DE102022121132A1 - Method for developing a technical component - Google Patents

Abstract

In a method for supporting a person in developing a technical component defined by a set of parameters, a computer-implemented agent (23) for machine learning through model-based reinforcement learning using a reward and an argumentation model (32) are used, and the argumentation model (32) is generated a linguistic description of proposed changes to the parameters to increase comprehensibility for a person (11).

Description

The invention relates to a method for supporting a person in the development of a technical component defined by a set of parameters and a device for carrying out such a method.

The WO 2020 / 236 665 A1 shows an identification method in which static content is provided which is used to generate dynamic content which is reproduced by voice output.

The US 2019 / 0 080 252 A1 shows voice control of a motorized robot, which is set up to have complex conversations with a human and to carry out tasks.

The US 2019 / 0 332 725 A1 shows a method for generating a design of an electrical machine.

The US 2021 / 0 402 611 A1 shows an analyzer that analyzes personal information related to a person.

The WO 2021 / 041 052 A1 shows a method for a conversational design bot user interface for accessing and manipulating system design information managed by an engineering design software application.

The WO 2018 / 071 779 A1 shows a method for computer-aided research of a similar image, text or language file with the formation of a ranking list.

This task is solved by the subject matter of the independent patent claims.

1. A) Einem computerimplementierten Agenten wird mindestens ein Zielwert einer ersten Eigenschaft eines technischen Bauteils vorgegeben, welcher computerimplementierte Agent zum maschinellen Lernen durch modellbasiertes bestärkendes Lernen unter Verwendung einer Belohnung eingerichtet ist;
2. B) ein computerimplementiertes Argumentationsmodell wird initialisiert;
3. C) der computerimplementierte Agent beobachtet seinen aktuellen Zustand, welcher aktuelle Zustand zumindest den aktuellen Parametersatz und einen aktuellen ersten Wert der ersten Eigenschaft für den aktuellen Parametersatz umfasst;
4. D) der computerimplementierte Agent führt eine Aktion am aktuellen Parametersatz aus, wobei er eine Modellvorhersage mit vorgeschlagenen Änderungen der Parameter berücksichtigt, falls diese vorhanden ist, ein sich durch die Aktion ergebender geänderter Zustand des computerimplementierten Agenten wird berechnet, und eine sich durch die Veränderung des aktuellen Zustands zum geänderten Zustand ergebende Belohnung wird berechnet;
5. E) dem Argumentationsmodell wird der geänderte Zustand des computerimplementierten Agenten zugeführt, und das Argumentationsmodell wird aktualisiert;
6. F) durch das Argumentationsmodell wird eine Modellvorhersage mit vorgeschlagenen Änderungen der Parameter für eine nachfolgende Aktion erzeugt, es wird eine linguistische Beschreibung der vorgeschlagenen Änderungen der Parameter und der hierdurch beeinflussten Zielgrößen erzeugt, und die linguistische Beschreibung wird durch eine Anzeigeeinheit oder akustisch ausgegeben, um einer Person ein Feedback über die geplante nachfolgende Aktion des computerimplementierten Agenten zu geben;
7. G) der geänderte Zustand wird zum aktuellen Zustand, und es erfolgt ein Sprung zum Schritt D).

A method for supporting a person in developing a technical component defined by a parameter set, which parameter set has a plurality of parameters, has the following steps:

1. A) A computer-implemented agent is given at least one target value of a first property of a technical component, which computer-implemented agent is set up for machine learning through model-based reinforcement learning using a reward;
2. B) a computer-implemented reasoning model is initialized;
3. C) the computer-implemented agent observes its current state, which current state includes at least the current parameter set and a current first value of the first property for the current parameter set;
4. D) the computer-implemented agent performs an action on the current set of parameters, taking into account a model prediction with proposed changes to the parameters if one exists, a changed state of the computer-implemented agent resulting from the action is calculated, and a resulting from the change in the The reward resulting from the current state to the changed state is calculated;
5. E) the reasoning model is fed the changed state of the computer-implemented agent and the reasoning model is updated;
6. F) the argumentation model generates a model prediction with proposed changes to the parameters for a subsequent action, a linguistic description of the proposed changes to the parameters and the target variables influenced thereby is generated, and the linguistic description is output by a display unit or acoustically to one to provide feedback to the person about the planned subsequent action of the computer-implemented agent;
7. G) the changed state becomes the current state and a jump to step D) occurs.

Machine learning through model-based reinforcement learning using a reward has proven to be very beneficial for the design of technical components. For the user, the process is a kind of black box, meaning it is difficult for them to understand whether the process is moving in the right direction and whether they should possibly intervene. By using the reasoning model and generating a linguistic description, the machine learning becomes much more transparent and comprehensible to the user, and the user can intervene if necessary. This leads to faster development and a better result.

• - ein mechanisches Bauteil,
• - ein elektrisches Bauteil, oder
• - ein elektrisches und mechanisches Bauteil ist.

According to a preferred embodiment, the method is carried out for a technical component, which

• - a mechanical component,
• - an electrical component, or
• - is an electrical and mechanical component.

These technical components can be optimized or developed very well using the process.

• - Antriebsstrang,
• - Kugellager,
• - Maschinenbauteil,
• - Wälzlager,
• - Kolben,
• - Kotflügel,
• - Gehäuse, und
• - Frontschürze.

According to a preferred embodiment, the method is carried out for a technical component consisting of a component group consisting of:

• - drive train,
• - Ball-bearing,
• - mechanical component,
• - Roller bearing,
• - Pistons,
• - fender,
• - housing, and
• - Front apron.

These technical components can be easily simulated.

According to a preferred embodiment, an update of the agent strategy and the argumentation model is carried out after step D. This is a preferred time to update the agent and reasoning model.

According to a preferred embodiment, in step E, images or simulation values of a simulation of the technical component with the changed parameter set are supplied to the argumentation model, in particular a deviation of the current first value from the target value. This allows the reasoning model to display an appropriate representation.

According to a preferred embodiment, a linguistic sentence generator is used in the argumentation model. Linguistic sentence generators are very suitable for generating a sentence that people can understand.

According to a preferred embodiment, the linguistic sentence generator is trained with a training data set. This can improve the quality of the sentences.

According to a preferred embodiment, the argumentation model generates a linguistic description which contains information about the planned change or non-change of at least one parameter of the parameter set and a reason as to why the at least one parameter is changed or not changed. This makes it easier for the user to understand and intervene if necessary.

According to a preferred embodiment, data input from a person is evaluated by the argumentation model, and the action on the current parameter set is influenced by the data input. This allows the person to intervene in the process and make targeted instructions.

According to a preferred embodiment, the data input is received in text form or as an acoustic input. The person can therefore work with sentences or language and come to a good result.

According to a preferred embodiment, the argumentation model is set up to enable both a description of the current state and a selection of a parameter of the current parameter set to be changed through data input. On the one hand, it is possible to intervene advantageously in the next iteration, and on the other hand, traceability and a list of the person's thoughts during development are possible.

• - Erhöhung des Werts mindestens eines vorgegebenen Parameters,
• - Erniedrigung des Werts mindestens eines vorgegebenen Parameters,
• - Nichtveränderung des Werts mindestens eines vorgegebenen Parameters.

According to a preferred embodiment, at least one action from the action group consisting of:

• - increasing the value of at least one specified parameter,
• - Reducing the value of at least one specified parameter,
• - No change in the value of at least one specified parameter.

The parameters can thus be changed in any direction, or it can be provided that no change takes place.

• - Abmessung an mindestens einer vorgegebenen Position des technischen Bauteils, insbesondere Durchmesser, Länge, Größe einer Kontaktfläche,
• - Verhältnis von mindestens zwei Abmessungen des technischen Bauteils,
• - Form des technischen Bauteils, insbesondere Form einer Kontaktfläche,
• - Material an mindestens einer vorgegebenen Position des technischen Bauteils,
• - auftretende Kraft, insbesondere Axialkraft,
• - auftretender Druck,
• - Drehzahl eines sich rotierenden Elements, und
• - dynamische Kapazität.

According to a preferred embodiment, the parameter set comprises at least one parameter from the parameter group consisting of:

• - Dimensions at at least one predetermined position of the technical component, in particular diameter, length, size of a contact surface,
• - ratio of at least two dimensions of the technical component,
• - Shape of the technical component, in particular the shape of a contact surface,
• - Material in at least one specified position of the technical component,
• - occurring force, especially axial force,
• - occurring pressure,
• - speed of a rotating element, and
• - dynamic capacity.

These parameters can be advantageously optimized.

According to a preferred embodiment, the reward can take on positive and negative values. In the event of a deterioration and resulting negative values, machine learning adapts to make such a step less likely.

• - der aktuelle erste Wert ist gleich wie oder besser als der Zielwert;
• - mit dem aktuellen Parametersatz ist ein Optimum für das technische Bauteil erreicht; und
• - eine vorgegebene Anzahl von weiteren Sprüngen zu Schritt D) hat keine Verbesserung des aktuellen ersten Werts ergeben.

According to a preferred embodiment, the method is ended in step G if at least one of the following conditions is met:

• - the current first value is equal to or better than the target value;
• - with the current parameter set, an optimum for the technical component has been achieved; and
• - a predetermined number of further jumps to step D) did not result in any improvement of the current first value.

• - maximale Drehzahl bei einem technischen Bauteil mit einem rotierenden Element,
• - auftretende mechanische Spannung unter vorgegebenen physikalischen Bedingungen,
• - auftretende Verformung unter vorgegebenen physikalischen Bedingungen,
• - Grenztemperatur für den Betrieb des technischen Bauteils,
• - erwartete Lebensdauer des technischen Bauteils,
• - Lebensdauer bei vorgegebener Verunreinigung eines Schmierstoffs des technischen Bauteils durch Verschmutzung,
• - Lebensdauer bei vorgegebener Verunreinigung eines Schmierstoffs des technischen Bauteils durch Feuchtigkeit,
• - Lebensdauer bei vorgegebenen mechanischen Schwingungen,
• - Schmierbedingung bei einem Lager, und
• - dynamische Tragzahl bei einem sich bewegenden technischen Bauteil.

According to a preferred embodiment, the first property comprises at least one property from a property group consisting of:

• - maximum speed for a technical component with a rotating element,
• - mechanical stress that occurs under given physical conditions,
• - deformation occurring under given physical conditions,
• - Limit temperature for the operation of the technical component,
• - expected service life of the technical component,
• - Service life with a given contamination of a lubricant of the technical component due to contamination,
• - Service life with a given contamination of a lubricant of the technical component by moisture,
• - Service life with given mechanical vibrations,
• - Lubrication condition for a bearing, and
• - dynamic load capacity for a moving technical component.

These are first properties that can be used advantageously for such a construction.

• - mindestens eine linguistische Beschreibung erzeugt und ausgegeben, oder
• - mindestens eine grafische Darstellung ausgegeben. Die Person kann damit den aktuellen Zustand besser beurteilen.

According to a preferred embodiment, step F becomes the current state

• - at least one linguistic description is generated and output, or
• - at least one graphical representation is output. This allows the person to better assess the current condition.

According to a preferred embodiment, the graphical representation is adjusted using a graphical neural network encoder depending on the model prediction. This also allows originally static images to be adapted to a planned change.

According to a preferred embodiment, the linguistic description of the changes in the parameters in step F is converted into the model prediction by a computer-implemented decoder. This enables the conversion of linguistic descriptions generated by the argumentation model as well as linguistic descriptions entered by the person and further processing by the agent.

A device is set up to carry out such a method.

• 1 in schematischer Darstellung eine Vorrichtung und eine Person,
• 2 eine Anzeige mit einer ersten Information,
• 3 ein Flussdiagramm für ein Verfahren, und
• 4 in schematischer Darstellung einen Aufbau eines Argumentationsmodells.

Further details and advantageous developments of the invention result from the exemplary embodiments described below and shown in the drawings, which are in no way to be understood as a limitation of the invention, and from the subclaims. It is understood that the features mentioned above and those to be explained below can be used not only in the combination specified in each case, but also in other combinations or alone, without departing from the scope of the present invention. It shows:

• 1 a schematic representation of a device and a person,
• 2 an ad with initial information,
• 3 a flowchart for a procedure, and
• 4 a schematic representation of the structure of an argumentation model.

In the following, identical or identically acting parts are provided with the same reference numbers and are usually only described once. The description builds on each other across characters in order to avoid unnecessary repetition.

1 shows a device 20 and a person 11.

The computer-implemented device 20 has an input device 21 and an output device 22. The input device 21 includes, for example, a keyboard, a mouse, a joystick and/or a microphone, and the output device 22 includes, for example, a display unit such as a monitor or screen and/or a speaker .

The device 20 has a computer-implemented agent 23, an environment 30 and a computer-implemented reasoning model 32. Data can be transmitted from the input device 21 to the environment 30 via a data line 24, data from the environment 30 to the argumentation model 32 via a data line 25, data from the environment 30 to the agent 23 via a data line 26, and 27 data via a data line from the argumentation model 32 to the agent 23, via a data line 28 data from the input device 21 to the argumentation model 32, via a data line 29 data from the argumentation model 32 to the output device 22 and via a data line 31 data from the agent 23 to the environment 30. The environment 30 can also be referred to as the state of the agent 23, which is transmitted to the agent 23 via the line 26.

The computer-implemented agent 23 is set up for machine learning through model-based reinforcement learning and is intended to support the person 11 in developing a technical component defined by a set of parameters. The technical component is preferably a mechanical and/or electrical component.

The models used in reinforcement learning are called simulators. For example, the behavior of the component and given conditions are simulated using the simulators. This can be done, for example, on a physical basis with physical/mathematical equations or on a data-supported basis, in particular with neural networks. Another approach can use a Markov decision process as a model. A preferred algorithm for model-based reinforcement learning is the DYNA algorithm.

The agent 23 has a first parameter set 41 and preferably a second parameter set 42, which define the technical component and are optimized by the agent 23.

• - Abmessung an mindestens einer vorgegebenen Position des technischen Bauteils, insbesondere Durchmesser, Länge, Größe einer Kontaktfläche,
• - Verhältnis von mindestens zwei Abmessungen des technischen Bauteils,
• - Form des technischen Bauteils, insbesondere Form einer Kontaktfläche,
• - Material an mindestens einer vorgegebenen Position des technischen Bauteils,
• - auftretende Kraft, insbesondere Axialkraft,
• - auftretender Druck,
• - Drehzahl eines sich rotierenden Elements, und
• - dynamische Kapazität.

The parameter set preferably contains at least one parameter from the parameter group consisting of:

• - Dimension at at least one predetermined position of the technical component, in particular diameter, length, size of a contact surface,
• - ratio of at least two dimensions of the technical component,
• - Shape of the technical component, in particular the shape of a contact surface,
• - Material in at least one specified position of the technical component,
• - occurring force, especially axial force,
• - occurring pressure,
• - speed of a rotating element, and
• - dynamic capacity.

The technical component can be, for example, a machine component, in particular a rolling bearing or a piston. The technical component can also generally be a technical vehicle component, e.g. a fender, a housing, a piston or a front apron.

2 shows an example of first information 61, which is output on the output device 22 designed as a display unit. The first information 61 has a virtual representation 62 of the technical component. In the exemplary embodiment, the technical component is a ball bearing or rolling bearing, which is to be designed for a given purpose. A first set of parameters defines a first state of the technical component, which does not necessarily have all the properties required for use.

A diagram 64 shows three curves 71, 72, 73. The abscissa marked x shows, by way of example, the ratio of the axial extent 81 of the ball bearing to the inner diameter 82 of the inner ring of the ball bearing. In the exemplary embodiment, the ordinate y shows the mechanical stress that occurs at a given operating point of the ball bearing (e.g. speed, temperature and axial forces). The curve 73 shows the specified target value Z, which should at least be achieved by the ball bearing in its final state. The curve 71 shows the maximum occurring mechanical stress at a parameter I1 for the axial extent 81 according to the first parameter set 41, and the curve 72 shows the corresponding maximum occurring mechanical stress at an axial extent 81 with the value I2 changed compared to I1.

The arrow 63 points to a parameter to be changed.

Further parameters P1 to P9 are displayed and can be selected by the person 11. The person can request a suggestion from the agent 23 for the other parameters via the user interface 21.

• - „Die Steifigkeit ist gut, und daher sollten die Parameter P3 und P8 nicht verändert werden.“
• - „Die Abweichung war groß im Bereich einer Drehzahl unter 1.000 min^-1, und daher sollte der Parameter P7 verringert werden.“
• - „Der Fehler wird höher, wenn ich die Werte des Parameters P5 um 10% erniedrige, und daher sollte eine Erniedrigung um 5% erfolgen.“

A field 65 contains a text 66, which text contains a linguistic description generated by the argumentation model 32 of information about an intended change or non-change of certain parameters of the parameter set. Examples for this are:

• - “The stiffness is good and therefore the parameters P3 and P8 should not be changed.”
• - “The deviation was large in the range of speeds below 1,000 ^rpm , and therefore the parameter P7 should be reduced.”
• - “The error will be higher if I decrease the values of the P5 parameter by 10%, and therefore a decrease of 5% should occur.”

The text 66 therefore preferably contains both a comment on the current situation and a suggestion for changing or not changing a parameter for improvement.

• - „Eine Erhöhung des Werts der Formel 1 führt zu einer höheren Lebensdauer des Lagers.“

Text 66 can also only contain a statement about the change, e.g.:

• - “Increasing the value of Formula 1 results in a longer bearing life.”

The text is generated by the device 20 or by the argumentation model 32 preferably with the help of a linguistic sentence generator.

The linguistic sentence generator is preferably trained at an earlier point in time with a training data set.

A network with a long short-term memory is particularly preferably used for this purpose (English: LSTM network or Long Short-Term Memory network), but other linguistic sentence generators are also possible.

For example, an increase, a decrease or a stay the same can be suggested for the respective parameter.

The strategy is adjusted in machine learning using a reward, and the reward can be chosen, for example, depending on the distance of the curve 72 to the curve 73 or depending on a comparison of the distance between the curve 71 and the curve 73 and the distance between the curve 72 and the curve 73. The agent 23 preferably minimizes the deviation between the current value and the specified target value. The reward can preferably take on positive and negative values, so in the event of a deterioration, the reward corresponds to a punishment that is received by the learning agent 23.

The person 11 can now either observe how the agent 23, together with the argumentation model 32, carries out the optimization described by the text 66, or he can intervene via the input device 21 and enter a comment and a suggestion for a change in text form in field 65 Enter parameters. The input can be made, for example, via a keyboard or by voice input. The device 20 preferably contains a decoder which is capable of decoding a text entered by the person 11 and from this generating information that can be understood by the agent for the desired modification of the selected parameter. For this purpose, the decoder preferably carries out a semantic analysis of the text.

These options enable the person 11 to clearly understand and understand the work of the agent 23 and the argumentation model 32 and to intervene if necessary. Using the linguistic description generated by the reasoning model allows for significantly better understanding compared to simply displaying a diagram. On the one hand, the device 20 develops the technical component and, on the other hand, it describes the incremental improvement using words and sentences that are understandable to humans. This greatly facilitates the required training of the person 11 in the use of the device 20, and the person 11 requires less specialist knowledge for the development. On the other hand, people 11 with appropriate specialist knowledge can understand the steps in the device 20 and, if necessary, influence the device 20. This can lead to a better technical result, and the technical result can be achieved more quickly.

The device 20 thereby enables an incremental improvement, and the agent 23 carries out corresponding actions on the environment 30, for example by changing parameters of a parameter set. This changes the overall state of the agent 23, and a reward can be calculated for the change and passed to the agent 23 for further processing.

• - Erhöhung des Werts mindestens eines vorgegebenen Parameters,
• - Erniedrigung des Werts mindestens eines vorgegebenen Parameters,
• - Nichtveränderung des Werts mindestens eines vorgegebenen Parameters.

At least one action from the action group consisting of:

• - increasing the value of at least one specified parameter,
• - Reducing the value of at least one specified parameter,
• - No change in the value of at least one specified parameter.

Preferably, the agent 23 can determine whether the current parameter set 42 is used as a result of the optimization. Alternatively or additionally, the person 11 can preferably determine by input via the input device 21 that the current parameter set 42 is accepted as the final parameter set. Person 11 has now reached the goal of development, and the technical component can then be built and used based on the final set of parameters. Depending on the final parameter set, CAD files can also be generated for the technical component. CAD means computer-aided design, and the data can be used, for example, for an additive printing process or for the design of an injection mold.

3 shows a flowchart with a method for supporting the person 11 in the development of a technical component defined by a parameter set, which parameter set has a plurality of parameters.

The method has steps A, B, C, D, E, F, G and preferably H.

In step A, the computer-implemented agent 23 is given at least one target value Z of a first property of the technical component. The computer-implemented agent 23 is set up for machine learning through model-based reinforcement learning using a reward.

• - maximale Drehzahl bei einem technischen Bauteil mit einem rotierenden Element,
• - auftretende mechanische Spannung unter vorgegebenen physikalischen Bedingungen,
• - auftretende Verformung unter vorgegebenen physikalischen Bedingungen,
• - Grenztemperatur für den Betrieb des technischen Bauteils,
• - erwartete Lebensdauer des technischen Bauteils,
• - Lebensdauer bei vorgegebener Verunreinigung eines Schmierstoffs des technischen Bauteils durch Verschmutzung,
• - Lebensdauer bei vorgegebener Verunreinigung eines Schmierstoffs des technischen Bauteils durch Feuchtigkeit,
• - Lebensdauer bei vorgegebenen mechanischen Schwingungen,
• - Schmierbedingung bei einem Lager, und
• - dynamische Tragzahl bei einem sich bewegenden technischen Bauteil.

The first property preferably comprises at least one property from a property group consisting of:

• - maximum speed for a technical component with a rotating element,
• - mechanical stress that occurs under given physical conditions,
• - deformation occurring under given physical conditions,
• - Limit temperature for the operation of the technical component,
• - expected service life of the technical component,
• - Service life with a given contamination of a lubricant of the technical component due to contamination,
• - Service life with a given contamination of a lubricant of the technical component by moisture,
• - Service life with given mechanical vibrations,
• - Lubrication condition for a bearing, and
• - dynamic load capacity for a moving technical component.

These properties enable advantageous optimization.

In step B, the argumentation model 32 is initialized. For example, it receives its own data, data from the agent and, if necessary, additional data that was generated by learning on learning data.

In step C, the computer-implemented agent 23 observes its current state, which current state includes at least the current parameter set and a current first value of the first property for the current parameter set. Your own state is stored in the environment 30, for example.

In step D, the computer-implemented agent 23 performs an action on the current set of parameters, taking into account a model prediction with proposed changes to the parameters, if present, a changed state of the computer-implemented agent resulting from the action is calculated, and a resulting from the Reward resulting from changing the current state to the changed state is calculated. Such an action can also be viewed in machine learning theory as an action by the agent 23 on the environment 30, and the impact on the environment 30 by this action as well as the reward are calculated.

In step E, the changed state of the computer-implemented agent 23 is fed to the argumentation model 32, and the argumentation model 32 is updated. The argumentation model 32 can therefore work with the changed state.

Preferably, in step E, the argumentation model 32 is supplied with images or simulation values of a simulation of the technical component with the changed parameter set, in particular a deviation of the current first value from Target value Z. These images can then be displayed or changed by the argumentation model 32 in step F.

In step F, the argumentation model 32 generates a model prediction with proposed changes to the parameters for a subsequent action, a linguistic description of the proposed changes to the parameters and the target variables influenced thereby is generated, and the linguistic description is output by a display unit or acoustically, to provide feedback to a person about the computer-implemented agent's planned subsequent action. By generating a model prediction, the argumentation model 32 thus makes a suggestion as to how the optimization of the parameters of the parameter set should proceed in the next step. The proposed changes are described linguistically, i.e. the changes are described using human language and are therefore easily understandable for the person 11.

• - mindestens eine linguistische Beschreibung erzeugt und ausgegeben, oder
• - mindestens eine grafische Darstellung ausgegeben.

Preference is given to the current state in step F

• - at least one linguistic description is generated and output, or
• - at least one graphical representation is output.

The person 11 can thus assess the current condition.

Preferably, the graphical representation is adapted using an encoder for a graphical neural network depending on the model prediction. This customization allows originally static graphical representations to be adjusted depending on planned changes.

Preferably, the linguistic description of the changes in the parameters in step F is converted into the model prediction by a computer-implemented decoder. It is therefore possible to convert a human-understandable text into a machine-understandable instruction for the next iteration, in particular also to convert a text entered by the person 11.

In step G, the changed state becomes the current state and a jump to step D occurs. The transfer from the changed state to the new current state can also take place earlier or - later in the program flow.

• - der aktuelle erste Wert ist gleich wie oder besser als der Zielwert Z,
• - mit dem aktuellen Parametersatz ist ein Optimum für das technische Bauteil erreicht, und
• - eine vorgegebene Anzahl von weiteren Sprüngen zu Schritt D hat keine Verbesserung des aktuellen ersten Werts ergeben.

The method is preferably ended in step G) by jumping to step H) if at least one of the following conditions is met:

• - the current first value is equal to or better than the target value Z,
• - with the current parameter set, an optimum for the technical component has been achieved, and
• - a predetermined number of further jumps to step D did not result in any improvement to the current first value.

These conditions enable the process to be terminated if it no longer makes sense to continue.

Data input from a person 11 is preferably evaluated by the argumentation model 32, and the action on the current parameter set is influenced by the data input. Person 11 can therefore influence the iterative process.

Data entry is preferably received in text form or as acoustic input. This allows intuitive use and leads to better technical results.

The argumentation model 32 is preferably set up to enable both a description of the current state and a selection of a parameter of the current parameter set to be changed through the data input.

4 shows a preferred structure of the argumentation model 32.

For example, four inputs 101, 102, 103, 104 are supplied to the argumentation model 32.

In the exemplary embodiment, the inputs 101, 102 are design formulas or development formulas, for example for the torque as a function of a parameter or for the torque-power ratio in a motor. In general, values of a simulation of a current development result or a current state of the agent (for example speed, torque, forces, strength) and, if necessary, parameters for development formulas, in particular a deviation between a target value and a current value, are supplied. Only some of the parameters mentioned can also be supplied.

The inputs 103, 104 are, for example, images or representations of a simulation of a current development result, in the exemplary embodiment a diagram in which the radial stiffness depends on the speed for the target value and for the current para meter is specified. The images can of course also concern other properties, for example the torque depending on a parameter.

The argumentation model preferably receives data sets with the images and an associated linguistic annotation. For example, the linguistic annotations are added by developers and they describe the current situation and the subsequent action that the developer wants to take. This could be, for example: "The deviation was large in the speed range below 1,000 ^rpm , and therefore the values in the F1 formula should be reduced." On the one hand, this enables a developer to have an influence, and on the other hand, the argumentation model can be based on this Learn input for future runs.

An encoder 120 with a graphic neural network is preferably provided, to which the inputs 101, 102 are fed via an agent encoder and the inputs 103, 104 are fed via an image encoder. A graphic neural network enables images to be influenced depending on the information available.

The output of the encoder 120 is fed to an encoder 122 as input. On the one hand, the encoder 122 generates a description of the proposed change in the agent's state via a linguistic sentence generator 122, for example "Increasing the values of the formula F1 leads to a longer life expectancy of the plain bearing."

The encoder 122 is preferably trained using the developer's linguistic notes. The description is fed to an action decoder 128. The action decoder 128 may also be supplied directly with parameters from the encoder 122, for example the initial state of the agent. The action decoder generates changed parameters for actions on the agent's current parameter set depending on the input.

An attention computation is preferably carried out in the action decoder 128 in order to generate an assignment to a dependency of events from the output of the linguistic sentence generator and to decode these into actions.

Naturally, a variety of variations and modifications are possible within the scope of the present invention.

QUOTES INCLUDED IN THE DESCRIPTION

This list of documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• WO 2020236665 A1 [0002]
• US 20190080252 A1 [0003]
• US 20190332725 A1 [0004]
• US 20210402611 A1 [0005]
• WO 2021041052 A1 [0006]
• WO 2018071779 A1 [0007]

Claims (19)

Method for supporting a person in the development of a technical component defined by a parameter set, which parameter set has a plurality of parameters, which method has the following steps: A) a computer-implemented agent (23) is given at least one target value (Z) of a first property of a technical component, which computer-implemented agent (23) is set up for machine learning through model-based reinforcement learning using a reward; B) a computer-implemented argumentation model (32) is initialized; C) the computer-implemented agent (23) observes its current state, which current state includes at least the current parameter set and a current first value of the first property for the current parameter set; D) the computer-implemented agent (23) performs an action on the current set of parameters, taking into account a model prediction with proposed changes to the parameters, if this exists, a changed state of the computer-implemented agent resulting from the action is calculated, and a resultant the reward resulting from the change from the current state to the changed state is calculated; E) the changed state of the computer-implemented agent (23) is fed to the argumentation model (32), and the argumentation model (32) is updated; F) the argumentation model (32) generates a model prediction with proposed changes to the parameters for a subsequent action, a linguistic description of the proposed changes to the parameters and the target variables influenced thereby is generated, and the linguistic description is output by a display unit or acoustically to provide feedback to a person about the computer-implemented agent's planned subsequent action; G) the changed state becomes the current state and a jump to step D) occurs. Procedure according to Claim 1 , which is carried out for a technical component, which is - a mechanical component, - an electrical component, or - an electrical and mechanical component. Procedure according to Claim 1 or 2 , which is carried out for a technical component consisting of a component group consisting of: - drive train, - ball bearings, - machine component, - roller bearings, - pistons, - fenders, - housing, and - front apron. Method according to one of the preceding claims, in which in step E) the argumentation model (32) is supplied with images or simulation values of a simulation of the technical component with the changed parameter set, in particular a deviation of the current first value from the target value (Z). Method according to one of the preceding claims, in which a linguistic sentence generator is used in the argumentation model (32). Procedure according to Claim 5 , in which the linguistic sentence generator is trained with a training data set. Method according to one of the preceding claims, in which a linguistic description is generated by the argumentation model (32), which contains information (66) about the planned change or non-change of at least one parameter of the parameter set and a reason as to why the at least one parameter was changed is or is not changed. Method according to one of the preceding claims, in which data input from a person is evaluated by the argumentation model (32), and in which the action on the current parameter set is influenced by the data input. Procedure according to Claim 8 , in which data input is received in text form or as acoustic input. Procedure according to Claim 8 or 9 , in which the argumentation model (32) is set up to enable both a description of the current state and a selection of a parameter of the current parameter set to be changed through data input. Method according to one of the preceding claims, in which at least one action from the action group consisting of: - increasing the value of at least one specified parameter, - Reducing the value of at least one specified parameter, - No change in the value of at least one specified parameter. Method according to one of the preceding claims, in which the parameter set comprises at least one parameter from the parameter group consisting of: - Dimension at at least one predetermined position of the technical component, in particular diameter, length, size of a contact surface, - Ratio of at least two dimensions of the technical component , - shape of the technical component, in particular shape of a contact surface, - material at at least one predetermined position of the technical component, - force occurring, in particular axial force, - pressure occurring, - speed of a rotating element, and - dynamic capacity. Method according to one of the preceding claims, in which the reward can take on positive and negative values. Method according to one of the preceding claims, in which the method is ended in step G) when at least one of the following conditions is met: - the current first value is equal to or better than the target value (Z); - with the current parameter set, an optimum for the technical component has been achieved; and - a predetermined number of further jumps to step D) did not result in any improvement of the current first value. Method according to one of the preceding claims, in which the first property comprises at least one property from a property group consisting of: - maximum speed for a technical component with a rotating element, - mechanical stress that occurs under given physical conditions, - deformation occurring under given physical conditions, - Limit temperature for the operation of the technical component, - expected service life of the technical component, - Service life with a given contamination of a lubricant of the technical component due to contamination, - Service life with a given contamination of a lubricant of the technical component by moisture, - Service life with given mechanical vibrations, - Lubrication condition for a bearing, and - dynamic load capacity for a moving technical component. Method according to one of the preceding claims, in which in step F) to the current state - at least one linguistic description is generated and output, or - at least one graphical representation is output. Procedure according to Claim 16 , in which the graphical representation is adjusted using a graphical neural network encoder depending on the model prediction. Method according to one of the preceding claims, in which the linguistic description of the changes in the parameters in step F) is converted into the model prediction by a computer-implemented decoder. Device (20) which is set up to carry out a method according to one of the preceding claims.

Images