DE102021213762A1 - Method for creating data sets for evaluating at least one lifetime parameter of a battery cell, device and software therefor. - Google Patents

Abstract

The invention relates to a method for creating data sets for evaluating at least one service life parameter of a battery cell, comprising the steps: ◯ Obtaining (100) a first data set, ◯ Determining (200) parameters for the synthetic data generation, ▪ a first parameter of a shift of measured values of the data set relative to a comparable data set, ▪ where a second parameter corresponds to a stretching of the measured values of the data set relative to the comparable data set, ▪ where a third parameter corresponds to an increase in the data set relative to the comparable data set, ◯ by means of the determined adjustment parameters and the comparable data set creating (300) virtual data values in addition to the first data set◯ providing (400) the created data set to an evaluation device for at least one service life parameter of the battery cell.Furthermore, the invention relates to a device for carrying out one of the methods and software for Programming a data processing system to carry out one of the methods.

Description

The invention relates to a method for creating data sets for evaluating at least one service life parameter of a battery cell, as well as a device and software for this.

background

Lithium-ion batteries are widely used in many areas of the economy, but especially in electromobility and energy storage, as they have a high gravimetric energy density and long service life. In addition, production costs are reduced. Such batteries are also referred to as battery cells.

However, like any electrochemical system, battery cells are subject to degradation during repeated charge/discharge cycles and storage, which becomes noticeable in the form of capacity and performance losses. Accurately predicting the lifetime of battery cells is therefore an important area of research.

Degradation insights enable OEMs, EV manufacturers and battery system designers to provide calculated battery cell warranties and ensure high performance and safety over the projected lifetime, while end users and battery system operators know in advance about can be informed when the batteries will be replaced.

However, reliable and accurate lifetime prediction remains a challenging task due to the nonlinear nature of such systems, which is due to internal responses and individual cell variability.

Several studies have proposed electrochemical models to characterize degradation mechanisms such as active material loss, solid-electrolyte interface growth, and lithium segregation. In principle, these models were able to demonstrate their ability to predict, but they have only been shown to be of limited use in capturing non-linear, application-related degradation trends and in parameterizing a wide range of degradation mechanisms. In addition, the electrochemical models are unable to account for the intrinsic differences of the cells during manufacture and therefore have limited accuracy in cell-specific predictions. In addition, the electrochemical models require the solving of complex algebraic differential equations, which is very computationally intensive for the existing practical applications.

In the meantime, semi-empirical models have also been proposed, in which a parameterization of cell aging data is derived and relationships between capacity and degradation and various stress factors such as temperature, depth of discharge and current intensity are obtained as empirical equations. However, these semi-/empirical models are highly dependent on the aging data, the test conditions and the empirical equations obtained. Therefore, the predictions for an unknown battery cell are less accurate, especially if its degradation path is outside the bounds defined by the obtained empirical equations.

Both the approach based on the electrochemical model and the approach based on the semi-/empirical model require thorough knowledge of mathematics and electrochemistry.

Against this background, other solutions have been sought in the past.

The availability of increasing amounts of data in the field and laboratory, advances in data processing and analysis, and the development of machine learning and deep learning models have paved the way for the use of data-driven models.

Instead of using degradation mechanisms, these methods are based on learning from battery data. In studies, machine learning models such as linear regression, Gaussian process regression and Bayesian approaches, as well as support and relevance vector machines were used. However, these learning models require feature engineering to extract features from the data, resulting in other important features being excluded (undetected) or even including features that may be detrimental to model performance. These models also allow only limited tuning of hyperparameters to optimize performance. In most cases, some of the feature generation inputs for these models are not directly available in existing field applications.

Deep learning models, e.g. Multi-layer neural networks, for example, can potentially overcome these hurdles. They provide a wider range of hyperparameters for model optimization, create len their own traits while learning from data, and therefore require less human intervention. Various deep learning models such as long short memory (LSTM), bidirectional LSTM, deep neural networks and hybrid models, combinations of different neural networks or deep and machine learning algorithms have also been developed by the applicant for lifetime prediction suggested.

However, data-driven models are hampered by insufficient data and poor data quality.

The main problems in creating datasets for degradation prediction models are insufficient number of aging scenarios and insufficient number of battery cells tested per aging scenario.

The rapid advances in battery cells and the discovery of new battery materials could lead to a possible redundancy of technologies that are only a few generations old. Therefore, investigating the performance characteristics of newly developed battery cell materials and the simultaneous development of data-based prediction models is of the highest priority.

In addition, cell testing in laboratories is resource, time and space intensive while at the same time there is an urgent need for research groups and OEMs to accelerate cell aging testing at lower cost.

The generation of synthetic battery data is a possible solution that has already been pursued in several studies. Markov chains were used in combination with a neural network structure (see "Generation of Synthetic Battery Data with Capacity Variation" by authors M. Pyne, B.J. Yurkovich, and S. Yurkovich, in IEEE CCTA 2019: 3rd IEEE Conference on Control Technology and Applications , pages 476-480, doi https://doi.org/10.1109/CCTA.2019.8920488) to generate synthetic midpoints of voltage curves as a function of capacitance values, without reference to an application in cell lifetime prediction.

A mechanistic framework based on electrochemical properties of battery cells was presented in the article " Benchmark Synthetic Training Data for Artificial Intelligence-based Li-ion Diagnosis and Prognosis” by authors M. Dubarry and D. Beck, DOI: 10.1016/j.jpowsour.2020.228806 , as well as the article " Big data training data for artificial intelligence-based Li-ion diagnosis and prognosis”, https://doi.org/10.1016/j.jpowsour.2020.228806 , which can parameterize the loss of active material and lithium stock to generate synthetic voltage and corresponding capacitance curves. However, this approach is extremely complex and offers only limited control over the generation process. Obtaining synthetic data requires a thorough analysis of a large number and combination of factors that may or may not be representative of the degradation pathways of a given battery cell, hence this approach requires further refinement and research. Repeating this method on different cell types would also be costly.

From the article " Generating Synthetic Time Series to Augment Sparse Datasets,” authors G Forestier, F Petitjain, H Anh Dau, GI Webb and E Keogh, published in in 17th IEEE International Conference on Data Mining: 18-21 November 2017, New Orleans, Louisiana : proceedings, New Orleans, LA, 2017, pages 865-870 a dynamic averaging approach is known which attempts to derive synthetic data based on weighted averages of existing real time series data. However, due to the different lifespans of lithium-ion cells, this method is not applicable. Furthermore, due to the averaging, the synthetic data generated in this way would always be within the limits of the real data.

Previous approaches have not been satisfactory for safely synthetically creating data.

Task

It would therefore be desirable to be able to specify a method with which data can be safely synthesized.

Summary of the Invention

The invention solves this problem through the independent claims. Advantageous developments are the subject of the dependent claims, illustrations and the description.

The invention is explained in more detail below with reference to a drawing and exemplary embodiments. The drawing is a schematic representation and not to scale. The drawing does not limit the invention in any way.

• eine Übersicht des Ansatzes.
• schematisch die Erzeugung von Datensätzen mit individuellen Parametern.
• RWTH-ISEA Stanford/MIT Datensätze nebst erfindungsgemäß erstellten Datensätzen.
• Parameterverteilungen für die Batteriezellen gemäß der RWTH-ISEA Datensätze.
• Parameterverteilungen für die Batteriezellen gemäß der Stanford/MIT Datensätze.
• Model Performanz bei Zugabe einer festen Anzahl von virtuelle Datenwerten. 30 und 36 synthetische Batteriezellen wurden entsprechend zu den Trainingsdatensätzen für die RWTH-ISEA (a, c, e) und die Stanford/MIT (b, d, f) Datensätze hinzugefügt. Mit jeder weiteren Zeile nimmt die Zahl der Trainings-Batteriezellen zu.
• Fehler des CNN-Modells bzw. des GPR Models mit der Zunahme erstellter virtueller Datenwerten.
• Fehler des Kniepunktes mit Zunahme erstellter virtueller Datenwerten,
• Ersatz von realen Daten durch synthetische Daten für Vorhersage der Lebensdauer EOL für den RWTH-ISEA Datensatz (a, b) und den Stanford/MIT Datensatz (c, d). Die Anzahl der realen Degradationskurven is durch eine Ganzzahl ‚n‘ gekennzeichnet. Die Anzahl der realen Degradationskurven, respektive synthetischen Degradationskurven, innerhalb der Trainingsdaten ist durch ,n'/‚m‘ angegeben.
• Mittlerer Fehler für den Oxford Datensatz (a) und den NASA Datensatz (b).
• ein Ablaufdiagramm von Ausführungsformen der Erfindung.

Show it:

• an overview of the approach.
• schematic of the generation of data sets with individual parameters.
• RWTH-ISEA Stanford/MIT data sets together with data sets created according to the invention.
• Parameter distributions for the battery cells according to the RWTH-ISEA data sets.
• Parameter distributions for the battery cells according to Stanford/MIT datasets.
• Model performance when adding a fixed number of virtual data values. 30 and 36 synthetic battery cells were added to the training data sets for the RWTH-ISEA (a, c, e) and the Stanford/MIT (b, d, f) data sets, respectively. With each additional row, the number of training battery cells increases.
• Error of the CNN model or the GPR model with the increase of created virtual data values.
• error of knee point with increase of created virtual data values,
• Replacement of real data with synthetic data for lifetime EOL prediction for the RWTH-ISEA dataset (a, b) and the Stanford/MIT dataset (c, d). The number of real degradation curves is denoted by an integer 'n'. The number of real degradation curves or synthetic degradation curves within the training data is indicated by 'n'/'m'.
• Mean error for the Oxford data set (a) and the NASA data set (b).
• a flowchart of embodiments of the invention.

Detailed presentation of the invention

In the following, the invention is presented in more detail with reference to the figures. It should be noted that different aspects are described, which can be used individually or in combination. That is, each aspect can be used with different embodiments of the invention, unless explicitly presented as purely alternative.

Furthermore, for the sake of simplicity, only one entity is generally referred to below. Unless explicitly noted, however, the invention can also have several of the entities concerned. In this respect, the use of the words “a”, “an” and “an” is only to be understood as an indication that at least one entity is used in a simple embodiment.

Insofar as methods are described below, the individual steps of a method can be arranged and/or combined in any order, unless something different is explicitly stated in the context. Furthermore, the processes can be combined with one another, unless expressly stated otherwise.

Insofar as standards, specifications or the like are named in this application, at least the standards, specifications or the like applicable on the filing date are referred to. This means that if a standard/specification etc. is updated or replaced by a successor, the invention can also be applied to this.

Various embodiments are shown in the figures.

In a method according to the invention for creating data sets for evaluating at least one service life parameter of a battery cell, a first data set is obtained in a step 100 . This data set can be sparse, but it can also be a larger one or even be so complete in terms of usability for evaluating a service life parameter of a battery cell that it provides satisfactory results (at least for some areas). In this respect, there are no upper limits to "completion" of the data set. The data set can come from a (known) aging test for a battery cell, for example.

In relation to the data set obtained, parameters for the synthetic data generation are now determined in a step 200 . A first parameter (offset) corresponds to a displacement of measured values of the data set relative to a comparable data set, a second parameter (elongation) corresponds to an extension of the measured values of the data set relative to the comparable data set, and a third parameter (slope) corresponds to an increase in the dataset relative to the comparable dataset.

The method according to the invention can thus be described as based on the modification of the geometries of existing real battery data on the basis of certain properties. These properties are implemented as parameters that can be fed as inputs to the synthetic data generation function.

It makes sense to identify the distinctive characteristic features - the initial capacity value and the cycle numbers of EOL and knee point for the capacity decay curve of a battery cell. These characteristics will then used to determine the parameters of a synthetic curve generation function (SCG), namely offset, elongation and slope parameters.

The first parameter "offset" can be described as the difference between the initial capacity values of two cells in Ah. The second parameter "elongation" can be described as the ratio of the endpoints of two curves, and can be a dimensionless quantity. The third parameter "slope" can be described by the variances in the slopes of the capacity degradation curves within the initial linear degradation phase (i.e. before the knee point). The third parameter can be calculated by subtracting the differences in capacitance at the first cycle and at the 'n'th cycle, up to which slight differences in slopes are observed. The third parameter has the unit Ah.

shows the parameters and the individual influence of these parameters on the type of synthetic curves obtained, where the three fitting parameters (for the SCG functions) were determined separately for each data set. The resulting exemplary sets of these parameters were presented in the form of histogram distributions in shown.

Once the parameters have been determined, it is now possible in step 300 to start creating virtual data values in addition to the first data set using the determined adjustment parameters and the comparable data set.

On the one hand, the initial scenario can be changed by means of random values, so that (expectable real) data values can also be generated for this changed initial scenario using the now enlarged data set that has been created. Equally, of course, the originally first data set - if it was sparse - can be supplemented with (expected real) data values.

The created data set or the now created data sets can then be made available in step 400 to an evaluation device for at least one service life parameter of the battery cell.

A data-driven model approach can also be used for the determination 200 without loss of generality.

In particular, it can be provided in embodiments of the invention that a neural network is used for the determination 200 .

In embodiments of the invention, this neural network can also be a convolutional neural network.

According to one embodiment of the invention, a machine-learning-based model, in particular a deep-learning-based model, is used for the determination in step 200 .

In yet another embodiment of the invention, the determination in step 200 is made for different data sets.

According to yet another embodiment of the invention, the first data set obtained is sparse.

In yet another embodiment of the invention, the first data sets are obtained from an aging experiment.

According to a further embodiment of the invention, a device for carrying out one of the methods according to the invention is also proposed. A device can be a computer, for example, with the requirements for computing power and memory size being low, so that a conventional desktop computer can also carry out the method according to the invention in a comparatively short time (compared to the measurement in the aging test). However, the device can also be implemented in a cloud application or in specialized hardware.

According to a further embodiment of the invention, software for programming a data processing system for carrying out one of the methods is also proposed. Such software can be offered and distributed to a suitable data processing system (e.g. a device for carrying out one of the methods according to the invention) by means of a storage medium or a data stream.

The invention has advantages over the current state of the art.

For one, the virtual data values created capture the exact characteristics of the data rather than approximations. Therefore, the created virtual data values are just as valid as real data values.

In addition, the invention is based on geometry and statistics, which makes it easy to precisely control the creation of the virtual data values (synthetic data). The invention ensures that the virtual data values are a possible practical degradation path of the cell.

No knowledge of complex battery aging mechanisms is required for the data generation itself.

As a result, the expenditure of time and money can be significantly reduced, since virtual data values can be easily generated given known adjustment parameters. This reduces the time and cost of repeating the process to create virtual data values for other cells and other data sets.

It is also possible to generate synthetic data that goes well beyond the real data limits, helping to obtain data that represent extreme degradation scenarios. Enormous amounts of synthetic data can be generated in the order of a few seconds.

This invention thus allows reducing the time and cost of cell aging tests while providing aging tests while improving model performance.

In particular, the invention can also be used when only sparse capacity degradation data sets are available in order to enrich them and thus reduce the time and resource expenditure for cell aging tests.

The effectiveness of this method could be validated using a dynamic deep learning convolutional neural network (CNN)-based prediction model.

To demonstrate the ability of our model to deal with the variability of the inputs, 4 datasets with different cell numbers and different degradation characteristics were used.

In errors of the CNN model and the GPR model are shown with the increase in created virtual data values in relation to the RWTH-ISEA dataset. The training data subset for a has 3 real cell capacity degradation curves, for b 15 real cell capacity degradation curves and for c 30 real cell capacity degradation curves. "N" denotes the ratio of the number of synthetic curves (virtual data sets) to the number of real cell capacity degradation curves within the training data set for the respective case.

In Knee point error shown with increase in generated virtual data values relative to the Stanford data set. The training data subset for a has 9 real cell capacity degradation curves, for b 18 real cell capacity degradation curves and for c 36 real cell capacity degradation curves. N'' denotes the ratio of the number of synthetic curves (virtual data sets) to the number of real cell capacity degradation curves within the training data set for the respective case.

ie in as well as 5 it can be observed that adding virtual data to the training data with fewer real datasets in show a huge improvement in performance compared to a high number of cells.

This underscores the effectiveness of the invention for generating synthetic data and combining it with training data that has few/minimal sets of data.

This highlights the effectiveness of our method of synthetic data generation and combination with training data having minimum data.

The results obtained for Gaussian Process Regression (GPR) show similar tendencies. The results for the GPR show similar trends as observed in the previous set of experiments for both the RWTH-ISEA and the Stanford/MIT datasets, demonstrating the combined effectiveness of the deep learning CNN model and our method for Generating synthetic data.

For both the RWTH-ISEA and the Stanford/MIT data sets and in the majority of the cases, it can be observed that the performance of the invention (cf 6 ) with a combination of synthetic and real data is either better or as good as with purely real data. In this invention, a method for generating synthetic data based on geometry and statistics is proposed, which can generate enormous amounts of data based on limited battery data. Existing sparse data sets with very little cell data or uneven data distribution can be enriched. Any time series measurement data of the cell whose synthetic data is to be generated can serve as input.

The generation of synthetic capacity data allows to increase the prediction accuracy of the model for battery lifetime.

The raw data can be used directly to generate synthetic data that exhibits the characteristics of the dataset. This can be done by a simple geometric manipulation based on the existing real world data certain defined parameters can be accomplished.

The method can be tailored to produce precise data or a set of randomly generated data, depending on the needs of different applications.

The obtained data can then be added to the real data to create e.g. training sets for predictive models.

The proposed method has several advantages over the current state of the art.

The generated synthetic data captures the exact characteristics of the data rather than approximations. Therefore, the generated data is just as valid as real data.

The proposed method is based on geometry and statistics, which makes it easier to precisely control the nature of the synthetic data.

The proposed method ensures that the synthetic data is a possible practical degradation path of the cell.

No knowledge of complex ones is required for the data generation itself.

This reduces the time and cost of repeating the proposed method to obtain synthetic data for different cells and datasets.

It is also possible to generate synthetic data that go (far) beyond the limits of real data, making it possible to obtain data that can represent extreme degradation scenarios.

Using the proposed method, it is possible to generate large amounts of synthetic data in the order of a few seconds.

The proposed method was also validated on several datasets representing potential application conditions, serving as evidence for improving the accuracy of the predictive models.

The proposed method allows to reduce the time and cost of cell aging tests while providing aging tests while improving model performance.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Non-patent Literature Cited

• Benchmark Synthetic Training Data for Artificial Intelligence-based Li-ion Diagnosis and Prognosis” by the authors M. Dubarry and D. Beck, DOI: 10.1016/j.jpowsour.2020.228806 [0018]
• Big data training data for artificial intelligence-based Li-ion diagnosis and prognosis”, https://doi.org/10.1016/j.jpowsour.2020.228806 [0018]
• Generating Synthetic Time Series to Augment Sparse Datasets,” authors G Forestier, F Petitjain, H Anh Dau, GI Webb, and E Keogh, published in 17th IEEE International Conference on Data Mining: 18-21 November 2017, New Orleans, Louisiana : proceedings, New Orleans, LA, 2017, pages 865-870 [0019]

Claims (10)

Method for creating data sets for evaluating at least one service life parameter of a battery cell, comprising the steps: ◯ obtaining (100) a first record, ◯ determining (200) parameters for synthetic data generation, ▪ where a first parameter corresponds to a shift of measured values of the data set relative to a comparable data set, ▪ where a second parameter corresponds to a stretching of the measured values of the dataset relative to the comparable dataset, ▪ where a third parameter corresponds to a slope of the dataset relative to the comparable dataset, ◯ using the determined adjustment parameters and the comparable data set, creating (300) virtual data values in addition to the first data set, ◯ Providing (400) the data set created to an evaluation device for at least one service life parameter of the battery cell. procedure after claim 1 , characterized in that for the determination (200) a data-driven-model approach is used. Method according to one of the preceding claims, characterized in that a neural network is used for the determination (200). Method according to one of the preceding claims, characterized in that a convolutional neural network is used for the determination (200). Method according to one of the preceding claims, characterized in that a machine-learning-based model, in particular a deep-learning-based model, is used for the determination (200). Method according to one of the preceding claims, characterized in that determination (200) is carried out for different data sets. Method according to one of the preceding claims, characterized in that the data set obtained is sparsely populated. Method according to one of the preceding claims, characterized in that the first data sets are obtained from an aging experiment. Device for carrying out one of the methods according to one of the preceding claims. Software for programming a data processing system for carrying out one of the methods according to one of the preceding Claims 1 until 8th .

Images