DE102020201599A1 - Method and device for determining a performance property of a product - Google Patents

Abstract

The invention relates to a method for manufacturing products by evaluating the manufactured products with the aid of a performance property (y); comprising the following steps: - obtaining measurement data indicating a characteristic of the manufactured product corresponding to or for deriving a plurality of product parameters given as a product parameter set item (u); To obtain product parameter set items (x), the normalization model (6) being trained to distribute the input product parameter set items (x) identically; - using a given physical model to calculate the performance property (y) on the basis of the normalized product parameter set item (x) .

Description

Technical area

The present invention relates to the manufacture of products in a production line and the assessment of a performance property of the product in an end-of-line stage.

Technical background

In order to assess the performance of a product that is manufactured in a production line, measurements are made and thus measurement data sets are obtained for each of the products. The measurement data sets contain easily available data that can be used to reflect the performance of the product in the subsequent end customer application. There are usually physical models that allow a performance characteristic to be derived on the basis of the measurement data records.

For example, in cases where the product corresponds to an electric motor, current and voltage measurements can be made which allow the mechanical performance of the electric motor to be assessed according to the physical model. However, due to manufacturing tolerances, the characteristics of the product deviate from the prediction of the physical model, so that an evaluation or classification of the finished product is generally inaccurate. If a required product specification is given, this could lead to products being discarded after the manufacturing process, although they meet the requirements of the specification based on the classification window considered.

Summary of the invention

According to the present invention, a method for producing products including an evaluation of a performance property of the product according to claim 1, a device and a production system according to further independent claims are provided.

Further embodiments are given in the dependent subclaims.

• - Erhalten von Messdaten, die eine Charakteristik des hergestellten Produkts entsprechend oder zum Ableiten mehrerer Produktparameter, gegeben als ein Produktparametersatzposten, angeben;
• - Anwenden eines adaptiven datenbasierten Normalisierungsmodells auf den erhaltenen Produktparametersatzposten, um einen normalisierten Produktparametersatzposten zu erhalten, wobei das Normalisierungsmodell dahingehend trainiert wird, die eingegebenen Produktparametersatzposten identisch zu verteilen;
• - Verwenden eines gegebenen physikalischen Modells, um die Performanceeigenschaft auf Basis des normalisierten Produktparametersatzpostens zu berechnen.

According to a first aspect, a method for manufacturing products by evaluating the manufactured products with the aid of a performance property is provided; comprising the following steps:

• Obtaining measurement data indicating a characteristic of the manufactured product corresponding to or for deriving a plurality of product parameters given as a product parameter set item;
• Applying an adaptive data-based normalization model to the obtained product parameter set items to obtain a normalized product parameter set item, the normalization model being trained to distribute the input product parameter set items identically;
• Using a given physical model to calculate the performance characteristic based on the normalized product parameter set item.

Products, in particular electromechanical components, for example electric motors, electric pumps and the like, which are manufactured in a production line, are subject to manufacturing tolerances, so that the performance properties of the products comprising mechanical and electrical characteristics can vary. During or after production, measurements can be made in order to obtain electrical and / or mechanical product parameters of the product, so that a performance property can be determined on the basis of the measured product parameters by using a given physical model. For example, the mechanical power output of an electric motor can be evaluated by measuring a current, a voltage, a speed, a temperature and the like at an operating point. Due to manufacturing tolerances, the performance property can often not be reproduced precisely by the physical model used.

One idea of the above method is thus to provide a manufacturing process in which a performance property of a manufactured product can be determined on the basis of a normalization function, which can be implemented as a data-driven, trainable normalization model such as a neural network or the like. The normalization model transfers each of the measurement data sets (sets of product parameters of each product) as manufacturing parameter data set items into a common feature representation that fulfills the generalization requirements for a larger number of products and performs an intuitive interpretation of the normalized product parameter set items. With the aid of the physical model provided, the normalized product parameter item is mapped to the observed performance property.

The normalization model has the advantage that the different behaviors (indicated by the product parameter sets) of the products due to manufacturing tolerances can be mapped to generalized product behaviors (indicated by normalized product parameter sets). The generalized product behavior can be determined by the physical model can be evaluated in order to obtain the exact performance property for this product.

The normalization model can be trained with the help of an Adversarial Training Framework on the basis of a number of arbitrarily selected products, the normalization model being designed as a domain-adaptive feature extractor, with a discriminator model being trained to map the normalized product parameter set items to a random variable, which as a The indicator of the original domain (ie the origin of the product parameter set used for the training) is effective.

The task of training the normalization model consists in transferring domain-specific subsets of arbitrarily selected product parameter sets (of random products) into a common feature representation (indicated by normalized product parameter set items). The subset of selected product parameter set items represents identically distributed subsets of a number of products on the test bench, so that a corresponding performance property can be calculated for each product parameter set item.

According to one embodiment, the discriminator model can be trained to find a discrimination between each of the selected products via the normalized product parameter set items, the normalization model with minimizing such discrimination being targeted by learning to extract a more generalizable representation in the form of the normalized product parameter set items.

This provides the discriminator model that assesses the generalization through adversarial training. The discriminator model is trained to map the normalized product parameter set items to a random variable which acts as an indicator of the original domain, i.e. it indicates from which product sample the normalized product parameter set items were obtained. In the Adversarial Training Framework, the discriminator model is trained to distinguish between the selected products via the normalized product parameter set items, while the normalization model with minimizing such discrimination is targeted by learning to extract a more generalizable representation in the form of the normalized product parameter set items.

In particular, the discriminator model can be designed as a data-driven model, in particular as a neural network, such as a multilayer perceptron, as a convolutional neural network or as a recurrent neural network.

Both the normalization model and the discriminator model can be trained end-to-end using back propagation gradients of a suitable approximation, such as a cross entropy or a Wasserstein distance, with regard to the parameters (weightings in a neural network) of the discriminator model or the normalization model.

The adversary is a result of the minimization of a selected approximation by the discriminator, while the normalization model aims to maximize the approximation. The normalization model becomes domain-adaptive after training, so that it maps given product parameter set items to a common representation of normalized product parameter set items that meets the generalization requirement for a large number of products and gives an intuitive interpretation of the product parameter set items. This is because the physical model is used for the intuitive interpretation of the generalized feature representation of the normalized product parameter set items.

In order to maintain stable training behavior and accelerate convergence, the optimization of the discriminator model is limited to a restrictive family of functions, such as the 1-Lipschitz continuous family for Wasserstein discriminators, which limits the learning capacity. Such a limitation can be achieved by spectral normalization of each linear transformation within the discriminator model and by restricting non-linear activations to the 1-Lipschitz family, e.g. ReLU or logistic sigmoidal activations.

It can be provided that the types of elements of the normalized product parameter set items correspond at least partially to the types of product parameter set items.

In addition, the normalization model can be designed as a data-driven model, in particular as a neural network, such as a multilayer perceptron, as a convolutional neural network or as a recurrent neural network.

It can be provided that the normalization model and the discriminator model end-to-end via back propagation gradients of an approximation, which in particular a cross entropy between the estimated values of the original domain indicator and the actual values or a Wasserstein distance is to be trained with respect to the parameters of the discriminator model or the normalization model.

In addition, the optimization of the discriminator model can be limited to the 1-Lipschitz continuous family for Wasserstein discriminators, which is achieved in particular by a spectral normalization of each linear transformation within the discriminator model or by restricting non-linear activations to the 1-Lipschitz family.

In addition, the manufactured product is classified or discarded according to the calculated performance property.

According to a further aspect, a method for training a normalization model is provided for use in the above method, the training being trained with the aid of an adversarial training framework on the basis of a number of arbitrarily selected products, the normalization model being designed as a domain-adaptive feature extractor wherein a discriminator model is trained to map the normalized product parameter set items to a random variable that acts as an indicator of the original domain.

• - Erhalten von Messdaten, die eine Charakteristik des hergestellten Produkts entsprechend oder zum Ableiten mehrerer Produktparameter, gegeben als ein Produktparametersatzposten, angeben;
• - Anwenden eines adaptiven datenbasierten Normalisierungsmodells auf den erhaltenen Produktparametersatzposten, um einen normalisierten Produktparametersatzposten zu erhalten, wobei das Normalisierungsmodell dahingehend trainiert wird, die eingegebenen Produktparametersatzposten identisch zu verteilen;
• - Verwenden eines gegebenen physikalischen Modells, um die Performanceeigenschaft auf Basis des normalisierten Produktparametersatzpostens zu berechnen.

According to a further aspect, a device for evaluating the manufactured products with the aid of a performance property is provided, the device being designed to carry out the following steps:

• Obtaining measurement data indicating a characteristic of the manufactured product corresponding to or for deriving a plurality of product parameters given as a product parameter set item;
• Applying an adaptive data-based normalization model to the obtained product parameter set items to obtain a normalized product parameter set item, the normalization model being trained to distribute the input product parameter set items identically;
• Using a given physical model to calculate the performance characteristic based on the normalized product parameter set item.

Figure list

• 1 shows a manufacturing system for manufacturing a product, including a block diagram of a rating system in order to obtain a performance characteristic for each of the products, which allows the manufactured product to be classified or discarded or further production steps to be controlled;
• 2 shows an Advesarial Training Framework for training the normalization model; and
• 3 Figure 13 is a flow diagram illustrating the method of training the Advesarial Training Framework used to train the normalization model.

Description of embodiments

1 shows a manufacturing system 1 with manufacturing stations 2 to manufacture a product 3 in a number of process steps. There is a measuring station at the end-of-line 4th (Test bench), where the manufactured product is tested or measured in order to determine a performance property y. For example, for electromechanical devices as products, such as electric motors, pumps and the like, the performance property y can correspond to or contain a mechanical power output, a power efficiency and / or the like under a given nominal condition.

During and / or after production, parameters can be obtained to characterize the product. For each measured product, these product parameters can be obtained and provided as a product parameter set. For electrical or electromechanical devices, product parameters can include current, voltage, speed (for electric motors), temperature and the like.

Due to process variations in the manufacturing process, the characteristics of the manufactured product 3 vary in such a way that for an identical kind / an identical type of products varying performance properties y can be obtained among different products of the same kind / of the same type.

Since the measurement of a performance property y is usually expensive, an evaluation system 5 used to the manufactured product 3 using the measurement data from the intermediate and / or end-of-line measurement in the measurement station 4th evaluate (to obtain a product parameter set item for each measured product) to determine the corresponding performance characteristic y. This can be done with the aid of a known physical model using one or more known physical relationships between the parameters of the product parameter set items u and the corresponding performance property y.

In the evaluation system 5 become the physical model 7th and the normalization model 6th used. The normalization model 6th allows mapping of product parameter set items u obtained as the measurement data of a manufactured product to normalized product parameter set items x. The normalization model 6th modifies the product parameter set items u into identically distributed normalized product parameter set items x corresponding to a nominal product with nominal characteristics.

The performance property y can be used to classify the manufactured products on the basis of the correspondingly measured product parameter set items or to control further manufacturing steps on the basis of the assessed performance property y. In addition, products can be discarded if the performance property y does not meet a given specification criterion.

The structure of the normalized product parameter set items x can correspond to the structure of input product parameter set items u. For example, if there is an input voltage and an input current at the measuring station 4th are measured, the normalized product parameter set items x can also correspond to an input voltage and an input current. However, it may also be possible that the structure of normalized product parameter set items differs from the structure of product parameter set items u depending on the physical model used.

Basically this is the normalization model 6th trained to normalize the product parameter set items u in such a way that the physical model 7th can be applied to the normalized product parameter set items x to obtain the performance property y.

For training the normalization model 6th can be an advesarial training framework 10 , as in 2 shown to be used. The functionality of the Advesarial Training Framework 10 is used in conjunction with a flow chart of 3 described in more detail.

The Advesarial Training Framework 10 includes the normalization model unit 11 , the physical model unit 12th and the discriminator unit 13th including a discriminator model f (; θ _f ). The normalization model and the discriminator model f (·; θ _f ) can be designed as neural networks or any other data-driven trainable model.

In order to obtain the product-invariant product parameter set item x, the physical model h (·) of the physical model unit 12th and the Advesarial discriminator model f (·; θ _f ) of the discriminator unit 13th applied simultaneously. The training takes place in a minimax process, with the objective function being _{maximized with respect to the normalization model parameters θ g of} the normalization model g (; θ _g ) and minimized with respect to the discriminator model parameters θ _{f of} the discriminator model f (; θ _f ) and using a physical model h () is regularized.

The rating system 5 performs a procedure as detailed with the flowchart of FIG 3 described by, wherein the method can be implemented as hardware or software and executed in a data processing unit of the evaluation device.

durch Messen von Sätzen von Produktparametersatzposten u₁, u₂, ..., u_L und entsprechende Performanceeigenschaft y₁, y₂, ..., y_L für eine Anzahl von willkürlich ausgewählten Produkten bereitgestellt, wobei jeder Messdatensatz eine Anzahl von Produktparametersatzposten u, ein Domain-Kennzeichen z ∈ [0; 1]^L, das das jeweilige individuelle Produkt anzeigt, und die entsprechende Performanceeigenschaft y enthält.In step S1 are measurement data sets for a number of L or products $u\in {ℝ}^{d_u}$

by measuring sets of product parameter set items u ₁ , u ₂ , ..., u _L and corresponding performance property y ₁ , y ₂ , ..., y _L for a number of arbitrarily selected products, each measurement data set being a number of product parameter set items u , a domain identifier z ∈ [0; 1] ^L , which indicates the respective individual product, and contains the corresponding performance property y.

In step S2 it is checked whether the process of training the normalization model 11 should be completed. This can be done if convergence has been determined or if a number of training cycles have been achieved. If it turns out that the process should not be completed (alternative: NO), the process is started with step S3 continued, otherwise (alternative: YES) the process is continued with step S10 continued.

In step S3 m training samples are generated from each product parameter set item U. The training samples contain the product parameter set items u ₁ , u ₂ , ..., u _L of one or more (number m) measurements for each product, a corresponding performance property y ₁ , y ₂ , ..., y _L and an identifier z, which indicates the product domain from which the product parameter set items originate, ie ${\left\{\left(u_i,y_i,z_i\right)\right\}}_{i=1}^{m\times \mathrm{L.}}.$

aus jeder der Anzahl L von Produkten unter Verwendung eines gemeinsamen Merkmalsextraktors zu extrahieren. Der Merkmalsextraktor kann als ein herkömmliches neuronales Netzwerk wie etwa ein MLP (Multilayer-Perceptron), ein Recurrent Neural Network bis zu ebenfalls in Betracht gezogenen zeitvarianten Abhängigkeiten oder dergleichen umgesetzt werden.In step S4 becomes the normalization model 11 used to normalized q product parameter set items ${\left\{x_i\right\}}_1^{m\times \mathrm{L.}}$

from each of the L number of products using a common feature extractor. The feature extractor can be implemented as a conventional neural network such as an MLP (multilayer perceptron), a recurrent neural network up to time-variant dependencies or the like that are also considered.

In step S5 a physical model _{loss L h} (z, h (x)) with respect to the normalization model parameters θ _{g is} minimized, so that the normalization model 11 can be updated. The minimization can be achieved using a back propagation process with a gradient descent based method.

auf Basis der m Abtastwerte von jedem Teilsatz der Produkte zu erhalten.In step S6 becomes the updated normalization model for each of the L products 11 used to normalized product parameter set items ${\left\{x_i\right\}}_1^{m\times \mathrm{L.}}$

based on the m samples of each subset of the products.

(z,f(x, θ_f) bezüglich der Diskriminatormodellparameter θ_f minimiert. Dies erfolgt ebenfalls durch einen Rückausbreitungsprozess, der ein Gradientenabstiegsverfahren anwendet. Das Diskriminatormodell 13 kann als ein herkömmliches neuronales Netzwerk wie etwa ein MLP oder dergleichen ausgebildet sein.In step S7 becomes the physical loss

(z, f (x, θ _f ) is minimized with respect to the discriminator _{model parameters θ f} . This is also done by a back propagation process using a gradient descent method 13th can be designed as a conventional neural network such as an MLP or the like.

In step S8 becomes the adversarial loss with respect to the parameters of the normalization model 11 also maximized based on a gradient increase based on a back propagation process. After that, the process returns to step S2 return.

In step S10 the training process can be paused or repeated for new smaller data sets of product parameter set items from other manufactured products.

As an example, the above method can advantageously be applied to primary variable displacement pumps for a hydrostatic drive train.

For a hydrostatic drive train, a primary variable displacement pump and a secondary variable displacement motor are connected via a hydraulic circuit. The primary pump, e.g. driven by an internal combustion engine, converts mechanical energy into hydraulic energy and the secondary motor vice versa.

In axial piston pumps with swash plate control, the pump volume flow is controlled by the swivel angle of the swash plate. The swash plate is operated by a differential hydraulic actuator cylinder. Both chambers of the differential cylinder can be actuated by pressure control valves. The pump is sensitive to the pressure differential at its input ports which reduce the pivot angle if the pressure differential is high. These forces acting on the pump are called reverse control forces. If the pump should maintain the swivel angle despite changes to the pressure difference, the actuation signal to the pump must be changed accordingly. Furthermore, the restoring forces depend on the speed and the pivot angle itself.

These reverse control forces can be calculated using a physical model, but only with a certain degree of accuracy, since the pumps differ due to production series deviations. The reverse control forces can be used as a performance property. This additional sensor information is not always available, and even when such information is available, good initial values are needed.

For all pumps, certain relevant points at the end-of-line are measured as product parameter set items, although the entire dimensional space cannot be measured. By applying the above procedure, the physical model can be used in combination with detailed measurements for a low number of pumps to find a generalization so that using only the available data points at the end-of-line in combination with the model the overall behavior the restoring forces is obtained. With the adversarial domain adaptation algorithm described above, a feature extractor of the normalization model is trained in such a way that the output can be foreseen in combination with the physical model. With the additional measurement points available at the end-of-line, the training process can be repeated for fine-tuning for the specific pump, taking into account all the detailed measurement information.

At the end of a successful training session, it is expected that the normalization model 11 becomes domain-invariant in the sense that it maps given product parameter set items u onto a common representation x that meets the above-mentioned requirements. However, further fine-tuning is possible using newly obtained data samples (for example from a new pump), since the training framework is not restricted to trained subsets of the same size.

Claims (13)

Method for manufacturing products by evaluating the manufactured products with the aid of a performance property (y); comprising the following steps: - Obtaining measurement data corresponding to or relating to a characteristic of the manufactured product Deriving a plurality of product parameters given as a product parameter set item (u); - Applying an adaptive data-based normalization model to the obtained product parameter set items (u) in order to obtain a normalized product parameter set item (x), the normalization model (6) being trained to distribute the input product parameter set items (x) identically; Using a given physical model to compute the performance characteristic (y) based on the normalized product parameter set item (x). Procedure according to Claim 1 , the structure of the normalized product parameter set item (x) partially corresponding to the structure of the product parameter set item (u). Procedure according to Claim 1 or 2 , the normalization model (6) being designed as a data-driven model, in particular as a neural network, such as a multilayer perceptron, as a convolutional neural network or as a recurrent neural network. Method according to one of the Claims 1 until 3 , the normalization model (6) being trained with the aid of an Adversarial Training Framework on the basis of a number of arbitrarily selected products, the normalization model (6) being designed as a domain-adaptive feature extractor, with a discriminator model (7) being trained to identify the map normalized product parameter set item (x) to a random variable (z) which acts as an indicator of the original domain. Procedure according to Claim 4 , the discriminator model (7) being trained to find a discrimination between the selected products via the normalized product parameter set items (x), while the normalization model (6) with minimizing such discrimination is targeted by learning to extract a more generalizable representation in Shape of the normalized product parameter set items (x). Procedure according to Claim 4 or 5 , wherein the discriminator model (7) is designed as a data-driven model, in particular as a neural network, such as a multilayer perceptron, as a convolutional neural network or as a recurrent neural network. Method according to one of the Claims 4 until 6th , the normalization model (6) and the discriminator model (7) being trained end-to-end using back propagation gradients of an approximation measure, which is in particular a cross entropy or a Wasserstein distance, with respect to the parameters of the discriminator model or the normalization model. Method according to one of the Claims 4 until 7th , whereby the optimization of the discriminator model (7) is limited to the 1-Lipschitz continuous family for Wasserstein discriminators, which is achieved in particular by a spectral normalization of each linear transformation within the discriminator model (7) and by limiting non-linear activations to the 1-Lipschitz -Family is achieved. Method according to one of the Claims 1 until 8th , the manufactured product being classified or discarded according to the calculated performance property (y). Method for training a normalization model (6) for use in a method according to one of the Claims 1 until 3 with the help of an adversarial training framework based on a number of arbitrarily selected products, the normalization model (6) being designed as a domain-adaptive feature extractor, a discriminator model (7) being trained to convert the normalized product parameter set items (x) to a random variable ( z), which acts as an indicator of the original domain. Device for evaluating the manufactured products with the aid of a performance property (y), the device being designed to carry out the following steps: - obtaining measurement data indicating a characteristic of the manufactured product corresponding to or for deriving a plurality of product parameters given as a product parameter set item (u); - Applying an adaptive data-based normalization model to the obtained product parameter set items (u) in order to obtain a normalized product parameter set item (x), the normalization model (6) being trained to distribute the input product parameter set items (x) identically; Using a given physical model to compute the performance characteristic (y) based on the normalized product parameter set item (x). A computer program product comprising a computer readable medium having thereon: computer program code means, when the program is loaded, for causing the computer to perform a procedure to perform all steps of the method according to any one of the Claims 1 until 10 to execute. Machine-readable medium with a program recorded thereon, the program to cause that the computer a method according to one of the Claims 1 until 10 executes.

Images