DE102022132797A1 - Computer-implemented method for calibrating an active system - Google Patents

Abstract

The invention relates to a computer-implemented method for calibrating an active system (1), wherein at least the following steps are carried out:a. providing measurement data from the operation of an active system (1);b. capturing and recording measurement values (3) from the measurement data;c. comparing recorded measurement values (3) and at least one associated target value, and finding a first change suggestion (16) and an associated justification (17) if a deviation from the associated target value is determined;d. based on the first change suggestion (16) and the justification (17), outputting the change suggestion (16) and the justification (17) in the form of a human-understandable phrase (5);e. based on the first change suggestion (16), generating a change value (10) for the operation of the active system (1);f. based on a phrase (8) provided by a calibration user (7), recognizing relevant information containing a second change proposal (19) and an associated justification (20); andg. based on the phrases (5,8), determining a similarity (13) between the change proposals (16,19) and justifications (17,20); andh. generating a weighting (22) based on the similarity (13), wherein the change value (10) is output to the active system (1) weighted by means of the weighting (22).With the method proposed here, an active system can be calibrated using AI and human control.

Description

The invention relates to a computer-implemented method for calibrating an active system. The invention further relates to a computer program comprising a computer program code, as well as a computer program product with such a computer program on which a computer program code is stored.

It is already known to calibrate highly complex active systems, such as an electrified drive train including an electric drive motor and traction battery of a motor vehicle, using so-called artificial intelligence [Kl]. This often results in very good and not infrequently unexpected solutions. So-called self-reinforcing learning has proven to be particularly effective, in which the algorithm receives a reward (or punishment) (for example as a factor between -1 [minus one] and +1 [plus one]) when a target value is reached or approached. Furthermore, so-called model-based self-reinforcing learning has proven to be an advantageous approach. This is also known as a simulator. For example, a dynamic environmental condition is simulated in great detail for the active system to be calibrated, and this can be done physically or purely data-driven. Another advantageous approach is to use the Markov decision process as a model.

The disadvantage of these approaches using AI is that the processes are not comprehensible to a human calibration user in detail and often also in the larger picture. This can lead to problems with regard to trust in the calibration results and during the calculations the AI can take the wrong path due to artifacts.

Based on this, the present invention is based on the object of at least partially overcoming the disadvantages known from the prior art. The features according to the invention arise from the independent claims, for which advantageous embodiments are shown in the dependent claims. The features of the claims can be combined in any technically reasonable manner, whereby the explanations from the following description and features from the figures can also be used, which include additional embodiments of the invention.

• - zumindest ein Maschinen-Lern-Algorithmus, welcher zumindest zum Erfassen und Aufzeichnen von ausgewählten Messwerten und zum Vergleichen mit Zielwerten, sowie zum Finden von Änderungsvorschlägen und zugehörigen Begründungen auf Basis von erfassten Abweichungen eingerichtet ist;
• - ein erster Natural-Language-Prozessor zum Erzeugen von menschenverständlichen ersten Phrasen aus bereitgestellten Änderungsvorschlägen und zugehörigen Begründungen;
• - ein zweiter Natural-Language-Prozessor zum Erkennen von relevanten computerverständlichen Informationen aus zumindest einer von einem menschlichen Kalibrier-Nutzer bereitgestellten menschlichen zweiten Phrase;
• - ein erster Software-Agent, welcher mittels Erzeugen von Änderungswerten zum Minimieren von Abweichungen zwischen Zielwerten und Messwerten eingerichtet ist;
• - ein zweiter Software-Agent, welcher mithilfe eines Natural-Language-Modells zum Ermitteln von Ähnlichkeiten zwischen Phrasen mit computerverständlichen Informationen eingerichtet ist; wobei zumindest die folgenden Schritte einer ersten Gruppe ausgeführt werden:
  1. a. mittels einer Messeinrichtung, Bereitstellen von komplexen Messdaten von einem Betrieb eines Aktivsystems;
  2. b. mittels zumindest eines der Maschinen-Lern-Algorithmen, Erfassen und Aufzeichnen von zumindest einer Teilmenge von Messwerten aus den in Schritt a. bereitgestellten Messdaten;
  3. c. mittels zumindest eines der Maschinen-Lern-Algorithmen, Vergleichen von aktuell aufgezeichneten Messwerten und zumindest einem zugehörigen vorbestimmten Zielwert, sowie Finden zumindest eines ersten Änderungsvorschlags und jeweils einer zugehörigen ersten Begründung, wenn eine Abweichung von dem zugehörigen vorbestimmten Zielwert festgestellt wird;
  4. d. mittels des ersten Natural-Language-Prozessors auf Basis des ersten Änderungsvorschlags und der zugehörigen ersten Begründung gemäß Schritt c., Ausgeben von dem zumindest einen ersten Änderungsvorschlag und der jeweils zugehörigen ersten Begründung in Form von zumindest einer menschenverständlichen ersten Phrase; und
  5. e. mittels des ersten Software-Agenten auf Basis des zumindest einen in Schritt d. ausgegebenen ersten Änderungsvorschlags, Erzeugen eines zugehörigen Änderungswerts zum Betrieb des zu kalibrierenden Aktivsystems, wobei weiterhin zumindest die folgenden Schritte einer zweiten Gruppe ausgeführt werden:
  6. f. mittels des zweiten Natural-Language-Prozessors auf Basis von zumindest einer durch einen Kalibrier-Nutzer bereitgestellten zweiten Phrase, Erkennen von zumindest einer für die bereitgestellten Messwerte relevanten computerverständlichen Information, enthaltend zumindest einen zweiten Änderungsvorschlag und jeweils eine zugehörige zweite Begründung;
  7. g. mittels des zweiten Software-Agenten auf Basis der zumindest einen in Schritt d. erzeugten ersten menschenverständlichen Phrase und der zumindest einen in Schritt f. bereitgestellten zweiten Phrase, Ermitteln von einer Ähnlichkeit von den und/oder von jeweils zwei gleichen enthaltenen Änderungsvorschlägen und jeweils zugehörigen Begründungen; und
  8. h. auf Basis von vorbestimmten Vorgaben und/oder mittels zumindest eines der Maschinen-Lern-Algorithmen, Erzeugen einer oder jeweils einer zugehörigen Gewichtung auf Basis der in Schritt g. ermittelten Ähnlichkeit,
  wobei der in Schritt e. erzeugte Änderungswert mittels der zugehörigen in Schritt h. erzeugten Gewichtung gewichtet an das zu kalibrierende Aktivsystem ausgegeben wird.

The invention relates to a computer-implemented method for calibrating an active system, in which at least the following software components are used:

• - at least one machine learning algorithm, which is set up at least to capture and record selected measured values and to compare them with target values, as well as to find suggestions for changes and associated justifications on the basis of recorded deviations;
• - a first natural language processor for generating human-understandable first phrases from provided change suggestions and associated justifications;
• - a second natural language processor for recognizing relevant computer-intelligible information from at least one human second phrase provided by a human calibration user;
• - a first software agent configured to minimize deviations between target values and measured values by generating change values;
• - a second software agent configured to use a natural language model to determine similarities between phrases containing computer-understandable information; wherein at least the following steps of a first group are carried out:
  1. a. by means of a measuring device, providing complex measurement data from the operation of an active system;
  2. b. by means of at least one of the machine learning algorithms, capturing and recording at least a subset of measurement values from the measurement data provided in step a.;
  3. c. by means of at least one of the machine learning algorithms, comparing currently recorded measured values and at least one associated predetermined target value, and finding at least one first change proposal and in each case an associated first justification if a deviation from the associated predetermined target value is detected;
  4. d. using the first natural language processor on the basis of the first change proposal and the associated first justification according to step c., outputting the at least one first change proposal and the respective associated first justification in the form of at least one human-intelligible first phrase; and
  5. e. by means of the first software agent on the basis of the at least one first change proposal issued in step d., generating an associated change value for the operation of the active system to be calibrated, wherein at least the following steps of a second group are also carried out:
  6. f. by means of the second natural language processor on the basis of at least one second phrase provided by a calibration user, recognizing at least one piece of computer-understandable information relevant to the measurement values provided, containing at least one second change proposal and an associated second justification;
  7. g. by means of the second software agent on the basis of the at least one first human-intelligible phrase generated in step d. and the at least one second phrase provided in step f., determining a similarity between the and/or two identical proposed changes and the respective reasons; and
  8. h. on the basis of predetermined specifications and/or by means of at least one of the machine learning algorithms, generating one or each associated weighting based on the similarity determined in step g.,
  wherein the change value generated in step e. is output to the active system to be calibrated in a weighted manner by means of the associated weighting generated in step h.

Unless explicitly stated otherwise, ordinal numbers used in the preceding and following descriptions are used only to clearly distinguish them and do not reflect any order or ranking of the designated components. An ordinal number greater than one does not necessarily mean that another such component must be present.

The computer-implemented method proposed here enables a calibration user to influence the course of the calculations for calibrating an active system. However, it cannot be ruled out that the method may lead the calibration user to take paths for calibrating the active system that may seem absurd. This is achieved by means of weighting.

A machine learning algorithm is an algorithm that is able to make certain decisions in a complex situation based on a previously learned (frozen) or evolving process. For example, at least one machine learning algorithm is used here, which is set up to collect and record measurement data and compare it with a target value. A machine learning algorithm is also provided here, which is set up to find suggestions for changes and an associated justification. A model-based algorithm is preferably used for this purpose.

Natural language processors are well known and already very reliable. For example, they are used to generate texts, understand spoken language and read texts aloud and simulate human phrases in a way that is perceived as extremely natural. Well-known applications include the search engine or email filter from Google®, the voice control from Apple® Siri® or Amazon® Alexa®. Here, a first natural language processor is used, which is set up to generate at least one (first) phrase from computer-based data that is understandable to a human, for example in the form of a written or spoken sentence or in the form of bullet points in a selected language. Furthermore, a second natural language processor is used, which is set up to extract information that is understandable for a computer or an algorithm from a (second) phrase in human language, for example in the form of a written or spoken sentence or in the form of bullet points in a selected language.

Software agents are machine learning algorithms that are capable of solving given tasks independently. For example, a so-called neural network, such as a graphical neural network [GNN], is used for this purpose. In one embodiment, such a software agent works on a model-based basis; that is, physical relationships are already firmly implemented, for example in the form of a look-up table [LuT] (with, for example, empirically determined values) and/or in the form of mathematical formulas that represent reality to a good approximation. Here, a first software agent is used, which is set up to advance the calibration process by generating a change value based on previously determined measurement data. Such a change value is, for example, a factor or a replacement for the previous value. Furthermore, a second software agent is used here, which is set up to determine a similarity between phrases or the information contained therein. In one embodiment, a natural language model is used. Part of the second software agent. Alternatively, the information of the phrases in question has already been prepared, in the case of the calibration user's second phrase using the second natural language processor or, in the case of the first phrase, at a stage before it was generated, i.e. on the basis of the intermediate results containing the desired information, which are themselves understandable to computers but not necessarily understandable to humans.

• In einem Schritt a. werden mittels einer Messeinrichtung, welche beispielsweise Bestandteil eines Aktivsystems ist oder durch Sensoren eines Aktivsystems gespeist ist, komplexe Messdaten von einem Betrieb eines Aktivsystems bereitgestellt. Solche Messdaten umfassen mehrere Datenarten als Messwerte, bei einem elektrifizierten Antriebsstrang beispielsweise betreffend eine Drehzahl, eine Spannung, einen Strom und anderes.
• In einem Schritt b. wird mittels zumindest eines der Maschinen-Lern-Algorithmen zumindest eine Teilmenge von Messwerten der in Schritt a. bereitgestellten Messdaten erfasst und aufgezeichnet. In einer Ausführungsform sind diese Messwerte bereits in einen physikalischen Zusammenhang gebracht, beispielsweise in Form von einer mathematischen Formel und/oder einem Diagramm.
• In einem Schritt c. werden mittels zumindest eines der Maschinen-Lern-Algorithmen aktuell aufgezeichnete Messwerte und zumindest ein zugehöriger vorbestimmter Zielwert miteinander verglichen. Weiterhin wird zumindest ein erster Änderungsvorschlags und jeweils eine zugehörige erste Begründung gefunden, wenn von dem betreffenden Maschinen-Lern-Algorithmus eine Abweichung von dem zugehörigen vorbestimmten Zielwert festgestellt wird. Dem betreffenden Maschinen-Lern-Algorithmus liegt dazu ein Begründungsmodell vor, beispielsweise aufgrund von einem physikalischen Modell und/oder einem erlernten Zusammenhang. Der zumindest eine (erste) Änderungsvorschlag ist für die Verwertung mittels des ersten Software-Agenten eingerichtet.
• In einem Schritt d. werden mittels des ersten Natural-Language-Prozessors auf Basis des ersten Änderungsvorschlags und der zugehörigen ersten Begründung gemäß Schritt c. der zumindest eine erste Änderungsvorschlag und die jeweils zugehörige erste Begründung in Form von zumindest einer menschenverständlichen ersten Phrase ausgegeben. In einer Ausführungsform wird diese einem Kalibrier-Nutzer wiedergegeben. In einer anderen Ausführungsform wird diese lediglich intern genutzt.
• In einem Schritt e. wird mittels des ersten Software-Agenten auf Basis des zumindest einen in Schritt d. ausgegebenen ersten Änderungsvorschlags ein zugehöriger Änderungswert zum Betrieb des zu kalibrierenden Aktivsystems erzeugt, Dies ist in einer Ausführungsform einfache Mathematik (beispielsweise Multiplizieren von einem Änderungsfaktor mit einem Eingangswert), wobei der erste Änderungsvorschlag bereits als mathematisch verwertbare Größe ausgegeben ist. Alternativ ist (zumindest davor) der erste Änderungsvorschlag erst noch zu interpretieren. Beispielsweise ist ein erster Änderungsvorschlag den Eingangswert leicht zu erhöhen, wobei dies in einen Faktor 1,05 (also eine Erhöhung um 5 %) umgewandelt wird. Es sei darauf hingewiesen, dass dies aber noch nicht der tatsächliche (geänderte) Eingangswert für die nächste Wiederholung des Kalibrierens ist. Vielmehr wird dieser (geänderte) Eingangswert noch gewichtet.

It is now suggested that the following steps be carried out:

• In a step a., complex measurement data from the operation of an active system are provided by means of a measuring device which, for example, is part of an active system or is fed by sensors of an active system. Such measurement data includes several types of data as measured values, for example, in the case of an electrified drive train relating to a speed, a voltage, a current and others.
• In a step b., at least a subset of measured values of the measurement data provided in step a. is detected and recorded using at least one of the machine learning algorithms. In one embodiment, these measured values are already put into a physical context, for example in the form of a mathematical formula and/or a diagram.
• In a step c., currently recorded measured values and at least one associated predetermined target value are compared with one another using at least one of the machine learning algorithms. Furthermore, at least one first change suggestion and an associated first justification are found if the machine learning algorithm in question detects a deviation from the associated predetermined target value. The machine learning algorithm in question has a justification model for this, for example based on a physical model and/or a learned relationship. The at least one (first) change suggestion is set up for use by the first software agent.
• In a step d., the first natural language processor uses the first change proposal and the associated first justification according to step c. to output the at least one first change proposal and the associated first justification in the form of at least one human-understandable first phrase. In one embodiment, this is reproduced to a calibration user. In another embodiment, this is only used internally.
• In a step e., an associated change value for the operation of the active system to be calibrated is generated by means of the first software agent on the basis of the at least one first change suggestion issued in step d. In one embodiment, this is simple mathematics (for example, multiplying a change factor by an input value), with the first change suggestion already being issued as a mathematically usable value. Alternatively, the first change suggestion still has to be interpreted (at least beforehand). For example, a first change suggestion is to slightly increase the input value, whereby this is converted into a factor of 1.05 (i.e. an increase of 5%). It should be noted that this is not yet the actual (changed) input value for the next repetition of the calibration. Rather, this (changed) input value is still weighted.

• Zunächst wird durch einen Kalibrier-Nutzer eine (zweite) Phrase bereitgestellt, beispielsweise durch gesprochene Sprache oder per Texteingabe in eine Computerschnittstelle. Die (zweite) Phrase wird durch Abfrage des Verfahrens an den Kalibrier-Nutzer ausgelöst, und bevorzugt ist eine Zugehörigkeit damit bereits klar. In einer anderen Ausführungsform entscheidet der Kalibrier-Nutzer selbst, wann und/oder ob dieser eine (zweite) Phrase abgibt. Eine Zugehörigkeit wird durch die Art der Schnittstelle bereits klar, wird mittels des zweiten Software-Agenten erkannt und/oder wird von dem Kalibrier-Nutzer selbst angegeben.

This represents the first group of steps. The second group further includes at least the following steps:

• First, a calibration user provides a (second) phrase, for example by spoken language or by entering text into a computer interface. The (second) phrase is triggered by querying the method to the calibration user, and preferably an affiliation is already clear. In another embodiment, the calibration user decides when and/or whether to provide a (second) phrase. An affiliation is already clear through the type of interface, is recognized by the second software agent and/or is specified by the calibration user himself.

In a step f., the second natural language processor is used to recognize at least one piece of computer-understandable information relevant to the measurement values provided from the (second) phrase of the calibration user. This computer-understandable information contains at least a second change suggestion with an associated second justification. In an advantageous embodiment, the first change suggestion and the (same) second change suggestion and/or the first justification and the (same) second justification are prepared in the same way from the perspective of the second software agent and are therefore easily comparable with one another. With a plurality of measurement values, there are often a plurality of input values for the calibration process. The first values (change suggestion and justification) are linked to the two ten values are equal if they relate to the same input value. In one embodiment, a (especially second) phrase is inaccurate or summarizes several input values. In a preferred embodiment, the individual (corresponding to second) change suggestions are recognized and treated separately from one another.

In a step g., the second software agent is used to determine a similarity between the and/or between two identical proposed changes and their respective associated justifications on the basis of the at least one first human-understandable phrase generated in step d. and the at least one second phrase provided in step f. It should be noted that the second software agent does not necessarily determine the similarity on the basis of the human-understandable phrases, but rather, for example, on intermediate data that is understandable to a computer and not necessarily understandable to a human. A high degree of similarity exists, for example, if the first proposed change and the second proposed change both provide for a slight increase in the input value and the justifications are the same. There is no similarity, for example, if the proposed changes contradict one another. A medium degree of similarity exists, for example, if the proposed changes point in the same direction, but the associated justifications are different from one another.

In a step h., a weighting or an associated weighting is generated based on the similarity determined in step g. on the basis of predetermined specifications and/or by means of at least one of the machine learning algorithms. The predetermined specifications are set in advance from outside, for example by a calibration user, and are unchangeable (at least on the software side). In an alternative embodiment, the machine learning algorithm determines a connection between the determined similarity and a weighting. For example, the machine learning algorithm has learned or is learning in the process that certain (second) phrases of a calibration user are more moderate, for example a slight increase means 10% to 15%, or more extreme, for example a significant increase means 5% to 10%. This also depends on the respective context, whereby the changes made previously are also taken into account. For example, previous changes were in the range of 0.1% to 0.5%, so a significant increase can then be set at 1% to 2%. Alternatively or additionally, the phrases contain concrete (e.g. numerical) values.

The change value generated in step e. is weighted using the associated weighting generated in step h. and output in this weighted manner to the active system to be calibrated. In one embodiment, the weighting makes it possible to block a (first) change proposal. In another embodiment, a (first) change proposal can be mitigated at most. In one embodiment, a (first) change proposal can also be increased. A weighting is associated with the first and second change proposals or justifications as above if it relates to the respective (i.e. the same) input value.

It is further proposed in an advantageous embodiment of the computer-implemented method that the first software agent is designed to be reinforcingly learning with the aid of such a reward, which is dependent on a deviation of at least one measured value from a respectively associated predetermined target value.

As already mentioned at the beginning, a (self-)reinforcing learning algorithm is particularly advantageous for efficiently finding a solution. Here, it is proposed that the first software agent should reward depending on the deviation between the measured value and the target value. The basis for the reward is preferably the comparison of a deviation present in a previous execution of a calibration loop (immediately or, for example, spaced apart by a predetermined interval) with the current deviation. If a deviation has decreased, a corresponding (for example, larger) reward is given. Alternatively, only the current deviation is always considered and the reward is made dependent solely on this, for example, a small reward or negative reward is given for a large deviation and a large (positive) reward is given for a small deviation. The reward is preferably not a factor that is multiplied by the change value, but only influences the decision-making process of the first software agent. The decision-making process is equipped, for example, with an optimization formula (for example, an evolutionary algorithm) in which a variable is influenced, i.e. changed, by means of the reward. For example, a strong reward closes off a decision path or, on the contrary, only a few decision paths, for example even only a single decision path, remain open.

It is further proposed in an advantageous embodiment of the computer-implemented method that the data provided in step a. provided complex measurement data are provided as image information, wherein the image information is prepared by means of an image converter for selecting a subset of measurement values from the measurement data and/or for identifying differences between the image information from temporally successive measurement values for the at least one machine learning algorithm and/or for the first natural language processor.

In this embodiment, the measurement data are not only complex, but the measurement values of the different data types are also related to one another, for example by means of a two-dimensional or three (or more) dimensional graph. The advantage of a graph is also that it is readable for a calibration user, who is then able to make decisions on the basis of which he can provide a (second) phrase. With the help of the image converter, the method and the calibration user now start from the same data. This means that similar, comparable decisions are made by both the method and the calibration user. The strengths of artificial intelligence, which can handle a large amount of data and can access previous values of any age, and of the human calibration user, who has experience coupled with an expectation horizon and usually reliably recognizes artifacts, are thus made synergistically usable in order to calibrate an active system quickly and efficiently. In an advantageous embodiment, the image converter is also an artificial intelligence, preferably a neural network. The image converter preferably works on a pixel-based basis, so that the usable image data requires little effort to prepare and any form of image data (can be converted) can be used.

• - einem Kalibrier-Nutzer;
• - dem in Schritt b. tätigen Maschinen-Lern-Algorithmus; und
• - einem weiteren eingesetzten Maschinen-Lern-Algorithmus.

It is further proposed in an advantageous embodiment of the computer-implemented method that the measured values acquired and recorded in step b. are selected from at least one of the following entities:

• - a calibration user;
• - the machine learning algorithm operating in step b.; and
• - another machine learning algorithm used.

Nowadays, many active systems have a significant amount of measurement data available. Even if a machine learning algorithm must fundamentally work with large amounts of data, efficiency can be significantly increased if only a few types of data are used, but with as much variation or temporal depth as possible. For example, it is particularly easy for a calibration user to understand if only the types of data taken into account in their model are recorded and recorded as measurement values. Alternatively or additionally, it is advantageous for efficiency or for openness to previously unconsidered effects if the machine learning algorithm described above itself or another machine learning algorithm determines the measurement values to be used or helps determine them in comparison with a calibration user.

• - einem Kalibrier-Nutzer;
• - dem in Schritt e. tätigen ersten Software-Agenten; und
• - einem weiteren eingesetzten Software-Agenten.

It is further proposed in an advantageous embodiment of the computer-implemented method that the at least one target value used in step c. is determined by at least one of the following entities:

• - a calibration user;
• - the first software agent operating in step e.; and
• - another software agent used.

It is often an easy task to set a target value (or a plurality of target values) from the outside, especially for an experienced calibration user. However, in some use cases, a rigid target value is detrimental to finding an unexpected solution and thus sets an unintended limit in the optimization space for artificial intelligence. It is therefore advantageous to change the target value (or a plurality of target values), for example by taking interim results into account or if conflicting suggestions for changes arise. For this purpose, it is advantageous if the (first) software agent itself or another software agent sets, changes or influences the target value (for example in coordination with the calibration user).

It is further provided in an advantageous embodiment of the
computer-implemented method, it is proposed that the first group of steps a. to e. is carried out at a higher frequency than the second group of steps f. to h.,
wherein preferably, before the second group is repeated, the at least one weighting is kept constant,
wherein preferably 1 to 100 thousand repetitions of the first group are carried out before the second group is repeated.

Another advantage of artificial intelligence that should be noted is the high speed with which the individual calibration loops can be carried out, compared to a human, but also compared to pure model calculations. It is therefore suggested here that the number of loops (i.e. repetitions) performed by the first group is higher, preferably significantly higher, than the number of repetitions performed by the second group at the same time. The frequency of repetition of the first group is therefore higher than the frequency of repetitions of the second group. This also means that the process is not slowed down by the calibration user needing a while, for example the duration of a large number of loops, to arrive at a result in the form of a second phrase.

In an advantageous embodiment, the weighting determined on the basis of the similarity of the phrases is kept constant between two repetitions of the second group. Alternatively, the weighting is increasingly weakened with increasing distance from the situation or purely in terms of time or the number of repetitions. This is advantageous if the weighting is set sharply, i.e. the second phrase of the calibration user has a large influence on the input value to be changed. Such a sharp weighting can lead to a realignment of the calibration process (as a result of a low similarity of the phrases).

In an advantageous embodiment, a thousand to a hundred thousand repetitions of the first group are carried out until the second group is repeated again. In one embodiment, only a fixed and/or manually adjustable number of repetitions, for example ten thousand or fifty thousand, of the first group are carried out until a second group is triggered, i.e. a calibration user is involved. In another embodiment, this is made dependent on a condition, for example if the direction of a previous change in an input value changes or a target value is exceeded. In one embodiment, the method is then stopped until a (second) phrase from a calibration user is provided.

It is further proposed in an advantageous embodiment of the computer-implemented method that the active system to be calibrated is a component of or an entire, preferably electrified, drive train of a motor vehicle.

The drive train of a motor vehicle is extremely complex and has a large number of individual active systems, such as pumps, valves, clutches, or in the case of an electrified drive train, for example the traction battery, the pulse inverter. In addition, the drive train is subject to a wide range of external conditions such as outside temperature, humidity and vibrations during its lifetime and/or in various regions of the world. Furthermore, drive trains nowadays have to meet the highest comfort requirements, such as noise emissions, reliability and predictability, as well as ease of use and low maintenance, but also consumption. It is therefore worthwhile to calibrate a motor vehicle drive train in detail in order to be able to meet these requirements. With the method proposed here, this can be designed to be both efficient and humanly understandable and controllable. This ultimately creates trust in the result of the calibration, which enables vehicle registration (which meets legal requirements) without the need for complex subsequent calculations or examinations.

• - in einen lokalen, bevorzugt mobilen, Rechner integriert ist; und/oder
• - zur Kommunikation mit einer Cloud eingerichtet ist.

According to a further aspect, a computer program comprising a computer program code is proposed, wherein the computer program code is executable on at least one computer such that the at least one computer is caused to carry out the method according to an embodiment according to the above description, wherein at least one of the computers:

• - is integrated into a local, preferably mobile, computer; and/or
• - is set up to communicate with a cloud.

A computer comprises one or more processors, for example a general purpose processor (CPU) or microprocessor, RISC processor, GPU and/or DSP. The computer has, for example, additional elements such as memory interfaces. Optionally or additionally, the terms refer to such a device which is capable of executing a provided or embedded program, preferably with a standardized programming language such as C++, JavaScript or Python, and/or controlling and/or accessing data storage devices and/or other devices such as input interfaces and output interfaces. The term computer also refers to a plurality of processors or a plurality of (sub)computers which are interconnected and/or otherwise communicatively connected and possibly share one or more other resources, such as memory.

A (data) storage device is, for example, a hard disk (HDD, SSD, HHD) or a (non-volatile) solid-state memory, for example a ROM- Memory or flash memory [Flash EEPROM]. The memory often comprises a number of individual physical units or is distributed across a large number of separate devices so that it can be accessed via (data) communication, for example package data service. The latter is a decentralized solution, whereby the memory and processors of a large number of separate computing units are used instead of a (single, uniform) central on-board computer or in addition to a central on-board computer.

In one embodiment, the computer program can be executed partially or completely on a server or a server unit of a cloud system, a handheld device (for example a smartphone) and/or on at least one unit of the computer device. The term server or server unit refers here to a computer which provides data and/or operational services or services for one or more other computer-based devices or computers and thus forms the cloud system.

• - in einen lokalen, bevorzugt mobilen, Rechner integriert ist; und/oder
• - zur Kommunikation mit einer Cloud eingerichtet ist.

According to a further aspect, a computer program product on which a computer program code is stored is proposed, wherein the computer program code is executable on at least one computer such that the at least one computer is caused to carry out the method according to an embodiment according to the above description, wherein at least one of the computers:

• - is integrated into a local, preferably mobile, computer; and/or
• - is set up to communicate with a cloud.

A computer program product comprising computer program code is, for example, a medium such as RAM, ROM, an SD card, a memory card, a flash memory card or a disc. Alternatively, a computer program product is stored on a server and can be downloaded. As soon as the computer program is made readable via a reading unit (for example a drive and/or an installation), the computer program code contained therein and the method contained therein can be executed by a computer or in communication with a plurality of computer-aided devices, for example as described above.

• 1 ein Ablaufdiagramm eines computerimplementierten Verfahrens zum Kalibrieren eines Aktivsystems;
• 2 das computerimplementiertes Verfahren gemäß 1 zum Kalibrieren eines Aktivsystems in einem Flussdiagramm; und
• 3 ein Kraftfahrzeug mit einem Aktivsystem in schematischer Draufsicht.

The invention described above is explained in detail below against the relevant technical background with reference to the accompanying drawings, which show preferred embodiments. The invention is in no way limited by the purely schematic drawings, whereby it should be noted that the drawings are not dimensionally accurate and are not suitable for defining size ratios. It is shown in

• 1 a flow chart of a computer-implemented method for calibrating an active system;
• 2 the computer-implemented method according to 1 for calibrating an active system in a flow chart; and
• 3 a motor vehicle with an active system in a schematic plan view.

In 1 is a flow chart of a computer-implemented method for calibrating an active system 1. Here, the active system 1 (for example a drive train 26 of a motor vehicle 27, cf. 3 ) using measured values 3 (for example torque, power and/or others) of a measuring device 15 and control values 31 (for example speed, electrical voltage, electrical current and/or others) and, if necessary, the method for calibrating the active system 1 is initialized, wherein a calibration user 7 (a human), for example a test engineer (for example by means of a human-machine interface of a computer 28), also analyzes the calibration of the active system 1 and checks its quality (deviation from a target value). In this embodiment, the measured values 3 are transmitted to the first machine learning algorithm 2 as image information 24 by means of an image converter 25.

If the measured values 3 of the active system 1 are outside predetermined target values during operation, the first machine learning algorithm 2 outputs a first change suggestion 16 with a first justification 17 to the calibration user 7 and to a first software agent 9, so that the state of the active system 1 can be changed by means of the first software agent 9 and the then at least one output specification 21. Based on the output first change suggestion 16 with the corresponding first justification 17 of the first machine learning algorithm 2, a first phrase 5 with the first change suggestion 16 and the corresponding first justification 17 is simultaneously output by means of a first natural language processor 4, so that the intention of the first machine learning algorithm 2 is made (human-) understandable for the calibration user 7.

The calibration user 7 analyses the data output by the first machine learning algorithm 2 and using the first natural language Processor 4 makes the first change proposal 16 understandable to humans with the corresponding first justification 17 or independently looks at the presented measured values 3 (provided here as image information 24) and gives its (second) change proposal 19 with a corresponding (second) justification 20 in the form of a (second) phrase 8. It should be explicitly pointed out that a first proposal compared to a second proposal does not indicate a ranking or chronological sequence. Rather, the first is the AI-side and the second is the user-side proposal and the justification accordingly.

This second, i.e. human phrase 8 is converted by a second natural language processor 6 into an input that is understandable for the algorithms used in the process, i.e. computer-readable (or computer-understandable). The two proposed changes 16, 19 are checked for their similarity 13 by a second software agent 11 within a natural language model 12, so that it can be determined whether the similarity 13 is high or medium or whether there is no similarity 13. A high similarity 13 exists, for example, if the first proposed change 16 and the second proposed change 19 both provide for a slight increase in an input value and the reasons 17, 20 are the same. No similarity 13 exists, for example, if the proposed changes 16, 19 contradict each other. A medium similarity 13 exists, for example, if the proposed changes 16, 19 point in the same direction, but the associated reasons 17, 20 are different from each other.

On the basis of the determined similarity 13, a weighting 22 is then calculated for the specification 21 of the first software agent 9 and the result of the second software agent 11 and multiplied by the specifications 21. The thus weighted specifications 21 (now referred to as change value 10) are used to operate the active system 1. The change values 10 cause a different operating point in the active system 1, which may also lead to at least one different control value 31. Alternatively, the at least one control value 31 is constant.

Another machine learning algorithm (not shown here) compares the determined measured values 3 with a corresponding target value and/or with the previous measured values 3 and, on the basis of this comparison, awards a reward 23 to the first software agent 9. The first software agent 9 is thus rewarded for successful changes and penalized for unsuccessful changes in a next repetition of the process, so that it is thus encouraged to find a correct way to solve the current calibration problem more quickly or efficiently.

In 2 is a computer-implemented method according to 1 for calibrating an active system 1 is shown in a flow chart. In a step a., complex measurement data from an operation of the active system 1 are provided by means of the measuring device 15, which is for example part of the active system 1 or is fed by sensors of the active system 1. Such measurement data comprise several types of data as measured values 3, for example relating to a speed, a voltage, a current and others in an electrified drive train 26.

In a step b., at least a subset of measured values 3 of the measurement data provided in step a. is detected and recorded using the first machine learning algorithm 2. In this exemplary embodiment, these measured values 3 are already put into a physical context, for example in the form of a mathematical formula and/or a diagram.

In a step c., currently recorded measured values 3 and at least one associated predetermined target value are compared with one another using the first machine learning algorithm 2. Furthermore, at least one first change suggestion 16 and an associated first justification 17 are found if the machine learning algorithm 2 in question detects a deviation from the associated predetermined target value. The machine learning algorithm 2 in question has a justification model for this, for example based on a physical model and/or a learned relationship. The at least one (first) change suggestion 16 is set up for use by the first software agent 9.

In a step d., the first natural language processor 4 uses the first change proposal 16 and the associated first justification 17 according to step c. to output the at least one first change proposal 16 and the respective associated first justification 17 in the form of at least one human-understandable first phrase 5. For example, this is reproduced to a calibration user 7.

• Zunächst wird durch den Kalibrier-Nutzer 7 eine zweite Phrase 8 bereitgestellt, beispielsweise durch gesprochene Sprache oder per Texteingabe in eine Computerschnittstelle. Die zweite Phrase 8 wird durch Abfrage des Verfahrens an den Kalibrier-Nutzer 7 ausgelöst. In diesem Ausführungsbeispiel ist eine Zugehörigkeit damit bereits klar.

In a step e., an associated change value 10 for the operation of the active system 1 to be calibrated is generated by means of the first software agent 9 on the basis of the at least one first change proposal 16 issued in step d. This represents the first group 14 of steps The second group 18 further comprises at least the following steps:

• First, the calibration user 7 provides a second phrase 8, for example by spoken language or by entering text into a computer interface. The second phrase 8 is triggered by querying the method to the calibration user 7. In this embodiment, affiliation is therefore already clear.

In a step f., the second natural language processor 6 is used to recognize at least one piece of computer-understandable information relevant to the provided measured values 3 from the second phrase 8 of the calibration user 7. This computer-understandable information contains at least a second change suggestion 19 with an associated second justification 20. In this exemplary embodiment, the first change suggestion 16 and the (same) second change suggestion 19 and/or the first justification 17 and the (same) second justification 20 are prepared in the same way from the perspective of the second software agent 11 and are therefore easily comparable with one another. With a plurality of measured values 3, there are often a plurality of input values for the calibration process. The first values (change suggestion and justification) are the same as the second values if they relate to the same input value.

In a step g., a similarity 13 of the and/or of two identical proposed changes 16, 19 and the respective associated justifications 17, 20 is determined by means of the second software agent 11 on the basis of the at least one first human-understandable phrase 5 generated in step d. and the at least one second phrase 8 provided in step f. It should be noted that the second software agent 11 does not necessarily determine the similarity 13 on the basis of the human-understandable phrases 5, 8, but rather, for example, on intermediate data that are understandable to computers and not necessarily understandable to humans.

In a step h., one or an associated weighting 22 is generated on the basis of predetermined specifications 21 and/or by means of at least one of the machine learning algorithms 2 on the basis of the similarity 13 determined in step g. The predetermined specifications 21 are set in advance from outside, for example by a calibration user 7, and are unchangeable (at least on the software side). In this exemplary embodiment, the machine learning algorithm 2 determines a connection between the determined similarity 13 and a weighting 22.

The change value 10 generated in step e. is weighted using the associated weighting 22 generated in step h. and is output in this weighted manner to the active system 1 to be calibrated. A weighting 22 is associated as above with regard to the first and second change proposals 16,19 or justifications 17,20 if it concerns the respective (i.e. the same) input value.

In addition, the dashed lines illustrate a different frequency of repetitions. In the first group 14, i.e. steps a. to e., 1,000 to 100,000 repetitions take place before the second group 18, i.e. steps f. to h., are repeated just once.

In 3 a motor vehicle 27 with an active system 1 is shown in a schematic plan view. The motor vehicle 27 here comprises, for example, as an active system 1, a drive train 26, which in turn comprises an electric drive motor 32, a transmission 33 and a traction battery 34. The motor vehicle 27 also comprises a computer 29 with a memory 35 and a processor 36, the computer 29 being set up for communication with a cloud 30 and/or an external computer 28. The computer 29 is designed to carry out the method described above (see 2 ). The drive train 26 is designed to generate and transmit a torque to a left drive wheel 37 and a right drive wheel 38. It should be noted that the computer 28 is a measuring computer which is used exclusively for an initial calibration of the motor vehicle 27 and is then used in another motor vehicle 27 to be calibrated, or is an on-board computer of the motor vehicle 27 which is permanently carried along. In the case of an on-board computer, the corresponding computer program code is deleted or deactivated after calibration or alternatively is used permanently over the life of the vehicle or sporadically (for example during maintenance measures or repairs).

1. a. Bereitstellen von Messdaten von einem Betrieb eines Aktivsystems;
2. b. Erfassen und Aufzeichnen von Messwerten aus den Messdaten;
3. c. Vergleichen von aufgezeichneten Messwerten und zumindest einem zugehörigen Zielwert, sowie Finden eines ersten Änderungsvorschlags und einer zugehörigen Begründung, wenn eine Abweichung von dem zugehörigen Zielwert festgestellt wird;
4. d. auf Basis des ersten Änderungsvorschlags und der Begründung, Ausgeben von dem Änderungsvorschlag und der Begründung in Form von einer menschenverständlichen Phrase;
5. e. auf Basis des ersten Änderungsvorschlags, Erzeugen eines Änderungswerts zum Betrieb des Aktivsystems;
6. f. auf Basis von einer durch einen Kalibrier-Nutzer bereitgestellten Phrase, Erkennen von relevanter Information, enthaltend einen zweiten Änderungsvorschlag und eine zugehörige Begründung; und
7. g. auf Basis der Phrasen, Ermitteln von einer Ähnlichkeit von den Änderungsvorschlägen und Begründungen; und
8. h. Erzeugen einer Gewichtung auf Basis der Ähnlichkeit,

wobei der Änderungswert mittels der Gewichtung gewichtet an das Aktivsystem ausgegeben wird.The invention relates to a computer-implemented method for calibrating an active system, wherein at least the following steps are carried out:

1. a. Providing measurement data from an operation of an active system;
2. b. Collecting and recording measured values from the measurement data;
3. c. comparing recorded measured values and at least one associated target value, and finding a first proposal for change and an associated justification, if a deviation from the corresponding target value is detected;
4. d. on the basis of the first amendment proposal and the justification, issuing the amendment proposal and the justification in the form of a human-understandable phrase;
5. e. based on the first change proposal, generating a change value for the operation of the active system;
6. f. based on a phrase provided by a calibration user, identifying relevant information containing a second change proposal and an associated justification; and
7. g. on the basis of the phrases, determining a similarity of the proposed changes and justifications; and
8. h. Generating a weighting based on similarity,

where the change value is output to the active system weighted by means of the weighting.

With the method proposed here, an active system can be calibrated using AI and human control.

List of reference symbols

1

Active system

2

first machine learning algorithm

3

Measured value

4

first natural language processor

5

first phrase

6

second natural language processor

7

Calibration users

8th

second phrase

9

first software agent

10

Change value

11

second software agent

12

Natural Language Model

13

similarity

14

first group

15

Measuring device

16

first amendment proposal

17

first justification

18

second group

19

second amendment proposal

20

second reason

21

Requirements

22

weighting

23

Reward

24

Image information

25

Image converter

26

Drivetrain

27

Motor vehicle

28

computer

29

calculator

30

Cloud

31

Control value

32

electric drive machine

33

transmission

34

Traction battery

35

Storage

36

processor

37

left drive wheel

38

right drive wheel

Claims (9)

Computer-implemented method for calibrating an active system (1), in which at least the following software components are used: - at least one machine learning algorithm (2), which is set up at least for detecting and recording selected measured values (3) and for comparing them with target values, as well as for finding suggested changes and associated reasons on the basis of detected deviations; - a first natural language processor (4) for generating human-understandable first phrases (5) from provided suggested changes and associated reasons; - a second natural language processor (6) for recognizing relevant computer-understandable information from at least one human second phrase (8) provided by a human calibration user (7); - a first software agent (9) which is set up to minimize deviations between target values and measured values (3) by generating change values (10); - a second software agent (11) which is set up using a natural language model (12) to determine similarities (13) between phrases (5, 8) with computer-understandable information; wherein at least the following steps of a first group (14): a. by means of a measuring device (15), providing complex measurement data from the operation of an active system (1); b. by means of at least one of the machine learning algorithms (2), capturing and recording at least a subset of measurement values (3) from the measurement data provided in step a.; c. by means of at least one of the machine learning algorithms (2), comparing currently recorded measurement values (3) and at least one associated predetermined target value, and finding at least one first change suggestion (16) and in each case an associated first justification (17) if a deviation from the associated predetermined target value is determined; d. by means of the first natural language processor (4) on the basis of the first change suggestion (16) and the associated first justification (17) according to step c., outputting the at least one first change suggestion (16) and the in each case associated first justification (17) in the form of at least one human-intelligible first phrase (5); and e. by means of the first software agent (9) on the basis of the at least one first change suggestion (16) issued in step d., generating an associated change value (10) for operating the active system (1) to be calibrated, wherein at least the following steps of a second group (18) are further carried out: f. by means of the second natural language processor (6) on the basis of at least one second phrase (8) provided by a calibration user (7), recognizing at least one piece of computer-intelligible information relevant to the provided measured values (3), containing at least one second change suggestion (19) and in each case an associated second justification (20); g. by means of the second software agent (11) on the basis of the at least one first change suggestion (16) issued in step d. generated first human-intelligible phrase (5) and the at least one second phrase (8) provided in step f., determining a similarity (13) of the and/or of two identical change suggestions (16, 19) contained therein and the respective associated reasons (17, 20); and h. on the basis of predetermined specifications (21) and/or by means of at least one of the machine learning algorithms (2), generating one or in each case one associated weighting (22) on the basis of the similarity (13) determined in step g., wherein the change value (10) generated in step e. is output to the active system (1) to be calibrated in a weighted manner by means of the associated weighting (22) generated in step h. Computer-implemented method according to Claim 1 , wherein the first software agent (9) is designed to learn in a reinforcement manner with the aid of such a reward (23), which is dependent on a deviation of at least one measured value (3) from a respectively associated predetermined target value. Computer-implemented method according to Claim 1 or Claim 2 , wherein the complex measurement data provided in step a. are provided as image information (24), wherein the image information (24) is prepared by means of an image converter (25) for selecting a subset of measurement values (3) from the measurement data and/or for identifying differences between the image information (24) from temporally successive measurement values (3) for the at least one machine learning algorithm (2) and/or for the first natural language processor (4). Computer-implemented method according to one of the preceding claims, wherein the measured values (3) acquired and recorded in step b. are selected by at least one of the following entities: - a calibration user (7); - the machine learning algorithm (2) operating in step b.; and - another machine learning algorithm used. Computer-implemented method according to one of the preceding claims, wherein the at least one target value used in step c. is determined by at least one of the following entities: - a calibration user (7); - the first software agent (9) active in step e.; and - a further software agent used. Computer-implemented method according to one of the preceding claims, wherein the first group (14) of steps a. to e. is carried out at a higher frequency than the second group (18) of steps f. to h., whereby preferably, before the second group (18) is repeated, the at least one weighting (22) is kept constant, whereby preferably 1 to 100 thousand repetitions of the first group (14) are carried out before the second group (18) is repeated. Computer-implemented method according to one of the preceding claims, wherein the active system (1) to be calibrated is a component of or an entire, preferably electrified, drive train (26) of a motor vehicle (27). Computer program comprising a computer program code, wherein the computer program code is executable on at least one computer (28) such that the at least one computer computer (28) is caused to carry out the method according to one of the preceding claims, wherein at least one of the computers (28): - is integrated into a local, preferably mobile, computer (29); and/or - is set up for communication with a cloud (30). Computer program product on which a computer program code is stored, wherein the computer program code is executable on at least one computer (28) such that the at least one computer (28) is caused to carry out the method according to one of Claim 1 until Claim 7 to carry out, wherein at least one of the computers (28): - is integrated into a local, preferably mobile, computer (29); and/or - is set up for communication with a cloud (30).

Images