Method for analyzing the results of calibration measurements of drive devices using a generative machine learning model

Abstract

The invention relates to a computer-implemented method for training a generative machine learning model, which is configured to analyze a calibration of a drive device of a motor vehicle, by supervised learning, wherein the generative machine learning model is trained using image data of the calibration measurements and text data associated therewith to analyze the calibration by predicting what could have happened before the calibration measurements, describing how good the calibration results are, and suggesting what should be analyzed further.

Description

[0001] The present invention relates to a method for analyzing the results of calibration measurements of drive devices by means of a generative machine learning model which is configured to analyze a calibration of a drive device of a motor vehicle.

[0002] In the context of the present invention, a computer-implemented method is understood in particular to mean that the method is executed by a computer, whereby interaction with a human user is possible. The computer can, for example, comprise a digital data memory and a processor. Instructions can be stored in the digital data memory which, when executed by the processor, cause the processor to execute the computer-implemented method.

[0003] It is known from the prior art to calibrate a drive device of a motor vehicle, in particular an internal combustion engine or an electric motor. For this purpose, one or more control units provided for controlling the drive device can be programmed accordingly. This allows, for example, two- or three-dimensional characteristic maps to be adapted for controlling the drive device.

[0004] To verify the effects of specific calibration measures, various calibration measurements are performed, the results of which are then analyzed accordingly. The results of the calibration measurements are typically presented graphically to an expert using a computer display. Such an analysis can be relatively difficult and, above all, time-consuming, depending on the number of measurement results and their complexity.

[0005] Furthermore, it is alternatively or additionally possible to carry out calibration simulations to simulate the effects of calibration measures.

[0006] Analyzing the results of calibration measurements or calibration simulations using artificial intelligence methods is already well-known in the art. However, with a frequently used deep learning approach, it is difficult to incorporate technical knowledge or expertise into the analysis task. Furthermore, an AI-based analysis of the results of calibration measurements or calibration simulations is not particularly intuitive. For example, artificial intelligence is not capable of predicting in a human-like manner what calibration measures may have been taken prior to the calibration measurements and what possible next steps in calibrating the drive device might be.

[0007] US 2021 / 0011475 A1 proposes a database for the continuous calibration of a motor vehicle's powertrain. This database links driving routes with specific drive settings using the navigation system and is increasingly refined through machine learning and trend analysis. The powertrain can thus be optimized according to the driver's preferences, particularly with regard to fuel efficiency, performance, or pollutant emissions.

[0008] DE 10 2021 131 742 A1 discloses a computer-implemented method for calibrating an electric powertrain of a vehicle. This method enables user inputs defining calibration tasks to be automatically transformed into a coded, machine-readable output using a transformer and made available to a machine learning agent. The outputs of the transformer represent technical goals for the machine learning agent to be achieved during the calibration of the electric powertrain, which includes an electric machine. The learning agent operates according to the principle of reinforcement learning to calibrate the vehicle's electric powertrain in an iterative, repeating process. The user interacts directly with the learning agent via the transformer, without having to program or configure it themselves.

[0009] The present invention aims to provide a method for analyzing the results of calibration measurements of drive devices by means of a generative machine learning model, which is designed to analyze a calibration of a drive device of a motor vehicle, which makes it easy to incorporate human expert knowledge into the training process.

[0010] The solution to this problem is provided by a method for analyzing the results of calibration measurements of drive devices by means of a generative machine learning model, which is configured to analyze a calibration of a drive device of a motor vehicle, with the features of claim 1. The subclaims relate to advantageous developments of the invention.

[0011] A method is provided for analyzing the results of calibration measurements of drive devices by means of a generative machine learning model which is set up to analyze a calibration of a drive device of a motor vehicle, - where a multi-task algorithm is used as the generative machine learning model, - where a ViLBERT algorithm is used as the multi-task algorithm, - wherein the generative machine learning model is trained by means of a computer-implemented method for training the generative machine learning model by supervised learning, comprising the steps A) Providing and visualizing a number of first image data representing calibration measurements and / or calibration simulations of the drive device, and annotating the first visualized image data with what can be recognized in the first image data by at least one user by means of a text input device and / or a voice input device of a computer, B) Providing and visualizing a number of second image data representing calibration measurements and / or calibration simulations of the drive device and enabling a further in-depth analysis of the first image data, and providing the second visualized image data with annotations representing the results of the in-depth analysis by the at least one user, by means of the text input device and / or the voice input device, C) Providing and visualizing a number of third image data representing calibration measurements and / or calibration simulations of the drive device, and providing the third visualized image data with annotations representing which calibration interventions were carried out before the calibration measurements and / or calibration simulations, by the at least one user by means of the text input device and / or the voice input device, D) storing the image data used in the previous steps and the associated text data as training data sets in a storage means, E) Repeating the previous steps until a predetermined or predeterminable number of training data sets has been obtained, F) Training the generative machine learning model using the training data sets, wherein the generative machine learning model is trained to analyze the calibration of the drive device by - based on the first image data and the associated text data, recognizes what is shown in the images in question and how good the calibration is, - based on the third image data and the associated text data, predicts which calibration interventions could have led to the measurement results and / or simulation results, and - based on the second image data and the associated text data, suggests which aspects of the calibration should be further analyzed, and wherein the trained generative machine learning model is used to analyze the results of calibration measurements of drive devices.

[0012] The invention provides a trained generative machine learning model which, after successful training, can analyze the measurement results resulting from specific calibration interventions. For this purpose, the generative machine learning model is intuitively trained based on the image data visualized by a display device, in particular a computer display device, and the text and voice inputs of a human expert to estimate and predict what the images show—in particular, how good the calibration is—which measures, in particular calibration interventions, could have led to the present measurement results or simulation results, and which other aspects—such as thermodynamic aspects—should still be investigated in subsequent analyses.The image data contains, in particular, information about calibration parameters and calibration results. The in-depth analysis performed by a human expert is a logical decomposition of the calibration measurements and / or calibration simulations of the visualized initial image data. The in-depth analysis can thus implicitly represent the search for causes or solutions (i.e., what should be specifically analyzed subsequently) in order to better understand the calibration results.

[0013] The multitasking algorithm, which is a ViLBERT algorithm, is designed to process the calibration results contained in the image data and the text data added to the image data.

[0014] When using the trained generative machine learning model to analyze the results of calibration measurements of drive devices, data from a control unit of a drive device is provided. The data can, for example, be recorded during operation of the control unit. The control unit can, for example, be a control unit of a motor vehicle. In particular, it can be an engine control unit of the motor vehicle. The data particularly comprise a plurality of calibration parameters and calibration results from which image data can be generated. After visualization, this image data is provided with at least one annotation by a user through voice and / or text input, analogous to the provision of the training data, and is then made available to the trained generative learning model.The trained generative learning model then outputs in text form what is represented in the image data (in particular, how good the calibration is), which calibration interventions may have led to the measurement results, and which aspects of the calibration should be further analyzed. Compared to training, the trained generative machine learning model can work with less image data. For example, it is sufficient to provide the trained generative machine learning model with a graph containing measurement results, a short label, and the labels of the axes.

[0015] Preferably, several hundred to several thousand training data sets generated in the manner described above can be used to train the generative learning model. The training of the generative machine learning model can be carried out, for example, using historical measurement data that is stored in a retrievable manner in a storage medium and from which the training data sets are generated.

[0016] In an advantageous development, it is possible that, before storing the training data sets, at least one transformer algorithm is used to expand the text data obtained in the preceding steps with a more extensive language. The transformer algorithm is configured to transform the words entered or spoken by the user into data readable by the generative machine learning model. The text data linguistically enriched in this way can advantageously be used to train the generative machine learning model. For improved functionality, in an advantageous embodiment, a transformer algorithm can be used which is configured to generate several synonymous word sequences from each word sequence entered by a user. Preferably, a generative, pre-trained transformer can be used.In particular, a GPT-X algorithm (GPT is the abbreviation for “Generative Pretrained Transformer”) can be used, where the “X” here represents the version number of the GPT language model.

[0017] Further features and advantages of the present invention will become clear from the following description of a preferred embodiment with reference to the attached Fig.1, which shows a highly simplified schematic representation of the basic sequence of a computer-implemented method for training a generative machine learning model, which is configured to analyze the calibration of a drive device of a motor vehicle, through supervised learning. The generative machine learning model uses a multitask algorithm that is configured to process image information and speech or text information as input variables and output text information. In this case, a ViLBERT algorithm is used as the multitask algorithm.

[0018] In a first method step 100, a number of first image data representing calibration measurements and / or calibration simulations of the drive device are provided. These first image data are displayed and thus visualized to a human expert (user) using a computer display device. In this context, the term "computer" refers in particular to stationary or portable personal computers, tablet computers, and mobile phones.

[0019] Subsequently, the first visualized image data is annotated (labeled) by the human expert using a computer's text and / or voice input device. These annotations represent what can be recognized in the first image data. These annotations include words and / or combinations of words and / or sentences that are preferably intuitive from a technical perspective and describe what is recognizable to the human expert in the visualized first image data. In particular, these annotations can also describe the current calibration of the drive device.

[0020] In a second step 101, a number of second image data are provided, representing calibration measurements and / or calibration simulations of the drive device and enabling a more detailed analysis of the first image data. These second image data are then visualized to the human expert using the computer's display device.

[0021] The second visualized image data is provided by the human expert using the text input device and / or the voice input device with annotations (labels), which represent the results of the in-depth analysis by the at least one expert. These annotations are in turn words and / or combinations of words and / or sentences that represent the in-depth analysis of the first graphical representations. The second image data provided in this method step 101 thus contain the information required for a further in-depth analysis of the first graphical representations and, if applicable, also of the annotations associated with them. In this second step 101, a more in-depth technical analysis of the calibration measurements and / or calibration simulations of the drive device is thus carried out.

[0022] In a third step 102a, a plurality of third image data representing calibration measurements and / or calibration simulations of the drive device are provided. These third image data are also displayed and thus visualized to the human expert via the computer's display device.

[0023] The expert uses the text input device and / or the voice input device to annotate the third visualized image data with annotations (labels) that represent the calibration interventions that occurred prior to the calibration measurements and / or calibration simulations. The annotations are words and / or combinations of words and / or sentences that represent the interventions, in particular calibration interventions, that occurred prior to the calibration measurements or calibration simulations. For example, a two- or three-dimensional characteristic map of the drive system was modified prior to the calibration measurements.

[0024] In a step 102b following method step 102a, which is an optional method step, at least one transformer is used to supplement and thereby enrich the annotations on the visualized image data generated by the expert in the preceding steps 100 to 102a with a more comprehensive language. Preferably, a generative pretrained transformer (GPT) can be used for this purpose.

[0025] In a subsequent step 103, the image data used in the previous steps and the associated text data are then stored as training data sets in a storage means.

[0026] Steps 100 to 103 are repeated until a predetermined or predeterminable number of training data sets required for training the generative machine learning model has been generated (step 104). Preferably, several hundred to several thousand training data sets are generated, which are subsequently used for the actual training of the generative machine learning model.

[0027] The training data sets obtained in method steps 100 to 102a and 100 to 102b, respectively, which comprise the first, second and third image data and the text data associated therewith, make it possible to describe what the human expert recognizes in the visualized image data.

[0028] In a method step 104 following method step 103, the generative machine learning model is trained on the basis of the training data sets obtained in method steps 100 to 102a or 100 to 102b and stored in method step 103, which comprise a plurality of first, second and third image data and the annotations associated with them and, if applicable, the annotations linguistically enriched in method step 102b.

[0029] The generative machine learning model is trained in method step 105 to analyze the calibration of the drive device by - based on the first image data and the associated text data, recognizes what is shown in the images in question and how good the calibration is, - based on the third image data and the associated text data, predicts which calibration interventions (e.g. map changes) could have led to the measurement results and / or simulation results, and - based on the second image data and the associated text data, suggests which aspects of the calibration should be further analyzed (for example, a more in-depth thermodynamic analysis).

[0030] The generative machine learning model trained as described above can then be used in production to analyze the results of calibration measurements of drive devices. For this purpose, the trained generative machine learning model receives, for example, image data from calibration measurements and the associated annotations as inputs. The generative machine learning model analyzes these inputs and generates corresponding outputs that include information about the current status of the calibration (i.e., what can be seen in the image data and, in particular, how good the results of the calibration measurements are), as well as information about which calibration interventions (e.g., map changes) could have led to the measurement results and which aspects of the calibration should be further analyzed. Compared to training, a smaller amount of image data and associated annotations is sufficient for this purpose.For example, it may be sufficient to provide the trained generative machine learning model with a graph with measurement results, a short description of what the graph shows, and, if applicable, the labels of the axes.

[0031] The method presented here proposes a generative machine learning model that is trained to analyze calibration results from graphical representations (images) and the associated speech or text inputs. The learning algorithm can predict, in a very intuitive way, what might have happened before the measurements and describe what can be seen in the graphical representations of the image data. Furthermore, it can suggest what the next analysis steps could be. The learning algorithm receives technical expertise from the expert's text or speech inputs during the generation of the training data. It is very intuitive for the expert to incorporate their speech- and text-based knowledge and expertise into the generative machine learning model.

Claims

[1] Method for analyzing the results of calibration measurements of drive devices by means of a generative machine learning model which is designed to analyze a calibration of a drive device of a motor vehicle, - where a multi-task algorithm is used as the generative machine learning model, - where a ViLBERT algorithm is used as the multi-task algorithm, - wherein the generative machine learning model is trained by means of a computer-implemented method for training the generative machine learning model through supervised learning, comprising the steps (100): providing and visualizing a number of first image data representing calibration measurements and / or calibration simulations of the drive device, and annotating the visualized first image data as to what can be recognized in the first image data, by at least one user using a text input device and / or a voice input device of a computer, (101): providing and visualizing a number of second image data representing calibration measurements and / or calibration simulations of the drive device and enabling a further in-depth analysis of the first image data, and annotating the visualized second image data,which represent the results of the depth analysis by the at least one user, by means of the text input device and / or the voice input device, (102a): providing and visualizing a number of third image data representing calibration measurements and / or calibration simulations of the drive device, and providing the visualized third image data with annotations representing which calibration interventions were carried out before the calibration measurements and / or calibration simulations by the at least one user by means of the text input device and / or the voice input device, (103): storing the image data obtained in the preceding steps and the associated text data as training data sets in a storage means, (104): Repeating the previous steps until a predetermined or predeterminable number of training data sets has been obtained, (105): Training the generative learning model using the training data sets, wherein the generative learning model is trained to analyze the calibration of the drive device by - based on the first image data and the associated text data, recognizes what is shown in the images in question and how good the calibration is, - based on the third image data and the associated text data, predicts which calibration interventions could have led to the measurement results and / or simulation results, and - based on the second image data and the associated text data, suggests which aspects of the calibration should be further analyzed, is trained, and wherein the trained generative machine learning model is used to analyze the results of calibration measurements of drive devices. [2] Method according to claim 1, characterized bythat before storing the training data sets in a method step (102b) at least one transformer algorithm is used to supplement the text data obtained in the preceding steps (100) to (102a) with a more extensive language. [3] Method according to claim 3, characterized by that a generative, pre-trained transformer algorithm is used.

Images