Method, apparatus, and computer program for training a machine learning model and generating training data

Abstract

Embodiments of the present invention relate to a method, apparatus, and computer program for training a machine learning model and for generating training data for such training, and more particularly to training a machine learning model to determine the position of a key device relative to a vehicle. The method for training a machine learning model includes: training a machine learning model based on data representative of at least two different vehicle environments. The machine learning model is trained to determine the position of the key device relative to the vehicle based on data from time-of-flight distance measurements of the distance between the key device and the vehicle.

Description

Technical Field

[0001] Embodiments of the present invention relate to a method, apparatus, and computer program for training a machine learning model and for generating training data for such training, in particular to the training of a machine learning model to determine the position of a key device relative to a vehicle. Background Art

[0002] A prerequisite for using a so-called keyless entry (keyless access) system (also known as a passive entry / passive start system) is the development of secure and at the same time robust methods for an authorized user to authenticate their identity to the vehicle.

[0003] This also includes a sufficiently accurate estimation of the position of the user or the identity authentication device (from when the vehicle is allowed to be unlocked or should be locked) and whether the vehicle key (smartphone or traditional key) is inside or outside the vehicle in order to grant or deny motor start permission.

[0004] In the case of a traditional key, this is done with very low radio frequencies. Currently, most passive entry systems are based on narrowband radio technology in the LF band (low frequency band, also long wave band). Once the distance between the key and the vehicle is small enough, the vehicle key establishes a connection with the vehicle. After the connection is established, positioning is carried out in another LF band. Here, a defined signal is emitted by the key and one or more receiving antennas receive signals with different signal strengths depending on the position relative to the vehicle. The reason for this is the attenuation of electromagnetic waves in different materials. Based on the received power of the signals at different receiving nodes, it can be determined whether the key is close enough to the vehicle or whether it is inside the vehicle.

[0005] When this function is implemented in a smartphone (programmable mobile phone), high-frequency radio frequencies can be used for positioning, which makes positioning very difficult. Summary of the Invention

[0006] There is a need to provide an improved method for using a smartphone as a vehicle key.

[0007] Embodiments of the present disclosure meet this need.

[0008] Embodiments of the present disclosure are based on the recognition that measurements in the high frequencies used by smartphones or other modern key devices are more associated with the environment than measurements in the low frequencies.

[0009] Due to the complex geometry of the vehicle, its electromagnetic characteristics cannot be simply calculated. Therefore, a machine learning process is used here, that is, for example, training data is measured at the vehicle and the training data is converted into a machine learning model (ML model) so that the vehicle can take over the classification of whether the authentication device is inside or outside the vehicle. Ideally, the entire solution space is covered so that the classification never fails (misclassification or blind spots, dead zones). In some cases, the process is that a person holds the key at defined positions inside, at, and around the vehicle. The person taking the measurements knows whether the key is currently inside the vehicle, outside the vehicle, or in the trunk. Then the power values of different receiving nodes are correlated with this knowledge (inside the vehicle, outside the vehicle, or in the trunk (hereinafter only further described as "inside" and "outside")). In some cases, the recording of this training data set is not carried out in a strictly specified environment. This is sufficient for measurements in the low-frequency range because the characteristics of the radio frequencies used for this are sufficiently independent of the environment, and thus this method is sufficient. Then the ML model is formed and it is checked whether the model works properly.

[0010] A new technology - Ultra-Wideband (UWB) - is used in some smartphones and other mobile devices used as key devices. The fundamental difference between this type of radio technology and LF radio is that instead of transmitting a narrowband signal with high power (i.e., low-frequency information is modulated onto a higher carrier frequency) in the SHF band (super high frequency, centimeter wave band, 3 - 30 GHz), a very wideband but low-power signal is transmitted.

[0011] Precise time-of-flight (ToF) measurements can be made by using a very wide spectrum (e.g., at least 500 MHz). By measuring the ToF (time-of-flight), the distance between the transmitter and the receiver can subsequently be calculated through the speed of light constant. In addition to ToF, the received power (RXP) can also be used as a second feature for distance calculation. In principle, the process of data processing and training a machine learning model can be applied to this radio technology.

[0012] Due to the high frequency and thus low wavelength (about 3 - 7 cm) of the electromagnetic wave, the interference effect of metal objects in the environment of the transmitter and receiver is much greater than that in the low - frequency band. The electromagnetic wave interacts very strongly with conductive geometric bodies on the order of the wavelength, that is, the wave is reflected, scattered, and diffracted. The wavelength order of LF radio is in the range of several kilometers, so in this kind of radio, there is little interaction between metal objects that are small compared to the wavelength. But for wavelengths in the centimeter range, the body - motor group, other vehicles, walls made of reinforced concrete, grilles, etc. all have a very large interference effect. Another reason is the penetration depth of the wave in the conductor. The higher the frequency of the electromagnetic wave, the lower the penetration depth into the conductor or the lower the possibility of penetrating the conductor. In LF, this penetration depth is about 100 μm, while in SHF it is about 1 μm. That is to say, high frequencies are easier to shield than low frequencies. In addition, free - space attenuation is proportional to the frequency, which also means that UWB waves are attenuated more strongly just because of their higher frequency. Moreover, in opaque objects, such as water or moisture, high - frequency waves are attenuated more strongly than low - frequency waves.

[0013] All of these cause the UWB signal to be so affected by the environment that recording a training data set in a general, non - strictly specified environment does not necessarily ensure that the model can also work properly in other environments. In principle, this would lead to the necessity of recording training data in all possible environments in order to be able to cover the solution space in all environments, which is very complex in the best case and impossible in the real situation because it is impossible to reach all conceivable vehicle environments and record training data there.

[0014] Embodiments of the present disclosure relate to how the training workload can be reduced in order to be able to generate a reliable ML model. Thereby, it is possible to classify whether an authentication device (such as a key (also called a remote key) or a smart phone) is located inside or outside the vehicle interior space. Embodiments of the present disclosure are based on the recognition that the workload for generating training data can be reduced by training a machine - learning model based on two different vehicle environments that are as different as possible in terms of possible reflections. For example, a vehicle environment with as few reflections as possible and another vehicle environment where a large number of reflections are expected (such as a vehicle environment in a densely parked underground parking lot) can be selected. Vehicle environments located between these two extremes in terms of reflections can be derived from these two vehicle environments by the machine - learning model, or the data can be additionally augmented in order to extend the training of the machine - learning model beyond these two vehicle environments.

[0015] Embodiments of the present disclosure provide a computer-implemented method for training a machine learning model. The method includes training a machine learning model based on data representative of at least two different vehicle environments. The machine learning model is trained to determine the position of a key device relative to a vehicle based on data of a time-of-flight distance measurement of the distance between the key device and the vehicle. By using data representative of two different vehicle environments, the characteristics of other vehicle environments located between the two studied vehicle environments in terms of possible reflections can be modeled by the machine learning model without additional measurements in the other vehicle environments.

[0016] For example, the at least two different vehicle environments differ in terms of possible reflections on the surfaces in the respective vehicle environments. Here, for example, vehicle environments that are as different as possible in terms of reflections can be selected so as to also be able to cover other vehicle environments located between the extremes in terms of reflections.

[0017] In principle, there are multiple possibilities for generating data for different vehicle environments. On the one hand, data can be measured. For example, data representative of at least two different vehicle environments can include at least one first data set measured in a first vehicle environment and at least one second data set measured in a second vehicle environment. Thus, for example, the vehicle environments available for measuring data can be used for training.

[0018] Alternatively or additionally, data can be generated by physical simulation. In other words, data representative of at least two different vehicle environments can include at least one first data set based on a physical simulation of a first vehicle environment and at least one second data set based on a physical simulation of a second vehicle environment or measured in a second vehicle environment. Thus, for example, vehicle environments in which data cannot be generated by measurement can be used to train the machine learning model.

[0019] In addition, the measured data or the simulated data can be augmented. For example, the method can supplement at least one data set with a plurality of additional computational data units to obtain data representative of the at least two different vehicle environments. In other words, the measured or simulated data can be supplemented with other data based on variants of the physical simulation data or the measured data. Or rather, the measured data or the simulated data can be supplemented with other data based on variants of the physical simulation data or the measured data. The additional data units can be calculated, for example, by adding artificial noise based on the respective data set. Alternatively or additionally, the additional data units can be calculated based on a position-dependent error model based on the respective data set. Alternatively or additionally, the additional data units can be calculated by interpolation between data at two positions based on the respective data set. These methods can be used to automatically generate additional training data.

[0020] For example, the flight time distance measurement and / or received signal strength may be based on one or more signals transmitted by ultra-wideband signals. In the case of UWB signals, training the machine learning model in different vehicle environments may be particularly advantageous based on the wavelength used.

[0021] In some embodiments, a machine learning model is trained to determine the position of the key device relative to the vehicle based on data of a time-of-flight distance measurement of the distance between the key device and the vehicle and based on signal strength of a signal transmitted between the key device and the vehicle. This enables a more reliable determination of the relative position of the key device.

[0022] Embodiments of the present disclosure also include a computer-implemented apparatus for training a machine learning model. The apparatus includes one or more processors and one or more storage devices. The apparatus is configured to perform a method for training a machine learning model. Embodiments also provide a vehicle having a computing module configured to determine a position of a key device relative to the vehicle using a machine learning model.

[0023] Embodiments of the present disclosure also provide a method for generating a data set for training a machine learning model. The data sets respectively include a plurality of data units, each of which contains a position of a key device relative to a vehicle, a time-of-flight distance measurement between the key device and the vehicle, and / or a signal strength of a signal transmission between the key device and the vehicle. The method includes generating a first data set in a first vehicle environment. The method also includes generating a second data set in a second vehicle environment. The two vehicle environments differ in terms of possible reflections on surfaces in the respective vehicle environments. Thus, for example, training data for the previously proposed method can be generated.

[0024] Embodiments of the present disclosure also include a computer-implemented apparatus for generating a data set for training a machine learning model. The apparatus includes one or more processors and one or more storage devices. The apparatus is configured to perform a method for generating a data set for training a machine learning model.

[0025] Embodiments of the present disclosure also include a program having a program code for performing at least one of the methods when the program code is executed on a computer, a processor, a control module or a programmable hardware component. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Some examples of the device and / or method are described in more detail below with reference to the accompanying drawings, which are merely exemplary.

[0027] Figure 1a A flow chart illustrating an embodiment of a computer-implemented method for training a machine learning model;

[0028] Figure 1b A block diagram showing an embodiment of a computer-implemented apparatus for training a machine learning model;

[0029] Figure 2a A flowchart showing an embodiment of a method for generating a data set for training a machine learning model; and

[0030] Figure 2b A block diagram showing an embodiment of an apparatus for generating a data set for training a machine learning model. DETAILED DESCRIPTION

[0031] The different examples will now be described in more detail with reference to the accompanying drawings, in which some examples are shown. For clarity, the thickness of lines, layers, and / or regions in the drawings may be exaggerated.

[0032] It will be understood that when an element is referred to as being "connected" or "coupled" to another element, these elements can be directly connected or coupled, or connected or coupled through one or more intermediate elements. When two elements A and B are combined using "or", this can be understood as disclosing all possible combinations, i.e., only A, only B, and A and B, unless otherwise explicitly or implicitly defined. An alternative expression for the same combination is "at least one of A and B" or "A and / or B". This applies, mutatis mutandis, to combinations of more than two elements.

[0033] Unless otherwise defined, all terms (including technical and scientific terms) are used herein in their ordinary meaning in the field to which the examples belong.

[0034] Generally speaking, embodiments of the present disclosure relate to the positioning of an authentication device, such as a key device, in the interior or exterior space of a vehicle. Thus, embodiments of the present disclosure particularly relate to an authentication device or key device for a keyless entry system and a keyless driving system for a vehicle.

[0035] Figure 1a A flowchart showing an embodiment of a computer-implemented method for training a machine learning model. The method includes training 120 a machine learning model based on data representative of at least two different vehicle environments. The machine learning model is trained to determine the position of a key device relative to a vehicle based on data of a time-of-flight distance measurement of the distance between the key device and the vehicle.

[0036] Figure 1bA block diagram showing an embodiment of a corresponding computer-implemented apparatus 10 for training a machine learning model. The apparatus includes one or more processors 14 and one or more storage devices 16. Optionally, the apparatus further includes an interface 12, for example for receiving training data or for providing a trained machine learning model. The one or more processors are coupled to the optional interface and the one or more storage devices. Generally, the functions of the apparatus are provided by the one or more processors with the assistance of the one or more storage devices and / or the optional interface. The apparatus is configured to execute Figure 1a the method of.

[0037] The following description relates to both Figure 1a the method of and Figure 1b the corresponding apparatus of.

[0038] At least some aspects of the present disclosure relate to methods, apparatuses, and computer programs for training a machine learning model. Machine learning involves algorithms and statistical models that a computer system can use to perform a specific task without using explicit instructions, rather than relying on models and inferences. In machine learning, for example, instead of rule-based data transformation, data transformation can be used that can be derived from the analysis of process data (Verlaufsdaten) and / or training data. Machine learning is used in a large number of applications, such as for object recognition in image data, for prediction of time series, for pattern analysis, etc. Generally, the following fact is utilized here: in many cases, so-called training data, i.e., data representing examples of the expected transformation of a corresponding machine learning model, is sufficient as a basis for training a machine learning model that should perform a specific task (so-called model "training").

[0039] For example, a machine learning model or a machine learning algorithm can be used to analyze the content of an image. To enable the machine learning model to analyze the content of an image, training images can be used as inputs and training content information can be used as outputs to train the machine learning model. By training the machine learning model using a large number of training images and / or training sequences (such as words or sentences) and associated training content information (such as labels or annotations), the machine learning model "learns" to recognize the content of the image, so that the machine learning model can be used to recognize the image content not included in the training data. The same principle can also be applied to other types of sensor data: by training the machine learning model using training sensor data and the desired outputs, the machine learning model "learns" the transformation between the sensor data and the outputs, which can be used to provide outputs based on non-training sensor data provided to the machine learning model. This can be used to provide outputs based on non-training sensor data provided to the machine learning model. The provided data (such as sensor data, metadata, and / or image data) can be preprocessed to obtain a feature vector, which is used as an input for the machine learning model.

[0040] In the current case, the machine learning model is trained to determine the position of the key device relative to the vehicle based on data of time-of-flight distance measurement (also known as Time-of-Flight-Ranging in English) of the distance between the key device and the vehicle. In addition to the time-of-flight distance measurement, the signal strength of the signal transmission between the key device and the vehicle (such as the signal of the time-of-flight measurement) can also be used as an input value for the machine learning model. In other words, the machine learning model can be trained to determine the position of the key device relative to the vehicle based on data of time-of-flight distance measurement of the distance between the key device and the vehicle and based on the signal strength of the signal transmission between the key device and the vehicle. Here, the key device can be, for example, a radio key (also known as a remote control key) or a mobile device, such as a programmable mobile phone (smartphone) or a so-called wearable device (a mobile device that can be worn on the body). The time-of-flight distance measurement and / or the received signal strength can be based on one or more signals transmitted by ultra-wideband (UWB) signals. However, other high-frequency (HF, also known as radio frequency, RF) or low-frequency (LF) signal transmissions can also be used for time-of-flight measurement and received signal strength measurement.

[0041] The transit time distance measurement and optionally the received signal strength can be used here as input values for a machine learning model and information about the respective position of the key device relative to the vehicle can be provided by the machine learning model as output values. Here the input values are also referred to as so-called "features". Now in order to train the machine learning model (using the so-called supervised learning method), the respective input data and output data are used as training input data and training output data and the machine learning model is trained to provide a transformation that generates the respective output data from the training input data for all training data sets.

[0042] The training input data can be used to train the machine learning model. The above example uses a training method called supervised learning. In supervised learning, the machine learning model is trained using a plurality of training data units, each data unit comprising one or more training input data and one or more desired output values, i.e. an expected output value is assigned to a combination of the one or more training input data. By specifying the training input data and the desired output values, the machine learning model "learns" which output value to provide based on input data similar to the training input data provided during training. In the proposed method, the data of the transit time distance measurement of the distance between the key device and the vehicle and optionally the signal strength of the signal transmission between the key device and the vehicle constitute the input data, i.e. the training input data, and the position of the key device relative to the vehicle constitutes the output data, i.e. the training output data. In other words, the machine learning model is trained to output the position of the key device relative to the vehicle when the data of the transit time distance measurement of the distance between the key device and the vehicle and optionally the signal strength of the signal transmission are applied to one or more input terminals of the machine learning model. Here the position of the key device relative to the vehicle can be classified, for example, according to one of two or three categories, such as "inside the vehicle", "outside the vehicle" and optionally "boot". As an alternative, the position of the key device relative to the vehicle can be represented in a partition- or sector-based system relative to the vehicle.

[0043] A machine learning model is trained based on data representative of at least two (or exactly two) different vehicle environments. In other words, the training data units for training the machine learning model represent at least two (or exactly two) different vehicle environments. The at least two different vehicle environments may differ in terms of possible reflections on surfaces in the respective vehicle environments. For example, the first vehicle environment may be a so-called "open field" vehicle environment, i.e., a vehicle environment in which reflections, scattering, or diffraction of signals on objects outside the vehicle are reduced or minimized. In contrast, the second vehicle environment may be a vehicle environment in which a large number of reflections of the respective signals can occur. For this purpose, for example, a vehicle environment in a parking lot with a low ceiling and narrow side walls can be selected. The vehicle environment relates on the one hand to the vehicle environment outside the vehicle, i.e., objects, surfaces, etc. outside the vehicle. In addition, the vehicle environment may refer to objects inside the vehicle, such as the cargo or passengers being carried.

[0044] In principle, the corresponding training data has two sources: on the one hand, the corresponding training data can be measured in a "real" vehicle environment. On the other hand, the corresponding training data can be generated by physical simulation. In addition, as will be further explained in detail below, the measurement data or simulation data is supplemented by so-called augmentation by additional data units. For example, data representative of at least two different vehicle environments may include at least one first data set measured in a first vehicle environment. In addition, data representative of at least two different vehicle environments may include at least one second data set measured in a second vehicle environment. For this purpose, for example, a key device can perform time-of-flight measurements and optionally received signal strength measurements at multiple positions relative to the vehicle in a real vehicle environment, which can be used as training input data for the training data units. The respective positions relative to the vehicle can be used as the desired output.

[0045] Alternatively or additionally, simulation data can be used. In other words, data representative of at least two different vehicle environments may include at least one first data set based on a physical simulation of a first vehicle environment. This data set can be used together with another simulation data set or with a measurement data set. In other words, data representative of at least two different vehicle environments may include at least one second data set based on a physical simulation of a second vehicle environment or measured in a second vehicle environment. For example, the physical simulation may correspond to a field simulation (Feldsimulation). For this purpose, a vehicle model can be created in a simulation environment. A plurality of spatial points can be determined inside and outside the vehicle, at which synthetic measurements are calculated based on physical parameters such as distance-dependent attenuation, attenuation when penetrating materials, reflections at the vehicle and the environment, occlusion of signals by vehicle components and the environment, and increases in time-of-flight and attenuation caused by non-line-of-sight transmission. Here, in order to simulate one vehicle environment, reflections outside the vehicle can be disregarded (open field simulation). And in order to simulate another vehicle environment, a plurality of additional reflection surfaces outside and optionally inside the vehicle can be introduced into the model.

[0046] In addition to measurement data or simulation data, other data units can be generated, and the other data units represent (reasonable or credible) deviations from the measurement data or simulation data. In other words, the method can supplement at least one of the at least two data sets 110 with a plurality of additional calculated data units in order to obtain data representing the at least two different vehicle environments. In other words, the data set for training the machine learning model can be augmented with synthetically generated data units. For example, additional data units or at least some of the additional data units can be calculated by adding artificial noise based on the corresponding data set. In other words, additional data units can be generated by adding additional random or deterministic noise (i.e., pseudo-random deviations) to the simulated or measured data units. Alternatively or additionally, additional data units or at least some of the additional data units can be calculated by interpolation between data at two positions based on the corresponding data set. In other words, a third data unit can be calculated for a position between two positions from two data units calculated for two positions relative to the vehicle, and the value of the third data unit is between the values of the two data units.

[0047] In some embodiments, additional data units or at least some of the additional data units can be calculated based on a location-dependent error model according to the corresponding data set. Such a location-dependent error model is based on the fact that signals are not received at certain locations relative to the vehicle in an open field, while signals are received in a reflective environment. This characteristic can be modeled as an error model and used to generate such errors caused by environmental factors. In the location-dependent error model, the area around the vehicle can, for example, be divided into smaller areas. For each area, it can be determined which characteristic changes of the features occur in different environments. From this, a region-specific environmental interference model can be generated, that is, a model that specifically models the interference effects of different environments for each region. Its application is an augmentation that uses the recorded training data of one environment to generate additional training data describing other environments. Then, these additional training data can be used to train the machine learning model.

[0048] The interface 12 can, for example, correspond to one or more input terminals and / or one or more output terminals for receiving and / or transmitting information in the form of digital bit values based on code, for example, within a module, between modules, or between modules of different entities.

[0049] In an embodiment, the one or more processors 14 may correspond to any controller or processor or programmable hardware component. For example, the one or more processors 14 may also be implemented as software programmed for the corresponding hardware component. In this regard, the one or more processors 14 may be implemented as programmable hardware with corresponding adapted software. Any processor may be used herein, such as a digital signal processor (DSP). The embodiments are not limited to a particular type of processor. Any processor or processors may be contemplated for implementation.

[0050] The one or more storage devices 16 may be, for example, at least one element of a group including a computer-readable storage medium, a magnetic storage medium, an optical storage medium, a hard disk drive, a flash memory, a floppy disk, a random access memory, a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), and a network memory.

[0051] Machine learning algorithms are generally based on machine learning models. In other words, the term "machine learning algorithm" may refer to a set of instructions that can be used to create, train, or use a machine learning model. The term "machine learning model" may refer to a data structure and / or set of rules that represents the learned knowledge (e.g., based on training performed by a machine learning algorithm). In an embodiment, using a machine learning algorithm may implicitly imply using one underlying machine learning model (or multiple underlying machine learning models). Using a machine learning model may implicitly imply training the machine learning model via a machine learning algorithm and / or the data structure / set of rules that is the machine learning model.

[0052] For example, a machine learning model may be an artificial neural network (ANN). An artificial neural network is a system inspired by biological neural networks (such as the neural networks that can be seen in the retina or the brain). An artificial neural network includes a plurality of interconnected nodes and a plurality of connections between the nodes, i.e., so-called edges. There are generally three types of nodes, namely input nodes that receive input values, hidden nodes that are (only) connected to other nodes, and output nodes that provide output values. Each node may constitute an artificial neuron. Each edge may send information from one node to another node. The output of a node may be defined as a (non-linear) function of the input (such as the sum of its inputs). The input of a node may be used in the function based on the "weights" of the edges or nodes that provide the input. The weights of the nodes and / or edges may be adjusted during the learning process. In other words, training an artificial neural network may include adjusting the weights of the nodes and / or edges of the artificial neural network, i.e., in order to achieve a desired output for a particular input.

[0053] As an alternative, the machine learning model can be a support vector machine, a random forest model, or a gradient boosting model. A support vector machine (i.e., a support vector network) is a supervised learning model with a corresponding learning algorithm for analyzing data (e.g., in classification or regression analysis). A support vector machine can be trained by providing inputs with multiple training input values belonging to one of two classes. The support vector machine can be trained to assign new input values to one of the two classes. As an alternative, the machine learning model can be a Bayesian network, which is a probabilistic directed acyclic graphical model. A Bayesian network can use a directed acyclic graph to represent a set of random variables and their conditional dependencies. As an alternative, the machine learning model can be based on a genetic algorithm, which is a search algorithm and heuristic technique that mimics the process of natural selection.

[0054] Combined with the scenarios or examples described before or after (such as Figure 2a and 2b ), more details and aspects of the method and apparatus are given. The method and apparatus can include one or more additional optional features corresponding to one or more aspects of the proposed scenarios or the described examples as described before or after.

[0055] Figure 2a A flowchart showing an embodiment of a method (such as a computer-implemented method) for generating a data set for training a machine learning model is shown. These data sets each include a plurality of data units, and the data units include the position of the key device relative to the vehicle, the transit time distance measurement between the key device and the vehicle, and / or the signal strength of the signal transmission between the key device and the vehicle. The method includes generating 210 a first data set in a first vehicle environment. The method further includes generating 220 a second data set in a second vehicle environment. These two vehicle environments differ in terms of possible reflections on the surfaces in the respective vehicle environments (as already explained in combination with Figure 1a and / or 1b).

[0056] Figure 2b A block diagram showing an embodiment of a corresponding computer-implemented apparatus 20 for generating a data set for training a machine learning model is shown. The apparatus includes one or more processors 24 and one or more storage devices 26. Optionally, the apparatus further includes an interface 22, for example, for receiving training data or for providing a trained machine learning model. The one or more processors are coupled to the optional interface and the one or more storage devices. Generally, the functions of the apparatus are provided by the one or more processors with the assistance of the one or more storage devices and / or the optional interface. The apparatus is configured to execute Figure 2a the method.

[0057] The following description relates to both Figure 2a the method and Figure 2bThe corresponding device.

[0058] Some embodiments of the present disclosure relate to the generation of training data for training a machine learning model (such as Figure 1a and / or the machine learning model of 1b). The method is applicable to generating a data set for training a machine learning model. These data sets each include a plurality of data units, and the data units include the position of the key device relative to the vehicle (as the desired output value), the time-of-flight distance measurement between the key device and the vehicle (as the training input data), and / or the signal strength of the signal transmission between the key device and the vehicle (as the training input data).

[0059] As already combined with Figure 1a and / or as described in 1b, the training data can in principle be generated by two methods, namely by measurement and physical simulation.

[0060] For example, generating 210 the first data set in the first vehicle environment can include performing multiple measurements at a plurality of predetermined positions inside and outside the vehicle, for example, in order to determine the time-of-flight distance measurement and / or the signal strength of the signal transmission for a plurality of positions of the key device relative to the vehicle in the first vehicle environment. Similarly, generating 220 the second data set in the second vehicle environment can include performing multiple measurements at a plurality of predetermined positions inside and outside the vehicle, for example, in order to determine the time-of-flight distance measurement and / or the signal strength of the signal transmission for a plurality of positions of the key device relative to the vehicle in the second vehicle environment.

[0061] Alternatively or additionally, simulated data can be used. For example, generating 210 the first data set in the first vehicle environment can include performing physical simulations at a plurality of predetermined positions inside and outside the vehicle in the model, for example, in order to determine the time-of-flight distance measurement and / or the signal strength of the signal transmission for a plurality of positions of the key device relative to the vehicle in the model of the first vehicle environment. Similarly, generating 220 the second data set in the second vehicle environment can include performing physical simulations at a plurality of predetermined positions inside and outside the vehicle in the model, for example, in order to determine the time-of-flight distance measurement and / or the signal strength of the signal transmission for a plurality of positions of the key device relative to the vehicle in the model of the second vehicle environment.

[0062] The details of these two methods and the two vehicle environments (which differ in terms of possible reflections on the surfaces in the respective vehicle environments) have been mentioned in combination with the description of Figure 1a and 1b They will be discussed in more detail below.

[0063] In some embodiments, the method further includes training 230 a machine learning model based on the generated data set, for example, similar to training the machine learning model in 120 Figure 1a This can for example include supplementing 110 the data set.

[0064] The interface 22 can, for example, correspond to one or more input terminals and / or one or more output terminals for receiving and / or transmitting information in the form of digital bit values based on code, for example, within a module, between modules, or between modules of different entities.

[0065] In an embodiment, the one or more processors 24 can correspond to any controller or processor or programmable hardware component. For example, the one or more processors 24 can also be implemented as software programmed for the corresponding hardware component. In this regard, the one or more processors 24 can be implemented as programmable hardware with correspondingly adapted software. Any processor can be used herein, such as a digital signal processor (DSP). The embodiments are not limited to a specific type of processor. Any processor or multiple processors can be contemplated for implementation.

[0066] The one or more storage devices 26 can, for example, be at least one element of a group including a computer-readable storage medium, a magnetic storage medium, an optical storage medium, a hard disk drive, a flash memory, a floppy disk, a random access memory, a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), and a network memory.

[0067] In combination with the scenarios or examples described previously (such as Figure 1a and 1b ), more details and aspects of the method and apparatus are given. The method and apparatus can include one or more additional optional features corresponding to one or more aspects of the proposed scenarios or the described examples as described previously or subsequently.

[0068] Embodiments of the present invention are based on two environments (such as two vehicle environments) as a basis for training a machine learning model. Thus, in some embodiments, training data is recorded in exactly two environments rather than in all possible environments. These two environments, in terms of their characteristics, represent two extreme cases that may generally occur in the environment. The first environment should be an environment without any reflections. This environment is hereinafter referred to as the "open field". It should be noted here that there are no metal / reflective objects within a radius of at least 5 meters. This can ensure that the received UWB packets can only reach the corresponding anchor on the "line of sight" (LOS) path. This in turn results in a rather low availability of the anchor. In this regard, "availability" represents the ratio between the received packets and the transmitted packets.

[0069] As a second environment, an environment with very strong reflections is recommended, such as an underground parking lot with a low ceiling height. Additionally, the selected parking space can be surrounded by two reinforced concrete walls, so that reflections occur both above and on the sides of the vehicle. In this environment, the anchor point availability is significantly improved because the packets can also be exchanged on "non-LOS" paths.

[0070] With the training data from these two environments, a large part of the complete solution space can be covered (e.g., supermarket parking lots or roadside parking).

[0071] Training in multiple environments can be improved by other methods. Thus, in some embodiments, additional data processing algorithms and data augmentation are used to train a machine learning model in different environments.

[0072] Some embodiments are based on generating synthetic training data by means of in-situ simulation (augmentation / augmentation). To further reduce the effort of obtaining training data, instead of measurement data, synthetically generated data based on physical models can be used in the training algorithm. The so-called in-situ simulation generates spatial points in and around the vehicle. Based on the physical model, the values of the corresponding points can now be calculated. The challenge here lies in the modeling of the vehicle and the environment. Which parts of the vehicle cause complete signal occlusion? Which parts of the vehicle cause greater signal attenuation? Which areas of the vehicle cause the signal to be transmitted with an increased transit time after multiple reflections? In an exemplary implementation, the model is constructed as follows.

[0073] Spatial points outside the vehicle are generated with the following parameters:

[0074] minDistCar (representing the minimum distance from the vehicle from which points are generated)

[0075] maxDistCar (representing the maximum distance from the vehicle until which points are generated)

[0076] gridDist (representing the distance between points)

[0077] zPoints (representing the number of planes in the z direction in which points are generated)

[0078] Spatial points are generated around the vehicle contour according to the parameters. In a simplified calculation, the contour of the vehicle is assumed to be a cube with the maximum dimensions of the vehicle. To calculate features, received signal strength (RXP), and distance, the positions of the anchor points should also be known.

[0079] First, vectors can be set from each anchor point to each generated spatial point. The lengths of these vectors give a first approximation for calculating two features, distance and received signal strength. Then it can be determined at which positions these vectors leave the vehicle's contour and how long the path through the vehicle is. Additional attenuation can be estimated based on the lengths of these vectors within the vehicle's contour.

[0080] The simulation only includes data under open field conditions at this time (no reflections, only LOS connections). To also simulate data from other environments, the model can also be equipped with different reflecting surfaces, so that NLOS paths (non-line-of-sight paths) can also be achieved. In addition, the vehicle model can also be specified. Since only a limited number of connections are established and there are not infinitely many receiving nodes, in some embodiments, it is not necessary to fully simulate all beams.

[0081] In some embodiments, it is pursued to generate synthetic training data by means of (augmenting) random noise. As a second processing step, a process can be used that generates additional training data by changing existing training data with an interference process. Thus, additional data units can be added to the training data. Such an interference process can be random or deterministic. Such a perturbation applied, for example, simulates interference on the connection path, such as interference caused by body parts or bags containing items. It has proven advantageous here to have uniformly distributed noise up to a 10 dB limit for the received signal data.

[0082] In some embodiments, it is pursued to generate synthetic training data by means of (augmenting) an error model. Data analysis shows that in certain regions, certain anchor points do not provide signals in the open field but receive signals in a reflective environment. This characteristic can be modeled as an error model and used to generate such errors caused by environmental factors.

[0083] At least some embodiments thus provide for the determination of a region-specific environmental model. For this purpose, the area around the vehicle is, for example, divided into smaller areas. It can be determined for each area which characteristic changes occur in the features in different environments. From this, a region-specific environmental interference model can be generated, that is, a model that specifically models the interference effects of different environments for each area. Its application is an augmentation that uses the training data of one recorded environment to generate additional training data describing other environments. These additional training data can then be used to train a machine learning model.

[0084] Accordingly, the embodiments provide a method for recording measurement data in different environments. Here, for example, training data can be recorded in two different specified environments (open field, underground parking lot). The embodiments also provide a method for generating additional training data (augmentation) to improve the performance of the ML model. Here, additional training data based on the measured training data can be incorporated into the formation of the ML model. For example, existing training data can add a location-independent interference process as additional training data. For example, additional points between two points belonging to the same category in space can be added to the training for augmentation. For example, existing training data can add a location-dependent interference process as additional training data. Alternatively or additionally, additional synthetically generated training data can be incorporated into the formation of the ML model. In some embodiments, spatial points can be generated outside the vehicle and eigenvalue can be calculated based on a physical model.

[0085] Aspects and features described in conjunction with one or more of the previous detailed examples and figures can also be combined with one or more other examples to replace the same feature in another example or to introduce that feature additionally into another example.

[0086] The examples can also be a computer program with program code for performing one or more of the above methods or related thereto when the computer program is executed on a computer or a processor. The steps, operations or processes of the above methods can be executed by a programmed computer or processor. The examples can also cover a program storage device, such as a digital data storage medium, which is machine-readable, processor-readable or computer-readable and encodes an instruction program executable by a machine, a processor or a computer. These instructions perform some or all of the steps of the above methods or cause their execution. The program storage device can, for example, include or be a digital memory, a magnetic storage medium such as disks and tapes, a hard disk drive, or an optically readable digital data storage medium. Other examples can also include a computer, a processor or a control unit programmed to perform the steps of the above methods, or a (field) programmable logic array ((F)PLA) or a (field) programmable gate array ((F)PGA) programmed to perform the steps of the above methods.

[0087] The description and the figures merely illustrate the principles of the present disclosure. In addition, all the examples listed herein are in principle clearly only for illustrative purposes to help the reader understand the principles of the present disclosure and the design concepts contributed by the inventors to improve the technology. All statements herein regarding the principles, aspects and examples of the present disclosure and their specific examples include their equivalents.

[0088] The functions of the different elements shown in the drawings, including the functions of each functional block referred to as "means", "means for providing a signal", "means for generating a signal", etc., can be implemented in the form of dedicated hardware, such as "signal provider", "signal processing unit", "processor", "control means", etc., and can be implemented as hardware capable of executing software in combination with relevant software. When provided by a processor, the functions can be provided by a single dedicated processor, a single shared processor, or multiple independent processors, some or all of which can be shared. However, the terms "processor" or "control means" are in no way limited to hardware capable of only executing software, but can include digital signal processor hardware (DSP hardware), network processors, application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), read only memories (ROMs) for storing software, random access memories (RAMs), and non-volatile storage devices. Other hardware, conventional and / or customer specific, can also be included.

[0089] A block diagram can represent, for example, a rough circuit diagram implementing the principles of the present disclosure. Similarly, a flowchart, sequence diagram, state transition diagram, pseudocode, etc. can represent various processes, operations, or steps, which are, for example, substantially shown in a computer-readable medium and are thus executed by a computer or processor, whether or not such a computer or processor is explicitly shown. The methods disclosed in the specification or claims can be implemented by a component that includes means for performing each step of these methods.

[0090] It goes without saying that, unless explicitly or implicitly stated otherwise, for example for technical reasons, the multiple steps, processes, operations, or functions disclosed in the specification or claims should not be construed as being in a particular order. Thus, the disclosure of multiple steps or functions does not limit them to a particular order, unless these steps or functions cannot be interchanged for technical reasons. Additionally, in some examples, a single step, function, process, or operation can include and / or be decomposed into multiple sub-steps, functions, processes, or operations. These sub-steps can be included and form part of the disclosure of that single step, unless explicitly excluded.

[0091] Furthermore, the following claims are hereby incorporated into the detailed description, and each claim can stand alone as a separate example. Although each claim can stand alone as a separate example, it should be noted that while the dependent claims in the claims can relate to a particular combination with one or more other claims, other examples can also include combinations of the technical solutions of the dependent claims with any other dependent or independent claims. These combinations are explicitly set forth herein unless it is stated that a particular combination is not considered. Additionally, the features of one claim should also be included in any other independent claim, even if that claim does not directly refer to that independent claim.

Claims

1. A computer-implemented method for training a machine learning model, the method comprising: training (120) a machine learning model based on a data set representative of at least two different vehicle environments, wherein the machine learning model is trained to determine the position of a key device relative to a vehicle based on data of a time-of-flight distance measurement of the distance between the key device and the vehicle, and supplementing (110) at least one data set with a plurality of calculated additional data units to obtain data representative of the at least two different vehicle environments, wherein the additional data units are calculated based on a location-related error model based on the corresponding data set, wherein the location-related error model is based on not receiving a signal at certain positions relative to the vehicle in an open field while receiving a signal in a reflective environment, wherein, in the location-related error model, the area around the vehicle is divided into smaller zones, and for each zone, which characteristic changes of features occur in different environments are determined and thereby a zone-specific environmental interference model is generated.

2. The method according to claim 1, wherein, The at least two different vehicle environments differ in terms of reflection on the surface in the respective vehicle environment.

3. The method according to claim 1 or 2, wherein The data representative of the at least two different vehicle environments includes at least one first data set measured in a first vehicle environment and at least one second data set measured in a second vehicle environment.

4. The method according to claim 1 or 2, wherein The data representative of the at least two different vehicle environments includes at least one first data set based on a physical simulation of a first vehicle environment and at least one second data set based on a physical simulation of a second vehicle environment or measured in a second vehicle environment.

5. The method according to claim 1 or 2, wherein The time-of-flight distance measurement and / or the received signal strength is based on one or more signals of ultra-wideband signal transmission, and / or the machine learning model is trained to determine the position of the key device relative to the vehicle based on data of the time-of-flight distance measurement of the distance between the key device and the vehicle and based on the signal strength of the signal transmission between the key device and the vehicle.

6. The method according to claim 1 or 2, wherein The additional data units are calculated by adding artificial noise based on the corresponding data set, and / or by interpolation between data of two positions based on the corresponding data set.

7. A method for generating a dataset for training a machine learning model, wherein, The data sets each include a plurality of data units, the data units containing the position of the key device relative to the vehicle, the time-of-flight distance measurement between the key device and the vehicle, and / or the signal strength of the signal transmission between the key device and the vehicle, the method comprising: generating (210) a first data set in a first vehicle environment; and generating (220) a second data set in a second vehicle environment, the first vehicle environment and the second vehicle environment differing in terms of reflection on the surface in the respective vehicle environment, and Supplement (110) at least one data set with additional data units that are calculated, to obtain data representative of a first vehicle environment and a second vehicle environment, wherein the additional data units are calculated based on a position-dependent error model based on the respective data set, wherein the position-dependent error model is based on not receiving a signal at certain positions relative to the vehicle in an open field while receiving a signal in a reflective environment, wherein in the position-dependent error model, the area around the vehicle is divided into smaller zones, and for each zone it is determined which characteristic changes of features occur in different environments and thereby a zone-specific environmental interference model is generated.

8. The method according to claim 7, wherein Calculate the additional data units by adding artificial noise based on the respective data set, and / or by interpolating between data at two positions based on the respective data set.

9. A program product having program code for performing at least one method according to any one of claims 1 to 8 when the program code is executed on a computer, a processor, a control module or a programmable hardware component.

10. A computer-implemented apparatus (10) for training a machine learning model, the apparatus comprising one or more processors (14) and one or more storage devices (16), wherein, The device is configured to perform the method according to any one of claims 1 to 6.

Images