A method and device for detecting internal short circuit of lithium-ion battery based on improved GAF

Abstract

The present invention relates to a lithium-ion battery internal short circuit detection method based on improved GAF, which belongs to the field of lithium battery detection, including: amplifying voltage deviation by curve Manhattan distance, selecting a reference curve to divide all curves, and calculating the curve Manhattan distance of the internal short circuit curve and the reference curve through a sliding window; converting one-dimensional data into a two-dimensional image through an improved Gram angle field; dividing the severity of the short circuit into z categories according to the application scenario, correspondingly constructing z Gram angle field image data sets, and using a model to classify the image data sets, and obtaining the possibility of each category according to the probability corresponding to the classification result. The present invention uses a sliding window to calculate the Manhattan distance of the curve, which not only amplifies the voltage deviation feature, but also retains the time characteristics of the data; by improving the Gram angle field, images with different internal short circuit degrees present different contour features, which can effectively detect and track the internal short circuit fault of the battery.

Description

Technical Field

[0001] The present invention relates to the field of lithium battery detection, and in particular to a lithium ion battery internal short circuit detection method and device based on improved GAF. Background Art

[0002] Internal short circuit (ISC) is the main cause of thermal runaway in lithium-ion batteries. Research on the diagnosis method of ISC fault is of great significance for early warning of battery thermal runaway and avoiding major accidents. However, in the process of ISC evolution, the short-circuit resistance value continues to decrease. How to effectively use the characteristics to track the severity of ISC faults is a problem that needs to be considered at present. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method and device for detecting internal short circuits in lithium-ion batteries based on an improved GAF, thereby solving the deficiencies of the prior art.

[0004] The purpose of the present invention is achieved by the following technical solution: a method for detecting short circuits in lithium-ion batteries based on improved GAF, the short circuit detection method comprising:

[0005] Step 1: amplify the voltage deviation by curve Manhattan distance, select the reference curve to divide all the curves, and calculate the curve Manhattan distance between the internal short-circuit curve and the reference curve through a sliding window;

[0006] Step 2: Combine the Manhattan distances of all small windows into a one-dimensional time series D , one-dimensional data is converted into two-dimensional images through the improved Gram angle field;

[0007] Step 3: Classify the severity of the short circuit into z categories according to the application scenario, construct z Gram angle field image datasets accordingly, and use the model to classify the image datasets, and obtain the possibility of each category according to the probability corresponding to the classification result.

[0008] The step one comprises:

[0009] In order to amplify the voltage deviation, the Manhattan distance calculation formula is set as ,in, and The curves are l 1 and l 2 of l points;

[0010] Select a healthy battery charging voltage curve as the reference curve, and use the deviation of other charging curves from the reference curve as the basis for detecting internal short circuits. The time length of all curves is TThen the window of length T The window is divided into multiple lengths m A small window;

[0011] Use the sliding window to calculate the Manhattan distance between the internal short-circuit curve and the reference curve in each small window, and get the Manhattan distance of the curve in each small window: , d i Indicates i The Manhattan distance of the curve of a small window, V R,m(i-1)+1 , V R,m(i-1)+2 and V R,mi Respectively represent the first m ( i- 1) + 1. m ( i- 1) + 2 and my A sample voltage, V ISC,m(i-1)+1 , V ISC,m(i-1)+2 and V ISC,mi Respectively represent the internal short-circuit curve m ( i- 1) + 1. m ( i- 1) + 2 and my A sampling voltage.

[0012] The second step comprises:

[0013] The curve Manhattan distance of all small windows is combined into a one-dimensional time series , and scaled to the interval [0, 1], the time series D After normalization, we get the point , convert the scaled value into polar coordinates ,in, i i and r i Points The phase angle and radius, t i Indicate point The timestamp of n The number of data in the time series;

[0014] The time series D As radians, calculate the absolute value function of the cosine value and get the sequence ,in, ,in a is the gain factor in radians, and the phase angle is expressed as ;

[0015] Set an arc , through the formula Calculate the radian , converting the phase angle into ;

[0016] Finally, through the Gram matrix Convert one-dimensional data into a two-dimensional Gram angular field image.

[0017] The step three comprises:

[0018] The severity of the short circuit is divided into z categories according to the application scenario, and z Gram angle field image datasets are constructed accordingly;

[0019] The texture features of the Gram angular field image are extracted through the convolutional layer in the ResNet50 model, and then batch normalization is performed and activated, and then the pooling layer is used to reduce the dimension of the output feature map.

[0020] The output of the pooling layer is processed by the fully connected layer, which has two output feature numbers, 128 and z. Finally, the input data is divided into z categories. The classification result is expressed in the form of probability, which indicates the possibility that the input data belongs to each category.

[0021] A lithium-ion battery internal short circuit detection device based on improved GAF, the device comprising a curve Manhattan distance calculation module, a data conversion module and a model classification module;

[0022] The curve Manhattan distance calculation module is used to amplify the voltage deviation by the curve Manhattan distance, select the reference curve to divide all the curves, and calculate the curve Manhattan distance between the internal short-circuit curve and the reference curve through a sliding window;

[0023] The data conversion module is used to convert the Manhattan distance of all small windows into a one-dimensional time series. D , one-dimensional data is converted into two-dimensional images through the improved Gram angle field;

[0024] The model classification module is used to classify the severity of the short circuit into z categories according to the application scenario, construct z Gram angle field image data sets accordingly, and use the model to classify the image data sets, and obtain the possibility of each category according to the probability corresponding to the classification result.

[0025] The curve Manhattan distance calculation module specifically includes the following contents:

[0026] In order to amplify the voltage deviation, the Manhattan distance calculation formula is set as ,in, and The curves are l 1 and l 2 of l points;

[0027] Select a healthy battery charging voltage curve as the reference curve, and use the deviation of other charging curves from the reference curve as the basis for detecting internal short circuits. The time length of all curves is T Then the window of length T The window is divided into multiple lengths m A small window;

[0028] Use the sliding window to calculate the Manhattan distance between the internal short-circuit curve and the reference curve in each small window, and get the Manhattan distance of the curve in each small window: , d i Indicates i The Manhattan distance of the curve of a small window, V R,m(i-1)+1 , V R,m(i-1)+2 and V R,mi Respectively represent the first m ( i- 1) + 1. m ( i- 1) + 2 and my A sample voltage, V ISC,m(i-1)+1 , V ISC,m(i-1)+2 and V ISC,mi Respectively represent the internal short-circuit curve m ( i- 1) + 1. m ( i- 1) + 2 and my A sampling voltage.

[0029] The data conversion module specifically includes the following contents:

[0030] The curve Manhattan distance of all small windows is combined into a one-dimensional time series , and scaled to the interval [0, 1], the time series D After normalization, we get the point , convert the scaled value into polar coordinates ,in, i i and r i Points The phase angle and radius, t i Indicate point The timestamp of n The number of data in the time series;

[0031] The time series D As radians, calculate the absolute value function of the cosine value and get the sequence ,in, ,in a is the gain factor in radians, and the phase angle is expressed as ;

[0032] Set an arc , through the formula Calculate the radian , converting the phase angle into ;

[0033] Finally, through the Gram matrix Convert one-dimensional data into a two-dimensional Gram angular field image.

[0034] The model classification module specifically includes the following contents:

[0035] The severity of the short circuit is divided into z categories according to the application scenario, and z Gram angle field image datasets are constructed accordingly;

[0036] The texture features of the Gram angular field image are extracted through the convolutional layer in the ResNet50 model, and then batch normalization is performed and processed by the activation function, and then the pooling layer is passed to reduce the dimension of the output feature map.

[0037] The output of the pooling layer is processed by the fully connected layer, which has two output feature numbers, 128 and z. Finally, the input data is divided into z categories. The classification result is expressed in the form of probability, which indicates the possibility that the input data belongs to each category.

[0038] The present invention has the following advantages: a lithium-ion battery internal short circuit detection method and device based on improved GAF, using a sliding window to calculate the Manhattan distance of a curve, which not only amplifies the voltage deviation characteristics but also retains the time characteristics of the data; converts one-dimensional data into an image through the Gram angle field, and the characteristics of the image data are richer than traditional data; by improving the Gram angle field, images with different internal short circuit degrees present different contour features, which can effectively detect and track the internal short circuit fault of the battery. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the process of the present invention;

[0040] Figure 2 Schematic diagram of calculating CMD for a sliding window;

[0041] Figure 3 Schematic diagram of the conversion of the CMD curve. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below in conjunction with the drawings is not intended to limit the scope of protection of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.

[0043] like Figure 1 As shown, the present invention specifically relates to a lithium-ion battery internal short circuit detection method based on improved GAF, which converts battery voltage data into an image through Gramian Angular Field (GAF), uses cosine function to improve GAF, so that the generated image can reflect the fault degree of the internal short circuit, and finally combines ResNet50 neural network to classify the fault degree. It can track the severity of the fault during the evolution of ISC and provide a reference for the safety monitoring of the battery management system of electric vehicles, which specifically includes the following contents.

[0044] 1. A battery with ISC (internal short circuit) fault will have a voltage drop relative to a healthy battery, and the voltage drop increases as the short circuit resistance decreases. The voltage deviation of the battery can be measured by the Manhattan distance (MD). MD represents the sum of the horizontal and vertical distances between two points in two-dimensional space. The traditional MD is defined as:

[0045] (1),

[0046] in Represents the MD between the point (x1, y1) and the point (x2, y2).

[0047] In order to amplify the voltage deviation, the Curvilinear Manhattan Distance (CMD) formula is defined as follows:

[0048] (2),

[0049] in, and The curves are l 1 and l 2 of l points.

[0050] 2. Select a healthy battery charging voltage curve as the reference curve, and use the deviation of other charging curves from the reference curve as the basis for detecting ISC. The intercept time length for all curves is T Then the window of length T The window is divided into multiple lengths m small window.

[0051] 3. If Figure 2 As shown, taking the reference curve (solid line) and the voltage curve when ISC occurs (dashed line) as examples, a sliding window is used to calculate the CMD of the ISC curve and the reference curve in each small window. The calculation formula of CMD in each small window is expressed as:

[0052] (3),

[0053] in, d i Indicates i The Manhattan distance of the curve of a small window, V R,m(i-1)+1 , V R,m(i-1)+2 and V R,mi Respectively represent the first m ( i- 1) + 1. m ( i- 1) + 2 and my A sample voltage, V ISC,m(i-1)+1 , V ISC,m(i-1)+2 and V ISC,mi Respectively represent the internal short-circuit curve m ( i- 1) + 1. m ( i- 1) + 2 and myThe sampling frequency is set to 1Hz.

[0054] 4. For one-dimensional time series , the first step of GAF is to scale it to the interval [0,1], so that the inverse cosine of the data can be taken later. The time series data D is normalized as follows:

[0055] (4),

[0056] Among them, n is the number of data in the time series;

[0057] The second step is to convert the scaled values ​​into polar coordinates:

[0058] (5),

[0059] in, i i and r i Points The phase angle and radius, t i Indicate point timestamp, The value range of is [0,π]. r i The time relationship is preserved, the phase angle i i The numerical relationship is preserved.

[0060] 5. Two different CMD curves A and B experience the following Figure 3 In the process shown, when the battery suffers from different internal short-circuit faults, the original CMD curves generated have different trends. However, after scaling by formula (4), the trends of the curves tend to be consistent. Correspondingly, two similar polar coordinate curves will be further obtained through formula (5). The original CMD sequence D is unidirectionally increasing. No matter how large the amplitude is at the end, it maintains similar polar coordinate features after the GAF compression process. Similar features will make the final generated image difficult to distinguish. In order to solve the above problem, the time series D is taken as radians, and the absolute value function of the cosine value is calculated to obtain the sequence ,in:

[0061] (6),

[0062] in, a is the gain factor in radians.

[0063] Combining equations (5) and (6), we can see that when the input of GAF is replaced from sequence D to sequence K, the phase angle i i It is expressed as:

[0064] (7),

[0065] Define an arc ,radian It can be expressed by formula (8):

[0066] (8),

[0067] Combining formula (7) and (8) we get:

[0068] (9),

[0069] Formula (9) shows that the amplitude of CMD will change the phase angle i i The changing trend is specifically manifested as follows: Each time a length of The interval of i i After the cosine function transformation, for more serious ISC faults, the greater the Manhattan distance end value, the more frequent the data direction changes.

[0070] Finally, the one-dimensional data is converted into a two-dimensional image through the Gram matrix. The expression of the Gram matrix G is:

[0071] (10),

[0072] 6. According to the application scenario, the severity of the internal short circuit is divided into z categories, and z GAF image datasets are constructed accordingly. The image dataset is classified using the ResNet50 model, which includes convolution, batch normalization, activation function, maximum pooling, fully connected (FC) layer and calculation of classification probability. ResNet50 contains a total of 49 convolution layers and 1 FC layer. Except for the first convolution layer using a 7*7 convolution kernel, the size of the remaining convolution kernels is 1*1 or 3*3. The texture features of the GAF image are extracted using the convolution layer, which can be described as:

[0073] (11),

[0074] in, C j is the convolution kernel, is the output feature map of the previous layer, l -1 indicates the previous layer number, b j is the deviation, j Indicates j convolution kernels, wis the output feature map of the current convolutional layer.

[0075] The curve is then batch normalized and passed through an activation function:

[0076] (12),

[0077] in represents the batch normalization operation, It is the Relu activation function, which enables the network to represent complex features. The ResNet network uses the Relu function to speed up convergence and reduce overfitting of model training.

[0078] The pooling layer is used to reduce the dimensionality of the output feature map:

[0079] (13),

[0080] in, is the pooling area, is the maximum pooling output, and the combination of all pooling outputs is the final output.

[0081] Output the final pooling After being processed by the FC layer, the output vector of the FC layer h The calculation formula is:

[0082] (14),

[0083] in, W F and b F are the weight matrix and bias vector of the FC layer respectively.

[0084] The last layer of the original ResNet50 model is an FC layer with 1000 output features (corresponding to 1000 categories). In order to meet the classification requirements of the task, this FC layer is replaced by two FC layers with output feature numbers of 128 and z respectively. The model finally divides the input data into z categories, and the classification results are given in the form of probability:

[0085] (15),

[0086] in, h q It is q The FC layer output of each category, It is q The output is converted into a probability distribution for the final classification decision, and these probability values ​​indicate the likelihood that the input data belongs to each category.

[0087] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used for various other combinations, modifications and improvements, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not deviate from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for detecting internal short circuit of a lithium-ion battery based on improved GAF, characterized in that: The short circuit detection method comprises: Step 1: amplify the voltage deviation by curve Manhattan distance, select the reference curve to divide all the curves, and calculate the curve Manhattan distance between the internal short-circuit curve and the reference curve through a sliding window; Step 2: Combine the Manhattan distances of all small windows into a one-dimensional time series D , one-dimensional data is converted into two-dimensional images through the improved Gram angle field; Step 3: Classify the severity of the short circuit into z categories according to the application scenario, construct z Gram angle field image datasets accordingly, and use the model to classify the image datasets, and obtain the possibility of each category according to the probability corresponding to the classification result; The step one comprises: In order to amplify the voltage deviation, the Manhattan distance calculation formula is set as ,in, and The curves are l 1 and l 2 of λ points; Select a healthy battery charging voltage curve as the reference curve, and use the deviation of other charging curves from the reference curve as the basis for detecting internal short circuits. The time length of all curves is T Then the window of length T The window is divided into multiple lengths m A small window; Use the sliding window to calculate the Manhattan distance between the internal short-circuit curve and the reference curve in each small window, and get the Manhattan distance of the curve in each small window: , d i Indicates i The Manhattan distance of the curve of a small window, V R,m(i-1)+1 , V R,m(i-1)+2 and V R,mi Respectively represent the first m ( i- 1) + 1. m ( i- 1) + 2 and mi A sample voltage, V ISC,m(i-1)+1 , V ISC,m(i-1)+2 and V ISC,mi Respectively represent the internal short-circuit curve m ( i- 1) + 1. m ( i- 1) + 2 and mi A sampling voltage.

2. The method for detecting internal short circuit of a lithium-ion battery based on improved GAF ​​according to claim 1, characterized in that: The second step comprises: The curve Manhattan distance of all small windows is combined into a one-dimensional time series , and scaled to the interval [0, 1], the time series D After normalization, we get , convert the scaled value into polar coordinates ,in, θ i and r i Points The phase angle and radius, t i Indicate point The timestamp of n The number of data in the time series; The time series D As radians, calculate the absolute value function of the cosine value and get the sequence ,in, ,in is the gain factor in radians, and the phase angle is expressed as ; Set an arc , through the formula Calculate the radian , converting the phase angle into ; Finally, through the Gram matrix Convert one-dimensional data into a two-dimensional Gram angular field image.

3. The method for detecting internal short circuit of a lithium-ion battery based on improved GAF ​​according to claim 1, characterized in that: The step three comprises: The severity of the short circuit is divided into z categories according to the application scenario, and z Gram angle field image datasets are constructed accordingly; The texture features of the Gram angular field image are extracted through the convolutional layer in the ResNet50 model, and then batch normalization is performed and activated, and then the pooling layer is used to reduce the dimension of the output feature map. The output of the pooling layer is processed by the fully connected layer, which has two output feature numbers, 128 and z. Finally, the input data is divided into z categories. The classification result is expressed in the form of probability, which indicates the possibility that the input data belongs to each category.

4. A lithium-ion battery internal short circuit detection device based on improved GAF, characterized in that: The device comprises a curve Manhattan distance calculation module, a data conversion module and a model classification module; The curve Manhattan distance calculation module is used to amplify the voltage deviation by the curve Manhattan distance, select the reference curve to divide all the curves, and calculate the curve Manhattan distance between the internal short-circuit curve and the reference curve through a sliding window; The data conversion module is used to convert the Manhattan distance of all small windows into a one-dimensional time series. D , one-dimensional data is converted into two-dimensional images through the improved Gram angle field; The model classification module is used to classify the severity of the short circuit into z categories according to the application scenario, construct z Gram angle field image data sets accordingly, and use the model to classify the image data sets, and obtain the possibility of each category according to the probability corresponding to the classification result; The curve Manhattan distance calculation module specifically includes the following contents: In order to amplify the voltage deviation, the Manhattan distance calculation formula is set as ,in, and The curves are l 1 and l 2 of λ points; Select a healthy battery charging voltage curve as the reference curve, and use the deviation of other charging curves from the reference curve as the basis for detecting internal short circuits. The time length of all curves is T Then the window of length T The window is divided into multiple lengths m A small window; Use the sliding window to calculate the Manhattan distance between the internal short-circuit curve and the reference curve in each small window, and get the Manhattan distance of the curve in each small window: , d i Indicates i The Manhattan distance of the curve of a small window, V R,m(i-1)+1 , V R,m(i-1)+2 and V R,mi Respectively represent the first m ( i- 1) + 1. m ( i- 1) + 2 and mi A sample voltage, V ISC,m(i-1)+1 , V ISC,m(i-1)+2 and V ISC,mi Respectively represent the internal short-circuit curve m ( i- 1) + 1. m ( i- 1) + 2 and mi A sampling voltage.

5. The lithium-ion battery internal short circuit detection device based on improved GAF ​​according to claim 4, characterized in that: The data conversion module specifically includes the following contents: The curve Manhattan distance of all small windows is combined into a one-dimensional time series , and scaled to the interval [0, 1], the time series D After normalization, we get , convert the scaled value into polar coordinates ,in, θ i and r i Points The phase angle and radius, t i Indicate point The timestamp of n The number of data in the time series; The time series D As radians, calculate the absolute value function of the cosine value and get the sequence ,in, ,in is the gain factor in radians, and the phase angle is expressed as ; Set an arc , through the formula Calculate the radian , converting the phase angle into ; Finally, through the Gram matrix Convert one-dimensional data into a two-dimensional Gram angular field image.

6. The lithium-ion battery internal short circuit detection device based on improved GAF ​​according to claim 4, characterized in that: The model classification module specifically includes the following contents: The severity of the short circuit is divided into z categories according to the application scenario, and z Gram angle field image datasets are constructed accordingly; The texture features of the Gram angular field image are extracted through the convolutional layer in the ResNet50 model, and then batch normalization is performed and activated, and then the pooling layer is used to reduce the dimension of the output feature map. The output of the pooling layer is processed by the fully connected layer, which has two output feature numbers, 128 and z. Finally, the input data is divided into z categories. The classification result is expressed in the form of probability, which indicates the possibility that the input data belongs to each category.

Images