Method and control device for automatically selecting a data set for a method for machine learning

Abstract

A method for automatically selecting a data set for a method for machine learning to detect operating variables of a motor vehicle is proposed, in which a sequence of measurement signals for a specific range of operating variables (10.3.1, 10.3.2, A, B, C) of the motor vehicle is detected during operation of the motor vehicle. The method is characterized in that storage areas are assigned to the range of operating variables (10.3.1, 10.3.2, A, B, C), each storage area having a plurality of slots (30), each slot being arranged to store a data set containing the detected sequence of measurement signals, and in that the data sets already stored in the slots (30) can be overwritten separately with newly detected data sets for the same storage area. For each slot (30) of the storage area in which a data set is stored, a slot error estimate value is formed and stored together with the sequence of measurement signals. Data sets with relatively small slot error estimate values are preferentially overwritten.

Description

Field of the Invention

[0001] The present invention relates to a method for automatically selecting a data set for a method for machine learning to detect operating variables of a motor vehicle, in which a sequence of measurement signals for a specific operating variable range of the motor vehicle is detected during operation of the motor vehicle. In terms of equipment, the present invention relates to a control device according to the preamble of claim 11. Such a method and such a control device are premised on being known per se. Background Art

[0002] Data is the "fuel" of machine learning algorithms, in which functions for determining difficult-to-detect variables are modeled based on large databases. In technical systems, signals are recorded by a data logger and stored locally or in a distributed memory (such as the cloud) for further processing later.

[0003] In this way, a very large amount of data will accumulate, and this data is usually not easily usable for machine learning. In addition to data preprocessing, most machine learning methods require generating training data sets, validation data sets, and test data sets from the available data.

[0004] As a specific example, the following scenario should first be considered. A computational model for machine learning should be trained, validated, and tested before the start of mass production of motor vehicles, and this computational model can determine the water film depth on the road surface of the driving road. The water film depth is an example of an operating variable of a road motor vehicle to be detected. The input variables of the computational model are driving speed, ambient temperature, and the sound spectrum from tire rolling noise. During a series of measurement drives, measurement data is collected over hundreds of hours. If training data sets, validation data sets, and test data sets are now clumsily generated from these measurement data, the following situation may occur: The selected sequence of measurement signals (for example, for wet road conditions) always randomly has a speed within a specific range. The method for machine learning will learn this non-causal relationship during training. The computational model learned before the start of mass production is taken over by the control device during subsequent mass production and does not change anymore. Then, the non-causal relationship learned during the pilot production phase may lead to incorrect results during the operation of motor vehicles produced in subsequent mass production.

[0005] To avoid this situation, the collected data is usually accurately analyzed before being divided into training data sets, validation data sets, and test data sets. The collected data may include several terabytes. So in most cases, for practical reasons, not all available data, but only a subset of this data, is analyzed and used. In addition, techniques such as cross-validation are used when validating and testing the learned model, in which such an "unfavorable" distribution of the data set is identified. Summary of the Invention

[0006] The difference between the present invention and the prior art lies in the characteristic features of the independent claims of the present invention. The method according to the present invention is characterized in that storage areas are allocated to ranges of operating variables, each storage area having a plurality of slots, each slot being arranged to store a data set containing a measurement signal sequence, and wherein the data sets already stored in the slots can each be overwritten with a data set newly detected for the same storage area at present. For each slot of the storage area in which a data set is stored, a slot error estimate value of the measurement signal sequence contained in the data set is formed and stored together with the measurement signal sequence, and when overwriting, the following data sets are preferentially overwritten, the slot error estimate value of which is less than the slot error estimate values of other slots formed for the same storage area. The control device according to the present invention is characterized in that the control device is arranged to control the process of the method according to the present invention and thus execute the method according to the present invention.

[0007] When detected, the data set is directly subdivided into a data set to be stored and a data set not to be stored. This saves the time-consuming analysis of possibly very large amounts of data. In addition, the storage of unnecessarily large redundant signal regions that do not provide additional information content is avoided, so less storage space is required. Thus, the amount of data and the analysis effort required for machine learning can be significantly reduced, and computing time, storage capacity, and data volume can be saved.

[0008] The selection of relevant measurement signal sequences already carried out when storing measurement data in a vehicle according to the present invention keeps the costs required for temporarily storing data and the computing time required for learning in the vehicle and transmitting data for computing in the cloud if necessary at a low level.

[0009] During machine learning, the individual slot error estimate value of each slot is calculated from the data in the slot, which corresponds to the error of the measurement signal sequence or the associated data set stored in the slot. The larger the slot error estimate value, the more interesting or instructive and thus valuable the data set for subsequent machine learning iterations. Therefore, when overwriting the slot content with a new data set (especially a measurement signal sequence and the associated slot error estimate value), the following data sets are preferentially used, i.e., overwritten, the slot error estimate value of which is less than the slot error estimate values of similar slots. Similar slots are those assigned to the same or similar measurement signal sequences. Thereby, in particular, the following data sets are retained in the memory, the measurement signal sequences of which have relatively large errors. This is advantageous for machine learning because training can be carried out more effectively using data sets with relatively large errors than using data sets with relatively small errors. Thereby, memory, computing time can be saved, and the amount of data to be transmitted can be saved in the case of transmission to an external server.

[0010] A preferred design is characterized in that a computing model is used to calculate the operating characteristic parameters of the motor vehicle from the currently detected sequence of measurement signals, and in parallel, the operating characteristic parameters are measured with a reference value sensor, and the slot error estimate value is formed based on the deviation between the calculated operating characteristic parameters and the measured operating characteristic parameters.

[0011] Another preferred design is characterized in that the computing model is a computing model to be trained by a method for machine learning.

[0012] Also preferably, only those data sets for which a slot error estimate value has been calculated during machine learning are released for overwriting.

[0013] Further preferably, when all the slot contents released for overwriting have been overwritten before the machine learning process can start, the next data set to be overwritten again is selected according to a random principle.

[0014] By randomly overwriting multiple times, it can be ensured that not only the latest data is available for training, but also some much older data is available for training. This increases the probability of providing particularly interesting new data in the memory.

[0015] Also preferably, only a predetermined portion of the data sets for which a slot error estimate value has been calculated during machine learning in the driving cycle of the motor vehicle is released for overwriting.

[0016] Another preferred design is characterized in that the data set is transmitted to a memory located outside the motor vehicle.

[0017] Also preferably, the data sets stored in the motor vehicle are sorted in descending order according to their slot error estimate values, and the transmission is performed in this sorted order and starting from the measurement signal sequence with the largest slot error estimate value.

[0018] Thereby, the most interesting data sets are always provided to the server and other vehicles for their learning algorithms, and thereby, compared with randomly selecting measurement data, the memory requirements and computing time required for learning can be significantly reduced.

[0019] Ideally, data sets of special driving maneuvers or reference drives are also pre-stored in the vehicle, and these data sets cannot be overwritten and are used to verify the machine learning. Thereby, it can be ensured that the model maintains the necessary robustness, even if the results of each individual learning process are not monitored by humans.

[0020] A computer program according to the present invention is characterized in that the computer program comprises computer-readable instructions which, when executed on a computer, run a method according to one of the above methods. A computer program according to the present invention is characterized in that a machine-readable storage medium stores thereon the computer program.

[0021] The design of a control device according to the present invention is characterized in that the control device is arranged to control the process of one of the above methods and thus execute the method. Description of the Drawings

[0022] Other advantages result from the following description, the drawings and the dependent claims. It is understood that the above features and the features to be explained below can be used not only in the combinations shown respectively, but also individually, without departing from the scope of the present invention.

[0023] Embodiments of the present invention are shown in the drawings and are explained in more detail in the following description. Herein, in a schematic form respectively:

[0024] Figure 1 A functional block diagram showing the method aspect and the device aspect of an embodiment of the present invention is shown; and

[0025] Figure 2 An embodiment of a selected measurement signal sequence having the features of the present invention is shown, and the selection can be made in Figure 1 the recording block. Detailed Description of the Invention

[0026] Specifically, Figure 1 Block 10 is shown, which represents a set of different sensors of a motor vehicle. In one embodiment, the set has an ambient temperature sensor, a sensor for detecting the vehicle speed, and a sensor for detecting the sound spectrum originating from the rolling noise of the tires. The set may also have additional and / or other sensors which detect operating variables of the motor vehicle.

[0027] Operating characteristic parameters of the motor vehicle are calculated in the calculation model 12 based on the measurement signal sequence detected by these sensors. The operating characteristic parameters are, for example, operating characteristic parameters which are difficult to measure, time-consuming to measure or costly to measure during operation, such as the water film depth on a wet driving road. The calculation of the calculation model is performed in the control device 13 of the motor vehicle. These calculations are preferably performed in two operating modes of the calculation model.

[0028] In a first operating mode, the calculation model 12 operates as a basic calculation model in which operating variables are calculated by simplified calculations and thus with limited precision. This calculation is performed during the driving operation based on the detected operating variables. The calculation model 12 is preferably part of the control device 13 of a motor vehicle, for example in the form of a program stored in the memory of the control device 13 and processed by the processor of the control device 13. This similarly applies to further functional blocks of the control device, which represent both the device aspect and the method aspect in this regard.

[0029] In a second operating mode activated when the motor vehicle is stopped, the calculation model is trained using the data set stored in the aforementioned first operating mode.

[0030] In the data acquisition phase, the operating characteristic parameters are measured by means of a reference sensor 14 in parallel with the calculation of the operating characteristic parameters. In the example of a water film, this may be an expensive laser sensor. In some cases, the reference sensor is only used in the data acquisition phase and does not exist when mass-producing motor vehicles later. The data acquisition phase is a phase in which data is collected using one or relatively few test vehicles (such as pre-production vehicles). The collected data is used to train the calculation models, which, in the trained state, will later be used for a larger number of vehicles, for example after the start of mass production of the vehicles.

[0031] In the difference formation block 16 of the control device 13, the difference between the operating characteristic parameters calculated in the basic calculation model 12 and the operating characteristic parameters measured in parallel therewith by the reference sensor 14 is calculated. In a subsequent functional block 18 of the control device 13, the difference is squared to produce a positive value. Instead of squaring, the absolute value of the difference can also be formed.

[0032] In the present application, the difference or its square or absolute value is regarded as an error estimate characterizing the error of the measurement signal sequence, and the operating characteristic parameters are calculated from the measurement signal sequence using the basic calculation model 12.

[0033] In the selection and recording block 20, the measurement signal sequences detected by the sensors 10 of the motor vehicle and / or the data sets contained in these measurement signal sequences are selected. For this purpose, the measurement signal sequences detected by the sensors 12 of the motor vehicle, the slot error estimate, and, if necessary, the corresponding associated values of the reference sensor 14 are fed to the recording block 20. This selection is performed in the control device 13 of the motor vehicle, so that the recording block 20 also represents both the method aspect and the device aspect.

[0034] The selected data is stored as new training data in the training data memory 22 and is used for training a trainable computational model for operating characteristic parameters (such as water film depth) separately from the acquisition phase time. The training data memory 22 can be the memory of a control device of a motor vehicle. As an alternative thereto, the memory 22 can also be a distributed memory (such as a cloud), which is located outside the motor vehicle and in which the training data selected according to the invention can be stored, said training data originating from a plurality of motor vehicles. This data exchange can take place, for example, via a mobile radio connection.

[0035] The trainable computational model can be the basic computational model already mentioned, and the basic computational model is then run in the training phase.

[0036] Figure 2 An embodiment of a selected measurement signal sequence with the features of the invention is illustrated, and this selection can be made in Figure 1 the recording block 20. In terms of its input signals and its output signals, the recording block 20 is located between the recording interface 21 and the data storage interface 23. Via the recording interface 21, the measurement signal sequence of the sensor 10, the error estimate of the block 18, and the signal of the reference sensor 14 are fed to the recording block. The recording interface 21 represents a data recorder for these data. The data storage interface 23 represents, for example, an interface to the training data memory, which is located in the motor vehicle, in particular in the control device 13 of the motor vehicle that executes the computational model. Alternatively or additionally, the data storage interface 23 is a connection to a server-side database in the cloud.

[0037] In Figure 2 the example shown, the measurement signal sequences from two sensors 10.1, 10.2 are fed to the selection and recording block 20 via the path 24, the error estimate is fed to the selection and recording block 20 via the path 26, and the measured value of the reference sensor 14 is fed to the selection and recording block 20 via the path 28.

[0038] The first sensor 10.1 has a value range 10.3, for example, from 0 to 100, and the second sensor 10.2 has a value range 10.4, for example, with three discrete values A, B, C.

[0039] The value range 10.3 of the first sensor 10.1 is divided into an operating variable range 10.3.1 corresponding to the interval [0...50] and an operating variable range 10.3.2 corresponding to the interval [50...100].

[0040] The value range 10.4 of the second sensor 10.2 is divided into operating variable ranges A, B, and C. Thus, in the example shown, six combination possibilities result. For each possible combination of the operating variable ranges 10.3.1, 10.3.2 of the first sensor 10.1 and the operating variable ranges A, B, C of the second sensor 10.2, there are n slots 30 available in the memory. The number n of slots is, for example, a number between 10 and 100.

[0041] In the preconfigured slots, each slot stores a measurement signal sequence of the sensors 10.1, 10.2 together with an associated error estimate value, and optionally also together with an associated reference value.

[0042] The mechanism using the slots is mainly used to collect as heterogeneous and diverse data as possible (i.e., not only data from a single narrow speed range, for example).

[0043] If now, during the data acquisition phase for the combination classes, the free slots are exhausted due to the progressive occupation, it is decided based on the error estimate value of the measurement error sequence stored for this combination which of the occupied slots can be overwritten. A small error estimate value means that the computational model has been able to process the stored measurement signal sequence well. Therefore, these measurement signal sequences are less interesting and can be overwritten by new measurement signal sequences with a larger slot error estimate value. Conversely, sequences with a larger assigned slot error estimate value should be retained (i.e., not overwritten), so that this information can be used later to retrain the computational model.

[0044] By directly selecting during the recording the data (measurement signal sequence, slot error estimate value, reference value sensor value) to be stored and overwritten, the laborious analysis of possibly very large amounts of data is saved. In addition, the storage of unnecessarily many redundant measurement signal sequences that do not provide additional information content is avoided, and thus less storage space is required.

[0045] Thus, the amount of data required for machine learning and the analysis effort can be drastically reduced, and computing time and memory capacity can be saved. When transferring the selected data between the external memory and the motor vehicle control device, the amount of data that can be transferred via a mobile radio connection, if necessary, is drastically reduced compared to the amount of data to be transferred without such a selection.

[0046] In all embodiments, the training is not performed in parallel with the data detection. During the data detection, in a first operating mode, the computational model is used to evaluate whether the detected data already largely corresponds to the data detected by the reference sensor, and in the corresponding case, the detected data is stored. Then, after the data detection session (e.g., test drive), the detected data is used to retrain the computational model.

[0047] The calculation model 16 can also be triggered by the control device of the motor vehicle to be executed on an external server. This has the following advantages: Many vehicles with the same structure can simultaneously collect data for the same calculation model. In this case, storage and overwriting are performed on the external server. The vehicle only provides the external server with the real-time data of the operating variables and the real-time data of the measured values of the reference sensors.

Claims

1. A method for automatically selecting a data set for a method for machine learning to detect operating variables of a motor vehicle, in which a sequence of measurement signals for a specific operating variable range of the motor vehicle is detected during operation of the motor vehicle, characterized in that, Assign a storage area to the operating variable range, each storage area having a plurality of slots (30), each slot being arranged to store a data set containing a detected sequence of measurement signals, and wherein the data sets already stored in the slots (30) can be respectively overwritten by newly detected data sets for the same storage area, and for each slot (30) of the storage area in which a data set is stored, form a slot error estimate value of the measurement signal sequence contained in the data set and store it together with the measurement signal sequence, and when overwriting, preferentially overwrite the following data sets, the slot error estimate values of which are less than the other slot error estimate values formed for the same storage area.

2. The method according to claim 1, wherein Calculate the operating characteristic parameters of the motor vehicle from the currently detected sequence of measurement signals using a calculation model (12), and in parallel measure the operating characteristic parameters with a reference value sensor (14), and form the slot error estimate value according to the deviation between the calculated operating characteristic parameters and the measured operating characteristic parameters.

3. The method according to claim 2, wherein The calculation model (12) is a calculation model (12) to be trained by a method for machine learning.

4. The method according to claim 3, wherein Only those data sets for which a slot error estimate value has been calculated during machine learning are released for overwriting.

5. The method according to claim 4, wherein During the driving cycle of the motor vehicle, only a predetermined portion of the data sets for which a slot error estimate value has been calculated during machine learning are released for overwriting.

6. The method according to claim 5, wherein When all the data sets released for overwriting have been overwritten before the machine learning process can start, select the next data set to be overwritten again according to the random principle.

7. The method according to claim 5, wherein Transmit the data set to a memory (22) located outside the motor vehicle.

8. The method according to claim 7, wherein The data sets stored in the motor vehicle are sorted in descending order according to their slot error estimate values, and the transmission is carried out in this sorted order and starting from the data set with the largest slot error estimate value.

9. A computer program product, characterized in that, The computer program product has a computer program including computer-readable instructions, which, when executed on a computer, run the method according to any one of claims 1 to 8.

10. A machine-readable storage medium, on which a computer program is stored, the computer program including computer-readable instructions, which, when executed on a computer, run the method according to any one of claims 1 to 8.

11. A control device (13) for a motor vehicle, the control device being configured to automatically select a data set for a method for machine learning to detect operating variables of the motor vehicle, in which method a sequence of measurement signals for a specific operating variable range of the motor vehicle is detected during operation of the motor vehicle, characterized in that, Assign a storage area of the control device (13) to the operating variable range, each storage area having a plurality of slots (30), each slot being arranged to store a data set containing a detected sequence of measurement signals, and wherein the data sets already stored in the slots (30) can be respectively overwritten by newly detected data sets for the same storage area, and the control device (13) is arranged to, for each slot (30) of the storage area in which a data set is stored, form a slot error estimate value of the stored measurement signal sequence, and when overwriting, preferentially overwrite the following measurement signal sequences, the slot error estimate values of which are less than the other slot error estimate values formed for the same storage area.

12. The control device (13) according to claim 11, characterized in that, The control device is arranged to control the operation of the method according to any one of claims 2 to 8.

Images