CN116306214A

CN116306214A - Method and device for providing an ageing state model for determining the ageing state of an energy store

Info

Publication number: CN116306214A
Application number: CN202211564400.6A
Authority: CN
Inventors: P·克里施南
Original assignee: Robert Bosch GmbH
Current assignee: Robert Bosch GmbH
Priority date: 2021-12-08
Filing date: 2022-12-07
Publication date: 2023-06-23
Also published as: DE102021213948A1

Abstract

The invention relates to a computer-implemented method for training a data-based aging state model with a probabilistic regression model, having the steps of: providing training data sets, wherein each training data set distributes operation characteristic points of a plurality of operation characteristics, which are determined from the operation variable change process, to an aging state serving as a label; dividing the training data set into a training set and a verification set; training a probabilistic regression model based on the training set; iteratively adapting at least one hyper-parameter of the regression model according to the validation set, thereby avoiding over-fitting and/or under-fitting; determining an extrapolated training data set from the validation set as a training data set in the validation set that is outside of the running feature space of the trained regression model; iteratively adapting at least one hyper-parameter of the probabilistic regression model based on the extrapolation metric; the regression model is trained based on all of the training data sets and the set at least one determined hyper-parameter.

Description

Method and device for providing an ageing state model for determining the ageing state of an energy store

Technical Field

The invention relates to an electrical system-independent electrical device, in particular an electrically drivable motor vehicle, in particular an electric vehicle or a hybrid vehicle, having an electrical energy store, and also to measures for determining the State of aging (SOH) of the electrical energy store. Furthermore, the invention relates not only to mobile electrical energy stores but also to stationary electrical energy stores.

Technical Field

The energy supply of electrical devices and machines (for example electrically drivable motor vehicles) operating independently of the electrical network takes place by means of an electrical energy store (typically a device battery or a vehicle battery). These electrical energy stores provide electrical energy for operating the device. However, a fuel cell system including a hydrogen tank is also considered as the electric accumulator.

The electrical energy storage or energy converter may degrade during its service life depending on its load or use. This so-called aging results in a continuous decrease in maximum performance and storage capacity. The aging state corresponds to a measure for illustrating the aging of the energy store. Conventionally, new energy storage devices have a 100% ageing state (related to capacity) that gradually decreases over their service life. The measure of the ageing of the energy store (change in the ageing state over time) depends on the individual load of the energy store, i.e. in the case of the vehicle battery of the motor vehicle, on the behavior of the driver in use, the external environmental conditions and the type of vehicle battery.

Although the instantaneous state of aging of the energy store can be determined on the basis of a history of operating state changes by means of a physical state of aging model, this model is inaccurate in certain situations. This inaccuracy of the conventional aging state model makes it difficult to predict the aging state change process. However, the prediction of the course of the aging state change of the energy store is an important technical variable, since the residual value of the energy store can be evaluated economically using this prediction.

Disclosure of Invention

According to the present invention, a method for providing a data-based ageing state model for determining an ageing state of an electrical energy storage device according to claim 1 and a corresponding device according to the parallel independent claim are provided.

Further designs are specified in the dependent claims.

According to a first aspect, a computer-implemented method for training a data-based aging state model with a probabilistic regression model for modeling the aging state of an electrical energy storage having at least one electrochemical cell, in particular a battery cell, is provided, the method having the steps of:

-providing training data sets, each training data set assigning operating feature points of a plurality of operating features determined from the operating variable change process to an aging state as a label;

-dividing the training data set into a training set and a validation set;

-training the probabilistic regression model based on the training set;

-iteratively adapting at least one hyper-parameter of the regression model according to the validation set, thereby avoiding overfitting and/or underfilling;

-determining an extrapolated training dataset from the validation set as a training dataset in the validation set that is outside the running feature space of the trained regression model;

-iteratively adapting at least one hyper-parameter of the probabilistic regression model based on an extrapolation metric, the extrapolation metric being determined from the extrapolated training dataset and a pre-given prior of the regression model to be created;

-training the probabilistic regression model based on all training data sets and the set at least one determined hyper-parameter.

An energy store within the meaning of the present description includes a device battery, an energy converter system with an electrochemical energy converter and an energy carrier store, for example a fuel cell system with a fuel cell and an energy carrier store.

The aging state of an electrical energy store, in particular of a device battery, is generally not measured directly. This would require a series of sensors to be provided within the accumulator which would make the manufacture of such an accumulator cost intensive and complex and would increase space requirements. Furthermore, there is no commercially available measurement method for the direct determination of the aging state in the energy store. The current aging state of the electrical energy store is therefore usually determined by means of a physical aging model in a control device separate from the energy store. Such physical aging state models are inaccurate under certain conditions and typically have model deviations of up to 5% or more.

The State of aging (SOH) is a key variable for explaining the remaining battery capacity or the remaining battery power in the case of the device battery. The aging state is a measure of the aging of the device battery. In the case of a device battery or battery module or battery cell, the aging state may be given as a capacity retention rate (Capacity Retention Rate, SOH-C). The capacity retention SOH-C is given as the ratio of the measured instantaneous capacity to the initial capacity of the fully charged battery. Alternatively, the aging state may be given as an increase in internal resistance (SOH-R) relative to the internal resistance at the beginning of the service life of the device battery. This relative change in internal resistance SOH-R increases as the battery ages.

Furthermore, due to the inaccuracy of the physical aging model, the physical aging model can only account for the instantaneous aging state of the energy store to a certain extent with accuracy. What is expected is a solution for modeling and predicting the aging state of an electrical energy store on the basis of an aging state model, which uses the time course of the operating variables from the start of operation for the time-step-by-time adaptation of the aging state from the start of operation. Such an aging model may be implemented purely on a data basis, but may also be implemented as a hybrid data-based aging state model. Such an aging state model may be implemented, for example, in a central unit (cloud) and may be parameterized or trained by means of operating variables of a large number of device batteries of different devices communicatively connected to the central unit.

The aging state model for determining the aging state of the electrical energy storage may be provided, for example, in the form of a hybrid aging state model, i.e. a combination of a physical aging model and a data-based correction model corresponding to a probabilistic regression model. In the case of a hybrid model, the physical aging state can be determined by means of a physical or electrochemical aging model, and the correction variables derived from the data-based correction model are applied to the physical aging state, in particular by addition or multiplication. The physical aging model is based on electrochemical model equations that characterize the electrochemical states of a system of nonlinear differential equations, which are continuously calculated using a time integration method and mapped into the physical aging state for output as SOH-C and/or SOH-R. These calculations may typically be performed in the cloud, for example once a week.

Furthermore, a correction model of the data-based hybrid aging state model may be constructed with a probabilistic regression model or an artificial intelligence-based regression model, in particular a gaussian process model, and may be trained to correct the aging state obtained by the physical aging model. For this purpose, a data-based correction model of the aging state is therefore present for correcting the capacity-based aging state SOH-C and/or at least one further data-based correction model for correcting the resistance-based aging state SOH-R. Possible alternatives to the gaussian process are further supervised learning methods, such as methods based on random forest models, adaBoost models, support vector machines or bayesian neural networks.

The prediction of the aging state is helpful, for example, when the remaining service life of the energy store should be determined and evaluated, for example, according to a warranty condition or a predetermined value of the CO2 fleet. For this purpose, a data-based aging state model can be continuously checked in connection with a predefined usage pattern, which specifies the type of use and operation of the electrical energy store.

In order to model the aging state of the energy store by means of a physical or electrochemical aging state model and by means of an optional specification of a data-based correction model (hybrid aging state model), it is necessary to provide a time-dependent course of the operating variables relatively frequently. In addition, these time-dependent processes of the operating variables must be provided as completely as possible in order to meet the necessary accuracy requirements, i.e. to determine the aging state of the device battery at the current time, it is necessary to provide time-dependent processes of the battery temperature, of the battery current, of the battery voltage and of the state of charge, in particular at the battery cell level.

The calculation of the physical aging model together with the correction model preferably takes place outside the device, since this calculation is very complex and the required processing power is often insufficient in or near the hardware of the battery-operated device or should not be maintained for cost reasons. The time-dependent course of the operating variables is thus transmitted to a central unit outside the installation and the aging state is determined there from the physical aging model and the correction model.

The data-based correction model uses at least one operating characteristic from the time-varying course of the operating variable and maps the operating characteristic to the correction variable.

The at least one operating characteristic may comprise a characteristic averaged over an evaluation period and/or a cumulative characteristic and/or a statistical characteristic determined over the entire lifetime to date, and may in particular comprise:

electrochemical state, in particular SEI layer thickness in case of battery as an energy storage, cyclizable lithium change due to anode/cathode side reaction, rapid absorption rate of electrolyte solvent, slow absorption rate of electrolyte solvent, lithium deposition rate, anode active material loss rate, cathode active material loss rate, internal resistance, etc.,

histogram features such as temperature versus state of charge, charge current versus temperature or discharge current versus temperature, current throughput, accumulated total charge (Ah), average capacity increase during charging, extremum (e.g. local maximum) of charge capacity and smoothed differential capacity dQ/dU, or accumulated distance travelled.

Providing the aging state model with a probabilistic regression model as a purely data-based model or as a hybrid model requires proper training of the data-based regression model. The probabilistic regression model can be constructed in particular as a gaussian process model and optimized by means of a bayesian optimization method in order to find, as model parameters, an optimized combination of superparameters, i.e. signal variances and length scales, for the gaussian process, and a covariance matrix with a minimum number of functional evaluations. This process may lead to generalization problems in the event of noise level fluctuations in the training data.

Furthermore, bayesian optimization does not provide the flexibility to adapt the regression model for extrapolation and interpolation scenarios. For the mixed aging state model, the correction model implemented as a gaussian process should so far only improve the interpolation of the physical aging model, i.e. make sufficient corrections in the feature space of the correction model, but impair the physical aging model for the extrapolation case as little as possible. Due to the multi-objective optimization, the general approach of optimizing the hyper-parameters of the regression model by means of a grid search algorithm is unsuitable, because the grid search algorithm cannot take into account the interpolated quality, extrapolated quality and generalization requirements at the same time.

In order to take these aspects into account, a method is proposed which considers both generalization requirements and interpolation and extrapolation requirements by correcting the model by adapting the super-parameters in advance. To this end, the available training data set may first be subdivided into a training set and a validation set, e.g. by random selection or by domain and expert knowledge, and a regression model may be trained based on the training set, e.g. in the form of a gaussian process model. The resulting hyper-parameters are now iteratively adapted according to the model quality, e.g. according to a fitting metric determining under-fitting or over-fitting from the validation set and the training set.

The training of the probabilistic regression model may be performed by means of a bayesian optimization method. Furthermore, at least one hyper-parameter of the regression model may be iteratively adapted by means of a grid search algorithm.

As with each probabilistic regression model, the gaussian process model provides a confidence interval at each model evaluation in addition to the model evaluation, wherein the confidence interval must be comparable or similar if the validation set is well mapped in the training set. The hyper-parameters may be adapted by means of a grid search algorithm when the fitting metrics indicate a significant under-fit and/or a significant over-fit, i.e. e.g. when the corresponding fitting metrics exceed or fall below a predetermined threshold. The overfitting and/or underfilling may be determined by means of a validated set of training data sets.

The overfitting may be identified using fitting metrics derived from model evaluations of the training set and the validation set. An average error or accuracy of the model output is determined for the training data points of the training set and for the data points of the validation set. If the deviation between the average error of the training data points related to the training set and the average error of the data points related to the verification set is negative, or the accuracy of the training set is better than the accuracy of the verification set, there is a tendency for the fit. The fitting metric may generally account for differences in accuracy of model evaluations of data points of the training set and the validation set.

The under-fit may be identified using a fit metric that accounts for the accuracy of a model used to map the data points of the training set with which the training model has been trained. The fit metric may correspond to an average error or accuracy of model outputs of training data points of the training set and the validation set.

By training the gaussian process model by means of bayesian optimization, the hyper-parameter space that has to be searched by a grid search algorithm can be significantly reduced.

Iterative adaptation of at least one hyper-parameter of the regression model to avoid overfitting and/or underfilling may be performed by means of a predefined bias/variance trade-off criterion and optimization of the grid search method. The grid search algorithm is performed until the fit metric determined by means of the validation set indicates neither a under-fit nor an over-fit. This may be indicated by a threshold query for fit metrics that are over-fit and/or under-fit.

The following training data sets may then be selected from the validation set of training data sets as extrapolated training data sets for which the generated model output is extrapolated. Extreme feature points in the validation set that are not present in the training set may be assumed to be extrapolated. Model quality is then evaluated for different extrapolation conditions, wherein model predictions of the regression model are evaluated for the different extrapolation conditions.

The extrapolated training data set may be selected from the validation set by evaluating the distance between each data point of the validation set and the training data point of the training set, respectively, e.g., as an average or minimum euclidean distance. If the average or minimum Euclidean distance is above a predetermined threshold, then the training data set from the validation set is assumed to be an extrapolated training data set.

The extrapolation metric may be determined as an average deviation between the respective aging state of the extrapolated training data set and the predetermined priori.

Iterative adaptation of at least one hyper-parameter of the regression model may be based on the extrapolated measure in a direction to minimize the extrapolated measure.

An extrapolation metric is thus determined that describes at what speed or at what distance from the feature space of the previously trained regression model (i.e., from the feature space of the training set) a model output value corresponding to a priori is derived for the extrapolated training dataset. In the case of a correction model of the hybrid aging state model, the prior is selected in such a way that it does not change the model output of the physical aging model and corresponds, for example, to correction value 0 when applied additively and correction value 1 when applied multiplicatively.

If the extrapolation metric does not meet the predetermined extrapolation criterion, adapting the hyper-parameters iteratively towards that model evaluation of the gaussian process using the extrapolation training dataset may return to the prior faster, i.e. the extrapolation metric moves towards meeting the extrapolation criterion, e.g. by a threshold comparison. This process is repeated as long as the extrapolation metric does not meet the predefined extrapolation criteria.

Further, an interpolated training dataset from the validation set may be determined as a training dataset in the validation set that is within the running feature space of the trained regression model. At least one hyper-parameter of the regression model is then adapted based on an interpolation metric determined from the interpolation training data set and the respective modeled aging state of the regression model evaluated at the operating characteristic points of the corresponding interpolation training data set.

Thus, the following training data sets may be similarly selected from the validation set of training data sets as interpolated training data sets for which to interpolate the generated correction variables. Feature points in the validation set, i.e. the common operating range covered by the training set, are assumed to be interpolated training data sets.

The interpolated training data set may be selected from the verification set by evaluating the distance between each data point of the verification set and the training data point of the training set, respectively, e.g. as an average or minimum euclidean distance. If the average or minimum Euclidean distance is below a predetermined threshold, then the training data set from the validation set is assumed to be the interpolated training data set.

For example, an interpolation metric may be determined as an average difference between an aging state of an interpolation training dataset and a corresponding modeled aging state of a regression model evaluated at an operating feature point of the corresponding interpolation training dataset, wherein at least one hyper-parameter of the regression model is iteratively adapted based on the interpolation metric towards minimizing the interpolation metric.

Thus, an interpolation metric is determined that accounts for the degree of consistency of the interpolated training dataset with the model predictions. To this end, the residuals of the labels (aging states) of the interpolated training dataset and the model predictions of the aging states may be determined and used to determine the interpolation metrics.

If the interpolation metric does not meet a predetermined interpolation criterion, adapting the hyper-parameters iteratively towards a direction that model evaluation using a gaussian process of interpolating the training dataset better corresponds to model prediction, i.e. the interpolation metric moves towards meeting the interpolation criterion, e.g. by a threshold comparison. This process is repeated as long as the interpolation metric does not meet the predefined interpolation criteria.

It may be desirable to adapt the superparameter according to the interpolation conditions, because the superparameter has deteriorated interpolation behavior due to bayesian optimization and adaptation based on extrapolation conditions. The finally obtained hyper-parameters are now used as set hyper-parameters for training a gaussian process model, which is trained using the entire training dataset. Only the covariance matrix is trained in this training. In this way, the requirements for extrapolation and, if necessary, interpolation of the gaussian process model as correction model for the aging state model can be combined to meet the bayesian optimization.

The aging state model as described above may be used to determine an aging state for an energy store in a device, such as a motor vehicle, an intelligent electric vehicle, an aircraft (particularly a drone), a machine tool, a consumer electronic device (such as a mobile phone), an autonomous robot, and/or a household appliance. For example, the aging state model may be used to determine remaining useful life and/or adapt the operating strategy to aging behavior resulting from an evaluation of the aging state model.

According to a further aspect, a device for training a probabilistic regression model of a data-based aging state model for modeling an aging state of an electrical energy storage having at least one electrochemical cell, in particular a battery cell, is provided, wherein the device is configured to:

-dividing the training data set into a training set and a validation set;

-training the probabilistic regression model based on the training set;

-iteratively adapting at least one hyper-parameter of the regression model based on an extrapolation metric, the extrapolation metric being determined from the extrapolated training dataset and a pre-given prior of the regression model to be created;

-training the regression model based on all training data sets and the set at least one determined hyper-parameter.

Drawings

Embodiments are explained in more detail below based on the drawings.

FIG. 1 shows a schematic diagram of a system for providing driver-specific and vehicle-specific operating variables to determine the aging state of a vehicle battery in a central unit;

FIG. 2 shows a schematic diagram of the functional structure of a hybrid aging state model;

FIG. 3 illustrates a flow chart showing a method for training a data-based aging state model.

Detailed Description

The method according to the invention is described below on the basis of a vehicle battery as an electrical energy store in a large number of motor vehicles as devices of the same type. As described below, the aging state model may be continuously updated or retrained in a central unit external to the vehicle based on operating variables from the vehicle battery of the fleet. The aging state model runs in the central unit and is used for aging calculation and aging prediction.

The following examples represent a large number of stationary or mobile devices with a grid-independent energy supply, such as vehicles (electric cars, intelligent electric vehicles, etc.), facilities, machine tools, household appliances, internet of things devices, etc., which are connected via corresponding communication connections (e.g. LAN, internet) to a central unit (cloud) outside the device.

Fig. 1 shows a system 1 for collecting fleet data in a central unit 2 for creating and running and evaluating aging state models. The aging state model is used to determine an aging state of an electrical energy storage device (e.g., a vehicle battery or fuel cell in a motor vehicle). Fig. 1 shows a fleet 3 with a plurality of motor vehicles 4.

One of the motor vehicles 4 is shown in more detail in fig. 1. The motor vehicle 4 has a vehicle battery 41 as a rechargeable electric energy store, an electric drive motor 42 and a control unit 43, respectively. The control unit 43 is connected to a communication module 44, which communication module 44 is adapted to transmitting data between the respective motor vehicle 4 and the central unit 2 (so-called cloud).

The motor vehicle 4 transmits to the central unit 2 an operating variable F which describes at least the variables which influence the aging state of the vehicle battery 41. In the case of a vehicle battery, the operating variables F may account for instantaneous battery current, instantaneous battery voltage, instantaneous battery temperature and instantaneous State of Charge (SOC: state of Charge), at the battery pack level, at the module level and/or at the battery cell level. The operating variable F is detected in a rapid time grid from 2Hz to 100Hz and can be transmitted periodically to the central unit 2 in uncompressed and/or compressed form. For example, the time series may be transmitted to the central unit 2 block by block at intervals of 10 minutes to several hours, wherein a compression algorithm is used to minimize the data traffic to the central unit 2.

The central unit 2 has a data processing unit 21 in which the method described below can be performed, and a database 22 for storing data points, model parameters, states, etc.

An aging state model is implemented in the central unit 2, which aging state model is based in part on the data as a hybrid model. The aging state model can be used periodically, i.e. for example after the end of the respective evaluation duration, to determine the momentary aging state of the vehicle battery 41 in question of the associated vehicle fleet on the basis of the time course of the operating variables (respectively starting from the start of the respective vehicle battery operation) and the operating characteristics determined therefrom. In other words, the aging state of the vehicle battery 41 concerned can be determined on the basis of the course of the change in an operating variable of one of the vehicle batteries 41 of the motor vehicle 4 of the associated fleet 3 and of an operating characteristic derived from the course of the operating variable.

The State of aging (SOH) is a key variable for explaining the remaining battery capacity or the remaining battery power in the case of the device battery. The aging state is a measure of the aging of the device battery. In the case of a device battery, battery module or battery cell, the aging state may be given as a capacity retention rate (Capacity Retention Rate, SOH-C). The capacity retention SOH-C is given as the ratio of the measured instantaneous capacity to the initial capacity of the fully charged battery and decreases with increasing aging. Alternatively, the aging state may be given as an increase in internal resistance (SOH-R) relative to the internal resistance at the beginning of the service life of the device battery. This relative change in internal resistance SOH-R increases as the battery ages.

Fig. 2 shows schematically and exemplarily the functional structure of an embodiment of a data-based aging state model 9 constructed in a hybrid manner. The aging state model 9 includes a physical aging model 5 and a correction model 6.

The physical aging model 5 is a mathematical model based on differential equations. Evaluating the physical aging model of the aging state model using the operating variable course, in particular from the beginning of the service life of the device battery, results in the internal state of the equation set of the physical differential equation being established, which corresponds to the physical internal state of the device battery. Since the physical aging model is based on the laws of physics, model parameters of the physical aging model are variables that account for physical characteristics.

The time sequence of operating variables F therefore goes directly into a physical aging state model 5, which is preferably implemented as an electrochemical model and describes the corresponding internal electrochemical states, such as layer thickness (e.g. SEI thickness), changes in cyclizable lithium due to anode/cathode side reactions, rapid consumption of electrolyte, slow consumption of electrolyte, loss of anode active material, loss of cathode active material, etc., by means of nonlinear differential equations and multidimensional state vectors.

Thus, the physical aging model 5 corresponds to an electrochemical model of the battery cell and cell chemistry. Based on the operating variables F, the aging model 5 determines the internal physical battery state in the form of a state vector by means of time integration in order to calculate an at least one-dimensional physical-based aging state SOHph in the form of the above-described electrochemical states, which are mapped linearly or nonlinearly into a capacity retention rate (SOH-C) and/or an internal resistance increase rate (SOH-R) in order to provide the capacity retention rate or the internal resistance increase rate as aging states (SOH-C and SOH-R).

However, the model values of the physical aging state SOHph provided by the physical aging model are inaccurate under certain conditions, and therefore provision is made to correct these model values with a correction variable k. The correction variables k are provided by a data-based correction model 6 which is trained by means of a training data set from the vehicles 4 of the fleet 3 and/or by means of laboratory data.

The correction model 6 obtains on the input side an operating characteristic M which is determined from the course of the change in the operating variable F and which may also comprise one or more internal electrochemical states of the differential equation system of the physical model. Furthermore, the correction model 6 may obtain the physical aging state SOHph obtained from the physical aging model 5 on the input side. The operating characteristics M of the current evaluation period are generated in the characteristic extraction block 8 on the basis of the time series of the operating variables F. Furthermore, the operating characteristics M also include internal states of the state vector from the electrochemical physical aging model 5.

Depending on the operating parameters F, operating characteristics M relating to the evaluation period can be generated in the central unit 2 for each fleet 3 or in other embodiments already in the respective motor vehicle 4. To determine the aging state, the evaluation period may be several hours (e.g., 6 hours) to several weeks (e.g., 1 month). A common value for the evaluation period is one week.

The operating characteristics M may comprise, for example, characteristics relating to the evaluation period and/or cumulative characteristics and/or statistical variables determined over the entire service life up to now. In particular, the operating characteristics may include, for example: electrochemical state (e.g., SEI layer thickness, change in recyclable lithium due to anode/cathode side reactions, rapid absorption of electrolyte solvent, slow absorption of electrolyte solvent, lithium deposition, loss of anode active material and loss of cathode active material), information about impedance or internal resistance, histogram characteristics (e.g., temperature versus state of charge, charge current versus temperature and discharge current versus temperature), multi-bit histogram data, particularly for battery temperature distribution over state of charge, charge current distribution over temperature and/or discharge current distribution over temperature, current throughput in ampere hours, accumulated total charge (Ah), average capacity increase during charging (particularly for charging processes where the charge increase is greater than a threshold fraction of total battery capacity (e.g., 20%), differential capacity extremum (maximum) (dQ/dU: charge change divided by battery voltage change) during the measured charging process where the state of charge has a sufficiently large stroke), or accumulated mileage. These variables are preferably scaled such that they represent the best possible characterization of the actual usage behavior. The operating characteristics M may be used in whole or in part in the methods described below.

In order to determine the corrected state of aging SOH to be output, the outputs SOHph, k of the physical aging model 5 and the data-based correction model 6, which is preferably embodied as a gaussian process model, are applied to one another. In particular, these outputs may be added or multiplied (not shown) in a summing block 7 to obtain a modeled state of aging SOH to be output during the current evaluation period. In the case of addition, the confidence of the gaussian process can still be used as the confidence of the corrected aging value SOH to be output of the hybrid model.

Scaling and dimension reduction of the run features is done, and PCA (Principal Components Analysis, principal component analysis) can be used as necessary to correspondingly reduce redundant linear correlation information in feature space before training the correction model (unsupervised). Alternatively, kernel PCA may also be used to enable non-linear effects to be mapped out while reducing data complexity. The entire running feature space (or principal component space) is normalized, e.g., using a min/max scaling or Z-transform, both before and especially after dimension reduction.

Thus, for any type of energy storage having at least one electrochemical cell, such as a battery cell, the state of aging and the predicted state of aging may be calculated. The method may also be applied to the whole system of the energy store by rule-based and/or data-based mapping. Thus, taking the battery as an example, the aging prediction can be applied not only to the battery cell level but also directly to the module level and the battery pack level.

Fig. 3 shows a flow chart illustrating an exemplary method for training the correction model 6 in the hybrid ageing state model in the central unit 2. For this purpose, a training data set is defined, which assigns the course of the change in the operating variable to an empirically determined aging state as a label.

The determination of the aging state as a tag can be carried out in a manner known per se by evaluating the course of the operating variables with additional models in the vehicle or in the central unit 2 under defined load and environmental conditions of the tag production, for example on workshops, test benches or in diagnostic or tag production modes which represent an operating mode and ensure compliance with predetermined operating conditions of the vehicle battery, for example constant temperature, constant current, etc. To this end, other models may be used to determine the aging state, for example based on analysis of identified charge phases and/or discharge phases of battery usage.

Preferably, the SOH-C measurement is performed by coulomb counting or by forming a current integral over time during the charging process, which current integral is divided by the charge state stroke between the beginning and end of the charging and/or discharging phase concerned. In this case, the idle voltage characteristic is advantageously calibrated during the idle phase in order to calculate the charge state change process together in the central unit. For example, a sufficiently reliable description about the aging state used as a tag can be obtained if the vehicle battery changes from a defined relaxed state to a fully charged state from a fully discharged charged state under reproducible load conditions and environmental conditions during the charging process. The maximum amount of power thus detected may be related to an initial maximum charge capacity of the vehicle battery. The state of aging (SOH-R value) associated with the resistance can also be calculated from the voltage change associated with the current change. Typically, the aging state associated with the resistance is related to a defined time interval as well as a defined environmental condition and the direction of energy flow of the system.

The training data set for the vehicle battery is thus derived from the respective aging state determined at the point in time and the change in the operating variables of the vehicle battery in question from the point in time the vehicle battery in question was put into operation to the point in time. For a vehicle battery, multiple training data sets may be determined at different points in time, which may be described with respect to a point in time of commissioning (beginning of service life). These training data sets are collected and provided for a large number of vehicles.

Based on the assignment of the operating characteristics of each vehicle battery 41 determined from the course of the operating variable change at a specific point in time (calendar age) to the training data set of the corresponding aging state of the vehicle battery 41 in question, a training data-based correction model, which is constructed as a probabilistic regression model, in particular as a gaussian process model, is described in the following method. The focus of the training of the gaussian process model is that the training considers training data sets as optimally as possible and has improved performance in terms of extrapolation and interpolation.

In step S1, the running variable change process F is first detected as described above and collected in the central unit 2.

In step S2, the outliers are removed from the running variable change process F and the data gaps are filled in according to a known method.

Then, in step S3, a label in the form of an aging state SOH-C, SOH-R is determined for the training data set by means of suitable methods, in which a suitable measurement or evaluation of the operating variable course is carried out, so that the aging state of the individual vehicle batteries 41 is given as a label at a specific evaluation time.

In step S4, as described above, the operation variable change process F is summarized as the operation characteristic M for these vehicle batteries 41. From these operating characteristics, the feature space can be further reduced by means of principal component analysis to reduce the dimensionality of the training dataset. Corresponding generated operation characteristics M form operation characteristic points, and corresponding aging states SOH-C, SOH-R are allocated to the operation characteristic points, so that training data sets are respectively formed.

In step S5, the training data set of a large number of vehicles 4 in the fleet 3 obtained in this way is subdivided into a training set and a validation set. The subdivision is performed by the training set and the validation set having the same or similar distribution in the running feature space. Otherwise the selection of the training data set for the training set and the validation set may be done by application domain and expert knowledge or by random selection.

In step S6, the correction model/gaussian process model is trained by means of a bayesian optimization method based on the training set of training data sets. The bayesian optimization method selects the training data set that provides the best (i.e. best) contribution to the improvement of the correction model in the considered training set feature space as the training data set available in each iteration. This step is used to initialize the hyper-parameters of the gaussian process model.

In step S7, model quality is determined by comparing the aging state of the training dataset of the validation set with the modeled aging state obtained by model evaluation at the operating feature points determined using the training dataset of the validation set to determine one or more fitting metrics. The fit metric characterizes the quality of the model with respect to whether there is an under-fit or an over-fit.

The corresponding fitting measures can then be evaluated in step S8, for example by means of one or more threshold comparisons. In this regard, a fitting criterion is provided for evaluating the corresponding fitting metric. If the fit criterion is met, i.e. there is no under-fit or over-fit (alternative: yes), the method continues with step S9, otherwise the method continues with step S10.

In step S10, the hyper-parameters are adapted or optimized by means of a grid search algorithm for a predefined deviation/variance trade-off criterion, by means of which overfitting or underfilling can be avoided. By using a bayesian optimization method in step S6, the hyper-parameter space that should be searched by the grid search algorithm can be significantly reduced.

The method then continues with step S7. This adaptation may be performed iteratively until the bias/variance trade-off criteria are met.

By analyzing the training data sets of the verification set and the training set, the extreme operating range in the verification set that is not represented in the operating feature space of the training set may be determined as an extrapolated training data set in step S9.

The extrapolated training data set may be selected from the validation set by separately evaluating the distance between each data point of the validation set and the training data point of the training set, e.g., as an average or minimum euclidean distance. If the average or minimum Euclidean distance is above a predetermined threshold, then the training data set from the validation set is assumed to be the extrapolated training data set.

In step S11, the model quality is evaluated for these data sets in the form of an extrapolation metric by means of the extrapolated training data set. In particular, the model evaluation of the gaussian process model is evaluated in such a way that how fast the model evaluation of the gaussian process model takes place, i.e. at what distance from the running feature space covered by the training set of the training dataset or from the limits of the running feature space, a priori (tolerances are taken into account if necessary).

The factorily applied correction variable is one in the case of the aging state model or zero in the case of correction variables to be applied in addition. Determining a corresponding extrapolation metric that negatively evaluates a deviation of the model evaluation of the pre-trained gaussian process model from the desired prior extrapolation training dataset over an extrapolation range and setting the corresponding extrapolation metric. For example, the extrapolation metric for an extrapolated training data set may be evaluated in a manner that how much the model evaluation at the corresponding operating feature points of the extrapolated training data set deviates from a priori. In particular, the average deviation may be determined as an extrapolation metric.

In step S12, it is checked whether the extrapolation metric meets a predefined extrapolation criterion. In particular, it can be checked in a threshold comparison whether the extrapolation metric exceeds a predefined threshold value, so that the extrapolation criterion is not met.

Thus, if it is determined in step S11 that the extrapolation metric meets the predefined extrapolation criterion (alternative: yes), the method continues with step S13, otherwise the method continues with step S14.

In step S13, one or more hyper-parameters of the gaussian process model are changed towards changing, in particular reducing, the extrapolation metric. This process may be performed iteratively.

Returning to step S11.

In step S14, a regular running range in the validation set that is identical, similar or identical to the training data set in the training set is determined as the interpolated training data set by analyzing the training data sets of the validation set and the training set.

In particular, the interpolated training data set may be selected from the verification set by evaluating the distance between each data point of the verification set and the training data point of the training set, respectively, e.g. as an average or minimum euclidean distance. If the average or minimum Euclidean distance is below a predetermined threshold, then the training data set from the validation set is assumed to be the interpolated training data set.

In step S15, model quality is evaluated for these data sets in the form of interpolation metrics by means of the interpolated training data sets. In particular, the model evaluation of the gaussian process model is evaluated using the operating feature points of the interpolated training dataset in such a way that the operating feature points deviate from the respectively modeled aging state by a large extent. These deviations may be averaged in order to determine an interpolation metric.

In step S16, it is checked whether the interpolation metric meets a predefined interpolation criterion. In particular, it can be checked in the threshold comparison whether the interpolation metric exceeds a predefined threshold value, so that the interpolation criterion is not fulfilled.

Thus, if it is determined in step S16 that the interpolation metric meets the predefined interpolation criteria (alternatively: yes), the method continues with step S17, otherwise the method continues with step S18.

In step S17, one or more hyper-parameters of the gaussian process model are changed towards changing the interpolation metric, in particular decreasing the interpolation metric. In particular, the length scale of the super-parameters may be adapted at a lower rate than the variance parameter of the super-parameters.

The process may be performed iteratively, i.e. adapting the super parameters to meet the interpolation criterion is continued until the interpolation criterion is met.

The finally determined hyper-parameters are then used in step S19 to train the gaussian process model based on all training data sets using the finally determined fixed hyper-parameters to obtain an associated covariance matrix.

Claims

1. A computer-implemented method for training a data-based aging state model (9) with a probabilistic regression model for modeling the aging state of an electrical energy store (41) having at least one electrochemical cell, in particular a battery cell, the method having the steps of:

-providing (S3, S4) training data sets, each training data set assigning an operating characteristic point of a plurality of operating characteristics determined from the operating variable variation process to an aging state as a label;

-dividing (S5) the training data set into a training set and a validation set;

-training (S6) the probabilistic regression model based on the training set;

-iteratively (S10) adapting at least one hyper-parameter of the regression model according to the validation set, thereby avoiding over-fitting and/or under-fitting;

-determining (S9) an extrapolated training dataset from the verification set as a training dataset in the verification set that is outside the running feature space of the trained regression model;

-iteratively adapting (S11, S12) at least one hyper-parameter of the probabilistic regression model based on an extrapolation metric, the extrapolation metric being determined from the extrapolated training dataset and a pre-given prior of the regression model to be created;

-training (S19) the regression model based on all training data sets and the set at least one determined hyper-parameter.

2. The method according to claim 1, wherein at least one hyper-parameter of the regression model is iteratively adapted by means of a grid search algorithm.

3. The method according to claim 1 or 2, wherein the training of the probabilistic regression model is performed by means of a bayesian optimization method.

4. A method according to any one of claims 1 to 3, wherein the extrapolation metric is determined as an average deviation between the respective aging state of the extrapolated training data set and the pre-given a priori.

5. The method according to any one of claims 1 to 4, wherein the iterative adaptation of at least one hyper-parameter of the regression model is performed based on the extrapolation metric towards minimizing the extrapolation metric (S13).

6. The method according to any one of claims 1 to 5, further comprising the step of:

-determining (S14) an interpolated training dataset from the validation set as a training dataset within the running feature space of the trained regression model in the validation set;

-iteratively adapting (S15, S16) at least one hyper-parameter of the regression model based on an interpolation metric, the interpolation metric being determined from the interpolation training data set and the respective modeled aging state of the regression model evaluated at the operating characteristic points of the corresponding interpolation training data set.

7. The method of claim 6, wherein the interpolation metric is determined as an average difference between an aging state of the interpolation training dataset and a respective modeled aging state of a regression model evaluated at a running feature point of the corresponding interpolation training dataset, wherein at least one hyper-parameter of the regression model is iteratively adapted based on the interpolation metric toward minimizing the interpolation metric.

8. The method according to any one of claims 1 to 7, wherein at least one hyper-parameter of the regression model is iteratively adapted by optimization by means of a predefined bias/variance trade-off criterion and grid search method to avoid over-fitting and/or under-fitting.

9. The method according to any one of claims 1 to 8, wherein the hybrid aging state model is a physical aging model (5) based on an electrochemical model equation and configured to output a physical aging state (SOHph), and the regression model is configured as a trainable data-based correction model (6), preferably a gaussian process, wherein the correction model (6) is trained to correct the physical aging state (SOHph) and to provide the corrected physical aging state as a modeled aging State (SOH), in particular with a quantified uncertainty.

10. The method according to any one of claims 1 to 9, wherein the at least one operating variable (F) comprises battery current (I), battery temperature (T), battery voltage (U) and state of charge (SOC) with the battery as an energy storage (41).

11. Method according to any of claims 1 to 10, wherein the trained aging state model is used for an energy store (41) to run devices such as motor vehicles, smart electric vehicles, in particular unmanned aerial vehicles, machine tools, consumer electronics devices such as mobile phones, autonomous robots and/or household appliances.

12. An apparatus for training a probabilistic regression model of a hybrid aging state model for modeling an aging state of an electrical energy store (41) having at least one electrochemical cell, in particular a battery cell, wherein the apparatus is configured to:

-dividing the training data set into a training set and a validation set;

-training the probabilistic regression model based on the training set;

13. A computer program product comprising instructions which, when the program is executed by at least one data processing apparatus, cause the data processing apparatus to perform the steps of the method according to any one of claims 1 to 11.

14. A machine readable storage medium comprising instructions which, when executed by at least one data processing apparatus, cause the data processing apparatus to perform the steps of the method according to any one of claims 1 to 11.