Method and device for machine learning for predicting the aging of a component of a vehicle

Abstract

The invention relates to a method for machine learning for predicting the aging of a component of a vehicle, in which method: a vehicle receives (10) model parameters for an aging model that is locally trained by a plurality of vehicles, and receives (11) a specification of vehicle data suitable for predicting the aging of the component; time series of the specified vehicle data of the vehicle are ascertained (12); at least one statistical parameter for the ascertained time series of the specified vehicle data is calculated (13); the aging model (NN) is trained (14) on the basis of the at least one statistical parameter, wherein a change in the model parameters is calculated; and the changed model parameters (MP1A, MP1N, MP2A, MP2N) or information relating to the change in the model parameters is transmitted (15) to a central server (ZS).

Description

[0001] Description

[0002] Method and apparatus for machine learning for the aging prediction of a component of a vehicle

[0003] The present invention relates to a machine learning method for predicting the aging of a vehicle component. Furthermore, the invention relates to a corresponding device that can be used in a vehicle to implement the method.

[0004] In machine learning (ML), a statistical model is built using suitable self-adaptive algorithms and training data. This model can be used to recognize patterns and regularities and to make predictions for future data or decisions based on the collected data. Due to the complexity of the algorithms used and the often very large data sets, the training process for machine learning can be very computationally intensive.

[0005] Since the provision of the necessary computing power is now straightforward, the importance of machine learning is constantly increasing in many technical fields. This also applies to its use in vehicles, for example, in driver assistance systems for partially automated driving or safety systems for fully automated driving. A variety of ML applications are known for driver assistance systems for partially automated driving or safety systems for fully automated driving. For example, vehicle sensors are used to record the vehicle's surroundings and, based on the acquired sensor data, an environment model is created. This allows learned objects in the environment to be recognized, which are then taken into account for safe vehicle control depending on the detected objects.The recorded sensor data can be evaluated in control units in the vehicle itself, by a backend server, or, for example, in a cloud environment with the provision of sufficient hardware resources by a service provider.

[0006] If a large number of vehicles are involved in the initial and / or ongoing training of such vehicle-specific ML models, this can pose a problem if sensitive data is processed. For example, this data may contain personal data that cannot be shared without permission. Likewise, in cross-manufacturer or cross-company ML applications, sharing data among competitors may be undesirable, for example, because it contains competitively relevant details.

[0007] Such a scenario occurs, for example, when machine learning is used to make a manufacturer- or company-wide prediction about the wear and tear or aging of a vehicle component.

[0008] It is an object of the invention to provide a machine learning method for predicting the aging of a component of a vehicle and a corresponding device.

[0009] This object is achieved by the independent claims. Preferred embodiments of the invention are the subject of the dependent claims.

[0010] The inventive method for machine learning for the aging prediction of a component of a vehicle comprises the following steps performed by a vehicle:

[0011] - Receiving model parameters for an aging model that is decentralized trained by multiple vehicles;

[0012] - Receiving a specification of vehicle data suitable for predicting component aging;

[0013] - Collecting time series of the vehicle's specified vehicle data;

[0014] - Calculating at least one statistical parameter for the recorded time series of the specified vehicle data,

[0015] - training the aging model based on the at least one statistical parameter, whereby a change in the model parameters is calculated;

[0016] - Sending the changed model parameters or information about the change in the model parameters to a central server.

[0017] In this way, when training the aging model, local changes in the model parameters are calculated based on time series of the specified vehicle data recorded locally in the vehicle. These changes can be aggregated with the changes in the model parameters determined by other vehicles by a central server in order to determine a global aging model or to update this global aging model. The global aging model can thus be trained decentrally by a large number of vehicles, with the vehicle data used for this purpose being available exclusively locally in the respective vehicle and not leaving the vehicle. This prevents this data from being shared when predicting the aging of vehicle components, where the aging model used uses sensitive internal vehicle signals, thus ensuring data security.

[0018] In particular, after the central server has aggregated the changed model parameters, global model parameters generated by the central server through the aggregation of the changed model parameters can be received in the respective vehicle. This then enables a more accurate prediction of the aging of the components installed in the vehicle and captured by the aging model.

[0019] According to one embodiment of the invention, the model parameters and the vehicle data for the aging model are specified by several participating parties, wherein the several participating parties each train the aging model independently of one another for vehicles or vehicle components assigned to them and in doing so calculate a change in the model parameters for the vehicles or vehicle components assigned to them, the changed model parameters or information about the change in the model parameters are sent by the several participating parties to the central server and the central server aggregates the changes in the model parameters of the individual parties to form the global model parameters.

[0020] This allows the parties involved to access the recorded vehicle data in detail for the vehicles or vehicle components assigned to them, while also ensuring that the global ML model can contain knowledge derived from the sensitive data of all parties involved without the parties having to explicitly share their sensitive data.

[0021] Advantageously, the aging model is initially trained by one of the parties involved, with the resulting initial model parameters being sent to the other parties involved and the subsequent training of the aging model being carried out by all of the parties involved.

[0022] According to a further embodiment of the invention, changed model parameters of a vehicle are determined which differ significantly from the changed model parameters of the other vehicles and are given a higher weighting when aggregating the changed model parameters.

[0023] In this way, the aging model can also adequately account for aging behavior that only a few vehicles exhibit. For example, the aging model can incorporate premature aging caused by extreme driving behavior or inferior material in the affected component. Accelerated aging can also be considered, such as in the case of batteries, massive aging within a short period of time or due to a few load cycles, also known as "sudden death."

[0024] According to a further embodiment of the invention, the changed model parameters or information about the change in the model parameters is sent to the central server at predefined time intervals. Thus, the parameter updates for the global aging model can be sent regularly at longer intervals, for example, every one or more weeks or even months.

[0025] Furthermore, the changed model parameters or the information about the change in the model parameters are preferably sent with priority if they differ significantly from the changed model parameters of the other vehicles.

[0026] In particular, the vehicle data used for training the aging model can be generated by a control unit in the vehicle and be available as time series of CAN messages.

[0027] According to one embodiment of the invention, the aging model predicts the aging of a battery of the vehicle, wherein a statistical parameter is determined for at least one battery parameter over a defined number of charging and discharging cycles.

[0028] According to a further embodiment of the invention, the aging model predicts the aging of an exhaust gas recirculation cooler of the vehicle, wherein a statistical parameter is determined for at least one parameter of the exhaust gas recirculation cooler over a defined period of time in which the exhaust gas recirculation cooler is operated.

[0029] A device according to the invention for machine learning for the aging prediction of a component of a vehicle comprises: - a communication unit for receiving model parameters for an aging model that is decentrally trained by several vehicles, for receiving a specification of vehicle data that is suitable for predicting the aging of the component and for sending changed model parameters or information about the change in the model parameters to a central server;

[0030] - a recording unit for recording time series of the specified vehicle data of the vehicle;

[0031] - a computing unit for calculating at least one statistical parameter for the recorded time series of the specified vehicle data and training the aging model based on the at least one statistical parameter, wherein a change in the model parameters is calculated.

[0032] The invention also includes a computer program with instructions that cause a device in a vehicle to carry out the steps of one of the methods according to the invention.

[0033] Finally, the invention also includes a motor vehicle which is configured to carry out a method according to the invention or has a device according to the invention.

[0034] Further features of the present invention will become apparent from the following description and claims in conjunction with the figures.

[0035] Fig. 1 shows a schematic flow diagram for a method according to the invention for machine learning for the aging prediction of a component of a motor vehicle;

[0036] Fig. 2 shows a schematic example of predicting the aging of a battery using an aging model based on a neural network, with a central server that combines distributed aging models from several parties, each of which has been locally trained using several vehicles; and

[0037] Fig. 3 shows a schematic block diagram of a vehicle with a prediction unit for predicting battery aging using the method according to the invention. For a better understanding of the principles of the present invention, embodiments of the invention are explained in more detail below with reference to the figures. It is understood that the invention is not limited to these embodiments and that the described features can also be combined or modified without departing from the scope of the invention as defined in the claims.

[0038] A flowchart of a method according to the invention, which is executed in one of the vehicles involved in distributed machine learning, is shown in Figure 1. Different machine learning methods are considered, such as training deep neural networks (deep learning) or creating a random forest model.

[0039] This is preceded by an agreement between the parties involved on the architecture of the jointly used aging model. For a deep neural network, for example, the number of intermediate layers (hidden layers) between the input and output layers, the number of neurons in each layer, the activation function (threshold function), etc., are determined.

[0040] In a method step 10, model parameters for the aging model are then received from a central server in one of the vehicles participating in the distributed machine learning, for example, via a mobile network connection. The central server coordinates the participating vehicles, aggregates the model parameter updates by the participating vehicles, and communicates these model parameter updates back to the participating vehicles.

[0041] The received model parameters can be untrained or randomly initialized. However, model parameters resulting from a previous training of an initial ML model at a party can also be received. The data from an initial training can, for example, come from prototypes, endurance runners, or laboratory / test bench tests of the components or systems.

[0042] vehicles come from.

[0043] Likewise, in method step 11, a specification of vehicle data suitable for predicting the aging of the component and to be recorded by the vehicle for the aging model is received from the central server in the vehicle. In the case of a battery aging model, for example, this can be one or more operating parameters of the battery, such as the battery current, the battery voltage, the battery temperature, or the battery state of charge (SoC). Likewise, predefined parameters for determining the statistical parameters, such as in particular a predefined period of time within which a load is applied and the statistical parameters are determined, can be received. For example, a number of charging cycles after which the statistical parameters are to be determined can be defined, for example 25 or 50 average charging and discharging cycles of the battery.Likewise, the predefined parameters for subdividing the aggregation may also include a specified mileage, such as 1,000, 2,000 km or more, or a specified operating period, such as one month of ferry operation.

[0044] Similarly, for example, in an aging model for the aging of a vehicle's exhaust gas recirculation cooler, statistical parameters of the exhaust gas recirculation cooler can be determined. For example, the EGR valve position or the EGR mass flow can be recorded over a defined period of time during which the exhaust gas recirculation cooler is in operation. For example, corresponding statistical characteristics can be determined for each nine hours of use of the exhaust gas recirculation cooler by clustering the measured values.

[0045] Likewise, the statistical parameter can be arbitrarily fine, so that it approximates an unchanging time series of measured data. In this case, the statistical parameter is itself a time series and can serve as input, for example, to a Long Short-Term Memory (LSTM) model.

[0046] In the subsequent method step 12, the time series of vehicle data required for training the aging model are then collected from the vehicle. This can, in particular, be vehicle data generated by a control unit in the vehicle and present as a time series of CAN messages. Both univariate time series, in which a series of chronologically ordered measurement data points are available, for each of which the measured variable is recorded with associated time information, and multivariate time series of several different measurement signals can be recorded. The analysis of such multivariate time series makes it possible to identify relationship or dependency structures between the various measurement signals and thus to understand how changes to one variable can affect the other variables.The individual measurement data points of the time series can initially be temporarily stored locally in the vehicle until they are used for training, for example, in a central data storage. However, the local measurement data is not transmitted to the central server or other vehicles.

[0047] In the subsequent method step 13, at least one statistical parameter is determined for the recorded time series of the specified vehicle data. A set of sample data can be selected from the time series data contained in the recorded vehicle data set according to a predefined time interval, for example, using a sliding window. This can also be used, for example, to determine frequency distributions of certain characteristics and classify measurement data.

[0048] It can also be provided that when the time series data is temporarily stored, only the relevant measurement data is saved, thus saving storage space in the vehicle. Furthermore, the temporary storage can be carried out in this way by combining method steps 12 and 13 if not the time series data itself, but only an aggregation of this data, such as histograms or other statistical parameters, is used for training. In this case, method steps 12 and 13 can, in particular, be carried out alternately, so that the time series data recorded in method step 12 does not have to be permanently stored, but only the at least one statistical parameter calculated in method step 13. This can then be continued until method step 14 is triggered by a predefined parameter, such as a specified mileage (km performance), number of charging and discharging cycles, or similar.In this case, a loop of process steps 12 and 13 is executed until process step 14 is executed.

[0049] In the subsequent method step 14, the aging model is trained based on the at least one statistical parameter, with locally updated model parameters being calculated based on the recorded vehicle data. For example, if a deep neural network is used for a battery aging model, the respective weightings of the individual neurons are adjusted.

[0050] Likewise, a so-called random forest model can be created, for example, in an aging model for the aging of an exhaust gas recirculation cooler. This model is based on an algorithm that is well suited for classification and regression tasks and is based on combining the results of many different decision trees. A single decision tree consists of several branches that are created by assigning data to a class based on its properties using rules. Starting from an initial decision, new branches are created until a predefined result level is reached. The random forest algorithm specifies rules for how the decision trees are to be generated and combined to achieve an overall result.

[0051] It can also be intended to identify model parameters of a vehicle or changes to these model parameters that differ significantly from the model parameters or changes to the model parameters of the other vehicles. Since, especially for safety-relevant vehicle components, it is not sufficient for the aging model to only make a reliable prediction for the majority of the affected vehicles, the global aging model benefits from considering such aging information, i.e., the aging behavior exhibited by only a few vehicles.

[0052] Such outliers can be identified on the vehicles themselves by identifying values ​​of large changes in the model parameters. Identification can also be performed subsequently on the central server, which can cluster the model parameter changes of all vehicles. This can result in one or more larger groups of vehicles with average or typical aging behavior, as well as smaller groups and / or individual outliers with extreme aging behavior. For example, the aging model can specifically take into account premature aging caused by extreme driving behavior by the driver, for example, by assigning a higher weighting during aggregation. This can also be achieved by using only individual representatives of the groups for the larger groups.Furthermore, other types of aging can also be identified in this way, such as aging that only occurs in small groups.

[0053] The changed model parameters or information about the change in the model parameters is then sent to the central server in process step 15. These parameter updates are preferably sent at predefined time intervals. This means that the parameter updates for the global aging model can be sent regularly at longer intervals, for example, every week or several weeks or even months, since the aging of vehicle components occurs over longer periods of time and therefore, unlike, for example, environmental information about a new traffic sign or a new construction site, does not have an urgent temporal relevance.

[0054] As with the predefined parameters for data acquisition or subdivision of the aggregation, the transmission of the changed model parameters to the central server can also occur at specified mileages or a specified operating time, rather than at predefined time intervals. Transmission can also be triggered by certain events during operation, such as exceeding or falling below predefined values ​​for the battery's operating parameters, such as battery current, battery voltage, battery temperature, or the ambient temperature. Furthermore, a combination of the aforementioned transmission criteria can be specified, such as transmission when a specified mileage is reached or a predefined event is recorded.Finally, the criterion for transmission can also be to carry out the transmission as soon as a certain number of samples, for example 50, 100 or 1000 samples, are available.

[0055] The exact time at which the parameter updates are transmitted is not relevant for further processing, so that transmission can be postponed, for example, if the vehicle is in a tunnel, abroad, or outside of a home / fleet hub Wi-Fi network when the time interval expires. Likewise, the transmission of locally updated model parameters can first be checked in the vehicle to determine when this can occur without disrupting other, more prioritized data communication within the vehicle. Furthermore, the transmission of parameter updates can be prioritized if outlier behavior is detected. Furthermore, additional anonymization can be performed before transmission if necessary to further protect the privacy of the owners of the vehicles participating in the training.

[0056] The received parameter updates are then aggregated on the server side to create globally updated model parameters. In the simplest case, this is done by calculating an average of all local updates. However, the received parameter updates can also be weighted to enable the special consideration of identified outliers by assigning a higher weighting, as described above.

[0057] In method step 16, after the central server has aggregated the changed model parameters, global model parameters generated by the central server through the aggregation of the changed model parameters can be received in the respective vehicle. This then enables a more accurate prediction of the aging of the components installed in the vehicle and captured by the aging model.

[0058] The described aging prediction method can be continued indefinitely or it can be continued in an iterative process until a termination condition is reached. The iteration can be terminated when a predefined number of iterations has been completed. Likewise, the method can be terminated when fewer than a predefined number of vehicles are participating in the learning process.

[0059] Furthermore, when training a neural network, it is also possible to adapt and terminate the process depending on the convergence of suitable metrics or loss functions, which can be used to determine how the performance of the aging model is changing. This way, the repetition rate for updating the global model—that is, how often parameter updates are sent to the central backend, updates to the global model are determined, and then made available to the vehicles again—can be reduced if only small changes to the model parameters occur, meaning the model learns less and less. However, the process is not aborted based on a convergence metric, because even if a local / vehicle-specific model cannot currently contribute anything to the global model, this does not mean that this will continue to be the case in the future.

[0060] The method according to the invention implemented in the vehicles can, for example, be implemented as a computer program on a control unit. For this purpose, the computer program is transferred to a memory of the respective control unit during production and stored there. The computer program comprises instructions that, when executed by a processor of the control unit, cause the control unit to perform the steps according to the method according to the invention.

[0061] Figure 2 shows a schematic example of predicting the aging of a battery using an aging model based on a neural network.

[0062] In the example shown, key data on the current operating status of the installed battery BAT is determined in several vehicles F1A to F1N of a first participating party P1 and several vehicles F2A to F2N of a second participating party P2. Based on this data, the battery aging model is trained locally in each of the vehicles.

[0063] For example, the two parties may be two vehicle manufacturers who install the same battery in their vehicles and therefore have a common interest in an aging model for this battery that is as accurate as possible, but do not want to share information with each other about how their vehicles are used. In the example shown, the method according to the invention is described for only two parties involved, but more than two parties may also be involved. Other parties, such as suppliers of vehicle components or operators of larger vehicle fleets, may also be involved in the method according to the invention. As described above, laboratory or test bench data in particular can also be used here. Furthermore, multiple data from different development stages or generations of a component can also be taken into account.

[0064] The aging model is trained on each vehicle based on the locally available vehicle data. The local model parameters MP1A to MP1N and MP2A to MP2N resulting from the training, or changes to the model parameters, result in corresponding changes to the model parameters MP1 and MP2 for the participating parties, each of which can operate its own servers for the vehicles assigned to them. Each party sends these changes to the model parameters to a central server (ZS), where the changes to the model parameters of the individual parties are summarized, for example, as an average of the individual changes to the model parameters.The central server (ZS) can be operated as a backend server, for example, by a dedicated service provider that offers a distributed machine learning service to the participating parties for a usage fee, using the parameter updates learned from these parties' vehicles. Furthermore, based on the specifications of the participating parties, the central server can also limit the aging prediction to specific vehicle components, such as batteries from a specific battery manufacturer, or specific vehicle models.

[0065] Figure 3 shows a schematic block diagram of a vehicle F with various units that can be used to predict the aging of a battery by means of the method according to the invention. In the example shown, this can in particular be an electrically powered vehicle that comprises one or more electric motors (not shown in the example) that are powered by a drive battery BAT. The drive battery can be constructed from one or more battery modules or according to the cell-to-pack (CTP) approach and can in particular contain lithium-ion accumulators. The drive battery system can additionally comprise further components, such as a battery management system, a battery heating / cooling system and charging electronics.

[0066] The vehicle has a communication unit KE with which model parameters for the aging model and a specification of vehicle data suitable for predicting battery aging can be received and locally updated model parameters can be sent. Furthermore, the vehicle has an acquisition unit EE for acquiring time series of the specified vehicle data, i.e. in the example shown, the vehicle's battery data used for the battery aging model, as well as a computing unit RE for training the aging model NN. The computing unit RE and the acquisition unit EE are shown separately in the example, but can also be integrated into one unit. In this case, the computing unit RE and, if applicable, the acquisition unit EE can be part of a control unit, for example a central control unit or vehicle server.To carry out the method according to the invention, the control unit can comprise one or more processors, each having one or more processor units, for example microprocessors, digital signal processors or combinations thereof.

[0067] The acquired vehicle data can be stored in a memory unit SE until processed by the computing unit. Furthermore, a computer program for executing the method according to the invention can also be stored in the memory unit SE. The memory unit SE can have both volatile and non-volatile memory areas and can be configured, for example, as a semiconductor memory. The memory modules can be configured, for example, as random access memory (RAM), dynamic random access memory (DRAM), EPROM, or flash memory.

[0068] Furthermore, the drive battery BAT, the processing unit RE, the acquisition unit EE, the communication unit KE and the storage unit SE are connected to a digital data bus B, for example a CAN, MOST, FlexRay or Automotive Ethernet bus.

[0069] The method according to the invention can be carried out in any vehicle, such as passenger cars, but is not limited thereto.

[0070] 10 - 16 process steps

[0071] ZS central server

[0072] F1A, F1 N , F2 A , F2 N Vehicles involved in distributed learning

[0073] P1 , P2 server involved party

[0074] MP1A, MP1 N, MP2A, MP2N locally changed model parameters

[0075] MP1 , MP2 changed model parameters of a party involved

[0076] MP updated global model parameters

[0077] NN machine learning model

[0078] NNz's global machine learning model

[0079] F vehicle

[0080] KE communication unit

[0081] SE storage unit

[0082] RE computing unit

[0083] BAT battery unit

[0084] EE registration unit

[0085] B digital data bus

Claims

Patent claims 1. A machine learning method for predicting the aging of a component of a vehicle, wherein a vehicle performs the following steps: - receiving (10) model parameters for an aging model that is decentrally trained by several vehicles; - receiving (11) a specification of vehicle data suitable for predicting the aging of the component; - capturing (12) time series of the specified vehicle data of the vehicle; - calculating (13) at least one statistical parameter for the recorded time series of the specified vehicle data, - training (14) the aging model (NN) based on the at least one statistical parameter, wherein a change in the model parameters is calculated; - Sending (15) the changed model parameters (MP1A, MP1N, MP2A, MP2N) or information about the change in the model parameters to a central server (ZS).

2. The method according to claim 1, wherein global model parameters (MP) are received (16) which have been generated by the central server (ZS) by means of an aggregation of the changed model parameters.

3. The method according to claim 1 or 2, wherein - the model parameters and the vehicle data for the aging model are specified by several parties involved (P1, P2); - the several parties involved, each independently of each other, for vehicles assigned to them (F1 Ä , F1 N , F2 Ä , F2 N ) or vehicle components train the aging model (NN) and calculate a change in the model parameters for the vehicles or vehicle components assigned to them; - the changed model parameters or information about the change in the model parameters are sent by the multiple parties involved to the central server (CS); and - the central server aggregates the changes in the model parameters of the individual parties into the global model parameters (MP).

4. The method according to claim 3, wherein the aging model is initially trained by one of the parties involved and the resulting initial model parameters are sent to the other parties involved and the subsequent training (14) of the aging model (NN) is carried out by all of the parties involved.

5. Method according to one of the preceding claims, wherein changed model parameters of a vehicle are determined which differ significantly from the changed model parameters of the other vehicles and are given a higher weighting in the aggregation of the changed model parameters.

6. Method according to one of the preceding claims, wherein the changed model parameters or the information about the change in the model parameters are sent to the central server at predefined time intervals.

7. The method according to claim 6, wherein the changed model parameters or the information about the change in the model parameters are sent with priority if they differ significantly from the changed model parameters of the other vehicles.

8. Method according to one of the preceding claims, wherein the vehicle data are generated by a control unit in the vehicle and are present as time series of CAN messages.

9. Method according to one of the preceding claims, wherein the aging model predicts the aging of a battery (BAT) of the vehicle and a statistical parameter is determined for at least one battery parameter over a defined number of charging and discharging cycles.

10. The method according to any one of claims 1 to 9, wherein the aging model predicts the aging of an exhaust gas recirculation cooler of the vehicle and a statistical parameter is determined for at least one parameter of the exhaust gas recirculation cooler over a defined period of time in which the exhaust gas recirculation cooler is operated.

11. Device for machine learning for the aging prediction of a component of a vehicle, comprising - a communication unit (KE) for receiving model parameters for an aging model that is decentrally trained by several vehicles, for receiving a specification of vehicle data that is used to predict the aging of the component and for sending changed model parameters or information about the change in the model parameters to a central server; - a recording unit (EE) for recording time series of the specified vehicle data of the vehicle; - a computing unit (RE) for calculating at least one statistical parameter for the recorded time series of the specified vehicle data and training the aging model based on the at least one statistical parameter, wherein a change in the model parameters is calculated.

Images