Battery management system for classifying battery modules

Abstract

This invention relates to a battery management system (100) for classifying battery modules having a first battery cell and a second battery cell, wherein the battery management system (100) includes an interface (101) and a processor (103). The interface (101) is designed to obtain a first voltage waveform on the first battery cell and a second voltage waveform on the second battery cell, wherein the first voltage waveform includes a first voltage waveform portion and a second voltage waveform portion, and the second voltage waveform includes a third voltage waveform portion and a fourth voltage waveform portion. The processor (103) is designed to determine a reference voltage waveform based on the average voltage of the first voltage waveform and the second voltage waveform, wherein the reference voltage waveform includes a first reference voltage waveform portion and a second reference voltage waveform portion. Further, the processor (103) is designed to compare a first voltage waveform with a reference voltage waveform to obtain a first index indicating a first voltage deviation between a first voltage waveform portion and a first reference voltage waveform portion, and a second voltage deviation between a second voltage waveform portion and a second reference voltage waveform portion; and to compare the second voltage waveform with the reference voltage waveform to obtain a second index indicating a third voltage deviation between a third voltage waveform portion and a first reference voltage waveform portion, and a fourth voltage deviation between a fourth voltage waveform portion and a second reference voltage waveform portion. If the first index is greater than the second index, the processor (103) is designed to associate a first electrical feature with the first voltage waveform portion and a second electrical feature with the second voltage waveform portion to classify the battery module by the first and second electrical features. Alternatively, if the first index is less than the second index, the processor (103) is designed to associate a third electrical feature with the third voltage waveform portion and a fourth electrical feature with the fourth voltage waveform portion to classify the battery module by the third and fourth electrical features.

Description

Technical Field

[0001] This invention relates to a battery management system and a method for classifying battery modules including battery cells. In particular, this invention relates to a classification algorithm for automatically detecting abnormal conditions in battery modules. Background Technology

[0002] If the voltage waveform of the battery module under load does not conform to the typical normal distribution for the battery type and aging state, it can currently only be visually inspected by professionals. This is a time-consuming and error-prone activity due to its manual and monotonous nature.

[0003] Determining the achieved product quality is crucial for optimizing the battery system manufacturing process. To this end, in modern production facilities, measurement data is typically collected fully automatically in digital form. Machine learning methods can be used to evaluate this data because they enable comprehensive, low-cost, and rapid inspection. Furthermore, unlike humans, machine learning models do not show any signs of fatigue from monotonous work. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a battery management system and a classification method for battery modules. This system and method can detect abnormal conditions in battery modules and classify these abnormal conditions into various types. In particular, it will be shown that machine learning methods are suitable for detecting abnormal conditions in battery systems.

[0005] This invention is based on the understanding that abnormal conditions are visible from the voltage waveform of a battery module. Therefore, the effects of incorrect internal resistance, capacity, and offset voltage are derived from the electrical circuitry of the battery model. It was found that all observed effects can be described based on these three physical quantities. Based on this finding, different types of abnormalities and normal conditions are defined.

[0006] According to a first aspect of the invention, the technical problem is solved by a battery management system for classifying battery modules, wherein the battery module includes a first battery cell and a second battery cell. The battery management system includes an interface and a processor. The interface is designed to obtain a first voltage waveform on the first battery cell and a second voltage waveform on the second battery cell, wherein the first voltage waveform includes a first voltage waveform portion and a second voltage waveform portion, and the second voltage waveform includes a third voltage waveform portion and a fourth voltage waveform portion.

[0007] The processor is further designed to: determine a reference voltage waveform based on the average voltage from a first voltage waveform and a second voltage waveform, wherein the reference voltage waveform has a first reference voltage waveform portion and a second reference voltage waveform portion; compare the first voltage waveform with the reference voltage waveform to obtain a first index indicating a first voltage deviation between the first voltage waveform portion and the first reference voltage waveform portion, and a second voltage deviation between the second voltage waveform portion and the second reference voltage waveform portion; compare the second voltage waveform with the reference voltage waveform to obtain a second index indicating a third voltage deviation between a third voltage waveform portion and the first reference voltage waveform portion, and a fourth voltage deviation between a fourth voltage waveform portion and the second reference voltage waveform portion; and if the first index is greater than the second index, assign a first electrical feature association to the first voltage waveform portion and assign a second electrical feature association to the second voltage waveform portion to classify the battery pack according to the first and second electrical features; or if the first index is less than the second index, assign a third electrical feature association to the third voltage waveform portion and assign a fourth electrical feature association to the fourth voltage waveform portion to classify the battery module according to the third and fourth electrical features.

[0008] This achieves the following technical advantages: abnormal conditions in the battery module are detected, thereby enabling the differentiation between fault-free battery modules (i.e., normal conditions) and faulty battery modules.

[0009] According to an exemplary embodiment of the battery management system, the battery management system includes a memory designed to store a plurality of electrical features, wherein a processor is designed to read the corresponding electrical features from the memory.

[0010] Therefore, electrical characteristics can be stored and read out.

[0011] According to an exemplary embodiment of the battery management system, the processor is designed to determine a reference voltage waveform based on a first voltage waveform and a second voltage waveform.

[0012] This provides a technical advantage for efficiently calculating the reference voltage waveform of the battery module.

[0013] According to an exemplary embodiment of the battery management system, the reference voltage waveform includes the median or average of a first voltage waveform and a second voltage waveform.

[0014] This enables the technical advantage of effectively calculating the reference voltage waveform of the battery module.

[0015] According to an exemplary embodiment of the battery management system, the battery module includes a third battery cell, and the interface is designed to obtain a third voltage waveform on the third battery cell, wherein the processor is designed to determine a reference voltage waveform based on a first voltage waveform, a second voltage waveform, and a third voltage waveform.

[0016] According to an exemplary embodiment of the battery management system, the reference voltage waveform includes the median, average, or mode of a first voltage waveform, a second voltage waveform, and a third voltage waveform.

[0017] This achieves a technological advantage by generating a reference voltage waveform for the battery module, which is used to detect the battery cell that differs most from the other battery cells in the battery module in terms of voltage waveform.

[0018] According to an exemplary embodiment of the battery management system, the processor is designed to: determine a first electrical feature based on a first voltage waveform portion of a first voltage waveform using principal component analysis, and determine a second electrical feature based on a second voltage waveform portion of the first voltage waveform, wherein the first electrical feature represents the first voltage waveform portion of the first voltage waveform, and the second electrical feature represents the second voltage waveform portion of the first voltage waveform; and / or determine a third electrical feature based on a third voltage waveform portion of the second voltage waveform, and determine a fourth electrical feature based on a fourth voltage waveform portion of the second voltage waveform, wherein the third electrical feature represents the third voltage waveform portion of the second voltage waveform, and the fourth electrical feature represents the fourth voltage waveform portion of the second voltage waveform.

[0019] This allows electrical characteristics to be determined efficiently.

[0020] According to an exemplary embodiment of the battery management system, the processor is designed to determine a first electrical characteristic based on a first voltage waveform portion of a first voltage waveform, a second electrical characteristic based on a second voltage waveform portion of the first voltage waveform, a third electrical characteristic based on a third voltage waveform portion of the second voltage waveform, and a fourth electrical characteristic based on a fourth voltage waveform portion of the second voltage waveform.

[0021] According to an exemplary embodiment of the battery management system, a first electrical feature corresponds to the offset voltage of a first battery cell, a third electrical feature corresponds to an additional offset voltage of a second battery cell, wherein the second electrical feature corresponds to the internal resistance of the first battery cell, and a fourth electrical feature corresponds to an additional internal resistance of the second battery cell.

[0022] The technical advantage of this approach is that it can effectively extract the electrical characteristics of the battery cells in the battery module, thereby allowing the description of abnormal conditions occurring in the battery module using the offset voltage and internal resistance of each battery cell.

[0023] According to an exemplary embodiment of the battery management system, the processor is designed to classify battery modules using a classification algorithm, wherein the classification algorithm includes at least one of logistic regression, support vector machine, random forest, multilayer perceptron, and single-class support vector machine.

[0024] This provides a technological advantage in effectively classifying battery modules to determine whether a battery module is fault-free or faulty.

[0025] According to an exemplary embodiment of the battery management system, an interface is designed to obtain a plurality of first voltage waveforms and a plurality of second voltage waveforms, wherein each of the plurality of first voltage waveforms corresponds to a first battery cell of one of a plurality of battery modules, and each of the plurality of second voltage waveforms corresponds to a second battery cell of one of the plurality of battery modules. The interface is designed to operate the plurality of battery modules without failure, wherein the processor is designed to classify each of the plurality of battery modules using an additional classification algorithm to form a reference group based on the classification.

[0026] This achieves the technical advantage of effectively forming a reference group of fault-free battery modules and sets a strict threshold for fault-free battery modules in the feature space, according to which battery modules exceeding the threshold can be declared abnormal.

[0027] According to an exemplary embodiment of the battery management system, the processor is designed to classify multiple battery modules using an additional classification algorithm, wherein the additional classification algorithm includes at least one of logistic regression, support vector machine, random forest, multilayer perceptron, and single-class support vector machine.

[0028] This achieves the following technical advantages: it effectively forms a reference group of fault-free battery modules and establishes a strict threshold for fault-free battery modules in the feature space.

[0029] According to an exemplary embodiment of the battery management system, a first voltage waveform has a fifth voltage waveform portion and a second voltage waveform has a sixth voltage waveform portion, wherein the processor is designed to extract a fifth feature based on the fifth voltage waveform portion of the first voltage waveform and to extract a sixth feature based on the sixth voltage waveform portion of the second voltage waveform.

[0030] According to an exemplary embodiment of the battery management system, the fifth feature corresponds to the capacity of the first battery cell, and the sixth feature corresponds to an additional capacity of the second battery cell.

[0031] This achieves a technological advantage by effectively extracting the electrical characteristics of the battery cells in the battery module, thereby enabling the description of anomalies occurring in the battery module using offset voltage, battery capacity, and / or the internal resistance of individual battery cells.

[0032] According to an exemplary embodiment of the battery management system, the interface is designed to obtain the number of groups to be generated, a plurality of additional first voltage waveforms and a plurality of additional second voltage waveforms, wherein each of the plurality of additional first voltage waveforms corresponds to a first battery cell of one of a plurality of additional battery modules, and wherein each of the plurality of additional second voltage waveforms corresponds to a second battery cell of one of a plurality of additional battery modules, wherein the processor is designed to assign each of the plurality of additional battery modules to a plurality of groups by a classification algorithm, wherein the number of the plurality of groups is equal to the number of groups to be generated.

[0033] This enables the technological advantage of effectively classifying battery modules, thereby detecting abnormalities in battery modules and classifying these abnormalities into different types.

[0034] According to a second aspect of the present invention, the technical problem is solved by a battery management method comprising the following steps: obtaining a first voltage waveform on a first battery cell of a battery module and a second voltage waveform on a second battery cell of a battery module, wherein the first voltage waveform includes a first voltage waveform portion and a second voltage waveform portion, and wherein the second voltage waveform includes a third voltage waveform portion and a fourth voltage waveform portion; determining a reference voltage waveform based on the voltage average of the first voltage waveform and the second voltage waveform, wherein the reference voltage waveform has a first reference voltage waveform portion and a second reference voltage waveform portion; comparing the first voltage waveform with the reference voltage waveform to obtain a first index, the first index indicating a first voltage deviation between the first voltage waveform portion and the first reference voltage waveform portion and a second voltage waveform portion. A second voltage deviation between the waveform portion and the second reference voltage waveform portion; comparing the second voltage waveform with the reference voltage waveform to obtain a second index, which indicates a third voltage deviation between the third voltage waveform portion and the first reference voltage waveform portion and a fourth voltage deviation between the fourth voltage waveform portion and the second reference voltage waveform portion; if the first index is greater than the second index, a first electrical feature is associated with the first voltage waveform portion and a second electrical feature is associated with the second voltage waveform portion, so as to classify the battery module according to the first electrical feature and the second electrical feature; or if the first index is less than the second index, a third electrical feature is associated with the third voltage waveform portion and a fourth electrical feature is associated with the fourth voltage waveform portion, so as to classify the battery module according to the third electrical feature and the fourth electrical feature.

[0035] This achieves the following technical advantages: battery module anomalies are detected, thus enabling the differentiation between fault-free battery modules (i.e., normal conditions) and faulty battery modules.

[0036] All embodiments of the battery management system mentioned in the first aspect are also applicable to embodiments of the method according to the second aspect. Attached Figure Description

[0037] The present invention will now be described in more detail with reference to the embodiments and accompanying drawings. Wherein:

[0038] Figure 1 A schematic diagram of a battery management system according to one embodiment is shown, which is used to classify battery modules;

[0039] Figure 2 A table of exception condition types defined according to one embodiment is shown;

[0040] Figure 3 Examples of parameter settings for favorable and unfavorable choices are shown;

[0041] Figure 4 Examples of voltage waveforms and signal interpolation are shown;

[0042] Figure 5 Examples of signals with atypical signal waveforms are shown;

[0043] Figure 6 The result after Discrete Wavelet Transform (DWT) is shown. Figure 5 The signal in the example, and the qualitative waveform of the piecewise linear cost function;

[0044] Figure 7 An example of segmenting a voltage waveform is shown;

[0045] Figure 8 The voltage waveform of a battery module with a faulty battery cell in one embodiment was simulated;

[0046] Figure 9 Here is an example of a feature vector generated from the training dataset;

[0047] Figure 10 A table showing the training results of various classification algorithms in one implementation scheme is provided; and

[0048] Figure 11 A flowchart illustrating a method according to one implementation scheme is shown. Detailed Implementation

[0049] In the following detailed description, reference is made to the accompanying drawings, which form a part of this document, in which specific embodiments of the invention may be practiced are shown as figures. It will be understood that other embodiments may be used, and structural or logical changes may be made, without departing from the concept of the invention. Therefore, the following detailed description should not be construed as limiting. It should also be understood that the features of the various embodiments described herein may be combined with each other, unless specifically stated otherwise.

[0050] Various aspects and embodiments of the invention will be described with reference to the accompanying drawings, wherein similar reference numerals generally refer to similar elements. In the following description, numerous specific details are set forth for purposes of explanation in order to provide a deeper understanding of one or more aspects of the invention.

[0051] Figure 1 A schematic diagram of a battery management system 100 for classifying battery modules according to one embodiment is shown, wherein the battery modules include at least a first battery cell and a second battery cell. Figure 1 As shown, the battery management system 100 includes an interface 101, a processor 103, and a memory 105, the functions of which will be discussed in detail below.

[0052] In one embodiment, interface 101 is designed to obtain a first voltage waveform on a first battery cell and a second voltage waveform on a second battery cell, wherein the first voltage waveform includes a first voltage waveform portion and a second voltage waveform portion, and wherein the second voltage waveform includes a third voltage waveform portion and a fourth voltage waveform portion.

[0053] In one embodiment, processor 103 is designed to determine a reference voltage waveform based on the average voltage of a first voltage waveform and a second voltage waveform, wherein the reference voltage waveform includes a first reference voltage waveform portion and a second reference voltage waveform portion. In one embodiment, processor 105 is designed to determine a reference voltage waveform based on a first voltage waveform and a second voltage waveform, and the reference voltage waveform includes the median or average value of the first voltage waveform and the second voltage waveform.

[0054] Additionally, the processor 103 is designed to compare a first voltage waveform with a reference voltage waveform to obtain a first index indicating a first voltage deviation between a first voltage waveform portion and a first reference voltage waveform portion, and a second voltage deviation between a second voltage waveform portion and a second reference voltage waveform portion; and to compare the second voltage waveform with the reference voltage waveform to obtain a second index indicating a third voltage deviation between a third voltage waveform portion and a first reference voltage waveform portion, and a fourth voltage deviation between a fourth voltage waveform portion and a second reference voltage waveform portion.

[0055] If the first indicator is greater than the second indicator, the processor 103 is designed to associate a first electrical feature with a first voltage waveform portion and a second electrical feature with a second voltage waveform portion, so as to classify the battery module by the first electrical feature and the second electrical feature.

[0056] In one embodiment, the processor 103 is designed to: use principal component analysis to determine a first electrical feature based on a first voltage waveform portion of a first voltage waveform, and to determine a second electrical feature based on a second voltage waveform portion of the first voltage waveform, wherein the first electrical feature represents the first voltage waveform portion of the first voltage waveform, and the second electrical feature represents the second voltage waveform portion of the first voltage waveform.

[0057] Alternatively, if the first indicator is less than the second indicator, the processor 103 is designed to match a third electrical feature to the third voltage waveform portion and to match a fourth electrical feature to the fourth voltage waveform portion in order to classify the battery module by the third and fourth electrical features.

[0058] In one embodiment, the processor 103 is designed to use principal component analysis to determine a third electrical feature based on a third voltage waveform portion of the second voltage waveform and a fourth electrical feature based on a fourth voltage waveform portion of the second voltage waveform, wherein the third electrical feature represents the third voltage waveform portion of the second voltage waveform and the fourth electrical feature represents the fourth voltage waveform portion of the second voltage waveform.

[0059] In one embodiment, memory 105 is designed to store a plurality of electrical features, wherein processor 103 is designed to read the respective electrical features from memory.

[0060] In terms of classification, the processor 103 is designed to classify the battery modules using classification algorithms, including at least one of logistic regression, support vector machine, random forest, multilayer perceptron and single-class support vector machine.

[0061] In one embodiment, the battery module further includes a third battery cell, and the interface 101 is designed to obtain a third voltage waveform on the third battery cell, wherein the processor 103 is designed to determine a reference voltage waveform based on the first voltage waveform, the second voltage waveform, and the third voltage waveform. In this case, the reference voltage waveform includes the median, average, or mode of the first voltage waveform, the second voltage waveform, and the third voltage waveform.

[0062] In one embodiment, the processor is configured to determine a first electrical characteristic based on a first voltage waveform portion of a first voltage waveform, a second electrical characteristic based on a second voltage waveform portion of the first voltage waveform, a third electrical characteristic based on a third voltage waveform portion of the second voltage waveform, and a fourth electrical characteristic based on a fourth voltage waveform portion of the second voltage waveform, wherein the first electrical characteristic corresponds to an offset voltage of a first battery cell, the third electrical characteristic corresponds to an additional offset voltage of a second battery cell, the second electrical characteristic corresponds to an internal resistance of the first battery cell, and the fourth electrical characteristic corresponds to an additional internal resistance of the second battery cell.

[0063] In one embodiment, interface 101 is designed to obtain a plurality of first voltage waveforms and a plurality of second voltage waveforms, each of the plurality of first voltage waveforms corresponding to a first battery cell of one of a plurality of battery modules, and each of the plurality of second voltage waveforms corresponding to a second battery cell of one of a plurality of battery modules, wherein the interface is designed to operate the plurality of battery modules without failure.

[0064] In this case, the processor 103 is designed to classify each of the multiple battery modules using an additional classification algorithm to form a reference group based on the classification, wherein the additional classification algorithm includes at least one of logistic regression, support vector machine, random forest, multilayer perceptron, and single-class support vector machine.

[0065] In one embodiment, the first voltage waveform has a fifth voltage waveform portion, and the second voltage waveform has a sixth voltage waveform portion. The processor 103 is configured to extract a fifth feature based on the fifth voltage waveform portion of the first voltage waveform, and to extract a sixth feature based on the sixth voltage waveform portion of the second voltage waveform. The fifth feature corresponds to the capacity of the first battery cell, and the sixth feature corresponds to an additional capacity of the second battery cell.

[0066] In one embodiment, interface 101 is further designed to obtain the number of groups to be generated, a plurality of additional first voltage waveforms, and a plurality of additional second voltage waveforms, wherein each of the plurality of additional first voltage waveforms corresponds to a first battery cell of one of the plurality of additional battery modules, and wherein each of the plurality of additional second voltage waveforms corresponds to a second battery cell of one of the plurality of additional battery modules. Accordingly, processor 103 is designed to associate each of the plurality of additional battery modules with a plurality of groups using a classification algorithm, wherein the number of the plurality of groups is equal to the number of groups to be generated.

[0067] Through the embodiments of the present invention, the characteristics of the battery module can be analyzed completely automatically, and possible abnormal situations can be detected and classified.

[0068] This achieves the following advantages:

[0069] - Cost savings, because potential product non-compliance recalls can be prevented at an early stage (immediately after production).

[0070] - This method can be applied to a single battery module.

[0071] - Locate the affected modules

[0072] - Classify abnormal situations (indicate the reason)

[0073] - Perform 100% testing on all modules

[0074] The implementation schemes of this invention involve the following aspects: battery simulation, outlier detection and feature extraction, classification algorithms and anomaly detection, which will be discussed in detail below.

[0075] With the aid of a battery simulation model, realistic anomaly types are simulated and generated in a uniform distribution. Subsequent data preprocessing extracts the features (feature vectors) needed for subsequent classification and detects heterogeneous or previously unknown anomalies (outliers). The classification model associates each module with a known group (classification), based on which the downstream anomaly detector performs a more rigorous review of whether a situation is an anomaly. However, in these cases, a purely binary assignment (abnormal or normal) is performed.

[0076] The purpose of simulating anomalies is to obtain changes in physical properties relative to their "normal" characteristics, and to be able to simulate them according to their characteristics. The characteristics of this change must be reproduced repeatedly. This must occur within the threshold of the anomaly in order to be able to associate a cause for the anomaly.

[0077] In order to reliably produce the required accuracy of this changing characteristic, it is necessary to take three steps, which will be considered in detail below:

[0078] - Simulate the characteristics of physical systems

[0079] - Define the parameters that cause the exception.

[0080] - Distribution of generation parameters

[0081] Simulate the characteristics of physical systems

[0082] The characteristics of a real physical system must be simulated as accurately as possible by a simulation model of the variables being analyzed. This is achieved by adjusting the parameters of the simulation model to minimize the deviation between the real and simulated characteristics.

[0083] This adjustment is accomplished through a combination of optimization algorithms used in a prescribed order. These algorithms determine the combination of parameters that minimizes the deviation between the simulation and actual measurements. Here, this parameter adjustment is always based on a set of parameters that has already been initially adjusted for general characteristics. The reflection of the specific, real-world system's characteristics after this adjustment is also known as a "digital twin."

[0084] Define the parameters that cause the exception.

[0085] To generate anomalies in a targeted manner, the parameters that cause them must be identified. This assignment can be done through pure data evaluation or by a specialized department, which can associate characteristic patterns with causes, and thus with parameters. Three parameters have been defined to characterize anomalies. All possible combinations of these parameters define the situations shown below.

[0086] Figure 2 Table 200, showing the defined exception types in one implementation scheme, is illustrated. Figure 2 It can be seen that all the abnormal situations that occur can be described by offset voltage, battery cell capacity, and internal resistance.

[0087] Distribution of Abnormal Situations

[0088] By analyzing numerous real-world anomalies, the characteristics of the parameters leading to the anomalies can be identified. These characteristics are then used to determine the order of magnitude of the parameter thresholds for specific anomalies.

[0089] These established thresholds can now be used to independently generate outliers whose characteristics and behavior are consistent with those of real-world systems. The generated outliers can be created frequently as needed, with any combination of features, and can be used to train machine learning algorithms.

[0090] While simulation models offer numerous advantages, the generation of training datasets also presents challenges. The selection of simulation parameters is crucial for using the trained model with real data later. For example, there is a risk that parameters in the training dataset might be unfavorably chosen, causing certain scenarios to occur more frequently. Figure 3 This shows an example of parameter settings with favorable (top right) and unfavorable (top left) choices.

[0091] from Figure 3As can be seen in the upper left corner, the distance between the normal case 301c and the capacity-only fault 303 is very small, while the other two abnormal cases 305 and 307 are much larger than the normal case. Below is an example decision threshold defined by the classifier based on the training data. The classifier will match a new data point with both abnormal internal resistance and capacity to the capacity fault group because the unfavorable decision threshold is defined based on the training data. This can be remedied by cleverly setting the parameters (see...). Figure 3 (Top right). Since the centers of all outlier data points lie on an arc, they are equidistant from the centers of the normal data points. Therefore, no case is artificially favored, and correct classification is possible (see...). Figure 3 (bottom right).

[0092] Figure 3 It is also shown that, crucial for training the classification algorithm, in this application scenario, the parameter distribution is designed to be circular around the normal cases 301a-d. In this case, the parameters and their combinations are designed to create gaps between the cases. This ensures the boundaries between the various outliers. This is illustrated here using a two-dimensional space with two parameters. However, this relationship must also be followed for higher-dimensional relationships. The circular arrangement around the normal cases ensures that the evaluation weights for the specified outliers are equal, thus resulting in an equal number of normal and outlier cases.

[0093] Extensive data preprocessing is necessary before using machine learning methods to detect and classify anomalies. This occurs in several process steps, which are listed below.

[0094] Filtering time series

[0095] In one implementation, the battery to be examined consists of either 28 or 33 modules. However, regardless of the battery type, data for 33 modules is recorded. If the battery has only 28 modules, the remaining five modules use default values. Removing these modules and their default values ​​from the actual dataset is the first step in data preprocessing. To identify relevant modules, the individual battery cell voltages are read. If a module only has a default value as its battery cell voltage and its module number is greater than 28, then that module and its associated time series are removed.

[0096] Determine the trigger point

[0097] In addition to time series of varying lengths, real-world data may also contain time-shifted signals. These time shifts are compensated for by triggering at prominent edges of the signal. The first derivative of the signal and a predetermined trigger threshold are used to dynamically detect these edges. If the slope magnitude exceeds this trigger threshold, it is identified as the edge being sought.

[0098] Interpolating time series

[0099] The actual data used has different step sizes within a time series. Examining the signal without corresponding time values ​​will distort the signal waveform. However, for further processing, only voltage values ​​without time information should be examined. To make this possible, the signal must be interpolated. Figure 4 This step is illustrated. With the trigger point 401 defined above, the endpoint of the interpolation range 403 can be determined. The starting point is calculated from the difference between the endpoint and a predetermined time period. Linear interpolation within this range results in the normalization of the step size. Therefore, regardless of the time base, the individual signals can be compared with each other without distorting the signal waveform.

[0100] from Figure 4 As can be seen, this process also applies to filtering out default values ​​that may appear at the beginning of a record. Time series always have default values ​​when data is being recorded but no measurements are being taken. If such default values ​​appear in the actual signal waveform, they must not be deleted, as this would manipulate the measurement results and result in the loss of information about potential defects in the measurement unit.

[0101] Smoothing time series

[0102] To improve the ratio between the useful signal and the signal-to-noise ratio, a first-order low-pass filter was used. However, the sampling rate was clearly too low to significantly suppress noise without affecting the useful signal. In order to still be able to determine the cutoff frequency, a trade-off was made between the remaining useful signal and noise, and a cutoff frequency of 0.1 Hz was empirically determined.

[0103] Filter out values

[0104] To further process the data, the shape similarity of the measurements must be considered because triggering occurs on certain characteristics. All defined anomalies and normal conditions exhibit this shape similarity. Therefore, voltage waveforms with different characteristics must be filtered out beforehand. Figure 5 This shows a possible example of a signal 501 with a different, atypical waveform.

[0105] Dynamic Time Warp (DTW) can be used to detect such outliers. This method is suitable for comparing signal waveforms because the time offset or distortion is compensated for. However, tests show that using DTW requires considerable computational and time costs. Therefore, it is not a favorable option for applications with limited computing power. To still be able to identify this unusual waveform, the signal is transformed using Discrete Wavelet Transform (DWT). This halves the signal length without losing essential information. The detail coefficients are then examined, as these coefficients contain the high-frequency components of the original signal. It turns out that all generated cases differ only in the height of four prominent peaks. If a signal has other peaks besides these four, it is due to deviations in the signal shape. Using Euclidean distance, the transformed signal can be compared with a reference signal. However, since only the region outside the peaks is important for this comparison, an additional cost function f is defined. cost The following describes the calculation of shape similarity:

[0106]

[0107] p j q represents the j-th detail coefficient of the investigated signal. j This is the j-th detail coefficient of the reference signal. Then, the square of the difference between these two values ​​is multiplied by f. cost (i). This coefficient corresponds to the j-th value of the cost function. If the cost function has zero values ​​at the locations of the peaks, they are not considered when determining shape similarity. Since peaks can move along the X-axis in real data, a rectangular waveform of the cost function is unhelpful. The remedy is to provide a function that increases with distance from the peak. This gives more weight to deviations far from the peak, while tolerating deviations near the peak. d s□ape The higher the value, the more different the shape of the signal.

[0108] Figure 6 The diagram shows the qualitative waveforms of the DWT transformed signal and the piecewise linear cost function in the example above. To determine the cost function, a reference signal is first derived to pinpoint the peak location. All derived signal values ​​below a predetermined threshold are set to zero. This threshold is the maximum value of the derivative of the signal outside the peak region. Therefore, only the derivative of the peak remains.

[0109] The calculation methods for each value of the cost function are as follows:

[0110]

[0111] Parameters w and k are set by the programmer. The tolerance range can be changed using the weighting factor w. The exponent p determines the cost function based on the distance d.epak The rate of increase of d. Here, d peak This represents the distance to the nearest peak. The j-th derivative of the transformed signal p is represented by p. j 'express.

[0112] Feature extraction

[0113] The anomaly under investigation is described by only three physical quantities: offset voltage, battery cell capacity, and battery cell internal resistance. Decomposing the signal into three segments constitutes the first step in feature extraction (see...). Figure 7 Three edges are used to separate the various regions. These edges can be easily determined by differentiating the signals, since each signal has a similar waveform, and measurement noise is largely compensated for.

[0114] Assuming that the first segment 701 is particularly suitable for determining the offset voltage, the second segment 702 is used to detect low battery cell capacity, and the effect of increased internal resistance becomes more pronounced in the third segment 703, these effects also indicate that the signal waveform of the second segment 702 is steeper, not solely due to low battery cell capacity. The nonlinear relationship between battery cell voltage and state of charge (SoC) is also clearly visible during charging, causing deviations in the dynamic characteristics of the voltage waveform during charging. Since the battery system is not in a uniform state of charge before the End-of-Line Test (EoL test), the offset voltage of each battery may differ slightly. By measuring the battery voltage at the start of the test, all batteries outside the specified tolerance range are removed. For this reason, it can be assumed that most battery cells with voltages within the tolerance range are not abnormal. The different states of charge of the batteries do not allow for the use of a rigid reference signal to distinguish between abnormal and normal conditions. Therefore, any abnormalities that occur must be declared as context-dependent, as the state of the battery cell cannot be determined solely by the signal waveform. The battery system will be examined at the module level below.

[0115] Figure 8 This diagram shows an analog voltage waveform of a battery module with a faulty battery cell in one embodiment, and a calculated reference signal 801. The voltage waveform of the reference signal 801 is formed by calculating the median voltage of all battery cells in the module. The median is more suitable as a reference than the arithmetic mean because outliers have a much smaller impact on the median. If multiple voltage waveforms are considered, they can be represented as follows:

[0116] in 1≤i≤m and 1≤j≤n

[0117] The number of voltage waveforms is represented by m, and the number of measurement points is represented by n. Therefore, for the reference signal:

[0118] in

[0119] The generated reference signal is used to detect the battery cell that differs most from other battery cells in the module in terms of voltage waveform. The reference signal for each module is regenerated. Since it's possible that several battery cells in a module may exhibit different anomalies, the comparison of the battery cell voltage with the reference voltage is performed segment by segment. The steps for feature extraction in each segment will be explained below.

[0120] Offset voltage

[0121] Since both excessively high and low offset voltages indicate a fault, only the difference between the battery cell and the reference voltage is important. Therefore, the battery cell with the largest difference from the reference signal is the one with the largest deviation from the ideal.

[0122] Battery cell capacity

[0123] In the second section, information about the battery cell capacity is extracted. To compare charging characteristics, all voltages are first set to a uniform starting value. 0V is chosen as the starting value because this can be easily achieved by subtracting the first voltage value from the second section. Then, the most problematic battery cells are detected. This can be identified by the steepest voltage waveform.

[0124] Internal resistance

[0125] Offset voltage is also compensated in the third segment because only relative voltage spikes can lead to conclusions about the internal resistance of the battery cell. Currently, excessively high voltage spikes, which are a manifestation of high internal resistance, are considered abnormal, just like excessively low voltage spikes, because voltage differences are also considered in terms of magnitude.

[0126] To select the battery cell with the largest deviation from the reference battery cell in terms of voltage waveform, its segmented signal needs to be extracted. in:

[0127]

[0128] d sec,i,j The calculation is as follows:

[0129]

[0130] This yields a time series for each module and segment. Now, the goal is to reduce the data. This is accomplished using principal component analysis (PCA). This method is suitable here because it identifies the points with the highest information content for each segment. Therefore, redundant information is eliminated. It turns out that, after PCA, one point per segment is sufficient to preserve the essential features. This results in an eigenvector with three terms.

[0131]

[0132] vector item c k、l , where l∈□, 1≤l≤3, represents the principal components of the l-th segment and the k-th module after principal component analysis (PCA) transformation.

[0133] Since the numerical ranges of the different segments can vary considerably, standardization is then performed, resulting in a mean of 0 and an empirical variance of 1. This step significantly improves the performance of many machine learning algorithms. If we now consider the entire training dataset, we obtain the following matrix V, which consists of multiple feature vectors.

[0134]

[0135] Then, standardization is performed using the following transformation:

[0136]

[0137] For the standard deviation σ cl have:

[0138]

[0139] By standardizing features, differences can be highlighted more clearly, and the different orders of magnitude of each feature can be compensated for.

[0140] Then, the feature vector can be represented as points in three-dimensional space. Data points are represented by color according to their group. Figure 9 Showing Figure 2 The feature vectors for the cases in Table 200. Figure 9 The results of feature extraction using a sample dataset are shown. It is clear from this that there are significant differences between the different groups. Therefore, we can conclude that the feature extraction was successful.

[0141] To classify anomalies, several classification algorithms can be tested, such as logistic regression, support vector machine, random forest, and multilayer perceptron.

[0142] Figure 10Table 1000 shows the results of different classification algorithms when trained with k-Means in one implementation scheme. From Figure 10 As can be seen, the support vector machine (SVM) yields the best results, requiring only 3% of the time of the multilayer perceptron to correctly classify the data.

[0143] In one implementation, a single-class support vector machine is used after the classifier to allow for a more rigorous examination of anomalies and normalities. This involves setting a strict threshold for normalities in the feature space based on the normalities contained in the training data. Figure 11 The threshold of 1101, generated by a single-class support vector machine, is displayed. All data points outside this threshold are declared outliers, regardless of the classifier's result. This label is assigned the second-highest probability so it can be associated with the outlier type.

[0144] Figure 12 A flowchart of method 1200 according to one embodiment is shown.

[0145] A method 1200 for battery management includes, as a first method step, obtaining a first voltage waveform on a first battery cell of a battery module and a second voltage waveform on a second battery cell of a battery module, wherein the first voltage waveform has a first voltage waveform portion and a second voltage waveform portion, and the second voltage waveform has a third voltage waveform portion and a fourth voltage waveform portion.

[0146] The method 1200 includes, as a second method step, determining a reference voltage waveform 1203 based on the average voltage of a first voltage waveform and a second voltage waveform, the reference voltage waveform having a first reference voltage waveform portion and a second reference voltage waveform portion.

[0147] Method 1200 includes, as a third method step, comparing a first voltage waveform with a reference voltage waveform 1205 to obtain a first index indicating a first voltage deviation between a portion of the first voltage waveform and a portion of the first reference voltage waveform, and a second voltage deviation between a portion of the second voltage waveform and a portion of the second reference voltage waveform.

[0148] Method 1200 includes, as a fourth method step, comparing a second voltage waveform with a reference voltage waveform 1207 to obtain a second index indicating a third voltage deviation between a third voltage waveform portion and a first reference voltage waveform portion, and a fourth voltage deviation between a fourth voltage waveform portion and a second reference voltage waveform portion.

[0149] Method 1200 includes, as a fifth method step, as follows: if a first indicator is greater than a second indicator, then associating a first electrical feature with a first voltage waveform portion 1209 and associating a second electrical feature with a second voltage waveform portion, so as to classify the battery module by the first electrical feature and the second electrical feature. Alternatively, as a sixth method step, if the first indicator is less than the second indicator, then associating a third electrical feature with a third voltage waveform portion 1211 and associating a fourth electrical feature with a second voltage waveform portion, so as to classify the battery module by the third electrical feature and the fourth electrical feature.

[0150] List of reference numerals

[0151] 100 Battery Management System

[0152] 101 Interface

[0153] 103 processor

[0154] 105 Memory

[0155] 200 Exception Type Table

[0156] 300 Parameter Settings

[0157] 301a Normal situation

[0158] 301b Normal situation

[0159] 301c Normal situation

[0160] 301d is normal.

[0161] 303 Abnormal Case

[0162] 305 Abnormal Situation

[0163] 307 Abnormal Situation

[0164] 400 Voltage Waveform

[0165] 401 Trigger Point

[0166] 403 Interpolation Range

[0167] 500 Voltage Waveform

[0168] 501 Unconventional signal waveforms

[0169] Signal after Discrete Wavelet Transform (DWT)

[0170] 700 Voltage Waveform

[0171] 701 First paragraph

[0172] 702 Second paragraph

[0173] 703 Third paragraph

[0174] 800 Simulated Voltage Waveform

[0175] 801 Reference Signal

[0176] 900 eigenvectors

[0177] Table of results from 1000 different classification algorithms

[0178] 1100 eigenvectors

[0179] 1101 Threshold under normal circumstances

[0180] 1200 Methods for Battery Management

[0181] 1201 First method step: Obtain the first voltage waveform and the second voltage waveform

[0182] 1203 Second method step: Determine the reference voltage waveform

[0183] 1205 Third method step: Compare the first voltage waveform with the reference voltage waveform.

[0184] 1207 Fourth method step: Compare the second voltage waveform with the reference voltage waveform.

[0185] 1209 Fifth method step: Associate the first electrical characteristic with the first voltage waveform portion, and associate the second electrical characteristic with the second voltage waveform portion.

[0186] 1211 Sixth method step: Associate the third voltage waveform portion with the third electrical characteristic, and associate the fourth voltage waveform portion with the fourth electrical characteristic.

Claims

1. A battery management system (100) for classifying battery modules, wherein, The battery module has a first battery cell and a second battery cell, and the battery management system includes... Interface (101), the interface is used to obtain a first voltage waveform on a first battery cell and a second voltage waveform on a second battery cell, wherein the first voltage waveform includes a first voltage waveform portion and a second voltage waveform portion, and the second voltage waveform includes a third voltage waveform portion and a fourth voltage waveform portion; Processor (103), which is designed for A reference voltage waveform is determined based on the average voltage of a first voltage waveform and a second voltage waveform, the reference voltage waveform having a first reference voltage waveform portion and a second reference voltage waveform portion; and A first voltage waveform is compared with a reference voltage waveform to obtain a first index, the first index indicating a first voltage deviation between a first voltage waveform portion and a first reference voltage waveform portion and a second voltage deviation between a second voltage waveform portion and a second reference voltage waveform portion. The second voltage waveform is compared with the reference voltage waveform to obtain a second index, which indicates a third voltage deviation between the third voltage waveform portion and the first reference voltage waveform portion, and a fourth voltage deviation between the fourth voltage waveform portion and the second reference voltage waveform portion. If the first indicator is greater than the second indicator, then the first voltage waveform portion is associated with the first electrical characteristic, and the second voltage waveform portion is associated with the second electrical characteristic, so as to classify the battery module by the first electrical characteristic and the second electrical characteristic; or if the first indicator is less than the second indicator, then the third electrical characteristic is associated with the third voltage waveform portion, and the fourth electrical characteristic is associated with the fourth voltage waveform portion, so as to classify the battery module by the third electrical characteristic and the fourth electrical characteristic, wherein the first electrical characteristic corresponds to the offset voltage of the first battery cell, and the third electrical characteristic corresponds to another offset voltage of the second battery cell, wherein the second electrical characteristic corresponds to the internal resistance of the first battery cell, and the fourth electrical characteristic corresponds to another internal resistance of the second battery cell.

2. The battery management system (100) according to claim 1, wherein the battery management system includes a memory (105) designed for storing a plurality of electrical characteristics, wherein, The processor (103) is designed to read the corresponding electrical features from the memory (105).

3. The battery management system (100) according to claim 1 or 2, wherein, The processor (103) is designed to determine the reference voltage waveform based on the first voltage waveform and the second voltage waveform.

4. The battery management system (100) according to claim 3, wherein, The reference voltage waveform includes the median or average of the first voltage waveform and the second voltage waveform.

5. The battery management system (100) according to claim 1 or 2, wherein, The battery module includes a third battery cell, and the interface is designed to obtain a third voltage waveform on the third battery cell, wherein the processor (103) is designed to determine a reference voltage waveform based on the first voltage waveform, the second voltage waveform, and the third voltage waveform.

6. The battery management system (100) according to claim 5, wherein, The reference voltage waveform includes the median, average, or mode of the first voltage waveform, the second voltage waveform, and the third voltage waveform.

7. The battery management system (100) according to claim 1 or 2, wherein, The processor (103) is designed to: Using principal component analysis, a first electrical characteristic is determined based on a first voltage waveform portion of a first voltage waveform, and a second electrical characteristic is determined based on a second voltage waveform portion of the first voltage waveform, wherein the first electrical characteristic represents the first voltage waveform portion of the first voltage waveform, and the second electrical characteristic represents the second voltage waveform portion of the first voltage waveform, and / or Principal component analysis is used to determine the third electrical feature based on the third voltage waveform portion of the second voltage waveform, and the fourth electrical feature based on the fourth voltage waveform portion of the second voltage waveform. The third electrical feature represents the third voltage waveform portion of the second voltage waveform, and the fourth electrical feature represents the fourth voltage waveform portion of the second voltage waveform.

8. The battery management system (100) according to claim 1 or 2, wherein, The processor (103) is designed to determine a first electrical characteristic based on a first voltage waveform portion of a first voltage waveform, a second electrical characteristic based on a second voltage waveform portion of the first voltage waveform, a third electrical characteristic based on a third voltage waveform portion of the second voltage waveform, and a fourth electrical characteristic based on a fourth voltage waveform portion of the second voltage waveform.

9. The battery management system (100) according to claim 1 or 2, wherein, The processor (103) is designed to classify battery modules using a classification algorithm, wherein the classification algorithm includes at least one of logistic regression, support vector machine, random forest, multilayer perceptron and single-class support vector machine.

10. The battery management system (100) according to claim 1 or 2, wherein, The interface (101) is designed to obtain a plurality of first voltage waveforms and a plurality of second voltage waveforms, wherein each of the plurality of first voltage waveforms corresponds to a first battery cell of one of a plurality of battery modules, and each of the plurality of second voltage waveforms corresponds to a second battery cell of one of a plurality of battery modules, wherein the interface is designed to operate the plurality of battery modules without failure. The processor (103) is designed to classify each of the multiple battery modules to form a reference group based on the classification.

11. The battery management system (100) according to claim 10, wherein, The processor (103) is designed to classify multiple battery modules using an additional classification algorithm, wherein the additional classification algorithm includes at least one of logistic regression, support vector machine, random forest, multilayer perceptron and single-class support vector machine.

12. The battery management system (100) according to claim 1 or 2, wherein, The first voltage waveform has a fifth voltage waveform portion, and the second voltage waveform has a sixth voltage waveform portion, wherein the processor (103) is designed to extract a fifth feature based on the fifth voltage waveform portion of the first voltage waveform and extract a sixth feature based on the sixth voltage waveform portion of the second voltage waveform, wherein the fifth feature corresponds to the capacity of the first battery cell and the sixth feature corresponds to an additional capacity of the second battery cell.

13. The battery management system (100) according to claim 1 or 2, wherein, The interface (101) is designed to obtain the number of groups to be generated, a plurality of additional first voltage waveforms, and a plurality of additional second voltage waveforms, wherein each of the plurality of additional first voltage waveforms corresponds to a first battery cell of one of the plurality of additional battery modules, and wherein each of the plurality of additional second voltage waveforms corresponds to a second battery cell of one of the plurality of additional battery modules. The processor (103) is designed to assign each of the other multiple battery modules to multiple groups using a classification algorithm, wherein the number of multiple groups is equal to the number of groups to be generated.

14. A battery management method (1200), wherein, The method (1200) includes the following steps: Obtain a first voltage waveform on the first battery cell of the (1201) battery module and a second voltage waveform on the second battery cell of the battery module, wherein the first voltage waveform has a first voltage waveform portion and a second voltage waveform portion, and the second voltage waveform has a third voltage waveform portion and a fourth voltage waveform portion; A reference voltage waveform (1203) is determined based on the average voltage of the first voltage waveform and the second voltage waveform, wherein the reference voltage waveform includes a first reference voltage waveform portion and a second reference voltage waveform portion; The first voltage waveform is compared with the reference voltage waveform (1205) to obtain a first index, the first index indicating a first voltage deviation between the first voltage waveform portion and the first reference voltage waveform portion and a second voltage deviation between the second voltage waveform portion and the second reference voltage waveform portion. The second voltage waveform is compared with the reference voltage waveform (1207) to obtain a second index, which indicates a third voltage deviation between the third voltage waveform portion and the first reference voltage waveform portion and a fourth voltage deviation between the fourth voltage waveform portion and the second reference voltage waveform portion. If the first indicator is greater than the second indicator, then the first voltage waveform portion is associated with the first electrical feature, and the second voltage waveform portion is associated with the second electrical feature, so as to classify the battery module according to the first electrical feature and the second electrical feature; or if the first indicator is less than the second indicator, then the third voltage waveform portion is associated with the third electrical feature (1211), and the fourth voltage waveform portion is associated with the fourth electrical feature, so as to classify the battery module by the third electrical feature and the fourth electrical feature, wherein the first electrical feature corresponds to the offset voltage of the first battery cell, and the third electrical feature corresponds to another offset voltage of the second battery cell, wherein the second electrical feature corresponds to the internal resistance of the first battery cell, and the fourth electrical feature corresponds to another internal resistance of the second battery cell.

Images