Method for classifying the suitability of an automotive battery

Abstract

The invention relates to a method for classifying the suitability of an automotive battery for further use in a non-automotive application, wherein in a first method step an input vector for behavioral pattern models is aggregated from measurement points of the automotive battery; in a second method step an output vector is provided by means of a first behavioral pattern model based on the aggregated input vector and a learned behavior of the automotive battery; in a third method step a deviation is determined from a comparison of the input vector and the output vector; and in a fourth method step a suitability of the automotive battery for further use in a non-automotive application is classified by means of a second behavioral pattern model trained with categorized data based on the input vector and the deviation.

Description

State of the art:

[0001] The invention relates to a method for classifying the suitability of an automotive battery according to the preamble of the independent claim.

[0002] Lithium-ion batteries are increasingly used in motor vehicles. These automotive batteries age both over time, known as calendar aging, and through the stresses of charging and discharging, known as cyclic aging. This can lead to, for example, a loss of capacity and / or an increase in internal resistance. Currently, for automotive batteries, i.e., those used in motor vehicles, the end of their service life, or EOL, is defined as typically being reached when the remaining capacity is between 70% and 80% or the internal resistance reaches approximately 200%.

[0003] While such a battery would then be considered unsuitable for a motor vehicle, it could potentially still be used in less demanding applications, for example as a stationary energy storage device or similar.

[0004] Therefore, at the end of its service life in the motor vehicle, a decision must be made as to whether the battery in question is still suitable for further use in a non-automotive application, a so-called second-life application (or SLA for short), or whether it can be directly recycled mechanically and / or chemically.

[0005] For modular batteries, this assessment can also be performed at the module level. Not every battery or module is suitable for reuse in a non-automotive application. In such cases, even a single unsuitable battery cell can disqualify the entire module or pack, as individual battery cells are not usually replaceable.

[0006] Possible causes include, for example, unusual behavior that could indicate a safety problem or a very steep aging gradient, a so-called Sudden Death, which would only allow for very short continued use.

[0007] When an automotive battery at the end of its service life needs to be assessed for its suitability for reuse, it or its modules are typically evaluated using a measurement program. For this purpose, the battery is removed from the vehicle, possibly disassembled further into modules, and connected to a suitable tester.

[0008] Subsequently, measurement procedures such as capacity tests, pulse tests, or electrochemical impedance spectroscopy are performed. The evaluation of the measurement data allows for a relatively good assessment under standardized conditions and a classification of the batteries or modules.

[0009] However, this involves a considerable amount of work and equipment, making the overall process difficult to scale. Disclosure of the invention:

[0010] A method for classifying the suitability of an automotive battery for further use in a non-automotive application with the features of the independent claim offers the advantage that, by evaluating measurement points of an automotive battery, an assessment or classification of the battery with regard to further use in a non-automotive application can be achieved.

[0011] According to the invention, a method for classifying the suitability of an automotive battery for further use in a non-automotive application is provided.

[0012] In a first step of the process, an input vector for behavioral pattern models is aggregated from measurement points of the automotive battery.

[0013] Furthermore, in a second process step, an output vector is provided using a first behavioral pattern model based on the aggregated input vector and a learned behavior of the automotive battery. In a third process step, a deviation is determined by comparing the input vector and the output vector.

[0014] Furthermore, in a fourth process step, a second behavior pattern model trained with categorized data is used to assess the suitability of the automotive battery for further use in a non-automotive application based on the input vector and the deviation.

[0015] The measures listed in the dependent claims enable advantageous further developments and improvements of the device specified in the independent claim.

[0016] It should be noted that the first behavioral pattern model was learned using unsupervised learning based on a large number of different input vectors from various batteries in different vehicles. This first behavioral pattern model therefore independently recognizes patterns and correlations at the measurement points. Using a sufficiently large number of input vectors, particularly based on typical fleet data from the vehicle, the first behavioral pattern model can predict and learn normal or typical battery behavior even without knowledge of anomalies or errors.

[0017] Furthermore, it should be noted that the second behavioral pattern model was learned using supervised learning based on a large amount of categorized data. This second behavioral pattern model therefore utilizes categorized data. The categorized data includes an assignment to an input vector indicating whether or not the automotive battery is suitable for further use in a non-automotive application. Each input vector can thus be either assigned a suitability or denied one. The second behavioral pattern model is therefore a learning algorithm that is trained and validated with datasets that already contain a corresponding output value for each input, i.e., the suitability or unsuitability for further use for each input vector.

[0018] For example, the second behavioral pattern model could be a weighting function. This would allow suitable batteries to be separated from unsuitable ones based on a threshold value. The threshold value is trained using both suitable and unsuitable batteries.

[0019] Of course, other clustering algorithms can also be used, such as Support Vector Machines.

[0020] Preferably, the measurement points can be collected over the lifetime of the automotive battery. In particular, these can be recorded and transmitted via a telemetry data connection to the vehicle. The measurement points can include current, voltage, temperature, and / or state of charge.

[0021] It is advantageous if the input vector is a histogram of a dwell time over a state of charge of the battery and a temperature of the battery or a discharge rate of the battery, or if the input vector is selected as the occurrence and frequency of individual events, e.g. a charge equalization between battery cells, so-called balancing.

[0022] This allows an input vector to be aggregated from the collected measurement points, which were collected particularly over the lifetime of the automotive battery.

[0023] The first behavioral pattern module is expediently implemented as a neural network in an autoencoder configuration. Autoencoders essentially output a reconstruction of the input vector. These autoencoders consist of two smaller networks: an encoder and a decoder. During training, the encoder learns a set of features from the input data. These are referred to as the latent representation. Simultaneously, the decoder is trained to reconstruct the data based on these features. Subsequently, the autoencoder can be used to predict inputs that were previously unknown.

[0024] When using an autoencoder, the dimensions of the output vector correspond to those of the input vector. However, the output vector is merely a reconstruction of the input vector, and therefore the input and output vectors are not identical. In the third step of the process, the deviations between the input and output vectors are determined. If the deviation exceeds a threshold, an anomaly or, if applicable, an error is present.

[0025] The mere fact that an anomaly or defect exists does not necessarily mean that the suitability of the automotive battery for further use in a non-automotive application can be definitively assessed.

[0026] A rating can now be generated, preferably from the difference between the respective input value and the corresponding output value, or their magnitude (i.e., the amount of the difference), as well as the frequency of the respective deviation and the affected data point. This rating serves to classify the significance of the anomaly and is referred to as the severity score.

[0027] Even based solely on the value determined in this way, a suitability for further use in a non-automotive application cannot yet be conclusively assessed.

[0028] In the fourth step of the process, categorized data is used to train the second behavioral pattern model.

[0029] Preferably, the categorized data is determined based on a capacity and / or a resistance.

[0030] Preferably, the capacity and / or resistance of an automotive battery is measured, and then the respective input vector associated with this battery, aggregated from individual measurement points, is assigned a suitability for further use in a non-automotive application.

[0031] In other words, a specific input vector is linked to the information about whether the associated battery is suitable for further use in a non-automotive application or not.

[0032] In particular, the presented method offers the overall advantage that measured laboratory and field data from some batteries exhibiting problematic processes, such as sudden death or thermal runaway, can be used to train supervised learning for classifying the suitability of an automotive battery for further use in a non-automotive application.

[0033] Furthermore, such an assessment can be carried out at any time without any loss of time.

[0034] Overall, it should be noted that while some initial effort is required to configure the second behavioral pattern model, the presented solution can subsequently be scaled easily and cost-effectively. In particular, dedicated measurements at the end of the automotive batteries' service life can be dispensed with. Brief description of the drawings:

[0035] Exemplary embodiments of the invention are shown in the drawings and explained in more detail in the following description. It shows Fig. 1 a method for classifying the suitability of an automotive battery for further use in a non-automotive application.

[0036] In a first process step, an input vector for behavioral pattern models is aggregated from measurement points of the automotive battery. The measurement points are collected over the lifetime of the automotive battery and include current, voltage, temperature, and / or state of charge. The input vector is selected as a histogram of dwell time versus battery state of charge, battery temperature, and / or battery discharge rate, or it is selected as the occurrence and frequency of individual events, e.g., an equal charge between battery cells, over the lifetime.

[0037] In a second process step, an output vector is provided using a first behavioral pattern model based on the aggregated input vector and a learned behavior of the automotive battery. The first behavioral pattern model is implemented as a neural network in an autoencoder configuration.

[0038] In a third step, a deviation is determined by comparing the input vector and the output vector. In this third step, a score is generated for the deviation based on its magnitude, frequency, and the affected data point. This score is then used in the fourth step by the second behavioral pattern model.

[0039] In a fourth process step, a second behavioral pattern model, trained with categorized data, is used to classify the suitability of the automotive battery for further use in a non-automotive application based on the input vector and the deviation. The categorized data is generated based on capacitance and / or resistance. The categorized data includes an input vector to which a suitability for further use in a non-automotive application is assigned.

[0040] Fig. Figure 2 shows one possible classification method, where, for example, the determined severity score is plotted against the number of individual events. A weighting can then be used to differentiate between suitability for reuse in non-automotive applications, such as second-life applications, and recycling.

Claims

[1] Method for classifying the suitability of an automotive battery for further use in a non-automotive application, wherein in a first method step an input vector is derived from measuring points of the automotive battery is aggregated for behavioral pattern models; in a second process step by means of a first behavioral pattern model based on the aggregated input vector and An output vector is provided for a learned behavior of the automotive battery; In a third process step, a deviation is determined from a comparison of the input vector and the output vector; in a fourth process step by means of a trained with categorized data based on the second behavioral pattern model on the input vector and the deviation suitability of the automotive battery is classified for further use in a non-automotive application. [2] Method according to the previous claim 1, characterized by that the measurement points are collected over the lifetime of the automotive battery and include current, voltage, temperature and / or state of charge. [3] Method according to one of the preceding claims 1 or 2, characterized by , that the input vector is selected as a histogram of a dwell time over a state of charge of the battery, a temperature of the battery and / or a discharge rate of the battery, or that the input vector is selected as the occurrence and frequency of individual events, e.g. an equal charge between battery cells, over lifetime. [4] Method according to any one of the preceding claims 1 to 3, characterized by , that the first behavioral pattern model is designed as a neural network in autoencoder configuration. [5] Method according to any one of the preceding claims 1 to 4, characterized by , that in the third procedural step for the deviation, an evaluation value is further created on the basis of an amount and a frequency of the deviation as well as an affected data point, which is also used in the fourth procedural step by the second behavior pattern model. [6] Method according to any one of the preceding claims 1 to 5, characterized by that the categorized data is created based on a capacity and / or a resistance. [7] Method according to any one of the preceding claims 1 to 6, characterized by , that the categorized data includes an input vector to which a suitability for further use in a non-automotive application is assigned.

Images