Computer-implemented method for detecting anomalous charging cycles of a rechargeable battery

A multilayer autoencoder method normalizes and compresses battery cell voltage data to detect abnormal charging cycles, enhancing battery performance by identifying and correcting anomalies.

DE102022134464B4Active Publication Date: 2024-11-14DR ING H C F PORSCHE AG
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
DE102022134464
Authority / Receiving Office
DE · DE
Patent Type
Patents
Current Assignee / Owner
Filing Date
2022-12-22
Publication Date
2024-11-14
Estimated Expiration
2042-12-22

AI Technical Summary

Technical Problem

Existing methods fail to reliably detect abnormal charging cycles of rechargeable batteries, which can negatively impact battery performance.

Method used

A computer-implemented method using a multilayer autoencoder to analyze time series of battery cell voltages, normalizing and compressing data to identify anomalies, and calculating an anomaly index through denormalization and comparison with a threshold.

Benefits of technology

Enables simple and effective detection of abnormal charging cycles, improving battery performance by identifying and addressing potential issues.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000000_0000_ABST
    Figure 00000000_0000_ABST
Patent Text Reader

Abstract

Computer-implemented method for detecting anomalous charging cycles of a rechargeable battery (1) with a plurality of battery cells using a multilayer autoencoder (3) which is trained to recognize features of anomalous charging cycles from measured time series of the individual electrical cell voltages of the battery cells, comprising the steps: - Recording (100) the cell voltages of each of the battery cells of the rechargeable battery (1) as time series over the entire charging cycle of the battery (1) and generating cell-specific measurement data sets, - Normalizing (101) the measurement data sets and generating a set of normalized input data with a predetermined length N for each of the battery cells, - Applying (102) the autoencoder (3) to the normalized input data of one of the battery cells, wherein an encoder (32) compresses the normalized input data and thereby reduces its dimension, so that a compressed representation of the input data of the autoencoder (3) is obtained in a representation layer (33), and wherein a decoder (34) reconstructs the input data from the compressed representation, the reconstructed input data being output as output data, - Denormalize (103) the output data, - Calculating (104) an anomaly indicator for the battery cell using the denormalized output data, - Repeating the three preceding steps (102-104) for each of the battery cells of the rechargeable battery (1), characterized in that normalization parameters used for normalization are stored in a storage medium in a retrievable manner and that inverse normalization parameters are calculated from the stored normalization parameters, which are used for denormalization.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] The present invention relates to a computer-implemented method for detecting abnormal charging cycles of a rechargeable battery according to the preamble of claim 1.

[0002] In the context of the present invention, a computer-implemented method is understood to mean, in particular, that the method is carried out by means of a computer. The computer can, for example, comprise a digital data memory and a processor. Instructions can be stored in the digital data memory which, when executed by the processor, cause the processor to execute the computer-implemented method.

[0003] Rechargeable batteries are used in a variety of technical fields. Examples include batteries in motor vehicles, especially traction batteries or starter batteries, batteries in two-wheelers, and batteries in electronic devices.

[0004] The problem with rechargeable batteries is that anomalies can occur during the charging cycle, which, for example, have a negative impact on the battery's performance, but cannot be easily and reliably detected.

[0005] WO 2022 / 003208 A1 discloses a method for estimating the aging of a rechargeable battery based on battery data, wherein the method uses an artificial neural autoencoder network.

[0006] DE 10 2020 212 299 A1 discloses a computer-implemented method for predicting a modeled aging state of an electrical energy storage device with at least one electrochemical unit, in particular a battery cell, or by rule- and / or data-based mapping in an overall system. The method comprises the following steps: - Providing a data-based aging state model that is trained to assign a modeled aging state to the electrochemical energy storage device depending on the course of operating variables of the energy storage device, - Generating a history of at least one load variable depending on a provided usage pattern using a usage model; - generating the curves of operating variables depending on the at least one load variable using a given dynamic model; and - Determination of a predicted modeled aging state depending on the generated trends of the operating variables.

[0007] WO 2018 / 203170 A1 discloses an energy storage system with a neural network and a storage battery, wherein the neural network has an input layer, an output layer and one or more hidden layers between the input layer and the output layer, wherein a predetermined hidden layer is connected to the previous hidden layer or the previous input layer by a predetermined weight coefficient and wherein the predetermined hidden layer is connected to the next hidden layer or the next output layer by a predetermined weight coefficient.In the storage battery, an electric voltage of the storage battery and the time at which the voltage is obtained are measured as sets of data, the sets of data measured at different times are supplied to the input layer, and an operating state of the storage battery is changed in accordance with a signal output from the output layer.

[0008] From CN 1 15 469 227 A, a computer-implemented method for detecting abnormal charging cycles of a rechargeable battery with the essential features of the preamble of claim 1 is already known.

[0009] The present invention has for its object to provide a computer-implemented method for detecting abnormal charging cycles of a rechargeable battery, which can easily detect abnormal charging cycles of the battery.

[0010] The solution to this problem is provided by a computer-implemented method for detecting abnormal charging cycles of a rechargeable battery having the features of the characterizing part of claim 1. The subclaims relate to advantageous developments of the invention.

[0011] A computer-implemented method for detecting anomalous charging cycles of a rechargeable battery having a plurality of battery cells by means of a multi-layer autoencoder trained to detect characteristics of anomalous charging cycles from measured time series of the individual electrical cell voltages of the battery cells comprises the steps: - Recording the cell voltages of each of the battery cells of the rechargeable battery as time series over the entire charging cycle of the battery and generating cell-individual measurement data sets, - Normalizing the measurement data sets and generating a set of normalized input data with a given length N for each of the battery cells, - applying the autoencoder to the normalized input data of one of the battery cells, wherein an encoder compresses the normalized input data and thereby reduces its dimension, so that a compressed representation of the input data of the autoencoder is obtained in a representation layer, and wherein a decoder reconstructs the input data from the compressed representation, wherein the reconstructed input data is output as output data, - Denormalizing the output data, - Calculating an anomaly index for the battery cell using the denormalized output data, - Repeating the three preceding steps for each of the battery cells of the rechargeable battery. According to the invention, the method is characterized in that normalization parameters used for the normalization are stored in a retrievable manner in a storage medium, and inverse normalization parameters are calculated from the stored normalization parameters, which are used for the denormalization.

[0012] The method according to the invention enables the simple detection of abnormal charging cycles of the rechargeable battery. The encoder, which has several hidden layers, serves the purpose of compressing the normalized input data in the method presented here, thereby reducing its dimension, so that a compressed representation of the normalized input data of the autoencoder is obtained in the representation layer. The decoder, which also has several hidden layers, is designed to recover the input data from the compressed representation in the representation layer, so that this recovered input data can subsequently be output as output data and further processed. The lower dimensionality of the compressed state forces the autoencoder to learn the most frequently occurring features of the input variables during training.It is assumed that anomalous charging cycles differ from non-anomalous charging cycles in their morphology and frequency. In this context, it is therefore important to train the autoencoder primarily with training datasets that correspond to anomalous battery charging cycles. An autoencoder trained primarily with training datasets that correspond to non-anomalous battery charging cycles will be less able to predict anomalous charging cycles.

[0013] In one embodiment, it is proposed that the measurement data sets for each of the battery cells be normalized to a specified length N by interpolation and resampling. The length N represents the number of data sets fed to the autoencoder. For example, normalization can be performed to a length (number) N = 250.

[0014] In one embodiment, it can be provided that the measurement data sets for each of the battery cells are normalized to a defined value range. The value range of the measurement data is then normalized to a predetermined value range between minimum and maximum values, for example, to a value range [0.1; 0.9].

[0015] In one embodiment, the normalized input data of the battery cells can be fed to the encoder via an input layer. Preferably, an input layer with a number of neurons corresponding to the length (number) N of the normalized measurement data sets is used.

[0016] In one embodiment, it is proposed that the reconstructed input data be output as output data via an output layer. Preferably, an output layer with a number of neurons corresponding to the length (number) N of the normalized measurement data sets is used.

[0017] Preferably, a sigmoid function can be used as an activation function before the output layer.

[0018] In an advantageous embodiment, it is provided that the detected cell voltages are compared with the denormalized output data and the anomaly index is calculated therefrom.

[0019] Further features and advantages of the present invention will become clear from the following description of a preferred embodiment with reference to the accompanying drawings. Fig. 1 is a highly simplified schematic diagram showing details of a computer-implemented method for detecting abnormal charging cycles of a rechargeable battery, Fig. 2 is a schematically highly simplified representation illustrating the structure of an autoencoder used in the computer-implemented method.

[0020] To carry out a computer-implemented method for detecting abnormal charging cycles of a rechargeable battery 1 having a plurality of battery cells, the cell voltages of the individual battery cells as well as the start and end of the charging cycle are required.

[0021] In a first step 100, the cell voltages of each of the battery cells of the rechargeable battery 1 are recorded as a time series over the entire charging cycle of the battery 1 as cell-specific measurement data. The cell-specific measurement data (charging cycle voltages) obtained in this way differ for the individual battery cells of the battery 1 in their length (i.e., in the number of data sets) and in their value range, which is defined by the minimum and maximum values ​​of the electrical voltage of the respective battery cell during the charging cycle.

[0022] In a next step 101, the measurement data obtained in the previous step 100 are normalized using a normalization algorithm 2. The length of the cell-specific measurement data and thus the number of data sets per battery cell is normalized to a predetermined length, for example, to a length (number) of N = 250 data sets, by interpolation and resampling. The value range of the measurement data is normalized to a predetermined value range, in particular to a value range [0.1; 0.9]. The normalization parameters used to normalize the measurement data are stored in a retrievable manner in a storage medium.

[0023] The measurement data normalized in the manner described above, which are normalized charging cycle voltages of the battery cells of the battery 1, are fed to an autoencoder 3 as input data in a subsequent step 102, the function of which is described below with reference to Fig. 2 will be explained in more detail.

[0024] Autoencoder 3 is an artificial neural network with a multi-layer structure. In this exemplary embodiment, autoencoder 3 has a total of nine layers, whereby the number of layers is expressly intended as an example.

[0025] The autoencoder 3 has an input layer 30, via which the input data is fed to the autoencoder 3, an output layer 31, via which output data of the autoencoder 3 is output, an encoder 32, a representation layer 33 and a decoder 34.

[0026] Input layer 30 and output layer 31 can each have 250 neurons, for example. This corresponds to the normalized number N of data records per battery cell used for the evaluation. Encoder 32 forms a first artificial neural network of autoencoder 3 and, in this exemplary embodiment, has three encoder layers 320, 321, 322, each configured as hidden layers and having a decreasing number of neurons—viewed from input layer 30 toward representation layer 33. Representation layer 33 has the smallest number of neurons of autoencoder 3.

[0027] Likewise, decoder 34, which forms a second artificial neural network of autoencoder 3, has three layers 340, 341, 342, each configured as a hidden layer and containing an increasing number of neurons as viewed from representation layer 33 toward output layer 31. The three decoder layers 340, 341, 342 of decoder 34 form a mirror image of the three encoder layers 320, 321, 322 of encoder 32.

[0028] The encoder 32 serves the purpose of compressing the normalized input data and thereby reducing its dimension, so that a compressed representation of the input data of the autoencoder 3 is obtained in the representation layer 33. The decoder 34 is configured to recover the input data from the compressed representation in the representation layer 33, so that it is subsequently output as output data via the output layer 31.

[0029] So-called “Leaky ReLu” activation functions are preferably implemented in the encoder 32 and in the decoder 34 in order to impart non-linear properties to them. Before deployment in production, autoencoder 3 must be trained accordingly so that it is capable of detecting characteristics of anomalous charging cycles. Autoencoder 3 is preferably trained through supervised learning using a large number of training data sets. It first compresses its input variables, i.e., the normalized measurement data, to a state with lower dimensionality in representation layer 33 using encoder 32 and then restores it from the compressed representation using decoder 34.

[0030] The lower dimensionality of the compressed state forces autoencoder 3 to learn the most frequently occurring features of the input variables. It is assumed that anomalous charging cycles differ from non-anomalous charging cycles in their morphology and frequency. In this context, it is therefore important to train autoencoder 3 primarily with training datasets corresponding to anomalous charging cycles of battery 1. An autoencoder 3 trained primarily with training datasets corresponding to non-anomalous charging cycles of battery 1 will predict anomalous charging cycles less accurately.

[0031] The autoencoder 3 is trained using the time series of the cell voltages of the battery cells (charge cycle voltages) of a plurality of rechargeable batteries 1 during their charging cycle. If the battery 1 is a traction battery of a motor vehicle, a random subset from a plurality of available motor vehicles is preferably selected for training the autoencoder 3. The autoencoder 3 is preferably trained using supervised learning, in particular using so-called backpropagation.

[0032] The error signal for a backpropagation training of Autoencoder 3, which is based on mean square error calculations, is defined as follows: ∑i=1N(Inputi−Outputi)2N

[0033] In this formula, "Input;" denotes the input variables of autoencoder 3, and "Outputi" denotes the output variables of autoencoder 3, where N is the number of data sets considered. The goal of training autoencoder 3 is to minimize the mean square error, which, from a functional perspective, constitutes a reconstruction error of autoencoder 3.

[0034] Preferably, all charging cycles of the vehicles selected for training the autoencoder 3, which include at least 100 points in time and a change in the state of charge (“Delta-State of Charge”, in short: Delta-SoC) of at least 20% during the charging cycle, are taken into account.

[0035] With further reference to Fig.1, the output data of the autoencoder 3 obtained in the preceding step 102 are fed in a step 103 via the output layer 31 to a reconstruction algorithm 4, which is designed to denormalize and thereby reconstruct the output data of the autoencoder 3. Preferably, in the last processing step from layer 342 to the output layer 31, a sigmoid function is used as the activation function.

[0036] The denormalization or reconstruction of the output data is performed using inverse normalization parameters calculated from the normalization parameters stored in method step 102. The result is a reconstructed charging cycle voltage of one of the battery cells of the rechargeable battery 1. By repeatedly executing method steps 102 and 103, a reconstructed charging cycle voltage can be obtained for each of the battery cells of the rechargeable battery 1.

[0037] In a step 104, the cell voltages measured during the charging cycle (charging cycle voltages), which form the input variables of the autoencoder 3, are then compared with the denormalized output data of the autoencoder 3 using a comparison algorithm 5. The difference forms an anomaly characteristic for the respective charging cycle. The anomaly characteristic is compared with a predefined or predefinable threshold. Charging cycles with an anomaly characteristic above the threshold are considered anomalous.

[0038] Using the method presented here, it is possible to check for each of the battery cells of the rechargeable battery 1 by calculating the corresponding anomaly characteristic whether any anomalies may have occurred during the charging cycle.

Claims

[1] Computer-implemented method for detecting abnormal charging cycles of a rechargeable battery (1) having a plurality of battery cells by means of a multi-layer autoencoder (3) which is trained to detect features of abnormal charging cycles from measured time series of the individual electrical cell voltages of the battery cells, comprising the steps: - recording (100) the cell voltages of each of the battery cells of the rechargeable battery (1) as time series over the entire charging cycle of the battery (1) and generating cell-individual measurement data sets, - normalizing (101) the measurement data sets and generating a set of normalized input data with a predetermined length N for each of the battery cells, - applying (102) the autoencoder (3) to the normalized input data of one of the battery cells, wherein an encoder (32) compresses the normalized input data and thereby reduces its dimension, so that a compressed representation of the input data of the autoencoder (3) is obtained in a representation layer (33), and wherein a decoder (34) reconstructs the input data from the compressed representation, wherein the reconstructed input data are output as output data, - Denormalizing (103) the output data, - calculating (104) an anomaly index for the battery cell using the denormalized output data, - Repeat the three previous steps (102-104) for each of the battery cells of the rechargeable battery (1), characterized bythat normalization parameters used for the normalization are stored in a storage means in a retrievable manner and that inverse normalization parameters are calculated from the stored normalization parameters, which are used for the denormalization. [2] Computer-implemented method according to claim 1, characterized by that the measurement data sets for each of the battery cells are normalized to the specified length N by interpolation and resampling. [3] Computer-implemented method according to one of claims 1 or 2, characterized by that the measurement data sets for each of the battery cells are normalized to a defined range of values. [4] Computer-implemented method according to one of claims 1 to 3, characterized by that the normalized input data of the battery cells are fed to the encoder (32) via an input layer (30). [5] Computer-implemented method according to one of claims 1 to 4, characterized by that the reconstructed input data are output as output data via an output layer (31). [6] Computer-implemented method according to claim 5, characterized by that a sigmoid function is used as an activation function before the output layer (31). [7] Computer-implemented method according to one of claims 1 to 6, characterized by that the recorded cell voltages are compared with the denormalized output data and the anomaly index is calculated from this.

Citation Information

Patent Citations

  • Processing of condition data of a battery to estimate ageing

    WO2022003208A1

  • Lithium battery anomaly detection method integrating variational auto-encoder and dynamic normalization

    CN115469227A

  • Method and apparatus for operating a system for providing predicted aging states of electrical energy storage devices for a device using machine learning methods

    DE102020212299A1

  • Neural network, power storage system, vehicle, and electronic device

    WO2018203170A1

  • CN000115469227A