Abnormal battery detection system and method based on elastic time series similarity network

Abstract

The present invention relates to the technical field of battery anomaly detection, and relates to an abnormal battery detection system and method based on an elastic time series similarity network. The abnormal battery detection system has: an abnormal battery sequence library, which is used to record battery feature time series of different abnormal categories; an input unit, which is used to input the battery feature time series of the battery to be detected; a conversion function model, which is used to obtain the depth sequence; a distance function model, which is used to calculate the Euclidean distance between feature sequences; an elastic function model, which is used to calculate the binary classification results of the difference sequence; a similarity measurement function model, which is used to calculate the similarity score between the battery feature time series and the battery feature time series; and an output unit, which is used to output the abnormal category of the battery to be detected. The method is implemented based on the above system. The present invention can better realize the mining of deep potential features and the alignment of time series, so it has better accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of battery anomaly detection, and in particular to an abnormal battery detection system and method based on an elastic time series similarity network. Background Art

[0002] Battery safety is undoubtedly the top priority in the battery production and application process. How to discover potential safety hazards in the battery manufacturing and application process and identify abnormally high-risk batteries in advance is a key issue worthy of continuous research.

[0003] In the actual application of batteries, we can monitor in real time the changes of battery characteristics such as current, voltage, resistance value, etc. during the charging and discharging process. Therefore, we can compare the changes between the characteristic time series of the current battery and the characteristic time series of the abnormal battery to realize the abnormal battery identification one step faster.

[0004] Similarity metric learning is an established task in machine learning and has been widely studied in recent years. For example, in the context of image retrieval, similarity learning is used for features such as “search by image” in search engine applications. Existing similarity metrics for time series are often based on statistical methods or simple and straightforward deep neural network methods.

[0005] In deep neural network methods, learning is usually based on twin networks, that is, using the same network to obtain the representation results of a pair of instances. For example, the same neural network structure is directly used to calculate L1 or L2 distance, cosine similarity, etc.

[0006] Existing methods face the following two problems when applied to battery anomaly detection:

[0007] (1) Statistical methods or simple and direct deep neural network methods cannot further capture the deep potential changes of time series;

[0008] (2) Since the time dimensions of different sequences are often inconsistent, this complexity makes it impossible to effectively align the time dimensions of the sequences, so ideal results cannot be achieved in practical applications. Summary of the invention

[0009] The present invention provides an abnormal battery detection system and method based on an elastic time series similarity network, which can overcome the defects of the existing similarity measurement algorithm in battery anomaly detection, such as difficulty in capturing the deep potential changes of the time series and difficulty in aligning the time series.

[0010] According to the present invention, an abnormal battery detection system based on elastic time series similarity network has:

[0011] Abnormal battery sequence library P, which is used to record the battery feature time series corresponding to different abnormal categories r ;

[0012] An input unit is used to input the battery characteristic time series of the battery to be tested ;

[0013] Transfer function model , which is used to analyze the battery characteristic time series and battery characteristic time series Processing is performed to obtain the depth sequence of the battery to be tested and the depth sequence of abnormal categories ;

[0014] Distance function model , which is used to calculate the depth sequence The characteristic sequence at any i-th time point in , and the depth sequence The characteristic sequence at any j-th time point The Euclidean distance between

[0015] Elasticity Function Model , which is used to calculate any feature sequence With any characteristic sequence The binary classification result of the difference sequence;

[0016] Similarity metric function model S(*), which is used to calculate the battery characteristic time series and battery characteristic time series The similarity score between ,

[0017]

[0018] Where T is the length of the time series; and

[0019] The output unit is used to output the abnormality category of the battery with the highest similarity score with the battery to be detected.

[0020] In the system of the present invention, the constructed abnormal battery sequence library P can better record the characteristic time series of batteries of different abnormal categories, so that the abnormal category identification of the unknown abnormal category of the battery to be detected can be better realized.

[0021] In the present invention, the conversion function model And the distance function model It is possible to align the battery feature time series of the battery to be detected and the battery in the abnormal battery sequence library P and to discover the deep potential features, thereby making the recognition result more accurate.

[0022] In the present invention, the abnormal category of the battery to be detected can be preferably output through the similarity measurement function model S(*) and the output unit, so that the detection result is more intuitive and clear.

[0023] Preferably, the transfer function model include,

[0024] Transformation input layer, which is used to input battery feature time series or battery characteristic time series , and obtain the battery characteristic time series or battery characteristic time series The characteristic sequence at each time point and ;

[0025] The first conversion fully connected layer has D neurons and is used to input feature sequences or ;

[0026] The second conversion fully connected layer has K neurons and uses the output of the first conversion fully connected layer as input, and is used to output the feature sequence corresponding to each time point or ;

[0027] The transformation fusion output layer is used to transform each feature sequence in the order of the time dimension. or Integrate and output the depth sequence or ;

[0028] Where D is the number of feature dimensions.

[0029] In the present invention, the conversion function model constructed can be defined as, ; Where K is the transformed function model The number of feature dimensions after processing; that is, through the transformation function model Can better convert the original battery characteristic time series Express in new form In this process, due to the role of the first conversion fully connected layer and the second conversion fully connected layer, it is possible to fully explore the deep potential changes of the time series, so that the detection results are more reasonable.

[0030] As a preference, the distance function model The expression is as follows,

[0031]

[0032] in, Represents a feature sequence The kth element in Represents a feature sequence Therefore, the Euclidean distance can be used as a criterion for similarity scoring.

[0033] As a preference, the elasticity function model include,

[0034] Elastic input layer, which is used to calculate the feature sequence and The difference sequence of

[0035] The first elastic fully connected layer has K neurons and takes the difference sequence as input;

[0036] The second elastic fully connected layer has 1 neuron and uses the output of the first elastic fully connected layer as input;

[0037] The elastic output layer processes the output of the second elastic fully connected layer based on the Sigmod function and outputs 0 when the result is less than 0.5, otherwise it outputs 1.

[0038] In the present invention, the elasticity function model constructed can be defined as, Therefore, it is possible to better determine whether the context information of the i-th and j-th data in the two sequences are aligned, that is, it is possible to better realize the battery feature time series or battery characteristic time series It can realize the similarity judgment on the premise of time series alignment.

[0039] In addition, the present invention also provides an abnormal battery detection method based on an elastic time series similarity network, which realizes abnormal battery detection based on any of the above abnormal battery detection systems based on an elastic time series similarity network, thereby being able to better realize the recognition of battery abnormalities.

[0040] As a preference, the method specifically comprises the following steps:

[0041] Step S1, select battery characteristic time series from the abnormal battery sequence library P one by one ;

[0042] Step S2: Input the battery characteristic time series of the battery to be tested through the input unit ;

[0043] Step S3: Transform the function model , for the battery characteristic time series And the battery characteristic time series is Processing is performed to obtain the depth sequence of the battery to be tested and the depth sequence of the rth abnormal category ;

[0044] Step S4: Based on the distance function model , calculate the depth sequence The characteristic sequence at any i-th time point in , and the depth sequence The characteristic sequence at any j-th time point The Euclidean distance between

[0045] Step S5: Based on elasticity function model , calculate any feature sequence With any characteristic sequence The binary classification result of the difference sequence;

[0046] Step S6: Calculate the battery characteristic time series based on the similarity measurement function model S(*) and battery characteristic time series The similarity score between ,

[0047]

[0048] Where T is the length of the time series;

[0049] Step S7, repeating steps S1-S7, obtaining similarity scores between the battery to be detected and batteries of each abnormal category in the abnormal battery sequence library P, and outputting the abnormal category with the highest similarity score through the output unit.

[0050] Through the above, abnormality detection of the battery to be detected and output of the abnormality category can be preferably achieved.

[0051] Preferably, in step S1, , r=0 is the normal category.

[0052] Therefore, the identification of batteries without abnormalities can be better achieved.

[0053] As a preferred embodiment, the battery characteristic time series is and battery characteristic time series The samples are collected during the operation of the corresponding battery, so the construction of the battery characteristic time series can be better realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a schematic block diagram of the abnormal battery detection system in Example 1;

[0055] Figure 2 The curve showing the change of charging current of a normal battery on the time axis;

[0056] Figure 3 It is a curve showing the change of charging current on the time axis of a battery with abnormal constant current of the charging cell being too large;

[0057] Figure 4 It is a curve showing the change of charging current on the time axis of a battery with abnormally large internal resistance;

[0058] Figure 5 The curve showing the change of charging current on the time axis of a battery with a completely damaged battery cell. DETAILED DESCRIPTION

[0059] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0060] Example 1

[0061] Based on the defects of the existing similarity metrics, this embodiment provides an abnormal battery detection solution based on elastic time series similarity network technology. First, we construct an abnormal battery sequence change database, and then design an elastic network to simultaneously achieve: (1) align the time dimensions of the sequences of the battery to be detected and the abnormal library; (2) learn the aligned deep latent representation, and then perform similarity comparison based on this deep latent representation; ultimately, it is possible to more accurately determine whether the battery has an abnormal risk, thereby effectively reducing the safety risks of the battery.

[0062] The time series described in this embodiment, also called a dynamic series, refers to a series in which the values ​​of the same statistical indicator are arranged in the order of their occurrence time.

[0063] The similarity measure described in this embodiment is a measure for comprehensively evaluating the similarity between two objects. The closer the two objects are, the greater their similarity measure is, and the farther the two objects are, the smaller their similarity measure is.

[0064] In the solution provided in this embodiment, a system and method for detecting abnormal batteries based on an elastic time series similarity network are provided.

[0065] Combination Figure 1 As shown, in the abnormal battery detection system of this embodiment, it has:

[0066] Abnormal battery sequence library P, which is used to record the battery feature time series corresponding to different abnormal categories r ;

[0067] An input unit is used to input the battery characteristic time series of the battery to be tested ;

[0068] Transfer function model , which is used to analyze the battery characteristic time series and battery characteristic time series Processing is performed to obtain the depth sequence of the battery to be tested and the depth sequence of abnormal categories ;

[0069] Distance function model , which is used to calculate the depth sequence The characteristic sequence at any i-th time point in , and the depth sequence The characteristic sequence at any j-th time point The Euclidean distance between

[0070] Elasticity Function Model , which is used to calculate any feature sequence With any characteristic sequence The binary classification result of the difference sequence;

[0071] Similarity metric function model S(*), which is used to calculate the battery characteristic time series and battery characteristic time series The similarity score between ,

[0072]

[0073] Where T is the length of the time series; and

[0074] The output unit is used to output the abnormality category of the battery with the highest similarity score with the battery to be detected.

[0075] In the system of this embodiment, the constructed abnormal battery sequence library P can better record the characteristic time series of batteries of different abnormal categories, so that the abnormal category identification of the unknown abnormal category of the battery to be detected can be better realized.

[0076] In this embodiment, through the conversion function model And the distance function model It is possible to align the battery feature time series of the battery to be detected and the battery in the abnormal battery sequence library P and to discover the deep potential features, thereby making the recognition result more accurate.

[0077] In this embodiment, the abnormality category of the battery to be detected can be preferably output through the similarity measurement function model S(*) and the output unit, so that the detection result is more intuitive and clear.

[0078] The abnormal battery sequence library in this embodiment can cover battery feature time series under various battery abnormal conditions; it can be understood that the richer the data in the abnormal battery sequence library, the more conducive it is to battery abnormality detection.

[0079] In this embodiment, the battery characteristic time series is described by taking the time series change of a single charging current during the charging process as an example.

[0080] See Figure 2 , which is a curve of the change of charging current of a normal battery on the time axis. It is first a stable process. After the charging is completed, the current will drop rapidly and smoothly to a lower level in a short time, indicating that the charging is completed at this time.

[0081] See Figure 3 , which is a curve of the change of charging current on the time axis of a battery with an abnormal constant current of the charging cell being too large. The current fluctuates violently during the charging process, and the current oscillation amplitude is too large, indicating that the charging cell encounters a problem of excessive cross current, so it is abnormal.

[0082] See Figure 4 , which is a curve of the change of charging current on the time axis of a battery with abnormally large internal resistance. It is relatively smooth during the charging process, but there is a large current fluctuation at the end of charging, indicating that the resistance is too large when the current changes, that is, the internal resistance is too large, so it is abnormal.

[0083] See Figure 5 , which is a curve of the change of charging current on the time axis of a battery with a completely damaged battery cell. Whether charging or ending charging, it immediately jumps to another gear, and there is no smooth transition process when the current is switched, so it is abnormal.

[0084] The abnormal battery sequence library P constructed in this embodiment records the battery feature time series under several abnormal categories; for the rth abnormal category, the corresponding battery feature time series is , , is the number of time dimensions (i.e., the length of the time series), and D is the number of feature dimensions; where , r=0 is the normal category.

[0085] It can be understood that the battery characteristic time series in this embodiment is and have the same dimensions, i.e. .

[0086] It can be understood that the battery feature time series used in this embodiment has two dimensions, one is the time dimension and the other is the feature dimension; the length of the time dimension is T, which represents the total number of sampling points; the length of the feature dimension is D, which represents the number of features used. Among them, features such as the number of times the battery is used, the current data and voltage data that change over time, and the battery life can be selected in the feature dimension. When using this type of feature to construct a battery feature time series, it is necessary to perform data processing such as missing data and uneven data distribution. In this embodiment, the mean interpolation method can be used to process the missing data. Only when the data can be normalized, the standardized processing of the battery feature time series can be better achieved.

[0087] Those skilled in the art should be aware that the preprocessing of such data can be performed according to a conventional data preprocessing process, so it will not be described in detail in this embodiment.

[0088] The conversion function model of this embodiment include,

[0089] Transformation input layer, which is used to input battery feature time series or battery characteristic time series , and obtain the battery characteristic time series or battery characteristic time series The characteristic sequence at each time point and ;

[0090] The first conversion fully connected layer has D neurons and is used to input feature sequences or ;

[0091] The second conversion fully connected layer has K neurons and uses the output of the first conversion fully connected layer as input, and is used to output the feature sequence corresponding to each time point or ;

[0092] The transformation fusion output layer is used to transform each feature sequence in the order of the time dimension. or Integrate and output the depth sequence or ;

[0093] Where D is the number of feature dimensions.

[0094] In this embodiment, the constructed conversion function model can be defined as, ; Where K is the transformed function model The number of feature dimensions after processing; that is, through the transformation function model Can better convert the original battery characteristic time series Express in new form In this process, due to the role of the first conversion fully connected layer and the second conversion fully connected layer, it is possible to fully explore the deep potential changes of the time series, so that the detection results are more reasonable.

[0095] That is, in this embodiment, , .

[0096] The distance function model of this embodiment The expression is as follows,

[0097]

[0098] in, Represents a feature sequence The kth element in Represents a feature sequence Therefore, the Euclidean distance can be used as a criterion for similarity scoring.

[0099] The elasticity function model of this embodiment include,

[0100] Elastic input layer, which is used to calculate the feature sequence and The difference sequence of

[0101] The first elastic fully connected layer has K neurons and takes the difference sequence as input;

[0102] The second elastic fully connected layer has 1 neuron and uses the output of the first elastic fully connected layer as input;

[0103] The elastic output layer processes the output of the second elastic fully connected layer based on the Sigmod function and outputs 0 when the result is less than 0.5, otherwise it outputs 1.

[0104] In this embodiment, the elasticity function model constructed can be defined as, Therefore, it is possible to better determine whether the context information of the i-th and j-th data in the two sequences are aligned, that is, it is possible to better realize the battery feature time series or battery characteristic time series It can realize the similarity judgment on the premise of time series alignment.

[0105] For example, the feature sequence and feature sequence Elasticity Function Model The processed output value is 1, and the two feature sequences are aligned in the time dimension.

[0106] In this embodiment, the expression of the Sigmod function is: , which is a typical binary classification function, making the elastic function model The output of the function is a binary classification result of 0 or 1.

[0107] The similarity measurement function model S(*) in this embodiment can be defined as: , which converts all pairwise elastic distances into similarity scores that satisfy the value range of [0,1] through the exponential function of negative distance exp, and can better realize the elastic alignment process of generalizing all paired vectors globally.

[0108] In summary, in the solution of this embodiment, two feature time series of the same dimension (T×D) can be obtained first; then based on the conversion function model They can all be converted into T×K dimensional expressions. In this process, due to the use of the fully connected layer, the deep potential connections of the feature time series can be better characterized; then, the distance function model is used and elasticity function model , it is possible to evaluate the similarity under the premise of time series alignment, thus making the detection results more accurate.

[0109] In addition, the conversion function model in this embodiment and elasticity function model , which is essentially a fully connected neural network. The various parameters of the neural network need to be obtained through training. In this embodiment, conventional training methods such as cross entropy function can be used as loss function to realize the conversion function model. and elasticity function model of convergence.

[0110] Those skilled in the art should be aware that the above-mentioned training method for the neural network is an existing mature technology, so it will not be described in detail in this embodiment.

[0111] Based on the system provided in this embodiment, this embodiment also provides an abnormal battery detection method based on an elastic time series similarity network, which realizes abnormal battery detection based on the above-mentioned abnormal battery detection system based on an elastic time series similarity network. Therefore, it is possible to better realize the identification of battery abnormalities.

[0112] The method of this embodiment specifically includes the following steps:

[0113] Step S1, select battery characteristic time series from the abnormal battery sequence library P one by one ;

[0114] Step S2: Input the battery characteristic time series of the battery to be tested through the input unit ;

[0115] Step S3: Transform the function model , for the battery characteristic time series And the battery characteristic time series is Processing is performed to obtain the depth sequence of the battery to be tested and the depth sequence of the rth abnormal category ;

[0116] Step S4: Based on the distance function model , calculate the depth sequence The characteristic sequence at any i-th time point in , and the depth sequence The characteristic sequence at any j-th time point The Euclidean distance between

[0117] Step S5: Based on elasticity function model , calculate any feature sequence With any characteristic sequence The binary classification result of the difference sequence;

[0118] Step S6: Calculate the battery characteristic time series based on the similarity measurement function model S(*) and battery characteristic time series The similarity score between ,

[0119]

[0120] Where T is the length of the time series;

[0121] Step S7, repeating steps S1-S7, obtaining similarity scores between the battery to be detected and batteries of each abnormal category in the abnormal battery sequence library P, and outputting the abnormal category with the highest similarity score through the output unit.

[0122] Through the above, abnormality detection of the battery to be detected and output of the abnormality category can be preferably achieved.

[0123] In step S1 of this embodiment, , r=0 is a normal category. Therefore, it is possible to better identify batteries without abnormalities.

[0124] In this embodiment, the battery characteristic time series is: and battery characteristic time series The samples are collected during the operation of the corresponding battery, so the construction of the battery characteristic time series can be better realized.

[0125] In order to further illustrate S1-S7 of the present case, this embodiment simply gives the following exemplary description.

[0126] In this illustrative illustration:

[0127] In step S1, the selected battery characteristic time series is for,

[0128]

[0129] In step S2, the battery characteristic time series of the battery to be tested for,

[0130]

[0131] That is, the time dimension of the constructed battery characteristic time series is 4, and the characteristic dimension is 3; that is, T=4, D=3;

[0132] Among them, the three selected features are charging current, charging voltage and battery cycle times.

[0133] After processing in step S3, the depth sequence and depth sequence They are respectively,

[0134]

[0135]

[0136] That is, the transfer function model The feature dimension can be compressed into 2 dimensions; that is, K=2.

[0137] After processing in step S4, the Euclidean distance between any two feature sequences can be calculated, and a total of 16 results of 4*4 can be obtained, which can be expressed as follows:

[0138] 1.00 3.61 2.83 1.41 2.24 1.00 1.41 2.00 1.41 3.16 3.00 2.24 1.00 3.61 3.16 2.00

[0139] After step S5, the difference sequences of any two feature sequences can be classified into two categories. It can be understood that there are 16 difference sequences in total, which are:

[0140]

[0141]

[0142]

[0143]

[0144] Afterwards, the 16 difference sequences are analyzed using the elastic function model. After processing, 16 corresponding binary classification values ​​can be obtained, which are:

[0145] 0,0,0,0,

[0146] 0,0,1,0,

[0147] 0,0,0,1,

[0148] 0, 0, 0, 0;

[0149] After that, after processing in step S6, the similarity score can be obtained as 0.79; so far, the similarity score between the battery feature time series of the battery to be detected and one of the battery feature time series in the abnormal battery sequence library P has been obtained;

[0150] Afterwards, through step S7, all similarity scores can be obtained, the highest score is 0.96, and the corresponding abnormality category is normal. At this time, it is output that there is no abnormality in the battery to be tested.

[0151] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.

Claims

1. An abnormal battery detection system based on elastic time series similarity network, which has: Abnormal battery sequence library P, which is used to record the battery feature time series corresponding to different abnormal categories r ; An input unit is used to input the battery characteristic time series of the battery to be tested ; Transfer function model , which is used to analyze the battery characteristic time series and battery characteristic time series Processing is performed to obtain the depth sequence of the battery to be tested and the depth sequence of abnormal categories ; Distance function model , which is used to calculate the depth sequence The characteristic sequence at any i-th time point in , and the depth sequence The characteristic sequence at any j-th time point The Euclidean distance between Elasticity Function Model , which is used to calculate any feature sequence With any characteristic sequence The binary classification result of the difference sequence; Similarity metric function model S(*), which is used to calculate the battery characteristic time series and battery characteristic time series The similarity score between , ； in, T is the length of the time series; as well as An output unit, which is used to output the abnormal category of the battery with the highest similarity score with the battery to be detected; Among them, the conversion function model include, Transformation input layer, which is used to input battery feature time series or battery characteristic time series , and obtain the battery characteristic time series or battery characteristic time series The characteristic sequence at each time point and ; The first conversion fully connected layer has D neurons and is used to input feature sequences or ; The second conversion fully connected layer has K neurons and uses the output of the first conversion fully connected layer as input, and is used to output the feature sequence corresponding to each time point or ; The transformation fusion output layer is used to transform each feature sequence in the order of the time dimension. or Integrate and output the depth sequence or ; Where D is the number of feature dimensions.

2. The abnormal battery detection system based on elastic time series similarity network according to claim 1 is characterized in that: Distance function model The expression is as follows, ； in, Represents a feature sequence The kth element in Represents a feature sequence The kth element in .

3. The abnormal battery detection system based on elastic time series similarity network according to claim 1 is characterized in that: Elasticity Function Model include, Elastic input layer, which is used to calculate the feature sequence and The difference sequence of The first elastic fully connected layer has K neurons and takes the difference sequence as input; The second elastic fully connected layer has 1 neuron and uses the output of the first elastic fully connected layer as input; The elastic output layer processes the output of the second elastic fully connected layer based on the Sigmod function and outputs 0 when the result is less than 0.5, otherwise it outputs 1.

4. An abnormal battery detection method based on an elastic time series similarity network, which realizes abnormal battery detection based on any of the above-mentioned abnormal battery detection systems based on an elastic time series similarity network.

5. The abnormal battery detection method based on elastic time series similarity network according to claim 4 is characterized in that: The specific steps include: Step S1, select battery characteristic time series from the abnormal battery sequence library P one by one ; Step S2: Input the battery characteristic time series of the battery to be tested through the input unit ; Step S3: Transform the function model , for the battery characteristic time series And the battery characteristic time series is Processing is performed to obtain the depth sequence of the battery to be tested and the depth sequence of the rth abnormal category ; Step S4: Based on the distance function model , calculate the depth sequence The characteristic sequence at any i-th time point in , and the depth sequence The characteristic sequence at any j-th time point The Euclidean distance between Step S5: Based on elasticity function model , calculate any feature sequence With any characteristic sequence The binary classification result of the difference sequence; Step S6: Calculate the battery characteristic time series based on the similarity measurement function model S(*) and battery characteristic time series The similarity score between , ； Where T is the length of the time series; Step S7, repeating steps S1-S7, obtaining similarity scores between the battery to be detected and batteries of each abnormal category in the abnormal battery sequence library P, and outputting the abnormal category with the highest similarity score through the output unit.

6. The abnormal battery detection method based on elastic time series similarity network according to claim 5 is characterized in that: In step S1, , r=0 is the normal category.

7. The abnormal battery detection method based on elastic time series similarity network according to claim 5 is characterized in that: The battery characteristic time series is and battery characteristic time series The samples are collected during the operation of the corresponding batteries.

Images