Computer-implemented method for calibrating the control logic of an engine control unit

Abstract

The invention relates to a method for calibrating the control logic of an engine control unit, in particular an engine control unit of a battery-electric vehicle, comprising the steps: S0: Providing a library of image representations (10) of calibration tasks and associated pre-trained machine learning models, S1: Selection of one of these models by the user, S2: Selecting a part of the control logic to be calibrated from a graphical representation, S3: Determining a similarity between the image representations (10) of the calibration tasks and the image representation (11, 11') of the selected part of the control logic by means of an artificial intelligence (1), S4: Selecting the calibration task whose image representation (10) has the greatest similarity to the image representation (11, 11') of the selected part of the control logic to be calibrated, S5: Loading the program code of a calibration environment, which corresponds to the image representation (10) of the selected calibration task (S5), and S6: Processing the selected calibration task in the calibration environment with the selected pre-trained machine learning model, wherein an energy-based approach is used to determine the similarity between the image representations (10) of the calibration tasks used for pre-training the selected machine learning model and the image representation (11, 11') of the selected part of the control logic to be calibrated by means of the artificial intelligence (1) in method step S3, and wherein an artificial intelligence (1) is used which is designed as a joint embedding predictive architecture, image joint embedding predictive architecture or as a latent-variable energy-based model.

Description

[0001] The present invention relates to a computer-implemented method for calibrating a control logic of an engine control unit, in particular an engine control unit of a battery-electric vehicle.

[0002] It is already known in principle from the prior art to calibrate the control logic of an engine control unit, in particular an engine control unit of a battery-electric vehicle, using artificial intelligence methods in an automated process. For the calibration of the control logic, specific calibration tasks are defined, which are to be solved using pre-trained machine learning models, preferably formed by a multi-layer neural network. In practice, it has been shown that numerous calibration tasks are repeated because they exist in many different control logics to be calibrated.

[0003] The control logic of an engine control unit, especially an engine control unit of a battery-electric vehicle, is highly complex. Numerous control logic parameters, such as electrical voltage, must be calibrated using machine learning models. Within a control logic, there are numerous logical calculations and logical connections that are similar to one another and can be processed using the same algorithm of a pre-trained machine learning model. How well a pre-trained machine learning model performs on a new calibration task depends heavily on whether the new calibration task bears a certain similarity to the calibration tasks used to train the machine learning model in question.

[0004] DE 10 2020 106 880 A1 describes a computer-implemented method for calibrating functions of a control unit, which comprises inputting control signals into an actuator model, processing the control signals, and outputting actuator states. Furthermore, the method describes processing the output actuator states in a virtual unit and outputting result parameters, as well as comparing the result parameters with target values. The method also includes determining an observation and a reward from the result parameters, training a training actuator model taking the observation and reward into account through reinforcement learning, and generating adapted actuator states.In addition, an iterative processing of the adapted actuator states to result parameters in the virtual unit and an iterative comparison of the processed result parameters with the target values ​​until a final value is reached is described.

[0005] DE 10 2022 117 623 A1 describes a method for adapting calibration data of a control unit for an electric drive of a motor vehicle. This method comprises the use of parameters and calibration data by the control unit to calculate calculated values ​​and the measurement of measured values ​​during operation of the drive, wherein the measured values ​​depend on the parameters. A difference between the calculated values ​​and the measured values ​​is detected, and an algorithm is used to reduce this difference by changing the calibration data. An artificial intelligence creates an instruction for a procedure for changing the calibration data, which is then transmitted to the algorithm, and the artificial intelligence outputs a description of the instruction in human language.

[0006] DE 10 2020 001 655 A1 discloses a system with a library that provides a plurality of machine learning models. Each of these machine learning models has been trained to perform a specific / separate task. A user selects via a user input interface which task an automatically selected pre-trained neural network should perform. The user interface can include text input, voice input, or the selection of sub-areas of a graphical representation. Based on the user's selection or input in / for the image, a corresponding task is defined for a neural network.The neural network (and the corresponding parameters and weights) specifically trained to perform the task selected by the user is then automatically selected, loaded, and executed from the library, which contains the majority of stored neural networks. The automatic selection of the optimal neural network for the specific task to be executed is based, for example, on determining a similarity between the part of the graphical representation selected by the user and the image representations used for pre-training the selected machine learning model. The comparison and determination of the similarity is preferably carried out in a vector space or in an embedding space. A similar method is known from US 2023 / 0 273 826 A1.

[0007] In a computer-implemented method, as known from DE 10 2022 121 545 A1, a machine-learned model for processing microscope data is trained using a dataset containing microscope data. An embedding of the dataset into a feature space is calculated. The embedding is analyzed to determine training design specifications for training the model. The training is determined based on the training design specifications and then executed, thus enabling the model to calculate a processing result from the microscope data to be processed.

[0008] To date, there are no approaches to solving the problem of systematically identifying those calibration tasks that can be solved particularly easily with a given, pre-trained machine learning model. Furthermore, there are no solutions in the state of the art that can automatically identify which parts of different control logics are similar to each other.

[0009] The invention aims to provide a computer-implemented method for calibrating a control logic of an engine control unit, in particular an engine control unit of a battery-electric vehicle, which enables autonomous recognition of the most suitable calibration task, which can be solved with a predetermined, pre-trained machine learning model.

[0010] The solution to these objects is provided by a computer-implemented method for calibrating a control logic of an engine control unit, in particular an engine control unit of a battery-electric vehicle, having the features of claim 1. The subclaims relate to advantageous developments of the invention.

[0011] A computer-implemented method according to the invention for calibrating a control logic of an engine control unit, in particular an engine control unit of a battery-electric vehicle, comprises the steps: S0: Providing a library comprising a plurality of image representations of calibration tasks and a plurality of corresponding pre-trained machine learning models that have been trained with the calibration tasks, S1: Selecting one of the pre-trained machine learning models by user input, wherein the corresponding image representations of all calibration tasks used to train the selected pre-trained machine learning model and the pre-trained parameter values ​​of the machine learning model are loaded from the library, S2: Selecting a part of the engine control unit's control logic to be calibrated from a graphical representation of the control logic by user input, wherein the selected part of the control logic is saved as an image representation, S3: Determining a similarity between the image representations of the calibration tasks used for pre-training the selected machine learning model and the image representation of the selected part of the control logic to be calibrated by means of artificial intelligence, S4: Selecting the calibration task from all calibration tasks used for training the selected pre-trained machine learning model whose image representation has the greatest similarity to the image representation of the selected part of the control logic to be calibrated, S5: Loading a program code of a calibration environment which corresponds to the image representation of the calibration task selected in process step S4, S6: Processing the selected calibration task in the calibration environment with the selected pre-trained machine learning model, wherein an energy-based approach is used to determine the similarity between the image representations of the calibration tasks used for pre-training the selected machine learning model and the image representation of the selected part of the control logic to be calibrated by means of the artificial intelligence in method step S3, and wherein an artificial intelligence is used which is implemented as a joint embedding predictive architecture, image joint embedding predictive architecture or as a latent-variable energy-based model.

[0012] The computer-implemented method presented here advantageously enables the automatic selection of a most suitable calibration task for the pre-trained machine learning model selected by a user, in particular an engineer. Furthermore, the method enables autonomous detection of the similarity of control logic tasks by detecting similarities in the image representations. With the concept presented here, the engineer simply scrolls through the graphical representation of the control logic on a display device, selects the pre-trained model they wish to use, and the artificial intelligence determines, in an automated process, which calibration task is best suited for calibration with the selected machine learning model.In particular, when using the Image Joint Embedding Predictive Architecture, not every detail (pixel) of an image representation of a calibration task is learned, but only the most important features from that image representation. This advantageously saves computational resources.

[0013] In one embodiment, it is provided that an energy function is calculated from features of each of the image representations of the calibration tasks that were used for the pre-training of the selected machine learning model and features of the image representation of the selected part of the control logic to be calibrated by means of the artificial intelligence.

[0014] In one embodiment, it is possible to calculate a scalar energy value from the energy function. The scalar energy value can advantageously be used to select the most suitable calibration task to be processed with the selected pre-trained machine learning model.

[0015] In one embodiment, it is proposed that the scalar energy value is compared with a threshold value and the calibration task for the selected pre-trained machine learning model is selected if the scalar energy value is lower than the threshold value.

[0016] In one embodiment, the calibration task with the lowest scalar energy value is selected for the pre-trained machine learning model if several scalar energy values ​​are lower than the threshold.

[0017] In one embodiment, latent variables may be used to train artificial intelligence.

[0018] Further features and advantages of the present invention will become clear from the following description of a preferred embodiment with reference to the accompanying drawings. Fig. 1 a schematically highly simplified representation illustrating a similarity training of an artificial intelligence, Fig. 2 another schematically simplified representation illustrating the similarity training of artificial intelligence, Fig. 3 a schematic representation showing the basic sequence of a computer-implemented method for calibrating a control logic of an engine control unit, in particular an engine control unit of a battery-electric vehicle.

[0019] The computer-implemented method for calibrating the control logic of an engine control unit, in particular an engine control unit of a battery-electric vehicle, explained in more detail below, utilizes a plurality of pre-trained machine learning models, each formed by a multi-layer artificial neural network. The pre-trained machine learning models make it possible, for example, to minimize the mean square deviation of a calibration target value, which may be, for example, an electrical voltage, from a calibration value, which may be, for example, a current calibrated voltage value.

[0020] In the following, details of a pre-training of the machine learning models for calibrating the control logic of an engine control unit and the training data sets used in this process will be explained in more detail.

[0021] A user, particularly an engineer, has access to a display device that graphically displays several calibration tasks for calibrating the control logic. In a first step, the user can select the desired calibration tasks for training one of the machine learning models using appropriate user input. These selected calibration tasks are stored as image representations (pixel representations) in a library and then defined as training tasks.

[0022] The user then selects one of the machine learning models and trains it to solve the selected training tasks. The user can train the machine learning model using a variety of learning methods, such as supervised machine learning, reinforcement learning, etc.

[0023] For example, if the control logic of a battery-electric vehicle's motor controller needs to be calibrated, possible input variables for calibration using one of the machine learning models could be an electric current, an engine speed, a traction battery charge level, or a wheel torque. The goal of the calibration could be, for example, to minimize the mean square deviation between calibrated and measured acceleration or calibrated and measured electrical voltage.

[0024] The entire control logic, which is graphically visualized by the display device, is implemented as program code in a calibration environment. Consequently, each of the calibration tasks is also implemented as program code, whereby the program codes of the calibration tasks can be assigned to the image of the associated calibration task using a classification algorithm. The corresponding implementations of the control logic parts to be calibrated in program code, as well as the input variables, output variables, and setpoints, can be downloaded from a data server or a cloud server. The calibration environment can, in particular, be implemented as a "hardware-in-the-loop" calibration environment or as a "software-in-the-loop" calibration environment.

[0025] After training, the parameters (weights) of the neural network, which forms the pre-trained machine learning model, are stored in the library.

[0026] The pre-training described above is performed for each of the machine learning models.

[0027] With reference to Fig. 1 and Fig. 2, details of a similarity training of an artificial intelligence 1 will be explained in more detail below. The artificial intelligence 1 is designed to autonomously recognize the similarity between the image representations 10 stored in the library of the calibration tasks used for pre-training and an image representation 11, 11' of a possible next calibration task. For this purpose, an energy-based model approach can be used in the artificial intelligence 1, which is preferably designed as an artificial neural network, so that an energy function 15 is implemented in the artificial intelligence 1. For example, the artificial intelligence 1 can be implemented as a Joint Embedding Predictive Architecture (JEPA), as an Image Joint Embedding Predictive Architecture (I-JEPA), or as a Latent-Variable Energy-Based Model (LVEBM). The training of the artificial intelligence 1 can preferably be carried out in a self-supervised manner.

[0028] In the Fig. In the example shown in Figure 1, on the left side, an image representation 10 of a portion of the motor controller's control logic is shown, which constitutes one of the calibration tasks from the library and has already been used for pre-training at least one of the machine learning models. On the right side, an image representation 11 of a calibration task is shown that could be suitable for calibration with the pre-trained machine learning model. In this example, the image representations 10, 11 are identical, and a scalar energy value 14, which can be calculated from an energy function 15 that receives features 12, 13 extracted from the image representations 10, 11 as input variables, is low. Based on the similarity, a scalar energy value 14 is thus calculated, which later makes it possible to select the most suitable calibration task for a selected, pre-trained machine learning model.

[0029] In the Fig. In the example shown in Figure 2, on the left side, an image representation 10 of a part of the control logic is shown, which forms one of the calibration tasks (training tasks) from the library and was used for the pre-training of at least one of the machine learning models. On the right side, an image representation 11' of a calibration task, which could be suitable for calibration with the pre-trained machine learning model, is shown. In this example, however, the two image representations 10, 11' are not the same, so their similarity is low and the scalar energy value 14 is accordingly higher than that in the example according to Fig. 1 is.

[0030] For example, latent variables 16 can also be used for training artificial intelligence 1. Latent variables 16 are unknown, but some of them could represent, for example, the name of a calibration map, the number of product terms in the image, the names of the calibration map inputs, etc.

[0031] With reference to Fig. Section 3 will explain in more detail the basic process of a computer-implemented method for calibrating an engine control unit, in particular an engine control unit of a battery-electric vehicle. The method, which is executed by a computer, uses the pre-trained machine learning models described above as well as artificial intelligence 1 with the energy-based model approach for similarity detection.

[0032] In step S0, the library containing the image representations of the training tasks and the corresponding pre-trained machine learning models is first made available. Preferably, all parameters of the pre-trained machine learning models are initialized to their pre-training values.

[0033] The image representations of the pre-trained machine learning models are visualized on a graphical user interface using a display device. In step S1, a user, in particular an engineer, can then select a pre-trained machine learning model from the plurality of available machine learning models using an operator input. For example, a pre-trained machine learning model can be selected that the user knows has delivered excellent results on a specific training task. The corresponding image representations of all calibration tasks used for pre-training the selected pre-trained machine learning model and the pre-trained parameter values ​​of the selected pre-trained model are loaded from the library.More specifically, the program code of the selected pre-trained machine learning model and the pre-trained parameter values ​​(neural network weights) of the selected pre-trained machine learning model, as well as the image representations of the calibration tasks used for pre-training, are loaded from the library.

[0034] In step S2, the user selects, by means of appropriate user input, from a graphical representation of the control logic those parts of the control logic that they wish to calibrate with the selected machine learning model. The selected graphical representation of the control logic, which represents the calibration task, is saved.

[0035] In a subsequent step S3, the similarity between the image representation of the calibration task selected by the user and the images of the calibration tasks used for pretraining the selected pretrained machine learning model is calculated. The similarity is calculated using artificial intelligence 1 via the energy-based approach explained above. If the image representation of the selected calibration task and the image representation of one of the calibration tasks used to pretrain the selected machine learning model are similar to each other, a low scalar energy value 14 results.

[0036] In step S4, the scalar energy value 14 from step S3 is used to select the most suitable calibration task to be processed with the pre-trained machine learning model selected in method step S1. For this purpose, the scalar energy value 14 is compared with a threshold value. If the scalar energy value 14 is lower than the threshold value, the calibration task is selected and can be solved with the pre-trained model selected in step S1. If several scalar energy values ​​are lower than the threshold value, the calibration task with the lowest scalar energy value 14 is selected for the pre-trained machine learning model.

[0037] In a subsequent step S5, program code for a calibration environment corresponding to the image representation of the calibration task selected in process step S4 is loaded. The user also selects the input and output variables of the pre-trained machine learning model. The corresponding signal values ​​are downloaded from the library. The user also defines a calibration objective, for example, minimizing the calibration loss between measured and simulated torque, as well as a metric, for example, minimizing a mean square deviation.

[0038] In step S6, the selected pre-trained model is applied in the calibration environment to the calibration task selected in step S4 after the calibration objective has been defined, such as minimizing a mean square deviation between the current value of a quantity to be calibrated and a predetermined target value of that quantity. The calibration objective is solved using new data, for example, by fine-tuning the weights of the last layer of the multi-layer pre-trained machine learning model to minimize the mean square error between a target quantity, such as a measured voltage, and a calibration value, such as a current calibrated voltage value.

Claims

[1] Computer-implemented method for calibrating a control logic of an engine control unit, in particular an engine control unit of a battery-electric vehicle, comprising the steps: S0: Providing a library comprising a plurality of image representations (10) of calibration tasks and a plurality of corresponding pre-trained machine learning models that have been trained with the calibration tasks, S1: Selecting one of the pre-trained machine learning models by a user input, wherein the corresponding image representations (10) of all calibration tasks used for training the selected pre-trained machine learning model and the pre-trained parameter values ​​of the machine learning model are loaded from the library, S2: selecting a part of the control logic of the engine control unit to be calibrated from a graphical representation of the control logic by a user input, wherein the selected part of the control logic is stored as an image representation (11, 11'), S3: Determining a similarity between the image representations (10) of the calibration tasks used for pre-training the selected machine learning model and the image representation (11, 11') of the selected part of the control logic to be calibrated by means of an artificial intelligence (1), S4: Selecting the calibration task from all calibration tasks used for training the selected machine learning model whose image representation (10) has the greatest similarity to the image representation (11, 11') of the selected part of the control logic to be calibrated, S5: Loading a program code of a calibration environment which corresponds to the image representation (10) of the calibration task selected in method step S4, S6: Processing the selected calibration task in the calibration environment with the selected pre-trained machine learning model, wherein an energy-based approach is used to determine the similarity between the image representations (10) of the calibration tasks used for pre-training the selected machine learning model and the image representation (11, 11') of the selected part of the control logic to be calibrated by means of the artificial intelligence (1) in method step S3, and wherein an artificial intelligence (1) is used which is designed as a joint embedding predictive architecture, image joint embedding predictive architecture or as a latent-variable energy-based model. [2] Computer-implemented method according to claim 1,characterized by that an energy function (15) is calculated from features (12) of each of the image representations (10) of the calibration tasks used for the pre-training of the selected machine learning model and features (13) of the image representation (11, 11') of the selected part of the control logic to be calibrated by means of the artificial intelligence (1). [3] Computer-implemented method according to one of claims 1 or 2, characterized by that a scalar energy value (14) is calculated from the energy function (15). [4] Computer-implemented method according to claim 3, characterized by that the scalar energy value (14) is used to select the most appropriate calibration task to be processed with the selected pre-trained machine learning model. [5] Computer-implemented method according to claim 4, characterized bythat the scalar energy value (14) is compared with a threshold value and the calibration task for the selected pre-trained machine learning model is selected if the scalar energy value (14) is lower than the threshold value. [6] Computer-implemented method according to claim 5, characterized by that the calibration task with the lowest scalar energy value (14) is selected for the pre-trained machine learning model if several scalar energy values ​​(14) are lower than the threshold. [7] Computer-implemented method according to one of claims 1 to 6, characterized by that latent variables (16) are used for training artificial intelligence (1).

Images