A lithium-ion battery fault diagnosis method based on voltage prediction

Abstract

The present invention relates to the technical field of lithium-ion battery fault diagnosis, and discloses a lithium-ion battery fault diagnosis method based on voltage prediction. First, the actual vehicle data of the lithium-ion battery is obtained and data processing is performed, then a CNN-GRU hybrid model is built and the CNN-GRU hybrid model is trained and optimized, and the current, temperature, and state of charge are input into the trained CNN-GRU hybrid model, and the battery voltage is output to obtain the battery predicted voltage, and the battery residual is generated based on the predicted voltage, and the fault type and level are determined based on the battery residual. The lithium-ion battery fault diagnosis method based on voltage prediction uses the CNN-GRU method to realize the voltage prediction of the actual vehicle battery, and after comparing the predicted voltage with the actual collected voltage, it can provide graded warnings for the battery failures that are about to occur.

Description

Technical Field

[0001] The invention relates to the technical field of lithium ion battery fault diagnosis, and in particular to a lithium ion battery fault diagnosis method based on voltage prediction. Background Art

[0002] Electric vehicles, with their unique advantages of being green, energy-saving, lightweight, convenient and low-cost, have become a representative solution for short-distance travel and commuting in the electrification of transportation. As the main power source of electric vehicles, the battery system greatly affects the power, safety, reliability and cost-effectiveness of electric vehicles. At present, lithium-ion batteries have become the preferred energy storage material for electric vehicles due to their long cycle life, wide operating temperature range, high power density and energy density. However, due to the continuous occurrence of battery failure accidents caused by battery aging, inherent defects and complex service environment, the safety of lithium-ion batteries has been scrutinized unprecedentedly. Therefore, battery fault diagnosis technology has become a research focus to ensure and improve the safe operation of battery systems. Most of the existing methods are unable to diagnose and warn battery voltage faults in advance, which may cause a small fault to eventually develop into an uncontrollable thermal runaway or explosion. Summary of the invention

[0003] 1. Technical issues to be resolved

[0004] In view of the deficiencies in the prior art, the present invention provides a lithium-ion battery fault diagnosis method based on voltage prediction, which has the advantages of early warning and diagnosis of battery faults, and solves the above-mentioned technical problems.

[0005] (II) Technical solution

[0006] To achieve the above object, the present invention provides the following technical solution: a lithium-ion battery fault diagnosis method based on voltage prediction, comprising the following steps:

[0007] S1. Obtaining real vehicle data of lithium-ion batteries: Collecting real vehicle data of lithium-ion batteries of fault-free electric vehicles, and recording battery charge and discharge voltage, current, temperature, and state of charge in real time;

[0008] S2, data processing: pre-processing the temperature, current, state of charge and voltage data obtained in S1, including the following steps:

[0009] S2.1. Construct the temperature, current and state of charge data in step S1 as the model input data set ;

[0010] S2.2. Divide the model input data set into a training set and a test set. The specific expression is as follows:

[0011] Training set:

[0012]

[0013] Test set:

[0014]

[0015] in, represents the model input in the training set, Represents the model input in the corresponding training set The predicted voltage, represents the model input in the test set, Represents the model input in the corresponding test set The predicted voltage, represents the training set, represents the test set, Represents dimension;

[0016] S3. CNN-GRU hybrid model construction: Combine the CNN model and the GRU model to build a CNN-GRU hybrid model;

[0017] S3.1. Construct a CNN layer through a CNN model to extract spatial features of data, wherein the CNN layer consists of a one-dimensional convolution layer and a maximum pooling layer, wherein the one-dimensional convolution layer consists of a group of filters, and the filters can extract spatial information from the input data, and pass the features extracted by the one-dimensional convolution layer to the maximum pooling layer to complete the extraction of spatial features;

[0018] S3.2, after the CNN layer completes the extraction of spatial features, a GRU layer is constructed through the GRU model to extract temporal features from the model input data;

[0019] S3.3, set the dropout layer to alleviate overfitting. The specific expression is as follows:

[0020]

[0021] in, Represents probability based A Bernoulli random variable, represents the Bernoulli distribution, Represents the output value of the dropout layer, Represents the hidden state at the current moment and serves as the input value of the dropout layer;

[0022] S3.4. Set up a fully connected layer to predict voltage Output is performed to complete the establishment of the CNN-GRU hybrid model;

[0023] S4, CNN-GRU hybrid model optimization: Use the moth-to-flame optimization algorithm to optimize the hyperparameters of the CNN-GRU hybrid model obtained in S3 to obtain the optimized CNN-GRU hybrid model;

[0024] S5. CNN-GRU model training: Take the training set formed by the voltage and current, temperature and state of charge data obtained in step S2 to run the optimized CNN-GRU hybrid model, find the nonlinear mapping relationship between current, temperature, state of charge and voltage, and obtain the trained CNN-GRU hybrid model;

[0025] S6. Obtaining predicted battery voltage: Based on the trained CNN-GRU hybrid model obtained in step S5, input current, temperature, and state of charge, and output battery voltage to obtain predicted battery voltage;

[0026] S7. Generate battery residual: Based on the predicted voltage obtained in step S6, calculate the difference between the predicted voltage and the actual collected voltage, that is, obtain the battery residual ;

[0027] S8. Fault diagnosis: Based on the battery residual obtained in step S7 , determine the fault type and level, the specific steps are as follows:

[0028] S8.1. Determine the battery residual If the battery residual If it is negative, it is an overvoltage state. If it is positive, it is undervoltage state;

[0029] S8.2. At the same time, different thresholds are set according to the different charging and discharging stages of the battery. If the battery residual If the battery residual is higher than the set threshold, the undervoltage fault and corresponding alarm level will be triggered immediately to locate the faulty battery cell. If the voltage is lower than the set threshold, an overvoltage fault and corresponding alarm level will be triggered immediately to locate the faulty battery cell.

[0030] As a preferred technical solution of the present invention, the expression for extracting spatial information by the filter in the CNN layer is as follows:

[0031]

[0032] in, , , , Represent the input, output, weight and bias value of the convolutional layer respectively, represents the convolution operation, represents the activation function;

[0033] The expression of the feature extracted by the one-dimensional convolution layer in step S3.1 and passed to the maximum pooling layer is as follows:

[0034]

[0035] in, represents the output of the maximum pooling layer, represents the window size during the pooling process, express The output value when the input is dimensional.

[0036] As a preferred technical solution of the present invention, the specific expression of the GRU layer in step S3.2 is as follows:

[0037]

[0038] in, Indicates the reset gate at the current moment, represents the weight matrix of the reset gate, represents the hidden state at the previous moment, Indicates the current hidden state. represents the update gate, represents the weight matrix of the update gate, Indicates the candidate status at the current moment, represents the weight matrix of the candidate state, the subscript t represents the time, represents the sigmoid activation function, represents the hyperbolic tangent function, represents the element-wise product, Represents the output of the maximum pooling layer and serves as the input of the GRU layer;

[0039] In step S3.4, the voltage is predicted by the fully connected layer. The expression is as follows:

[0040]

[0041] in, Represents the output value of the dropout layer, represents the weight matrix of the fully connected layer, Represents the bias vector of the fully connected layer.

[0042] As a preferred technical solution of the present invention, the calculation expression of the battery residual in step S7 is as follows:

[0043]

[0044] in, Indicates the battery residual. Indicates the actual acquisition voltage, Represents the voltage output by the CNN-GRU hybrid model.

[0045] Compared with the prior art, the present invention provides a lithium-ion battery fault diagnosis method based on voltage prediction, which has the following beneficial effects:

[0046] The present invention first obtains real vehicle data of lithium-ion batteries and processes the data, then builds a CNN-GRU hybrid model and trains and optimizes the CNN-GRU hybrid model, inputs current, temperature, and state of charge into the trained CNN-GRU hybrid model, outputs battery voltage, and obtains battery predicted voltage, generates battery residual based on the predicted voltage, and determines the fault type and level according to the battery residual. The CNN-GRU method realizes voltage prediction of real vehicle batteries, and after comparing the predicted voltage with the actual collected voltage, it can provide graded warning for impending battery faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is the result of the present invention using the moth-to-flame algorithm to optimize the hyperparameters of the CNN-GRU hybrid model;

[0048] Figure 2 is the undervoltage fault diagnosis result obtained by implementing the present invention;

[0049] Figure 3 is the overvoltage fault diagnosis result obtained by implementing the present invention;

[0050] Figure 4 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0051] 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.

[0052] See also Figure 1-4 , a lithium-ion battery fault diagnosis method based on voltage prediction, comprising the following steps:

[0053] S1. Obtaining real vehicle data of lithium-ion batteries: Collecting real vehicle data of lithium-ion batteries of fault-free electric vehicles, and recording battery charge and discharge voltage, current, temperature, and state of charge in real time;

[0054] S2, data processing: pre-processing the temperature, current, state of charge and voltage data obtained in S1, including the following steps:

[0055] S2.1. Construct the temperature, current and state of charge data in step S1 as model input ;

[0056] S2.2. Input the model into the dataset Divided into training set and test set, the specific expression is as follows:

[0057] Training set:

[0058]

[0059] Test set:

[0060]

[0061] in, represents the model input in the training set, Represents the model input in the corresponding training set The predicted voltage, represents the model input in the test set, Represents the model input in the corresponding test set The predicted voltage, represents the training set, represents the test set, Represents dimension;

[0062] S3. CNN-GRU hybrid model construction: Combine the CNN model and the GRU model to build a CNN-GRU hybrid model;

[0063] S3.1. Construct a CNN layer through the CNN model to extract the spatial features of the data. The CNN layer consists of a one-dimensional convolution layer and a maximum pooling layer. The one-dimensional convolution layer consists of a group of filters. The filter can extract spatial information from the input data. The features extracted by the one-dimensional convolution layer are passed to the maximum pooling layer to complete the extraction of spatial features.

[0064] The expression of the filter extracting spatial information in the CNN layer is as follows:

[0065]

[0066] in, , , , Represent the input, output, weight and bias value of the convolutional layer respectively, represents the convolution operation, represents the activation function;

[0067] The expression of the features extracted by the one-dimensional convolutional layer in step S3.1 and passed to the maximum pooling layer is as follows:

[0068]

[0069] in, represents the output of the maximum pooling layer, represents the window size during the pooling process, express Output value when the input dimension is 2;

[0070] S3.2, after the CNN layer completes the extraction of spatial features, a GRU layer is constructed through the GRU model to extract temporal features from the model input data;

[0071] The specific expression of the GRU layer in step S3.2 is as follows:

[0072]

[0073] in, Indicates the reset gate at the current moment, represents the weight matrix of the reset gate, represents the hidden state at the previous moment, Indicates the current hidden state. represents the update gate, represents the weight matrix of the update gate, Indicates the candidate status at the current moment, represents the weight matrix of the candidate state, the subscript t represents the time, represents the sigmoid activation function, represents the hyperbolic tangent function, represents the element-wise product, Represents the output of the maximum pooling layer and serves as the input of the GRU layer;

[0074] S3.3, set the dropout layer to alleviate overfitting. The specific expression is as follows:

[0075]

[0076] in, Represents probability based A Bernoulli random variable, represents the Bernoulli distribution, Represents the output value of the dropout layer, Represents the hidden state at the current moment and serves as the input value of the dropout layer;

[0077] S3.4, set up a fully connected layer for predicting voltage Output is performed to complete the establishment of the CNN-GRU hybrid model;

[0078] In step S3.4, the voltage is predicted by the fully connected layer The output expression is as follows:

[0079]

[0080] in, Represents the output value of the dropout layer, represents the weight matrix of the fully connected layer, Represents the bias vector of the fully connected layer;

[0081] S4, CNN-GRU hybrid model optimization: Use the moth-to-flame optimization algorithm to optimize the hyperparameters of the CNN-GRU hybrid model obtained in S3 to obtain the optimized CNN-GRU hybrid model;

[0082] The moth-to-fire algorithm is a swarm intelligence optimization algorithm with the characteristics of strong parallel optimization capability, superior global search performance, simple structure and limited parameters. In the moth-to-fire algorithm, the position of the moth is described as:

[0083] M location = [ a 1 1 a 1 2 ⋯ a 1 g a 2 1 a 2 2 ⋯ a 2 g ⋮ ⋮ ⋮ a n M 1 a n M 2 ⋯ a n M g ]

[0084] in represents the moth position matrix, Indicates the specific position of the moth, that is, its position in the matrix. represents the number of moths, g represents the number of parameters that need to be optimized, and the corresponding fitness value is,

[0085] O M = [ O M 1 O M 2 ⋮ O M n M ]

[0086] Where OM represents the fitness value. In the moth-to-fire algorithm, each moth needs to use a unique flame to update its position to avoid falling into a local optimum. Therefore, in the search space, the number of flames at the initial iteration is consistent with the size of the moth population. The position of the flame and its corresponding fitness value are expressed as:

[0087] F = [ F 1 1 F 1 2 ⋯ F 1 g F 2 1 F 2 2 ⋯ F 2 g ⋮ ⋮ ⋮ F n M 1 F n M 2 ⋯ F n M g ] n M × g

[0088] O F = [ O F 1 O F 2 ⋮ O F n M ]

[0089] in, represents the flame position matrix, Represents the fitness value. In order to mathematically model the flight behavior of moths flying into the flame, the position update mechanism of each moth relative to the flame can be expressed as:

[0090]

[0091] In the formula Indicates A moth, Indicates A flame, represents the spiral function, specifically:

[0092]

[0093] in Indicates The moth and The distance between the flames, is a constant value that changes the shape of the spiral, is a path coefficient, which is randomly generated between -1 and 1. If each position update of the moth is based on a different position in the search space, the local development ability of the algorithm will be reduced. To overcome this defect, a flame number adaptive mechanism is proposed to adaptively reduce the number of flames during the iteration process, that is:

[0094]

[0095] in Indicates the number of flames in the current iteration, is the number of flames, represents the maximum number of iterations, Indicates the current iteration number, Denotes the rounding function. The fitness function used for the moth-to-fire algorithm optimization is the root mean square error between the model predicted voltage and the experimental voltage, which is:

[0096]

[0097] in represents the model predicted voltage, represents the experimental test voltage, RMSE represents the root mean square error, and k represents the kth step length. In this way, the optimal model parameters can be calculated.

[0098] S5. CNN-GRU model training: Take the training set formed by the voltage and current, temperature and state of charge data obtained in step S2 to run the optimized CNN-GRU hybrid model, find the nonlinear mapping relationship between current, temperature, state of charge and voltage, and obtain the trained CNN-GRU hybrid model;

[0099] S6. Obtaining predicted battery voltage: Based on the trained CNN-GRU hybrid model obtained in step S5, input current, temperature, and state of charge, and output battery voltage to obtain predicted battery voltage;

[0100] S7. Generate battery residual: Based on the predicted voltage obtained in step S6, calculate the difference between the predicted voltage and the actual collected voltage, that is, obtain the battery residual ;

[0101] Furthermore, the calculation expression of the battery residual in step S7 is as follows:

[0102]

[0103] in, Indicates the battery residual. Indicates the actual acquisition voltage, Represents the voltage output by the CNN-GRU hybrid model.

[0104] S8. Fault diagnosis: Based on the battery residual obtained in step S7 , determine the fault type and level, the specific steps are as follows:

[0105] S8.1. Determine the battery residual If the battery residual If it is negative, it is an overvoltage state. If it is positive, it is undervoltage state;

[0106] S8.2. At the same time, different thresholds are set according to the different charging and discharging stages of the battery. If the battery residual If the battery residual is higher than the set threshold, the undervoltage fault and corresponding alarm level will be triggered immediately to locate the faulty battery cell. If the voltage is lower than the set threshold, an overvoltage fault and corresponding alarm level will be triggered immediately to locate the faulty battery cell. For specific judgment and fault type, please refer to Table 1 below;

[0107]

[0108] Table 1

[0109] like Figure 1 As shown in the figure, the moth-to-fire algorithm achieves the optimal maximum absolute error value of 0.0891V when the number of iterations is 127. At this time, the hyperparameters of the optimized CNN-GRU hybrid model, including the number of filters, the number of GRU neurons, the number of epochs, the dropout rate, and the batch size are 4, 200, 50, 0.5, and 62 respectively;

[0110] like Figure 2 As shown, the voltage is abnormal at 1660s and 3150s, and the actual vehicle issues an alarm at 4010s. The method proposed in the present invention issues level 1, level 2 and level 3 undervoltage warnings at 1660s, 3150s and 4010s respectively;

[0111] like Figure 3As shown, the voltage was abnormal at 3150s and 3380s, but the actual vehicle issued an alarm at 4010s. The method proposed in the present invention issued a level 1 overvoltage warning, a level 1 overvoltage warning and a level 3 overvoltage warning at 3150s, 3380s and 4010s respectively.

[0112] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A lithium-ion battery fault diagnosis method based on voltage prediction, characterized in that: The following steps are involved: S1. Obtaining real vehicle data of lithium-ion batteries: Collecting real vehicle data of lithium-ion batteries of fault-free electric vehicles, and recording battery charge and discharge voltage, current, temperature, and state of charge in real time; S2, data processing: pre-processing the temperature, current, state of charge and voltage data obtained in S1, including the following steps: S2.1, construct the temperature, current and state of charge data in step S1 as model input {x}; S2.

2. Divide the model input data set {x} into a training set and a test set. The specific expression is as follows: Training set: Test set: Among them, x i represents the model input in the training set, y i Represents the model input x in the corresponding training set i The predicted voltage, x i * represents the model input in the test set, y i * Represents the model input x in the corresponding test set i * The predicted voltage, D train represents the training set, D test represents the test set, and m represents the dimension; S3. CNN-GRU hybrid model construction: Combine the CNN model and the GRU model to build a CNN-GRU hybrid model; S3.

1. Construct a CNN layer through a CNN model to extract spatial features of data, wherein the CNN layer consists of a one-dimensional convolution layer and a maximum pooling layer, wherein the one-dimensional convolution layer consists of a group of filters, and the filters extract spatial information from the input data, and pass the features extracted by the one-dimensional convolution layer to the maximum pooling layer to complete the extraction of spatial features; S3.2, after the CNN layer completes the extraction of spatial features, a GRU layer is constructed through the GRU model to extract temporal features from the model input data; S3.3, set the dropout layer to alleviate overfitting. The specific expression is as follows: Among them, r′ t represents a Bernoulli random variable based on probability p, Bernoulli(*) represents a Bernoulli distribution, y′ t Represents the output value of the dropout layer, h t Represents the hidden state at the current moment and serves as the input value of the dropout layer; S3.4, set the fully connected layer to predict the voltage U t Output is performed to complete the establishment of the CNN-GRU hybrid model; S4, CNN-GRU hybrid model optimization: Use the moth-to-flame optimization algorithm to optimize the hyperparameters of the CNN-GRU hybrid model obtained in S3, where the hyperparameters include the number of filters, the number of GRU neurons, the number of epochs, the dropout rate, and the batch size, to obtain the optimized CNN-GRU hybrid model; Specifically, a flame number adaptive mechanism is proposed to adaptively reduce the number of flames during the iteration process, namely: Where flame.no represents the number of flames in the current iteration, Nf is the number of flames, Mi represents the maximum number of iterations, l represents the current number of iterations, and round represents the rounding function. The fitness function used for the moth-to-flame algorithm optimization is the root mean square error between the model predicted voltage and the voltage in the experiment, which is: Where V sim represents the model predicted voltage, V Exp represents the experimental test voltage, RMSE represents the root mean square error, and k represents the kth step length, so the optimal model parameters can be calculated; S5. CNN-GRU model training: Take the training set formed by the voltage and current, temperature and state of charge data obtained in step S2 to run the optimized CNN-GRU hybrid model, find the nonlinear mapping relationship between current, temperature, state of charge and voltage, and obtain the trained CNN-GRU hybrid model; S6. Obtaining predicted battery voltage: Based on the trained CNN-GRU hybrid model obtained in step S5, input current, temperature, and state of charge, and output battery voltage to obtain predicted battery voltage; S7. Generate battery residual: Based on the predicted voltage obtained in step S6, calculate the difference between the predicted voltage and the actual collected voltage, that is, obtain the battery residual R; S8. Fault diagnosis: Based on the battery residual R obtained in step S7, determine the fault type and level. The specific steps are as follows: S8.

1. Determine the positive or negative of the battery residual error R. If the battery residual error R is negative, it is an overvoltage state. If the battery residual error R is positive, it is an undervoltage state. S8.

2. At the same time, different thresholds are set according to the different charging and discharging stages of the battery. If the battery residual error R is higher than the set threshold, an undervoltage fault and a corresponding alarm level are immediately triggered to locate the faulty battery cell; if the battery residual error R is lower than the set threshold, an overvoltage fault and a corresponding alarm level are immediately triggered to locate the faulty battery cell.

2. A lithium-ion battery fault diagnosis method based on voltage prediction according to claim 1, characterized in that: The expression of the filter extracting spatial information in the CNN layer is as follows: Among them, x t ,y t 、w c , b c Represent the input, output, weight and bias value of the convolutional layer respectively, represents the convolution operation, and f(*) represents the activation function; The expression of the feature extracted by the one-dimensional convolution layer in step S3.1 and passed to the maximum pooling layer is as follows: p t =max{y t (n×1) m(n,1)} Among them, p t represents the output of the maximum pooling layer, m(n,1) represents the window size during the pooling process, and y t (n×1) Represents the output value for n×1-dimensional input.

3. A lithium-ion battery fault diagnosis method based on voltage prediction according to claim 2, characterized in that: The specific expression of the GRU layer in step S3.2 is as follows: Among them, r t Represents the reset gate at the current moment, W r represents the weight matrix of the reset gate, h t-1 represents the hidden state at the previous moment, h t Indicates the current hidden state, z t represents the update gate, W s represents the weight matrix of the update gate, c t represents the candidate state at the current moment, W c represents the weight matrix of the candidate state, the subscript t represents the time, σ represents the sigmoid activation function, tanh represents the hyperbolic tangent function, ⊙ represents the element product, and p t Represents the output of the maximum pooling layer and serves as the input of the GRU layer; In step S3.4, the voltage U is predicted by the fully connected layer. t The expression is as follows: U t =W f y′ t +b f Among them, y′ t Represents the output value of the dropout layer, W f represents the weight matrix of the fully connected layer, b f Represents the bias vector of the fully connected layer.

4. The lithium-ion battery fault diagnosis method based on voltage prediction according to claim 1, characterized in that: The calculation expression of the battery residual in step S7 is as follows: Among them, R represents the battery residual, Indicates the actual collected voltage, U t Represents the voltage output by the CNN-GRU hybrid model.

Images