Method and apparatus for machine learning for aging prediction of components of a vehicle

Abstract

In a method for machine learning for aging prediction of a component of a vehicle according to the invention, model parameters for an aging model trained dispersedly by a plurality of vehicles and specifications of vehicle data suitable for inferring aging of the component are received (10, 11) by the vehicle. A time sequence of normalized vehicle data of the vehicle is acquired (12), and at least one statistical characteristic variable for the acquired time sequence of normalized vehicle data is calculated (13). Training (14) the aging model (NN) on the basis of the at least one statistical characteristic variable, wherein a change in a model parameter is calculated. The changed model parameter (MP1A, MP1N, MP2A, MP2N) or information about the change in the model parameter is sent (15) to a central server (ZS).

Description

Technical Field

[0001] The present invention relates to a method for machine learning for ageing prediction of components of a vehicle and to a corresponding device that can be used in a vehicle to carry out the method. Background Art

[0002] In machine learning (ML), statistical models are constructed based on training data using suitable adaptive algorithms. These models can identify patterns and regularities and make predictions about future data or decisions based on the collected data. Due to the complexity of the algorithms used and the often very large amounts of data, the training process for machine learning can be very computationally intensive.

[0003] Since the availability of the necessary computing power is now readily available, machine learning is gaining increasing importance in many areas of technology. 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. Consequently, numerous 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 capture the vehicle's surroundings, and a model of the surroundings is created based on the captured sensor data. This model identifies learned objects in the surroundings, which are then taken into account for safe vehicle control based on the captured objects. The captured sensor data can be evaluated by a controller in the vehicle itself, by a backend server, or, for example, in a cloud environment with sufficient hardware resources provided by a service provider.

[0004] If a large number of vehicles are involved in the initial and / or ongoing training of these vehicle-related ML models, processing sensitive data for this purpose can be problematic. This data could, for example, include personal data that should not be easily shared. Similarly, in cross-manufacturer or cross-company ML applications, sharing data between competitors may be undesirable, for example because the data contains competitively relevant details.

[0005] Such a scenario arises, for example, when a cross-manufacturer or cross-company prediction regarding the wear or aging of vehicle components is to be made with the aid of machine learning. Summary of the Invention

[0006] The object of the present invention is to provide a method for machine learning for ageing prediction of components of a vehicle and a corresponding device.

[0007] This object is achieved by the independent claim. Preferred embodiments of the invention are the subject matter of the dependent claims.

[0008] The method for machine learning for aging prediction of components of a vehicle according to the present invention comprises the following steps performed by the vehicle:

[0009] - receiving model parameters for an aging model, the aging model being decentralized trained for a plurality of vehicles;

[0010] - receiving a specification of vehicle data suitable for inferring degradation of the component;

[0011] - collecting a time series of normalized transport data of the transport;

[0012] - calculating at least one statistical characteristic variable for the time series of the acquired normalized vehicle data;

[0013] - training the aging model based on the at least one statistical characteristic variable, wherein changes in model parameters are calculated;

[0014] - Sending the changed model parameters or information about the changes in model parameters to a central server.

[0015] In this way, when training the aging model based on a time series of normalized vehicle data collected locally in the vehicle, local changes in model parameters are calculated. These changes can be aggregated by a central server with changes in model parameters determined for other vehicles to thereby determine a global aging model or enable the updating of the global aging model. A global aging model can thus be trained decentralized across a large number of vehicles, with the vehicle data used for this purpose existing only locally in the respective vehicle and never leaving the vehicle. This prevents the sharing of sensitive internal vehicle signals when using aging models for predicting vehicle component aging, thereby ensuring data security.

[0016] In particular, after the central server aggregates the modified model parameters, the global model parameters generated by the central server using the aggregation of the modified model parameters can be received in the respective vehicles. This then enables a more accurate inference of the aging of the components installed in the vehicles and detected by the aging model in the respective vehicles.

[0017] According to one embodiment of the present invention, the model parameters and the vehicle data for the aging model are standardized by multiple participants, wherein the multiple participants each independently train the aging model for the vehicles or vehicle components assigned to them and thereby calculate changes in the model parameters for the vehicles or vehicle components assigned to them, the changed model parameters or information about the changes in the model parameters are sent by the multiple participants to the central server, and the central server aggregates the changes in the model parameters of the individual participants into global model parameters.

[0018] This allows the participants to have detailed access to the collected vehicle data for the vehicles or vehicle components assigned to them, while ensuring that the global ML model can contain knowledge derived from the sensitive data of all participants without them having to explicitly share their sensitive data.

[0019] Advantageously, the aging model is initially trained by one of the participants, wherein the initial model parameters generated thereby are transmitted to the other participants, and the subsequent training of the aging model is carried out by all participants.

[0020] According to a further embodiment of the invention, modified model parameters of the means of transport are determined, which differ significantly from the modified model parameters of the remaining means of transport and are given a higher weighting when aggregating these modified model parameters.

[0021] In this way, the aging model can also appropriately account for aging behaviors that only a small number of vehicles exhibit. Thus, for example, the aging model can include premature aging caused by extreme driving styles or poor-quality materials in affected components. Similarly, it can account for accelerated aging, such as in batteries, which experience extensive aging within a short period of time or a few load cycles, also known as "sudden death."

[0022] According to another embodiment of the present invention, the changed model parameters or information about the changes in the model parameters are sent to the central server at predefined time intervals. Thus, parameter updates for the global aging model can be sent regularly at relatively large time intervals, for example, once a week, several weeks, or even several months.

[0023] Furthermore, if the changed model parameters or the information on the change in the model parameters are significantly different from the changed model parameters of the remaining transportation means, the changed model parameters or the information on the change in the model parameters are preferentially transmitted.

[0024] In particular, the vehicle data used to train the aging model may be generated by a controller in the vehicle and exist as a time series of CAN messages.

[0025] According to one embodiment of the invention, the aging model infers the aging of the battery of the vehicle, wherein a statistical characteristic variable is determined for at least one battery characteristic variable over a defined number of charging and discharging cycles.

[0026] According to a further embodiment of the invention, the aging model infers aging of the exhaust gas recirculation cooler of the vehicle, wherein a statistical characteristic variable is determined for at least one characteristic variable of the exhaust gas recirculation cooler over a defined period of operation of the exhaust gas recirculation cooler.

[0027] The apparatus for machine learning for aging prediction of components of a transportation vehicle according to the present invention comprises:

[0028] a communication unit for receiving model parameters for an aging model trained decentralized by a plurality of vehicles, for receiving specifications of vehicle data suitable for inferring aging of the component, and for sending changed model parameters or information about changes in model parameters to a central server;

[0029] - a collection unit, configured to collect a time series of normalized transport means data of the transport means;

[0030] A calculation unit for calculating at least one statistical characteristic variable for the time series of the acquired normalized vehicle data and for training the aging model based on the at least one statistical characteristic variable, wherein changes in model parameters are calculated.

[0031] The invention also comprises a computer program with instructions which cause a device in a vehicle to carry out the steps of one of the methods according to the invention.

[0032] Finally, the present invention also comprises a motor vehicle which is configured to carry out the method according to the present invention or has the device according to the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Additional features of the present invention will become apparent from the following description and claims taken in conjunction with the accompanying drawings.

[0034] Figure 1 A flow chart schematically shows a method according to the invention for machine learning for aging prediction of components of a vehicle;

[0035] Figure 2 Schematically illustrates an example for inferring battery aging using a neural network-based aging model, with a central server aggregating distributed learned aging models from multiple parties, each of which was locally trained using multiple vehicles; and

[0036] Figure 3 A block diagram of a vehicle with a prediction unit for inferring battery aging using the method according to the present invention is schematically shown. DETAILED DESCRIPTION

[0037] In order to better understand the principle of the present invention, the embodiments of the present invention will be explained in more detail below with reference to the accompanying drawings. It should be understood that the present invention is not limited to these embodiments, and the features described can also be combined or modified without departing from the scope of protection of the present invention as defined in the claims.

[0038] exist Figure 1 , a flow chart of the method according to the invention is shown, which is implemented in one of the vehicles participating in distributed machine learning. Various machine learning methods are considered, such as the training of deep neural networks (English: Deep Learning) or the creation of random forest models.

[0039] Prior to this, the parties agreed on the architecture of the jointly used aging model. In the case of deep neural networks, this included, 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), and so on.

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

[0041] The received model parameters can be untrained or randomly initialized. However, it is also possible to receive model parameters generated by a party in previously training an initial ML model. The data from the initial training can come from, for example, prototypes of components or vehicles, endurance test vehicles, or laboratory / test bench tests.

[0042] Similarly, in method step 11, specifications for vehicle data suitable for inferring the aging of the component and to be collected by the vehicle for the aging model are received from the central server in the vehicle. For example, in a battery aging model, these could be one or more battery operating parameters, such as battery current, battery voltage, battery temperature, or battery state of charge (SoC). Similarly, predefined parameters for determining statistical characteristic variables can be received, such as, in particular, a predefined time period within which loading is performed and the statistical characteristic variables are determined. Thus, a number of charging cycles (e.g., 25 or 50 average charging and discharging cycles of a battery) can be defined after which the statistical characteristic variables are to be determined. Similarly, predefined parameters for segmenting the aggregates can also include a fixed number of kilometers (or mileage), such as 1,000, 2,000 kilometers, or more, or a fixed operating duration, such as one month of driving.

[0043] Similarly, in an aging model for the aging of an exhaust gas recirculation cooler in a vehicle, for example, statistical characteristic variables of the exhaust gas recirculation cooler can be determined. Thus, the EGR valve position or the EGR mass flow can be recorded over a defined duration of operation of the exhaust gas recirculation cooler. For example, for every nine hours of operation of the exhaust gas recirculation cooler, corresponding statistical characteristics can be determined by clustering the measured values.

[0044] Likewise, the statistical characteristic variable can be arbitrarily refined so that it approximates an unchanged time series from the measurement data. In this case, the statistical characteristic variable itself is a time series and can be used as input to, for example, a long short-term memory (LSTM) model.

[0045] In the subsequent method step 12, a time series of vehicle data required for training the aging model is then collected from the vehicle. This can be, in particular, vehicle data generated by a controller in the vehicle and present as a time series of CAN messages. Both univariate time series (a series of time-ordered measurement data points, for which a measured variable with associated time information is acquired) and multivariate time series (multiple different measurement signals) can be recorded. Analysis of such multivariate time series allows the identification of correlations or dependency structures between different measurement signals and, thus, an understanding of how changes in one variable can influence other variables.

[0046] For this purpose, the individual measurement data points of the time series can first be temporarily stored locally in the vehicle until they are used for training, for example in a central data memory. However, no transmission of the local measurement data to the central server or other vehicles is performed.

[0047] In the subsequent method step 13, at least one statistical characteristic variable is then determined for the time series of the acquired, normalized vehicle data. Here, a set of sample data (English: samples) can be selected from the time series data contained in the acquired vehicle data set according to predefined time intervals, for example using a sliding window. For example, the frequency distribution of specific characteristics can also be determined, and the measurement data can be divided into categories.

[0048] It can also be arranged that only relevant measurement data is stored during the temporary storage of the time series data, thereby saving storage space in the vehicle. Furthermore, if it is not the time series data itself, but only an aggregation of this data (e.g., a histogram or other statistical characteristic variables) that is used for training, temporary storage can be performed in this manner by combining method steps 12 and 13. In particular, method steps 12 and 13 can be performed alternately, so that the time series data acquired in method step 12 does not need to be permanently stored, but only the at least one statistical characteristic variable calculated in method step 13. This can continue until method step 14 is triggered by predefined parameters (e.g., a fixed number of kilometers, a number of charging and discharging cycles, etc.). In this case, a loop of method steps 12 and 13 is thus executed until method step 14 is then executed.

[0049] In the subsequent method step 14, the aging model is trained based on the at least one statistical characteristic variable, wherein locally updated model parameters are calculated based on the acquired vehicle data. For example, if a deep neural network is used for the battery aging model, the corresponding weights of the individual neurons are adapted.

[0050] Similarly, in aging models for exhaust gas recirculation coolers, for example, a so-called random forest model can be created. This 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 multiple branches, which are generated by assigning data to a category based on its characteristics using rules. New branches are generated starting from the first decision until a predefined level of results is reached. The random forest algorithm specifies the rules for generating and combining decision trees to obtain the overall result.

[0051] In this context, it can also be provided that model parameters of vehicles or changes in these model parameters are identified that differ significantly from the model parameters or changes in model parameters of the remaining vehicles. Since, in particular with safety-relevant vehicle components, it is not sufficient for the aging model to produce reliable predictions only for the majority of the vehicles involved, a global aging model benefits from taking into account aging information, i.e., aging behavior exhibited by only a few vehicles.

[0052] The identification of such outliers can be performed on the vehicles themselves by identifying values ​​with significant changes in model parameters. Similarly, this identification can also be performed subsequently in a central server, which can cluster the changes in model parameters for all vehicles for this purpose. In addition to one or more larger groups of vehicles with average or typical aging behavior, smaller groups with extreme aging behavior and / or individual outliers are generated. Thus, the aging model can, for example, specifically take into account premature aging caused by a driver's extreme driving style, for example by assigning higher weights during aggregation. This can also be achieved by using only individual representatives of a group in a larger group. Furthermore, other types of aging can also be identified in this way, such as aging that occurs only in a small group.

[0053] The changed model parameters or information about the changes in the model parameters are then sent to the central server in method step 15. These parameter updates are preferably sent at predefined time intervals. Thus, parameter updates for the global aging model can be sent regularly at relatively long time intervals, for example, at intervals of one or more weeks or even months, since the aging of vehicle components occurs over a longer period of time and is therefore less time-sensitive than, for example, information about the surrounding environment, such as new traffic signs or new construction sites.

[0054] As with the predefined parameters for data acceptance or segmentation aggregation, the transmission of modified model parameters to the central server can also take place after a fixed number of kilometers or a fixed operating duration, instead of after a predefined time interval. Transmission can also be triggered by specific events during operation, such as exceeding or falling below predefined values ​​for battery operating parameters (such as battery current, battery voltage, battery temperature, or ambient temperature). In addition, a combination of the aforementioned criteria for transmission can also be set, such as transmission upon reaching a fixed number of kilometers or upon the detection of a predefined event. Finally, as a criterion for transmission, it can also be set so that transmission is performed once a specific number of samples (e.g., 50, 100, or 1000 samples) are available.

[0055] The exact timing of parameter update transmission is unimportant for further processing; thus, transmission can be postponed, for example, if the vehicle is in a tunnel, abroad, or outside of home / fleet hub Wi-Fi at the end of the time interval. Similarly, for the transmission of locally updated model parameters, a check can be made within the vehicle to determine when the transmission can occur without impacting other priority data communications of the vehicle. Furthermore, if outlier behavior is identified, parameter updates can be prioritized. Furthermore, if necessary, additional anonymization can be performed before transmission to further protect the privacy of the owners of the vehicles participating in the training.

[0056] Subsequently, the server then aggregates the received parameter updates into a globally updated model parameter. In the simplest case, this is achieved by forming the average of all local updates. However, it is also possible to weight the received parameter updates to give higher weight to the identified outliers as described above.

[0057] In method step 16, after the central server aggregates the modified model parameters, the global model parameters generated by the central server using the aggregation of the modified model parameters can be received in the respective vehicles. This then enables a more accurate inference of the aging of the components installed in the vehicles and detected by the aging model in the respective vehicles.

[0058] The described method for aging prediction can continue indefinitely, or it can proceed in an iterative process until a termination condition is reached. Thus, the iteration process can be terminated after a predefined number of iterations have been performed. Similarly, provision can be made for the method to be terminated if fewer than a predefined number of vehicles have participated in the learning process.

[0059] Furthermore, when training neural networks, it is also possible to terminate the process based on convergence of suitable metrics or loss functions, which can be used to determine how the performance of the aging model has changed. Thus, if only small changes in model parameters occur, meaning the model learns fewer and fewer new things, the repetition rate of global model updates—that is, how often parameter updates are sent to the central backend, the global model update is determined, and then provided again to the vehicles—can be reduced. However, the method is not terminated based on convergence metrics, as even if a local / vehicle-specific model currently does not contribute to the global model, this does not necessarily mean that this will remain the case in the future.

[0060] The method according to the present invention implemented in a vehicle can be implemented, for example, as a computer program on a controller. To this end, the computer program is transferred to and stored in the memory of the corresponding controller during its manufacture. The computer program includes instructions that, when executed by the controller's processor, cause the controller to carry out the steps of the method according to the present invention.

[0061] Figure 2 An example for inferring battery aging using a neural network-based aging model is schematically shown.

[0062] In the example shown, a plurality of transport means F1 of a first party P1 A to F1 N and multiple means of transport F2 of the second party P2 A to F2 N Characteristic data about the current operating state of the respective installed battery BAT are determined in the vehicle. Based on this, the battery aging model is trained locally in each vehicle.

[0063] For example, these two parties could be two vehicle manufacturers that install the same battery in their vehicles and therefore share a common interest in developing the most accurate aging model possible for these batteries, but do not wish to share information about how their vehicles are used. In the example shown, the method according to the present invention is described only as an example for two parties, but more than two parties are equally possible. Similarly, other parties may also be involved in the method according to the present invention, such as suppliers of vehicle components or even operators of large vehicle fleets. As mentioned above, laboratory or test bench data can also be used in this context. Furthermore, multiple data from different development stages or generations of components can also be considered.

[0064] On each vehicle, the aging model is trained based on the locally available vehicle data. The local model parameters MP1 obtained from the training are A To MP1 N and MP2 A to MP2 NA change in the model parameters results in a corresponding change in the model parameters MP1 and MP2 for each participant, who can each operate their own server for their assigned vehicle. Each participant sends these model parameter changes to a central server ZS, where the changes in the model parameters of each participant are aggregated, for example, as an average of the individual changes in the model parameters. This central server ZS can be operated as a backend server, for example, by a specialized service provider that provides distributed machine learning services to the participants for a fee and, for this purpose, calls upon parameter updates learned from the parties' vehicles. Furthermore, based on the participant's specifications, the central server can also limit aging predictions to specific vehicle components, such as batteries from a specific manufacturer or a specific vehicle model.

[0065] Figure 3 A block diagram of a vehicle F is schematically shown, which includes various units that can be used to infer battery aging using the method according to the present invention. In the example shown, this may be an electrically driven vehicle, comprising one or more electric motors (not shown in this example), which are driven by a drive battery BAT. The drive battery may consist of one or more battery modules or be constructed according to the cell-to-pack (CTP) method, and may contain lithium-ion batteries in particular. The drive battery system may also include further components, such as a battery management system, battery heating / cooling, and charging electronics.

[0066] The vehicle has a communication unit KE, which can be used to receive model parameters for the aging model and specifications of vehicle data suitable for inferring battery aging; and to send locally updated model parameters. In addition, the vehicle has an acquisition unit EE for acquiring a time series of standardized vehicle data, that is, in the example shown, battery data of the vehicle for the battery aging model, and a calculation unit RE for training the aging model NN. The calculation unit RE and the acquisition unit EE are shown separately in the example, but can also be integrated into a single unit. In this case, the calculation unit RE and the acquisition unit EE, if necessary, can be part of a controller, such as a central controller or a vehicle server. In this case, the controller can include one or more processors to implement the method according to the invention, each of which has one or more processor units, such as microprocessors, digital signal processors, or a combination thereof.

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

[0068] Furthermore, the drive battery BAT, the computing 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 present invention can be implemented in any means of transport, such as a passenger car, but is not limited thereto.

[0070] Reference Signs List

[0071] 10-16 Method Steps

[0072] ZS Central Server

[0073] F1 A ,F1 N ,F2 A ,F2 N Transportation for participating in distributed learning

[0074] P1, P2 participants' servers

[0075] MP1 A ,MP1 N ,MP2 A ,MP2 N Locally changed model parameters

[0076] Changed model parameters of MP1, MP2 participants

[0077] Global model parameters updated by MP

[0078] NN machine learning models

[0079] NN ZS Global Machine Learning Model

[0080] F. Transportation

[0081] KE Communication Unit

[0082] SE storage unit

[0083] RE calculation unit

[0084] BAT battery cell

[0085] EE acquisition unit

[0086] B Digital Data Bus

Claims

1. A method for machine learning for aging prediction of components of a vehicle, wherein the vehicle performs the following steps: - receiving (10) model parameters for an aging model decentralized trained on a plurality of transport vehicles; - receiving (11) specifications of vehicle data suitable for inferring the aging of said component; - collecting (12) a time series of normalized transport data of the transport; - calculating (13) at least one statistical characteristic variable for the time series of the acquired normalized means of transport data, - training (14) the aging model (NN) based on the at least one statistical characteristic variable, wherein changes in model parameters are calculated; - The changed model parameters (MP1 A ,MP1 N ,MP2 A ,MP2 N ) or information about changes in said model parameters is sent (15) to a central server (ZS).

2. The method according to claim 1, wherein global model parameters (MP) are received (16), which are 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 - Standardizing model parameters and vehicle data for the aging model by a plurality of parties (P1, P2); - the plurality of participants independently provide the transport means (F1) assigned to them A ,F1 N ,F2 A ,F2 N ) or vehicle components and thereby calculating changes in the model parameters for the vehicles or vehicle components assigned to them; - sending of the changed model parameters or information about the changes in the model parameters by the plurality of participants to the central server (ZS); as well as - The central server aggregates the changes of the model parameters of each party into the global model parameter (MP).

4. The method according to claim 3, wherein the aging model is initially trained by one of the participants and the initial model parameters generated therefrom are sent to the other participants and subsequent training (14) of the aging model (NN) is performed by all participants.

5. A method according to any one of the preceding claims, wherein Modified model parameters of the transport means are determined which differ significantly from the modified model parameters of the remaining transport means and are weighted more highly in the aggregation of the modified model parameters.

6. A method according to any one of the preceding claims, wherein The changed model parameters or information about the changes in the model parameters are sent to the central server at predefined time intervals.

7. The method according to claim 6, wherein: If the changed model parameters or the information about the change of the model parameters differ significantly from the changed model parameters of the remaining transport means, the changed model parameters or the information about the change of the model parameters are sent with priority.

8. A method according to any one of the preceding claims, wherein The vehicle data is generated by a controller in the vehicle and exists as a time series of CAN messages.

9. A method according to any one of the preceding claims, wherein The aging model infers aging of the battery (BAT) of the vehicle and determines a statistical characteristic variable for at least one battery characteristic variable 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 infers aging of an exhaust gas recirculation cooler of the vehicle and determines a statistical characteristic variable for at least one characteristic variable of the exhaust gas recirculation cooler over a defined period of operation of the exhaust gas recirculation cooler.

11. A device for machine learning used for aging prediction of components of a transportation vehicle, comprising a communication unit (KE) for receiving model parameters for an aging model trained decentralized by a plurality of vehicles, for receiving specifications of vehicle data suitable for inferring aging of the component, and for sending changed model parameters or information about changes in said model parameters to a central server; an acquisition unit (EE) for acquiring a time series of normalized vehicle data of the vehicle; A calculation unit (RE) for calculating at least one statistical characteristic variable for the time series of the acquired normalized vehicle data and for training the aging model based on the at least one statistical characteristic variable, wherein changes in model parameters are calculated.

Images