Method for providing at least one aging parameter of a battery

Abstract

Method (100) for providing at least one aging parameter of a battery (1), wherein the method (100) comprises the following steps: - applying (101) a physical model (3) based on at least one sensor signal to describe an aging process of the battery (1), wherein the at least one sensor signal results from a measurement on the battery (1), - applying (102) a machine learning model (2) to determine at least one battery parameter of the battery (1), wherein the battery parameter is determined at least on the basis of an output of the physical model (3), and wherein the machine learning model (2) is based on training with extended data resulting from an interpolation between individual measurement points of a reference signal for data extension, wherein the reference signal is specific for the at least one aging parameter, - Providing (103) the at least one aging parameter of the battery (1) on the basis of the determined battery parameter, wherein the physical model (3) and the machine learning model (2) are applied sequentially and characterized in that the machine learning model (2) comprises a transformer model.

Description

The invention relates to a method for providing at least one aging parameter of a battery. The invention further relates to a computer program, an apparatus and a storage medium for this purpose.Prior ArtDegradation of lithium-ion batteries (i.e., in particular, loss of capacity) can be a major problem for battery electric vehicle (BEV) owners and often are difficult to model and estimate. Battery-in-the-cloud (BitC) preferably uses data connectivity and cloud-based computing power to execute advanced battery aging models that use physical and machine learning (ML) components. In particular, existing hybrid model architectures combine a physical aging model with a Gaussian process regression (GPR) that corrects the capacity results of the physical aging model using the delta model. However, this can lead to difficulties in the detection of different aging scenarios.DE 10 2021 203 729 A1 describes a computer-assisted method for modelling or predicting the current state (SOH) of an energy store. It uses a data-based model that determines the SOH based on the temporal profile of operating parameters. Should there be a data gap in the course of the operating parameters, this is filled with a usage pattern model before the SOH is calculated.Disclosure of the InventionThe subject matter of the invention is a method with the features of claim 1, a training method with the features of claim 6, a computer program with the features of claim 7, an apparatus with the features of claim 8 and a computer-readable storage medium with the features of claim 9. Features and details which are described in connection with the method according to the invention naturally also apply in connection with the training method according to the invention, the computer program according to the invention, the device according to the invention and the computer-readable storage medium according to the invention, and vice versa, so that with regard to the disclosure reference is or can always be made to the individual aspects of the invention reciprocally.The invention relates in particular to a method for providing at least one aging parameter of a battery, wherein the method comprises the following steps, which are preferably carried out successively and / or repeatedly:applying a physical model based on at least one sensor signal to describe an aging process of the battery, wherein the at least one sensor signal results from a measurement at the battery,applying a machine learning model for determining at least one battery parameter of the battery, wherein the battery parameter is determined at least on the basis of an output of the physical model, and wherein the machine learning model is based on training with extended data which result from interpolation between individual measurement points of a reference signal for data extension, wherein the reference signal is specific to the at least one aging parameter,providing the at least one aging parameter of the battery on the basis of the ascertained battery parameter.Here, the physical model and the machine learning model may be sequentially applied. The battery can be embodied as a Li-ion cell and / or as a rechargeable battery (rechargeable battery). The aging parameter may be specific to a remaining life of the battery.For example, the aging parameter is embodied as a state-of-health (SoH). A physical model may include at least one mathematical equation to describe the aging process with the at least one sensor signal as input. For this purpose, the sensor signal can be represented by digital sensor data and provided as the input, for example by means of an analog-to-digital conversion of an analog measurement signal during the measurement. Applying the physical model can be understood to mean a parameterization of the physical model, wherein the parameterization can be, expressed in simplified terms, a filling of parameters of the physical model, for example equations of the physical model, with the measured values of the sensor signal. The extended data can also be digital data and the reference signal can also be represented by digital data. A sensor signal can describe, for example, a voltage, current or temperature of the battery. The machine learning model is preferably configured as a neural network, and more preferably, as a deep neural network. A deep neural network, also referred to as deep neural network (DNN), is in particular a special type of artificial neural network (ANN) that consists of or comprises multiple layers, in particular at least two layers, of neurons. An advantage of a deep neural network may be that it is able to learn complex and hierarchical features from the input data. Each layer of the network may be specialized for different types of information and generate progressively more abstract representations of the input. This allows the network to advantageously recognize and learn deep and complex patterns and relationships in the data. A battery parameter is, for example, the capacity of the battery or a parameter which correlates with the capacity of the battery. The term sequential may be used to mean that the physical model is first applied and then the machine learning model is applied. The interpolation may be a linear interpolation.For example, it may be provided that the physical model comprises at least one non-linear, usual differential equation which describes the aging process of the battery on the basis of a development of aging states of the battery, wherein the output of the physical model represents the aging states. An ageing state corresponds in particular to a specific value of the ageing parameter and / or of the at least one battery parameter at a specific point in time.Advantageously, it can be provided within the scope of the invention that the physical model describes at least one physical feature from the at least one sensor signal in order to provide the description of the at least one physical feature as the output of the physical model. The at least one physical feature can be specific to the aging parameter and / or the at least one battery parameter and predict these, for example, in the form of equations for specific points in time.It may be possible that the method further comprises:calculating at least one expert feature on the basis of the at least one sensor signal and the at least one physical feature,wherein the machine learning model is based on the training with the augmented data and the at least one expert feature. The expert features are in particular individually constructed by a person skilled in the art and can be based on empirical values, for example on empirical values with respect to the aging process of the battery.It may be further possible that the machine learning model comprises a transformer model. In particular, unlike other architectures that may be used for processing sequences, such as recurrent neural networks (RNNs), transformer architectures do not use loops and recursive operations to process the input sequence. Instead, they preferably use a mechanism called "self-attention" or "attention" to model and take into account the relationships between all elements of the input sequence. Specifically, self-attenuation is that each element of the input sequence is related to each other element or that each element is related to all preceding elements to determine which elements are most relevant to the calculation of the output value. This can be achieved by a matrix multiplication between the input vectors and an attention matrix calculated from the similarities between the input vectors. The attention matrix is preferably then used to calculate weighted sums of the input vectors, which are then used as input for the next step. The transformer architecture may consist of multiple layers of self-attention modules and fully connected layers that calculate the output of the model. By using self-attenuation, transformer architectures can take into account long dependencies in the input sequence and thus achieve better results in processing sequence data.According to an advantageous development of the invention, it can be provided that the application of the physical model further comprises the following step:performing a physics-based preprocessing of input data for the machine learning model, in particular based on physical properties of the battery, preferably for removing noise from input data for the machine learning model by the application of a filter based on characteristics of the noise and physical properties of the battery, wherein the input data digitally represent the sensor signal.Physics-based preprocessing (engl. Physics-based preprocessing) may be a machine learning approach in which physical principles and models are used to process data before further processing or modelling. This approach is used in particular when the underlying data are characterized by physical laws or relationships. Physics-based preprocessing may use the existing physical knowledge to clean, transform, or interpret the data before feeding it to a machine learning process. Within the scope of the present invention, an advantage of physics-based preprocessing can be that long-term effects are better mapped. This may help reduce the influence of disturbances, noise or unwanted effects and to enhance the relevant information. An example of physics-based preprocessing is the removal of noise from data by the application of a suitable filter based on the characteristics of the noise and the physical properties of the system that generates the data. Another example is normalization of data to adjust units or scales based on known physical constants or relationships. By integrating the physical knowledge into the physics-based preprocessing, an improved data quality and a better understanding of the underlying phenomena can be achieved, which can lead to better results and models in machine learning.The invention likewise relates to a training method for training a machine learning model for providing at least one aging parameter of a battery, comprising:providing training data, wherein the training data represent at least one sensor signal and comprise reference data, wherein the reference data are extended data which result from interpolation between individual measurement points of a reference signal, wherein the reference signal is specific to the at least one aging parameter, wherein the at least one sensor signal results from a measurement at the battery,training the machine learning model on the basis of the provided training data.Thus, the training method according to the invention brings with it the same advantages as have been described in detail with reference to the method according to the invention.It is possible for the method according to the invention to be used in an at least partially autonomous robot or vehicle. The vehicle can be designed, for example, as a motor vehicle and / or passenger motor vehicle and / or at least semi-autonomous or autonomous vehicle. The vehicle may have a vehicle device, for example for providing an autonomous driving function and / or a driver assistance system. The vehicle device can be designed to at least partially automatically control and / or accelerate and / or brake and / or steer the vehicle.The invention likewise relates to a computer program, in particular a computer program product, comprising instructions which, when the computer program is executed by a computer, cause the computer program to execute the method according to the invention. Thus, the computer program according to the invention brings with it the same advantages as have been described in detail with reference to a method according to the invention.The invention likewise relates to a device for data processing which is set up to carry out the method according to the invention. The device can be, for example, a computer which executes the computer program according to the invention. The computer may include at least one processor for executing the computer program. A non-volatile data memory can also be provided, in which the computer program can be stored and from which the computer program can be read out by the processor for execution.The invention can likewise be a computer-readable storage medium which has the computer program according to the invention and / or comprises instructions which, when executed by a computer, cause the computer to execute the method according to the invention. The storage medium is embodied, for example, as a data memory such as a hard disk and / or a nonvolatile memory and / or a memory card. The storage medium may be integrated into the computer, for example.In addition, the method according to the invention can also be embodied as a computer-implemented method.Further advantages, features and details of the invention will become apparent from the following description, in which exemplary embodiments of the invention are described in detail with reference to the drawings. The features mentioned in the claims and in the description can be essential to the invention individually or in any combination. The following are shown: FIG. 1 shows a schematic visualization of a battery, a method, a device, a storage medium and a computer program according to exemplary embodiments of the invention, FIG. 2 shows a schematic illustration of a training method according to exemplary embodiments of the invention, FIG. 3 shows a schematic illustration of a hybrid modeling with a physical model and a data-based model according to exemplary embodiments of the invention, FIG. 4 shows a schematic illustration of a transformer model according to exemplary embodiments of the invention.FIG. 1 schematically illustrates a battery 1, a method 100, an apparatus 10, a storage medium 15 and a computer program 20 according to exemplary embodiments of the invention.According to one embodiment of the method 100 according to the invention, in a first step 101, a physical model 3 based on at least one sensor signal is applied in order to describe an aging process of a battery 1, wherein the at least one sensor signal results from a measurement at the battery 1. In a second step 102, a machine learning model 2 is applied for determining at least one battery parameter of the battery 1, wherein the battery parameter is determined at least on the basis of an output of the physical model 3, and wherein the machine learning model 2 is based on training with extended data which results from interpolation between individual measurement points of a reference signal for data extension, wherein the reference signal is specific for the at least one aging parameter.In a third step 103 of the at least one aging parameter of the battery 1 on the basis of the determined battery parameter. The physical model 3 and the machine learning model 2 may be sequentially applied.FIG. 2 shows a training method 200 according to exemplary embodiments of the invention. In a first step 201, training data are provided, wherein the training data comprise at least one sensor signal and reference data, wherein the reference data are extended data which result from interpolation between individual measurement points of a reference signal, wherein the reference signal is specific to the at least one aging parameter, wherein the at least one sensor signal results from a measurement at the battery 1. In a second step 202, the machine learning model 2 is trained on the basis of the provided training data.According to embodiments of the invention, hybrid modeling (HYM) may be used to combine a physical model 3 with a data-based model 2, wherein the data-based model 2 may be a machine learning model 2. Preferably, the HYM design pattern "physics-based preprocessing" according to FIG. 3 is applied, i.e. the physical model 3, which is marked as P in FIG. 3, is used in particular first for extraction of features from the raw input. A data-based model 2, labeled D in FIG. 3, may then be trained on the combined raw and preprocessed input. As the data-based model 2, a deep neural network 2 is preferably used, and a transformer model is further preferably used (FIG. 4 ). Furthermore, a simple but effective strategy for data expansion can be used, for example by linear interpolation.In comparison to known hybrid models, the new formulation in particular (i) does not use a delta / residual approach for combining physics and data and (ii) replaces the stationary Gaussian process with a more powerful transformer model. Both innovations preferably add more weight to the data and thus can reduce model assumptions. The approach works particularly well if sufficient data is provided to the transformer model, which can be ensured by the newly introduced data expansion scheme.Outputs within the framework of training the machine learning model, or the deep neural network, can be battery parameters, in particular capacitance values, which can also be determined by carrying out reference power tests or by estimating the capacitance on the basis of the stored data.Inputs are, in particular, field data (e.g., temperature, current, voltage, SOC) and may be provided by the battery management system (BMS) of electric vehicles connected to the cloud. Preprocessing steps (e.g., data garbage collection) may be performed in the cloud before applying an algorithm according to an embodiment of the present invention.The machine learning (ML) model may be configured to determine a regression result.An embodiment of the present invention preferably relates to a method for training an ML model, which method can improve the performance of the ML model of an existing hybrid model. The method comprises intelligent reuse of existing data by (i) combining a physical model 3 with transformers and (ii) using an intelligent data expansion strategy.The method according to exemplary embodiments of the invention is based, for example, on the following aspects:Use of an existing physical model 3. in short, the physical model 3 preferably models the aging process of batteries, which may be Li-ion cells within the scope of the present invention. Model 3 may be computationally efficient and may be controlled by measurable inputs such as cell voltage, current, and temperature. It consists in particular of a series of non-linear, usual differential equations (ODE) which describe the development of the ageing states. In an emission model, the aging states are preferably coupled by a series of output equations to predict the estimated capacity.combining the physical model 3 with a data-based model 2, wherein the data-based model 2 is preferably a machine learning model and more preferably a deep neural network 2, using a hybrid modelling design pattern ("physics-based preprocessing"; cf. FIG. 3 ). This design pattern may sequentially apply the physical model 3 and the data-based model 2 therein. Some parts of the data are preferably initially preprocessed by the physical model 3, which are then fed (together with the original data) as input into the data-based model 2.Use of a transformer architecture as data-based model 2.expanding training data by linear interpolation between the measurement points.Model Formulation: It can be assumed that the ageing dynamics can be modeled by the differential function ẋ=f(x(t ), u(t)), where x(t) ∈ R Dx represents the state of the usual differential equation (ODE) system, u(t) ∈ R Du is the external inputs at time t, and f: R Dx+Du → R Dx is the temporal differential function which controls the dynamics development. The state solution at any time t n may be characterized by the initial value at time t 0 and the following differential function:For a particular observation time, the output can then be calculated from the latent states as follows: where G: R K(Dx+Du) → R Dy is in particular the emission function that takes into account the latent states and inputs of the last K time steps.The physical model 3 can be interpreted as such a model, whereinThe transition and emission functions are determined in particular by physics. The long-term effects are preferably incorporated into the model via the ODE states, while the emission model takes into account only the latent state of the current point in time.In [2], where the references are given in brackets at the end of the description, the corresponding model family may be referred to as neural ODEs when f and G are parameterized by neural networks.The weights of the neural networks are estimated in particular only from data during training in the prior art. According to advantageous embodiments of the invention, a hybrid model is proposed in which the function f is taken over from physics, while a neural network is used as the emission model G.This approach may have several advantages compared to purely data- or physics-based approaches. Compared to the physical model 3, the emission model may be more meaningful, whereby the accuracy of the predictions may be increased. Training can be simplified compared to standard neural ODEs, since the long-term effects are already detected in the ODE states in particular. As a result, it may be sufficient during training to take into account only discrete sub-trajectories of the size K. Compared to standard neuralODEs, learning of only the emission model may also be interesting from the point of view of validation and verification, since the learned part cannot "spoil" the inherent properties of the ODE model (e.g. stability).The description of the invention is presented in three parts below. In the first part, the scheme for data expansion according to an embodiment will be explained. In the second part, the design pattern of physics-based preprocessing ("physics-based preprocessing") according to one exemplary embodiment is described. In the third part, the transformer architecture used according to one exemplary embodiment will be discussed in more detail.Capacitance measurements may only be sparsely sampled, while the measurement inputs and the outputs of the physical model may be densely sampled. The capacitance measurements can be extended according to an advantageous embodiment of the invention by linear interpolation between the observations. In an example dataset, this augmentation scheme may increase the number of data points from 192 to 1.2 million, for example, which enables the later use of deep learning technologies, in particular.The physical model 3 and the data-based model 2 may be sequentially applied. Some parts of the data are preferably initially preprocessed by the physical model 3, which can then be fed as input into the data-based model 2.A capacity of a Li-Ion battery can be estimated according to a possible embodiment of the present invention as follows on the basis of steps 1 to 4: 1. using a model based on first principles to extract features from the raw input. Here, the model based on first principles corresponds in particular to the physical model 3.2. collecting data to train D: storing both the raw data u_t and the result of the physical preprocessing p(t n) = [ x P( t n), y P( t n)], where x P( t n) are the latent states and y P( t n) are preferably the outputs of the physical model at the time t n. The raw data preferably correspond to the sensor signals. 3. constructing a set of expert features which can be calculated from the input signals and the physical features. 4. In the following, U=(u(t n),..., u(t n-K)) denotes the sequence of raw inputs for the preceding K time steps. Analogously, the vectors P=(p(t n),..., p(t n-K)) and E=(e(t n),..., e(t n-K)) may be introduced. In one step, in particular, the capacity is approximated by learning a data-based model D(U,P,E) on the combined raw and processed inputs. D(U,P,E) can be a slight extension of the emission model G D in which the results of the physical model 3 and expert features can be used as additional inputs.In an advantageous embodiment, the data-based model 2 consists of a transformer architecture, which is described in particular below.The transformer preferably acts on time-series inputs, i.e. for an input sequence 4 Z=(z(t n),..., z(t n-K)) with z(t n-k)=( u(t n-k), p(t n-k), e(t n-k)) the task can be to predict the target sequence 7 Y=(y(t n),..., y(t n-K)). The input z(t n-k) is preferably an F-dimensional vector 5. it comprises in particular measured input signals 51, outputs of the physical models 52, expert features 53 which can be constructed by a person skilled in the art, and the position coding 6 which uses the sequence order.The data are recorded, for example, at irregular intervals which can be predetermined by the expert's people.A neural network (NN) may be used to predict the capacity:The NN consists in a possible embodiment of two layers: in the first layer temporal features are preferably extracted from the input sequence with the aid of L=3 parallel transformer blocks 8. The second layer consists in particular of a simple linear transformation in order to infer the output space from the feature space. See FIG. 4 for a representation of the model.For the transformer block 8, for example, the standard implementation of pytorch "TransformerEncoderLayer", which is based on [1], can be used. The main constituent is in particular the (self) etching mechanism which is discussed below.First, preferably take the input sequence Z and apply three linear projections thereto, e.g., Q=ZW Q( queries), K=ZW K( keys), V=ZW V( values). The projection matrices W Q, W K and W V have the size FxF, for example, and can be learned during training. In addition, a mask M of size KxK may be defined as follows: M[i,j]=0 if i<=j and M[i,j]=∞ otherwise. The task of the mask is to prevent look-ahead in time, i.e., to ensure that only information from z(t n-K),..., z(t n-k) can be used to predict y(t n-k).Next, an attention block is specifically applied:The results of the attention block are in particular a weighted sum of the values, wherein the weights can be calculated on the basis of the similarity between the queries Q and the keys K. The mask M sets the weight [i,j] preferably to zero when j>i. By learning a different set of weights for each parallel block, different temporal features can be extracted.During training, the data are in particular broken down into pieces with a maximum length of K=200. The weights of the network can be trained by minimizing the mean square error between the output of the neural network and the target value.To predict the capacity, the architecture described above may be used. Since it is assumed in particular that the capacitance changes only slowly, the output can additionally be smoothed 11 (see FIG. 4 ), for example over the last N=200 steps, in order to provide the prediction 12 in a smoothed manner. In particular, a loss 13 and a target 14 are determined.The foregoing explanation of the embodiments describes the present invention solely by way of examples. Of course, individual features of the embodiments can be freely combined with one another, insofar as technically expedient, without departing from the scope of the present invention.[1] Ash Vaswani et al, Attention is all you need, Neural Information Processing Systems, 2018, https: / / arxiv.org / abs / 1706.03762.[2] Chen et al., Neural Ordinal Differential Equations, Neural Information Processing Systems, 2018, https: / / arxiv.org / pdf / 1806.07366.pdf.

Claims

Method (100) for providing at least one ageing parameter of a battery (1), wherein the method (100) comprises the following steps: - applying (101) a physical model (3) on the basis of at least one sensor signal in order to describe an ageing process of the battery (1), wherein the at least one sensor signal results from a measurement at the battery (1), - applying (102) a machine learning model (2) for determining at least one battery parameter of the battery (1), wherein the battery parameter is determined at least on the basis of an output of the physical model (3), and wherein the machine learning model (2) is based on training with extended data which results from an interpolation between individual measurement points of a reference signal for data extension, wherein the reference signal is specific to the at least one ageing parameter, providing (103) the at least one aging parameter of the battery (1) on the basis of the determined battery parameter, wherein the physical model (3) and the machine learning model (2) are applied sequentially and characterized in that the machine learning model (2) comprises a transformer model.Method (100) according to claim 1, characterized in that the physical model (3) comprises at least one non-linear ordinary differential equation describing the aging process of the battery (1) based on a development of aging conditions of the battery (1), wherein the output of the physical model (3) represents the aging conditions.Method (100) according to one of the preceding claims, characterized in that the physical model (3) describes at least one physical feature from the at least one sensor signal in order to provide the description of the at least one physical feature as the output of the physical model (3).The method (100) according to claim 3, characterized in that the method (100) further comprises: - calculating at least one expert feature based on the at least one sensor signal and the at least one physical feature, wherein the machine learning model (2) is based on the training with the extended data and the at least one expert feature.The method (100) according to any of the preceding claims, characterized in that applying the physical model further comprises the step of: - performing a physics-based pre-processing of input data for the machine learning model (2), in particular based on physical properties of the battery (1), preferably for removing noise from input data for the machine learning model (2) by applying a filter based on characteristics of the noise and physical properties of the battery (1), wherein the input data digitally represent the sensor signal.Training method (200) for training a machine learning model (2) for providing at least one ageing parameter of a battery (1), characterized in that the machine learning model (2) comprises a transformer model, comprising: - providing (201) training data, wherein the training data represent at least one sensor signal and comprise reference data, wherein the reference data are extended data resulting from an interpolation between individual measurement points of a reference signal, wherein the reference signal is specific for the at least one ageing parameter, wherein the at least one sensor signal results from a measurement at the battery (1), - training (202) the machine learning model (2) on the basis of the provided training data.A computer program (20) comprising instructions which, when the computer program (20) is executed by a computer (10), cause the computer program to execute the method (100) according to any one of the preceding claims.Data processing device (10) configured to carry out the method (100) or training method (200) according to any one of claims 1 to 6.A computer readable storage medium (15) comprising instructions which, when executed by a computer (10), cause the computer to carry out the steps of the method (100) or training method (200) according to any one of claims 1 to 6.

Images