A method for detecting abnormal individual battery packs based on convolutional neural network

Abstract

The present invention discloses a method for detecting abnormal individuals in a battery pack based on a convolutional neural network, comprising the following steps: step one, obtaining training data; step two, constructing a cyclic random hole convolutional network; step three, battery pack detection; the cyclic random hole convolutional network constructed by the present invention uses an end-to-end parallel processing mode, which reduces the complexity of the entire network, and can obtain the health evaluation of all cells in the battery pack through one calculation, thereby improving the detection efficiency. The method is not easily affected by external interference and has high detection accuracy and reliability; the present invention can balance speed and accuracy for the scoring estimation of the health of cells in the battery pack, and the highest accuracy can reach more than 85%. For the detection of abnormal cells in the battery pack, it is only necessary to detect whether the state of the cell is abnormal, and no specific health score is required, and the accuracy can reach more than 96%, which can effectively balance the network complexity and accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of battery detection, and in particular to a method for detecting abnormal individual battery packs based on a convolutional neural network. Background Art

[0002] The state of health of the power battery, namely SOH (State Of Health), is a quantitative indicator used to evaluate the degree of aging of the power battery. It is closely related to the aging of the power battery, and the aging of the power battery is a key factor affecting the safety of the power battery system. Therefore, the estimation of the battery health (SOH) can not only optimize the operation of the power battery system, but also has important significance for ensuring the safe and reliable operation of the power battery system.

[0003] At present, SOH estimation methods can be divided into two categories. One is the experimental analysis method, that is, by analyzing the collected test data such as power battery current, voltage, temperature, etc., the SOH value of the power battery can be calibrated. According to the selected parameters, it can be divided into direct measurement method and indirect analysis method; the other is the model-based method, that is, using the power battery model to estimate the selected battery parameters, so as to calibrate the SOH value of the power battery. According to the modeling method, it can be divided into adaptive algorithm and data-driven method; the data-driven SOH estimation method does not rely on precise mathematical models to describe the aging principle and evolution process of the power battery. It only relies on historical data, that is, through a specific learning algorithm to find the statistical laws in the historical data as the basis for evaluating SOH.

[0004] However, most of the existing data-driven SOH estimation methods are based on traditional machine learning algorithms, such as decision trees, EM and other statistical learning methods. Or they are based on recursive neural network learning methods to estimate the SOH of single cells. This method is based on time series data, so only one cell can be detected each time. Multiple tests are required for the entire battery pack, resulting in low detection efficiency and extremely time-consuming. In addition, external interference has a great impact on the detection accuracy of this method. Therefore, this method has high requirements for the external environment. The complex and changeable production environment will greatly reduce the detection accuracy of this method, resulting in low practicality. At the same time, although battery health can be used to evaluate whether the battery is abnormal, its representation method is still not simple and intuitive enough. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing methods and propose a new network structure based on a convolutional neural network according to the characteristics and operating conditions of a power energy storage battery pack. According to the fact that a convolutional neural network can extract translation invariant features in information, in order to use a convolutional neural network to calculate the characteristics of each battery cell, a method for detecting abnormal individuals in a battery pack based on a convolutional neural network is proposed to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for detecting abnormal individual battery packs based on a convolutional neural network, comprising the following steps: step 1, obtaining training data; step 2, constructing a cyclic random hole convolutional network; step 3, battery pack detection;

[0007] In the above step 1, a data pair consisting of a measurement parameter and a battery SOH is obtained and used as a data set for building a network;

[0008] In the above step 2, constructing a cyclic random hole convolutional network includes the following steps:

[0009] 1) Perform data labeling and matrix processing on the data set obtained in the above step 1;

[0010] 2) Using the labeled and matrixed datasets, we construct a cyclic random dilated convolutional network.

[0011] In the above step three, the battery pack parameter data measured in real time at the work site is sent to the constructed cyclic random void convolutional network for detection, and a score estimate of the health of each battery cell is given in real time. Whether the state of the battery cell is abnormal is judged based on the output value of the network.

[0012] Preferably, in step one, the data is obtained in two ways: one is to collect measurement data from the actual work site and then calibrate the battery SOH value; the other is to collect data in a laboratory simulated work site, select a batch of single cells with calibrated SOH values ​​to form a battery pack, and collect current, voltage, ambient temperature and charge and discharge times.

[0013] Preferably, in step 1) of step 2, when labeling the data set, 0 or 1 is used to indicate that the battery health status is normal, and 1 or 0 is used to indicate that the battery health status is abnormal.

[0014] Preferably, in step 2), in the process of constructing a cyclic random hole convolutional network, it can be implemented through PyTorch or TensorFlow in Python language, or it can be directly implemented using C / C++.

[0015] Preferably, in step 2 1), the method for matrix processing of the data set is specifically as follows: assuming that the total number of single cells in the energy storage battery pack is N, each single cell is numbered ci, and 1≤i≤N, the N single cells are divided into a matrix of N=m*n, each matrix element corresponds to a single cell, each measurable parameter of the single cell, such as voltage, current and temperature, can constitute a matrix, for N, there are usually multiple m*n combinations that meet the conditions, and the selected m and n should satisfy two conditions, that is, N=m*n, and abs(mn) takes the minimum value. Under ideal conditions, N is a square number, and m=n at this time. The division of the matrix has nothing to do with the series-parallel connection mode between the single cells in the battery pack, but only with the total number of battery cells, that is, as long as the number of battery cells in the battery pack is the same, the division method of the matrix is ​​also the same, each measurable parameter of the single cell is used as a feature, and each parameter can constitute a feature matrix. If there are multiple measurable parameters, each measurable parameter constitutes a feature matrix, and for c measurable parameters, they can be combined into an m*n*c feature tensor.

[0016] Preferably, in the process of constructing the cyclic random hole convolution network in step 2), the size of the convolution kernel depends on the values ​​of m and n, and needs to meet the following conditions: for m=n, that is, the number of battery cells can form a matrix of square numbers, the convolution kernel can choose a size of k*k, where k≤m, and k is an odd number; for m!=n, the convolution kernel can choose a size of k*k, where k≤min(m,n), that is, k is less than or equal to the smaller one between m and n, and is an odd number.

[0017] The preferred k value depends on the total number of cells N in the battery pack. For a cell number N≤300, it can be selected according to the following formula:

[0018] k=min(m,n)-(min(m,n)+1)%2 (1)

[0019] For the case where the number of monomers N ≥ 500, the following formula can be used for selection:

[0020] k=floor(r*min(m,n)–(r*min(m,n)+1)%2) (2)

[0021] Among them, floor means rounding down, % means finding the remainder, min means the minimum value between the two, r is a hyperparameter, 0.5≤r≤1, that is, k is less than or equal to the smaller one between m and n, and a suitable odd number between half the value and the full value is selected;

[0022] For the case where the number of monomers 300 < N < 500, the selection can be made by combining formula (1) and formula (2). The selection of the k value needs to be combined with the dilated convolution parameters. At the same time, in order to facilitate the determination of the convolution center and calculation, k needs to be an odd number.

[0023] Preferably, in 2) of the second step, the random dilated convolution kernel in the cyclic random dilated convolution network is specifically: the number of dilations in the convolution kernel is controlled by a probability parameter p. For each element in the convolution kernel, whether this position participates in the convolution calculation is controlled by the probability p; the selection of the specific probability parameter p needs to be set in combination with the convolution kernel parameter k. For a battery pack with N single cells, p, k, and N should satisfy the following formula:

[0024] 2 * k * k * p ≥ N (3)

[0025] That is, the number of elements participating in the convolution calculation should be more than half of the number of single cells.

[0026] For the input matrix I, the convolution kernel K, and the output matrix O, there is the following calculation formula:

[0027] O m，n = Σ ij I m+i，n+j * K ij (4)

[0028] Construct an abstract closed-loop data structure Ω from the input matrix I. Ω satisfies the following conditions: the dimensions, size, and number of elements in Ω are the same as those of I; for integers -k < i < 2k and -k < j < 2k, the elements in Ω satisfy the following formula:

[0029] Ω i，j = I i％m，j％n (5)

[0030] Calculate the convolution value at a certain point I m，n of the input matrix I, and then construct an intermediate matrix I' through Ω and satisfy the following conditions: the size of I' is the same as that of the convolution kernel K; for integers -k / 2 < i < k / 2 and -k / 2 < j < k / 2, the elements in I' satisfy the following formula:

[0031] I' i，j = Ω m-i，n-j (6)

[0032] Finally, use formula (4) to calculate the convolution of I' and K, and use the obtained calculation result as the output matrix O; for the tensor form with multiple characteristic parameters, when filling the data in a loop, only the elements at the corresponding positions are replaced by vectors, and the remaining processing methods are similar to those of matrix operations; for the convolution kernel with random dilations, the construction method of the intermediate matrix I' remains unchanged, and the calculation at the dilated points is ignored during the convolution calculation.

[0033] Preferably, a battery pack abnormal individual detection system based on a convolutional neural network comprises:

[0034] The first building module is used to obtain a data pair consisting of a measurement parameter and a battery SOH and use it as a data set for building a network;

[0035] The first processing module is used to perform data labeling and matrix processing on the acquired data set;

[0036] The second building module is used to build a cyclic random hole convolutional network using the labeled and matrixed data set;

[0037] The real-time measurement module is used to send the battery pack parameter data measured in real time at the work site into the constructed cyclic random void convolutional network for detection, give a real-time score estimate of the health of each battery cell, and judge whether the state of the battery cell is abnormal based on the output value of the network.

[0038] Preferably, the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, they are used to implement a method for detecting abnormal individual battery packs based on a convolutional neural network.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows: the cyclic random hole convolutional network constructed by the present invention uses an end-to-end parallel processing mode, which reduces the complexity of the entire network. The health evaluation of all cells in the battery pack can be obtained through one calculation, thereby improving the detection efficiency. The method is not easily affected by external interference and has high detection accuracy and reliability. The present invention can balance speed and accuracy for the scoring estimation of the health of cells in the battery pack, and the highest accuracy can reach more than 85%. For the detection of abnormal cells in the battery pack, it is only necessary to detect whether the state of the cell is abnormal, and no specific health score is required. The accuracy can reach more than 96%, which can effectively balance the network complexity and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a flow chart of the method of the present invention;

[0041] Figure 2 This is an example diagram of battery cells that are first connected in series and then in parallel;

[0042] Figure 3 This is an example diagram of a battery matrix representation that is first connected in series and then in parallel;

[0043] Figure 4 This is an example diagram of battery cells that are first connected in parallel and then in series;

[0044] Figure 5This is an example diagram of a matrix representation of battery cells connected in parallel first and then in series;

[0045] Figure 6 This is an example diagram of a 14×16 feature matrix;

[0046] Figure 7 This is an example graph of a 14×16×3 feature tensor;

[0047] Figure 8 This is an example diagram of convolution calculation with padding;

[0048] Fig. 9 This is an example diagram of convolution calculation with holes;

[0049] Fig.10 is an example graph of the input matrix I;

[0050] Fig.11 is an example diagram of the intermediate matrix I';

[0051] Fig.12 This is an example diagram of a neural network structure. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] See also Figure 1-12 , an embodiment provided by the present invention: a method for detecting abnormal individual battery packs based on a convolutional neural network, comprising the following steps: step 1, obtaining training data; step 2, constructing a cyclic random hole convolutional network; step 3, battery pack detection;

[0054] In the above step 1, the data pair consisting of the measurement parameters and the battery SOH is used as the data set for constructing the network; the data pair is obtained in two ways: one is to collect the measurement data at the actual work site and calibrate the battery SOH value at the same time; the other is to collect data at the laboratory simulation work site, select a batch of single cells with calibrated SOH values ​​to form a battery pack, and collect current, voltage, ambient temperature and charge and discharge times;

[0055] In the above step 2, constructing a cyclic random hole convolutional network includes the following steps:

[0056] 1) The data set obtained in the above step 1 is labeled and matrixed. When labeling the data pair, 0 is used to indicate that the battery health state is normal, and 1 is used to indicate that the battery health state is abnormal. In the process of constructing a cyclic random hole convolutional network, the number of layers of the convolutional network, the number of convolution kernels, the regularization method, and the selection of the activation function can follow the general principles of neural networks. The convolutional neural network can be implemented by PyTorch and TensorFlow in Python language, or directly implemented by C / C++.

[0057] The method of matrixing the characteristic data of the battery pack is as follows: assuming that the total number of single cells in the energy storage battery pack is N, then each single cell is numbered ci, and 1≤i≤N, and the N single cells are divided into a matrix of N=m*n, each matrix element corresponds to a single cell, and each measurable parameter of the single cell, such as voltage, current and temperature, can constitute a matrix; for N, there are usually multiple m*n combinations that meet the conditions. At this time, the selected m and n should meet two conditions, namely N=m*n, and abs(mn) takes the minimum value. Under ideal conditions, N is a square number, and m=n at this time. The division of the matrix has nothing to do with the series and parallel connection mode between the single cells in the battery pack, but only with the total number of battery cells, that is, as long as the number of battery cells in the battery pack is the same, the division method of the matrix is ​​also the same (such as Figure 2-4 ), each measurable parameter of the monomer is regarded as a feature, and each parameter can constitute a feature matrix. If there are multiple measurable parameters, each measurable parameter constitutes a feature matrix. For c measurable parameters, they can be combined into an m*n*c feature tensor; Figure 6 It is a battery pack composed of 224 cells, which are constructed into a 14×16 matrix, and the voltage forms a 14×16 characteristic matrix. Figure 7 It is a battery pack composed of 224 cells, which are constructed into a 14×16 matrix, and the voltage, current, and temperature form a 14×16×3 feature tensor;

[0058] 2) Using the labeled and matrixed dataset, a cyclic random dilated convolutional network is constructed. The construction process is the same as the general neural network training and detection method.

[0059] In the process of constructing the cyclic random hole convolution network, the size of the convolution kernel depends on the values ​​of m and n, and needs to meet the following conditions: for m = n, that is, the number of battery cells can form a square matrix, the convolution kernel can choose the size of k*k, where k≤m, and k is an odd number; for m! = n, the convolution kernel can choose the size of k*k, where k≤min(m, n), that is, k is less than or equal to the smaller one between m and n, and is an odd number;

[0060] The value of k depends on the total number of cells N in the battery pack. For the case where the number of cells N ≤ 300, it can be selected according to the following formula:

[0061] k = min(m, n) - (min(m, n) + 1) % 2 (1)

[0062] For the case where the number of cells N ≥ 500, it can be selected according to the following formula:

[0063] k = floor(r * min(m, n) – (r * min(m, n) + 1) % 2) (2)

[0064] Where floor means rounding down, % means finding the remainder, min means the minimum of the two, r is a hyperparameter set manually according to the actual situation. Usually, 0.5 ≤ r ≤ 1, that is, k is less than or equal to the smaller one between m and n, and an appropriate odd number between half value and full value is selected; for example, for a battery pack composed of N = 480 cells, if m = 20 and n = 24 are selected, then k can be selected from 11, 13, 15, 17, 19;

[0065] For the case where the number of cells 300 < N < 500, it can be selected by combining formula (1) and formula (2). The selection of the value of k needs to combine the dilated convolution parameters. At the same time, in order to facilitate the determination of the convolution center and calculation, k needs to be kept odd;

[0066] The random dilated convolution kernel in the cyclic random dilated convolution network is specifically as follows: The number of dilations in the convolution kernel is controlled by a probability parameter p. For each element in the convolution kernel, whether this position participates in the convolution calculation is controlled by the probability p; the selection of the specific probability parameter p needs to be set in combination with the convolution kernel parameter k. For a battery pack with the number of single cells N, p, k, and N should satisfy the following formula:

[0067] 2 * k * k * p ≥ N (3)

[0068] That is, the number of elements participating in the convolution calculation should be more than half of the number of single cells;

[0069] For the input matrix I, convolution kernel K, and output matrix O, there is the following calculation formula:

[0070] O m，n = Σ ij I m+i，n+j * K ij (4)

[0071] Construct an abstract closed-loop data structure Ω from the input matrix I. Ω satisfies the following conditions: The dimensions, sizes, and number of elements in Ω are the same as those of I; for integers -k < i < 2k and -k < j < 2k, the elements in Ω satisfy the following formula:

[0072] Ω i，j = I i％m，j％n (5)

[0073] Calculate the convolution value at a certain point I of the input matrix I, and then construct an intermediate matrix I' through Ω and satisfy the following conditions: The size of I' is the same as the convolution kernel K; for integers -k / 2 < i < k / 2 and -k / 2 < j < k / 2, the elements in I' satisfy the following formula: m，n At, the convolution value is calculated, and then an intermediate matrix I' is constructed through Ω and satisfies the following conditions: The size of I' is the same as the convolution kernel K; for integers -k / 2 < i < k / 2 and -k / 2 < j < k / 2, the elements in I' satisfy the following formula:

[0074] I' i，j = Ω m-i，n-j (6)

[0075] Finally, use formula (4) to calculate the convolution of I' and K, and take the obtained calculation result as the output matrix O; for the tensor form with multiple characteristic parameters, when filling data cyclically, only replace the elements at the corresponding positions with vectors, and the remaining processing methods are similar to those of matrix operations; for the convolution kernel with random holes, the construction method of the intermediate matrix I' remains unchanged, and the calculation at the hole points is ignored during convolution calculation; such as Figure 10-11 , is an example diagram of converting the input matrix I to the intermediate matrix I'; for example, for a battery pack composed of N = 244 monomers, according to the method introduced above, select m = 16, n = 14, k = 13, and take the voltage as the measurement signal, then the input is a 16*14 matrix. This matrix undergoes multiple convolution calculations and channel expansion and contraction, and finally outputs a 16*14 estimated matrix. The specific network structure is as Fig.12 ;

[0076] Among them, in the above step three, the parameter data of the battery pack measured in real time at the work site is sent into the constructed cyclic random hole convolution network for detection, and the health score estimate of each battery monomer is given in real time. According to the output value of the network, it is judged whether the state of the battery monomer is abnormal.

[0077] Based on the above, the advantages of the present invention are that the detection method designed by the present invention is divided into two stages: training and detection. In the training stage, the collected historical parameter data of various battery packs or the parameter data of laboratory simulation are manually annotated and then sent to the neural network for training; in the detection stage, the parameter data measured in real time at the work site are sent to the trained neural network for calculation, so as to give the health evaluation value of each battery cell in real time, so as to timely find the abnormal SOH cell in the battery pack; the detection method designed by the present invention is based on the convolutional neural network design. Because the convolutional neural network is a parallel network structure, the number of units that can be processed simultaneously is not limited. Specifically, in the present invention, the number of single cells is not limited, so only one operation is required for the trained neural network to give the real-time evaluation result of the health of each cell of the battery pack; the method can be applied to battery packs containing any number of cells, but for each battery pack with a given structure, it is necessary to start training from scratch, and then detect the battery pack with the same structure, that is, the battery packs trained and tested are the same in physical structure. According to the convolutional neural network, the translation invariance features in the information can be extracted. In order to use the convolutional neural network to calculate the features of each battery cell, the present invention proposes a new convolution calculation method, Taking into account the independence of each battery cell in the battery pack, the convolution matrix is ​​constructed by cyclic filling. At the same time, a method for constructing the convolution kernel is proposed, and convolution calculation is performed by randomly setting holes in the convolution kernel. The random holes achieve a function similar to dropout, which not only reduces the overall calculation amount, but also effectively prevents overfitting. Because this type of convolutional neural network does not use the pooling layer for downsampling, in the process of data flowing from input to output, only the number of data channels changes. The same input and output size of each layer facilitates the data processing of the entire network without the need for too many auxiliary structures, thus The complexity of the entire network is reduced, end-to-end output is achieved, and training is also more convenient. At the same time, considering that the actual production environment has much higher requirements for safety than the attention to battery health, although the health can be used to evaluate whether the battery is abnormal, it is more complicated than directly detecting abnormal cells in the battery pack. Therefore, directly detecting abnormal cells in the battery pack also has certain practical significance. In order to achieve the purpose of directly detecting abnormal cells, the present invention uses 1 and 0 to indicate whether the battery state is abnormal or not, simplifies the representation of battery health, improves the scalability, accuracy and efficiency of the detection method, and makes it better suitable for production environments.

[0078] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A method for detecting abnormal individual battery packs based on a convolutional neural network, comprising the following steps: Step 1, obtaining training data; Step 2, constructing a cyclic random hole convolutional network; Step 3, battery pack detection; It is characterized by: In the above step 1, a data pair consisting of a measurement parameter and a battery SOH is obtained and used as a data set for building a network; In the above step 2, constructing a cyclic random hole convolutional network includes the following steps: 1) Perform data labeling and matrix processing on the data set obtained in the above step 1; 2) Using the labeled and matrixed datasets, we construct a cyclic random dilated convolutional network. In step 1) of step 2, the method of matrix processing of the data set is specifically as follows: assuming that the total number of single cells in the energy storage battery pack is N, each single cell is numbered ci, and 1≤i≤N, the N single cells are divided into a matrix of N=m*n, each matrix element corresponds to a single cell, each measurable parameter of the single cell, such as voltage, current and temperature, can constitute a matrix, for N, there are usually multiple m*n combinations that meet the conditions, and the selected m and n should meet two conditions at this time, that is, N=m*n, and abs(mn) takes the minimum value, in an ideal condition, N is a square number, and m=n at this time, the division of the matrix has nothing to do with the series-parallel connection mode between the single cells in the battery pack, but only with the total number of battery cells, that is, as long as the number of battery cells in the battery pack is the same, the division method of the matrix is ​​also the same, each measurable parameter of the single cell is used as a feature, and each parameter can constitute a feature matrix, if there are multiple measurable parameters, each measurable parameter constitutes a feature matrix, for c measurable parameters, they can be combined into a m*n*c feature tensor; In step 2), in the process of constructing the cyclic random hole convolution network, the size of the convolution kernel depends on the values ​​of m and n, and needs to meet the following conditions: for m=n, that is, the number of battery cells can form a square matrix, the convolution kernel can be selected to be k*k in size, where k≤m, and k is an odd number; for m!=n, the convolution kernel can be selected to be k*k in size, where k≤min(m,n), that is, k is less than or equal to the smaller one between m and n, and is an odd number; The k value depends on the total number of cells N in the battery pack. For a cell number N≤300, it can be selected according to the following formula: k=min(m,n)-(min(m,n)+1)%2(1) For the case where the number of monomers N ≥ 500, the following formula can be used for selection: k=floor(r*min(m,n)-(r*min(m,n)+1)%2)(2) Among them, floor means rounding down, % means finding the remainder, min means the minimum value between the two, r is a hyperparameter, 0.5≤r≤1, that is, k is less than or equal to the smaller one between m and n, and a suitable odd number between half the value and the full value is selected; When the number of monomers is 300<N<500, the selection can be made by combining formula (1) and formula (2). The selection of k value needs to be combined with the dilated convolution parameters. At the same time, in order to facilitate the determination of the convolution center and calculation, k needs to be kept as an odd number. In step 2) of the second step, the random hole convolution kernel in the cyclic random hole convolution network is specifically: the number of holes in the convolution kernel is controlled by a probability parameter p, and for each element in the convolution kernel, the probability p is used to control whether the position participates in the convolution calculation; the selection of the specific probability parameter p needs to be set in combination with the convolution kernel parameter k. For a battery pack with N single cells, p, k, and N should satisfy the following formula: 2*k*k*p≥N(3) means the number of elements involved in the convolution calculation must be greater than half of the number of single cells; For the input matrix I, convolution kernel K, and output matrix O, there is the following calculation formula: O m，n =∑ ij I m+i，n+j *K ij (4) An abstract closed loop data structure Ω is constructed from the input matrix I. Ω satisfies the following conditions: the dimension, size and number of elements in Ω are the same as I; for integers -k<i<2k and -k<j<2k, the elements in Ω satisfy the following formula: Oh i，j =I i％m，j％n (5) Calculate the input matrix I at a certain point I m，n The convolution value at Ω is used to construct the intermediate matrix I′ and satisfy the following conditions: the size of I′ is the same as the convolution kernel K; for integers -k / 2<i<k / 2 and -k / 2<j<k / 2, the elements in I′ satisfy the following formula: I' i，j =Oh m-i，n-j (6) Finally, the convolution of I′ and K is calculated using formula (4), and the calculated result is used as the output matrix O. For tensor forms with multiple feature parameters, the data loop is filled by simply replacing the elements at the corresponding positions from numbers to vectors, and the rest of the processing methods are similar to the matrix operation methods. For convolution kernels with random holes, the construction method of the intermediate matrix I′ remains unchanged, and the calculation at the hole points is ignored when calculating the convolution. In the above step three, the battery pack parameter data measured in real time at the work site is sent to the constructed cyclic random void convolutional network for detection, and a score estimate of the health of each battery cell is given in real time. Whether the state of the battery cell is abnormal is judged based on the output value of the network.

2. According to claim 1, a method for detecting abnormal individual batteries based on a convolutional neural network is characterized in that: In step 1, there are two ways to obtain data pairs: one is to collect measurement data from the actual work site and then calibrate the battery SOH value; the other is to collect data at a simulated work site in the laboratory, select a batch of single cells with calibrated SOH values ​​to form a battery pack, and collect current, voltage, ambient temperature and charge and discharge times.

3. According to claim 1, a method for detecting abnormal individual batteries based on a convolutional neural network is characterized in that: In step 1) of step 2, when labeling the data set, 0 is used to indicate that the battery health status is normal, and 1 is used to indicate that the battery health status is abnormal.

4. The method for detecting abnormal individual batteries based on a convolutional neural network according to claim 3 is characterized in that: In step 2) of the present invention, in the process of constructing a cyclic random hole convolutional network, it can be implemented by PyTorch or TensorFlow in Python language, or it can be directly implemented by C / C+++.

5. A battery pack abnormal individual detection system based on convolutional neural network, the system comprising: The first building module is used to obtain a data pair consisting of a measurement parameter and a battery SOH and use it as a data set for building a network; The first processing module is used to perform data labeling and matrix processing on the acquired data set; The second building module is used to build a cyclic random hole convolutional network using the labeled and matrixed data set; The real-time measurement module is used to send the battery pack parameter data measured in real time at the work site to the constructed cyclic random void convolutional network for detection, give a real-time score estimate of the health of each battery cell, and judge whether the state of the battery cell is abnormal based on the output value of the network; The method for matrix processing of the data set is specifically as follows: assuming that the total number of single cells in the energy storage battery pack is N, each single cell is numbered ci, and 1≤i≤N, the N single cells are divided into a matrix of N=m*n, each matrix element corresponds to a single cell, each measurable parameter of the single cell, such as voltage, current and temperature, can constitute a matrix, for N, there are usually multiple m*n combinations that meet the conditions, at this time, the selected m and n should meet two conditions, that is, N=m*n, and abs(mn) takes the minimum value, in an ideal condition, N is a square number, at this time m=n, the division of the matrix has nothing to do with the series-parallel connection mode between the single cells in the battery pack, but only with the total number of battery cells, that is, as long as the number of battery cells in the battery pack is the same, the division mode of the matrix is ​​also the same, each measurable parameter of the single cell is used as a feature, each parameter can constitute a feature matrix, if there are multiple measurable parameters, each measurable parameter constitutes a feature matrix, for c measurable parameters, they can be combined into a m*n*c feature tensor; In the process of constructing the cyclic random hole convolution network, the size of the convolution kernel depends on the values ​​of m and n, and needs to meet the following conditions: for m = n, that is, the number of battery cells can form a square matrix, the convolution kernel can choose the size of k*k, where k≤m, and k is an odd number; for m! = n, the convolution kernel can choose the size of k*k, where k≤min(m, n), that is, k is less than or equal to the smaller one between m and n, and is an odd number; The k value depends on the total number of cells N in the battery pack. For a cell number N≤300, it can be selected according to the following formula: k=min(m,n)-(min(m,n)+1)%2(1) For the case where the number of monomers N ≥ 500, the following formula can be used for selection: k=floor(r*min(m,n)-(r*min(m,n)+1)%2)(2) Among them, floor means rounding down, % means finding the remainder, min means the minimum value between the two, r is a hyperparameter, 0.5≤r≤1, that is, k is less than or equal to the smaller one between m and n, and a suitable odd number between half the value and the full value is selected; When the number of monomers is 300<N<500, the selection can be made by combining formula (1) and formula (2). The selection of k value needs to be combined with the dilated convolution parameters. At the same time, in order to facilitate the determination of the convolution center and calculation, k needs to be kept as an odd number. The random hole convolution kernel in the cyclic random hole convolution network is specifically: a probability parameter p is used to control the number of holes in the convolution kernel. For each element in the convolution kernel, the probability p is used to control whether the position participates in the convolution calculation; the selection of the specific probability parameter p needs to be set in combination with the convolution kernel parameter k. For a battery pack with N single cells, p, k, and N should satisfy the following formula: 2*k*k*p≥N (3) That is, the number of elements involved in the convolution calculation must be more than half of the number of single cells; For the input matrix I, convolution kernel K, and output matrix O, there is the following calculation formula: O m，n =∑ ij I m+i，n+j *K ij (4) An abstract closed loop data structure Ω is constructed from the input matrix I. Ω satisfies the following conditions: the dimension, size and number of elements in Ω are the same as I; for integers -k<i<2k and -k<j<2k, the elements in Ω satisfy the following formula: Oh i，j =I i％m，j％n (5) Calculate the input matrix I at a certain point I m，n The convolution value at Ω is used to construct the intermediate matrix I′ and satisfy the following conditions: the size of I′ is the same as the convolution kernel K; for integers -k / 2<i<k / 2 and -k / 2<j<k / 2, the elements in I′ satisfy the following formula: I' i，j =Oh m-i，n-j (6) Finally, the convolution of I′ and K is calculated using formula (4), and the calculated result is used as the output matrix O. For tensor forms with multiple feature parameters, the data loop is filled by simply replacing the elements at the corresponding positions from numbers to vectors, and the rest of the processing methods are similar to the matrix operation methods. For convolution kernels with random holes, the construction method of the intermediate matrix I′ remains unchanged, and the calculation at the hole points is ignored when calculating the convolution.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 4 when executed by a processor.

Images