be transmitted.

[0091] Specifically, this concerns the suction-side Mach number distribution. But SS for the points: (0.2 | But SS (0.2)), (0.4 | But SS (0.4)), (0.6 | But SS (0. 6)), (0.8 | But SS (0.8)) and ( X max | But SS ( X max )). In the print-side Mach number distribution But PS: These are the points: (0.2 | But PS (0.2)), (0.4 | But P S (0.4)), (0.6 | But PS (0.6)) and (0.8 | But PS (0.8)). Alternatively or additionally, other points of the suction-side Mach number distribution can be used. But SS and / or other points of the pressure-side Mach number distribution But PS as an input parameter to the neural network system 4 FIG 1 be transmitted.

[0092] FIG 5 shows a schematic representation of a neural network 10 for determining one or more blade parameters.

[0093] The neural network 10 from FIG 5 has an input layer 42 comprising several artificial neurons 40. Furthermore, this neural network 10 comprises an ensemble 52 of several neural subnetworks 54, each linked to the input layer 42. The ensemble 52 of neural subnetworks 54 is preferably composed of neurons of one or more (in FIG 5 (not shown) intermediate layers of the neural network 10 are formed.

[0094] Input parameters are received from input layer 42, or rather from its neurons 40, and transmitted to each of the neural subnetworks 54 of the ensemble 52. Each neural subnetwork 54 of the ensemble 52 determines a blade parameter value 56 for the same blade parameter (for example, for the maximum profile thickness D) based on the input parameters. The individual blade parameter values ​​56 are averaged (weighted or unweighted), resulting in an averaged blade parameter value 58. The averaged blade parameter value 58 is then used by neural network 10 via its (in FIG 5 (Output layer not shown) output as a blade parameter.

[0095] Optionally, the neural network 10 can include at least one further such ensemble of neural subnetworks, which serves to determine another blade parameter. In such a case, the averaged blade parameter value 58 can be transmitted as an additional parameter to the further ensemble or its neural subnetworks, alongside the previously mentioned input parameters, so that the averaged blade parameter 58 can be included in the determination of the further blade parameter.

[0096] One or more of the four neural networks 12-18 of neural network system 4 from FIG 1 can have the same structure and functionality as the neural network 10 from FIG 5 have.

[0097] The description of the following embodiment is primarily limited to the differences from the preceding embodiment, to which reference is made with regard to identical features and functions. Identical and / or corresponding elements are, where appropriate, designated with the same reference numerals, and features not mentioned are adopted in the following embodiment without being described again.

[0098] FIG 6 schematically shows a second computer unit 60 for designing a turbomachine blade.

[0099] Just like computer unit 2 from FIG 1 This also includes computer unit 60. FIG 6 a neural network system 4 with a first, a second, a third and a fourth neural network 12-18.

[0100] In the present embodiment, one-dimensional parameters are transmitted as input parameters 6 to the four neural networks 12-18 of the neural network system 4. In this case, the input parameters 6 comprise several flow variables (here, for example, an approach Mach number). But in , an outflow Mach number But out , an inflow angle β 1 , a deflection angle Δβ and a planned blade ring speed n 1). In addition, the input parameters include 6 geometric quantities (here, for example, a specified trailing edge radius). R TE and a planned division ratio t / s).

[0101] As in the previous embodiment, geometric blade parameters 20 are determined and output by the first and second neural networks 12, 14 in the present embodiment.

[0102] The first neural network 12 determines curvature parameters 22 (here, for example, a step angle λ and, for various blade sections between a hub-side end and a remote end of the turbomachine blade, the angle γ1 between the chord line and the skeleton line at the blade trailing edge) and outputs them via its output layer. The second neural network 14, on the other hand, determines thickness parameters 24 (here, for example, the maximum profile thickness D and the thickness offset N for various blade sections between the hub-side end and the remote end of the turbomachine blade) and outputs them via its output layer.

[0103] The indices given as percentages for some of the blade parameters 8 in FIG 6Each indice represents the position of the blade section to which the respective blade parameter refers. The meaning of these indices (expressed as percentages) is defined as follows: An index of 0% corresponds to the hub-side end of the turbomachine blade, while an index of 100% corresponds to the far end of the turbomachine blade. An index of 50% thus corresponds to a position midway between the hub-side and far-side ends of the turbomachine blade.

[0104] Although the invention has been further illustrated and described in detail by the preferred embodiments, the invention is not limited by the disclosed examples and other variations can be derived from them without leaving the scope of protection of the invention.

Claims

1. A method for designing a turbomachine blade (26) performed by means of a computer unit, in which predetermined input parameters (6) are transmitted to a neural network system (4) and blade parameters (8) are determined and output by the neural network system (4) based on the transmitted input parameters (6), wherein the neural network system (4) has multiple separate neural networks (10, 12-18) with one output layer (44) each, each of which determines one or more of the blade parameter(s) (8) and outputs it / them via its output layer (44), wherein a first neural network and a second neural network belong to the separate neural networks (10, 12-18) of the neural network system (4) and the blade parameter(s) which is / are determined by the first neural network and output via its output layer (44) is / are different from the blade parameter(s) which is / are determined by the second neural network and output via its output layer (44), and wherein the respective neural network of the neural network system is trained using training data before determination of the blade parameters, wherein the blade parameters (8) determined by the neural network system (4) comprise geometric blade parameters (20), wherein the geometric blade parameters (20) - comprise one or more curvature parameters (22) and one or more thickness parameters (24), wherein the curvature parameter(s) (22) is / are determined by a different neural network (10, 12-18) of the neural network system (4) and is / are output via its output layer (44) than the thickness parameter(s) (24) and, that the input parameters (6) transmitted to the neural network system (4) are two-dimensional parameters, wherein the input parameters (6) represent at least one Mach number distribution (MaSS, MaPS) and / or at least one pressure distribution.

2. The method according to claim 1, characterized in that prior to determination of the blade parameters (8), the respective neural network (10, 12-18) of the neural network system (4) is trained using training data that are a false or true subset of a basic training data set, wherein it is checked whether the basic training data set contains redundant parameters, and the training data which are used for training the neural networks (10, 12-18) correspond to the basic training data set or, if the basic training data set contains redundant parameters, to the basic training data set adjusted for the redundant parameters.

3. The method according to claim 1 or 2, characterized in that prior to determination of the blade parameters (8), the respective neural network (10, 12-18) of the neural network system (4) is trained using training data, wherein after training of the neural networks (10, 12-18) of the neural network system (4) and prior to determination of the blade parameters (8), one or more existing neuron connections (50) are removed in at least one of the neural networks (10, 12-18) by means of a pruning method.

4. The method according to any one of the preceding claims, characterised in that the blade parameters (8) determined by the neural network system (4) comprise a blade efficiency (η) of the turbomachine blade (26) and / or a logarithmic decrement (Ξ) of the turbomachine blade (26).

5. The method according to any one of the preceding claims, characterised in that a third neural network belongs to the separate neural networks (10, 12-18) of the neural network system (4), wherein the first neural network determines geometrical blade parameters (20) and outputs these via its output layer, the second neural network determines a blade efficiency (η) and outputs it via its output layer (44), and the third neural network determines a logarithmic decrement (Ξ) of the turbomachine blade (26) and outputs it via its output layer (44).

6. The method according to any one of the preceding claims, characterised in that at least one of the neural networks (10, 12-18) has a plurality of intermediate layers (46, 48) in addition to its said output layer (44).

7. The method according to any one of the preceding claims, characterised in that at least one of the neural networks (10, 12-18) comprises an ensemble (52) of multiple neural subnetworks (54), wherein - every neural subnetwork (54) of this ensemble (52) determines one blade parameter value (56) each for the same blade parameter, and - the blade parameter values (56) determined by the neural subnetworks (54) of the ensemble (52) are averaged weighted or unweighted, wherein the averaged blade parameter value (58) obtained in this manner by said neural network (10, 12-18) to which the ensemble (52) belongs is output as a blade parameter via its output layer (44).

8. A method for producing a turbomachine blade (26), comprising the steps: - designing a turbomachine blade (26) according to a method according to any one of the preceding claims, wherein blade parameters (8) are determined, - optionally modifying one or more of the blade parameters (8), taking into account at least one optimisation criterion, - producing the turbomachine blade (26) in accordance with the determined, optionally modified blade parameters (8).

9. A computer unit (2, 60) for performing a method according to any one of claims 1 to 7.

10. A machine-readable storage medium with programme code which causes a computer unit (2, 60) to execute a method according to any one of claims 1 to 7 when the programme code is executed by the computer unit (2, 60).