Lithium ion battery health state evaluation modeling method based on transfer learning

Abstract

The invention provides a lithium ion battery health state evaluation modeling method based on transfer learning. The lithium ion battery health state evaluation modeling method comprises the following steps: determining performance parameters of a lithium ion battery energy storage system having a change trend in a life cycle; taking the discharge capacity of the lithium ion battery energy storage system as a health state parameter, dividing the state parameter into four levels and setting corresponding labels; the method comprises the following steps: acquiring performance parameter data and label data of a lithium ion battery energy storage system in development and test stages, and constructing a source domain data set FS; using the source domain data set FS to train a mapping relation model between six performance parameters and state tags established by using a stacked auto-encoder algorithm to obtain a model parameter set; and constructing a target domain data set FT, optimizing the model parameter set, verifying the precision of the mapping relation model, and obtaining the lithium ion battery health state evaluation model. The technical problem of performance information fusion of multiple performance parameters in the use process of the lithium ion battery energy storage system can be solved, and health state grading of the lithium ion battery energy storage system is achieved.

Description

Technical Field

[0001] The invention relates to a lithium-ion battery health status assessment modeling method based on transfer learning, and belongs to the field of lithium-ion battery fault prediction and health management technology (PHM). Background Art

[0002] Lithium-ion battery energy storage system is a typical secondary energy source that can be charged and discharged multiple times and is widely used in various fields of industry and society. In the process of lithium-ion battery research and development, in order to fully improve the product performance indicators, it is necessary to continuously carry out various types of research and development tests, cycle charge and discharge tests and accelerated life tests. The product research and development department has accumulated a large amount of test and use data, covering the life cycle of the battery. However, the current battery health status classification is mostly roughly determined by the number of cycles, service life, etc., and multi-source data has not been fully mined and utilized, making it difficult to determine the health level of the battery energy storage system and support usage decisions. In the context of multi-source data, how to mine multi-stage and multi-type performance data of lithium-ion battery energy storage systems, establish a mapping relationship between life cycle data and status labels, and migrate the model from the high-rate charge and discharge and high-stress conditions of the research and development test to the actual use conditions, so as to evaluate the health status of the lithium-ion battery energy storage system, classify the health status, and assist the equipment user department in the use and maintenance decisions of the energy system, has become an urgent problem to be solved by the current lithium-ion battery energy storage system development and equipment operation departments. Summary of the invention

[0003] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art and provide a lithium-ion battery health status assessment modeling method based on transfer learning. According to the idea of ​​"model parameter migration + fine-tuning", the life cycle data accumulated by the lithium-ion battery energy storage system in various types of tests is fully integrated to solve the problem of accurately assessing the health status during use.

[0004] The technical solution of the present invention is: a lithium-ion battery health status assessment modeling method based on transfer learning, comprising:

[0005] Determine the performance parameters of lithium-ion battery energy storage systems that have changing trends during their life cycle;

[0006] After each discharge cycle is completed, the discharge capacity of the lithium-ion battery energy storage system is used as the health status parameter, which is divided into four levels, and a corresponding label is set for each level;

[0007] Obtain the performance parameter data and label data of the lithium-ion battery energy storage system during the development and testing phase, sort them in time sequence, and construct the source domain dataset F S , including the input set X S and the output set Y S ;

[0008] Using the source domain dataset F S The mapping relationship model between the six performance parameters and state labels established by the stacked autoencoder algorithm is trained to obtain the source domain F S The model parameter set

[0009] Construct the target domain dataset F T , and divide it into a target domain normal state dataset and a target domain abnormal state dataset, and use the target domain normal state dataset to adjust the model parameter set Optimize and obtain the target domain F T The model parameter set Then use the target domain abnormal state dataset to calibrate the model parameter set Conduct tests to verify the accuracy of the mapping relationship model;

[0010] The mapping relationship model after verification is a lithium-ion battery health status assessment model, which can perform health assessment on the lithium-ion battery energy storage system.

[0011] Preferably, the performance parameters include direct measurement values ​​and indirect measurement values, specifically:

[0012] Direct measurements include: battery internal temperature, charge and discharge internal resistance, and maximum voltage difference of single cells;

[0013] Indirect measurement values ​​include: discharge median voltage, constant current charging time, constant voltage charging voltage value;

[0014] The direct measurement value records the maximum value within a discharge cycle in real time as a data sample, and the discharge capacity and indirect measurement value are recorded directly after a discharge cycle is completed.

[0015] Preferably, the discharge capacity is divided into four levels, which are specifically defined as:

[0016] If the discharge capacity is less than 85%, it is considered "abnormal" and the label is set to "3";

[0017] When the discharge capacity is between 85% and 90%, it is “accelerated decay” and the label is set to “2”;

[0018] When the discharge capacity is between 90% and 95%, it is “fading” and the label is set to “1”;

[0019] A discharge capacity greater than or equal to 95% is defined as good and the label is set to "0".

[0020] Preferably, the source domain dataset F S The construction process is as follows:

[0021] Construct source domain dataset FS The input set X S When: Get six performance parameters of the lithium-ion battery energy storage system with decay characteristics at different times, namely: battery internal temperature x1, charge and discharge internal resistance x2, single cell maximum voltage difference x3, discharge median voltage x4, constant current charging time x5, constant voltage charging voltage value x6; express the input set X in matrix form S :

[0022]

[0023] Where k is the discharge cycle, and t represents the time when the parameter is recorded in the first discharge cycle;

[0024] Construct source domain dataset F S The output set Y S When: During the development and testing of the lithium-ion battery energy storage system, the system status label is set according to the discharge capacity each time; the output set Y is represented in matrix form S :

[0025]

[0026] Among them, Z1, Z2, Z3, and Z4 are set to labels corresponding to the ranges of the four discharge capacities: when the battery discharge capacity is labeled "0", Z1 is set to 1, and Z2, Z3, and Z4 are set to 0; similarly, when the battery discharge capacity labels are "1", "2", and "3" respectively, Z2, Z3, and Z4 are set to 1 respectively, and the rest are set to 0.

[0027] Preferably, the source domain dataset F is used S Training the mapping relationship model between the six performance parameters and the state labels established by the stacked autoencoder algorithm includes the following steps:

[0028] S5-1 Source domain dataset F S Input set X S The data is normalized so that the value range of the sample data is between [0,1];

[0029] S5-2 Using the normalized input set X S The mapping relationship model between performance parameters and state labels is trained through two processes: forward unsupervised learning and backward supervised fine-tuning to obtain the source domain F S The model parameter set The model parameter set includes the weights and biases of the stacked autoencoders.

[0030] Preferably, in forward unsupervised learning, a stacked autoencoder network is used to obtain the input set X S The reconstruction value z is based on the input set X SThe mean square error between the reconstructed value z and the original value z is used as the objective function J. The minimum value of J is used as the target for training layer by layer to obtain the optimal value of weights and biases. The model parameter set is denoted as θ S ; The objective function J is specifically:

[0031]

[0032] Where k is the discharge period, z i is the performance parameter reconstruction value obtained by the self-stacked encoder, x i is the performance parameter in the input set.

[0033] Preferably, a backward-propagation algorithm is used for backward supervised fine-tuning, specifically:

[0034] State classification results using stacked autoencoder outputs during feed-forward unsupervised learning and the source domain output set Y S The corresponding state label y establishes the minimum cost function J B :

[0035]

[0036] By minimizing the cost function J B To optimize and update the model parameter set θ S The weights and biases in are as follows:

[0037]

[0038] Among them, L is the number of stacked autoencoder stacking layers, Represents the model parameter set θ S The weight between the i-th input layer neuron and the j-th hidden layer neuron in the n-th stacked autoencoder; Represents the model parameter set θ S In the n-th layer stacked autoencoder, the bias of the i-th input layer neuron, α is the learning rate;

[0039] After completing the supervised fine-tuning, we obtain the source domain F S The model parameter set

[0040] Preferably, the source domain F is constructed according to S The target domain dataset F is constructed by using the dataset method T , when dividing it into the target domain normal state data set and the target domain abnormal state data set, the data labels are divided according to the battery capacity:

[0041] The performance parameters corresponding to the discharge capacity labels "0" and "1" are used as the target domain normal state input set;

[0042] The performance parameters corresponding to the discharge capacity labels “2” and “3” are used as the target domain abnormal state input set.

[0043] In a second aspect, a terminal device is provided, including:

[0044] a memory for storing instructions executed by at least one processor;

[0045] The processor is used to execute instructions stored in the memory to implement the above-mentioned lithium-ion battery health status assessment modeling method based on transfer learning.

[0046] In a third aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and when the computer instructions are executed on a computer, the computer executes the above-mentioned lithium-ion battery health status assessment modeling method based on transfer learning.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] (1) The present invention innovatively proposes to integrate the life cycle data under high rate and high stress conditions in the development and testing of lithium-ion battery energy storage systems, and divide the battery into healthy, degraded, accelerated degradation and abnormal states according to the discharge capacity, and use stacked autoencoders to characterize the mapping relationship between performance parameters and health status, thereby achieving accurate description of the health status from the battery internal temperature, charge and discharge internal resistance, maximum voltage difference of single cells, discharge median voltage, constant current charging time, and constant voltage charging voltage value performance parameters;

[0049] (2) The present invention innovatively proposes to transfer the model parameters trained during the development and testing of the lithium-ion battery energy storage system to the actual operation process, and combines the characteristics of low rate and conventional stress in actual use to transfer the model parameters, fine-tune the model parameters, and improve the generalization ability of the model, thus realizing the model migration from knowledge accumulation in the product development process to operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the method of the present invention;

[0051] Figure 2 Schematic diagram of the stacked autoencoder model of the present invention;

[0052] Figure 3 This is a schematic diagram of transfer learning of the present invention;

[0053] Figure 4 Schematic diagram of the confusion matrix calculated by the present invention. DETAILED DESCRIPTION

[0054] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art and propose a technical method for health status assessment modeling of a lithium-ion battery energy storage system based on transfer learning. According to the idea of ​​"model parameter migration + fine-tuning", the life cycle data accumulated by the lithium-ion battery energy storage system in various types of tests is fully integrated to solve the problem of accurately assessing its health status during use.

[0055] A health status classification modeling method for lithium-ion battery energy storage system based on transfer learning, such as Figure 1 As shown, it includes the following main steps:

[0056] S1. Determine the performance parameters of the lithium-ion battery energy storage system that have a changing trend during its life cycle, and use the discharge capacity as the basis for state division. The performance parameters include two types of direct measurement values ​​and indirect measurement values, totaling 6 parameters. Direct measurement values ​​include: battery internal temperature, charge and discharge internal resistance, and maximum voltage difference of single cells. Indirect measurement values ​​include: discharge median voltage, constant current charging time, and constant voltage charging voltage value. Direct measurement values ​​are obtained by real-time recording to obtain the maximum value within a discharge cycle as a data sample, and indirect measurement values ​​are directly recorded after a discharge cycle is completed;

[0057] S2. After each discharge cycle is completed, the discharge capacity of the lithium-ion battery energy storage system is used as the health status parameter;

[0058] S3, the discharge capacity is divided into four levels: "good", "decline", "accelerated decline", and "abnormal", and the capacity below 85% is defined as "abnormal", the capacity between 85% and 90% is defined as "accelerated decline", the capacity between 90% and 95% is defined as "decline", and the capacity greater than or equal to 95% is defined as good;

[0059] S4, in the source domain F S , obtain the data of the development and test phase of the lithium-ion battery energy storage system. Since high-rate and high-environmental stress charge and discharge tests are usually carried out during the development phase, the full life cycle data can be obtained. After sorting in time sequence, the health degradation state data label is set according to the discharge capacity defined in S3 to obtain a labeled data set;

[0060] Source domain F S The dataset construction process is as follows:

[0061] Source domain F S Input set construction: Obtain six performance parameters of the lithium-ion battery energy storage system with degradation characteristics at different times, namely: x1 battery internal temperature, x2 charge and discharge internal resistance, x3 single cell maximum voltage difference, x4 discharge median voltage, x5 constant current charging time, x6 constant voltage charging voltage value. Represent the source domain F in matrix form S The input set X S, the form is as follows.

[0062]

[0063] Wherein, k is the discharge cycle, and t represents the time when the parameter is recorded in the first discharge cycle.

[0064] Source domain F S Output set construction: In the process of lithium-ion battery energy storage system development and testing, the system status label is set according to each discharge capacity. "0" represents the battery discharge capacity greater than 95%, "1" represents the discharge capacity between 90% and 95%, "2" represents the discharge capacity between 85% and 90%, and "3" represents the discharge capacity below 85%. The source domain F is represented in matrix form. S The output set Y S , of the following form:

[0065]

[0066] Among them, k is the discharge cycle. When the battery discharge capacity is labeled "0", Z1 is set to 1, and Z2, Z3, and Z4 are set to 0. Similarly, when the battery discharge capacity is labeled "1", "2", and "3", Z2, Z3, and Z4 are set to 1 respectively, and the rest are set to 0.

[0067] S5. Using the stacked auto-encoders (SAE) algorithm, a mapping relationship model between six performance parameters and state labels is established, which includes four steps:

[0068] (1) For the source domain F S The six performance parameters mentioned in step S1, namely, the internal temperature of the battery, the charge and discharge internal resistance, the maximum voltage difference of the single battery, the median discharge voltage, the constant current charging time, and the constant voltage charging voltage value, are normalized so that the value range of the sample data is between [0,1];

[0069] (2) Using the source domain input set X S The mapping relationship model between performance parameters and state labels is trained, including two processes: forward unsupervised learning and backward supervised fine-tuning.

[0070] In the forward unsupervised learning process, the source domain input set X is used S Perform model training.

[0071] A single autoencoder consists of an input layer and a hidden layer, such as Figure 2 As shown in Figure 1, the stacked autoencoder is composed of multiple autoencoders stacked together. The input of the first autoencoder of the stacked autoencoder is 6 performance parameters, which are then stacked layer by layer. The hidden layer of the previous autoencoder is the input layer of the next autoencoder, and so on.

[0072] The weight parameter W and bias b are used in each autoencoder to represent the transfer relationship between the input layer and the hidden layer. The stacked autoencoder includes two parts: encoding and decoding. The encoding process is from the input layer to the hidden layer, and the decoding process is from the hidden layer to the output layer.

[0073] Assume that the number of stacked autoencoder stacking layers is L, and express the weight matrix parameter W and bias vector b of the stacked autoencoder in matrix form:

[0074] W=[w (1) ,w (2) ,L,w (n) ,L,w (L) ]

[0075] b=[b (1) ,b (2) ,L,b (n) ,L,b (L) ]

[0076] by represents the element in W, that is, the weight between the i-th input layer neuron and the j-th hidden layer neuron in the n-th stacked autoencoder. Represents the elements in b, that is, the bias vector of the i-th input layer neuron in the n-th stacked autoencoder.

[0077] Stack the autoencoder network to take the input set X S The mean square error J between the reconstructed value z and the original value z is used as the objective function to find the optimal values ​​of the weight matrix parameters W and the bias vector b:

[0078]

[0079] Where k is the discharge period, z i is the performance parameter reconstruction value obtained by the self-stacked encoder, x i is the performance parameter in the input set.

[0080] Take the loss function J to get the minimum value as the goal of layer-by-layer training, obtain the weight matrix parameter W and the bias vector b, and record the weight matrix parameter W and the bias vector b model parameter set obtained in this step as θ S .

[0081] (3) Using the source domain output set Y S , based on the back propagation algorithm (BP) S Perform reverse supervised fine-tuning.

[0082]

[0083] In the formula, is the state classification result of the output of the stacked autoencoder input performance parameter in the forward unsupervised learning process, and y is the source domain output set Y S Corresponding state labels, BP reverse fine-tuning algorithm minimizes the cost function J B to optimize the model parameters.

[0084] Update θ S The weight matrix parameter W in s and the bias vector b s The rules are as follows:

[0085]

[0086] Represents the model parameter set θ S The weight between the i-th input layer neuron and the j-th hidden layer neuron in the n-th stacked autoencoder. Represents the model parameter set θ S In the n-th layer stacked autoencoder, the bias of the i-th input layer neuron is α, and α is the learning rate.

[0087] After fine-tuning, we obtain the source domain F S The model parameter set

[0088] S6. Set the model parameters Migrate to target domain F T , and in the target domain F T Model parameter learning is performed within Figure 3 As shown. During use, lithium-ion battery energy storage systems are usually shallowly charged and discharged, and the discharge rate is small to ensure service life. Therefore, the accumulated data are mostly healthy data, and there is less fault data.

[0089] In the target domain F T Model parameter optimization learning consists of four steps:

[0090] (1) In the target domain F T In the example, according to the rules specified in S3, the data labels are divided by battery capacity, and the performance parameters corresponding to the labels "0" and "1" are used as the normal state input set, denoted as X T-normal , the label is recorded as the normal state output set, recorded as Y T-normal ; The performance parameters corresponding to the labels "2" and "3" are taken as the abnormal state input set, denoted as X T-failure , the label is recorded as the abnormal state output set, recorded as Y T-failure ;

[0091] (2) Based on the SAE algorithm, using the target domain F T The normal state input set and output set data in the data set are the model parameter set θs * Optimize and obtain the target domain F T The weight matrix parameter W T and the bias vector b T The model parameter set θ T * ;

[0092] (3) Using the target domain F T The abnormal state input set and output set in are tested, and the confusion matrix of state classification is calculated, such as Figure 4 As shown, the accuracy of the transfer learning health status assessment model based on SAE lithium-ion battery energy storage system under different working modes is verified;

[0093] (4) After the verification is completed, the model with a certain degree of accuracy can be used to conduct health assessment of the lithium-ion battery energy storage system. At the same time, as the target domain data continues to accumulate, (1) to (4) can be repeated to continuously optimize the model parameters.

[0094] In a second aspect, the present invention provides a terminal device, comprising:

[0095] a memory for storing instructions executed by at least one processor;

[0096] The processor is used to execute instructions stored in the memory to implement the above-mentioned lithium-ion battery health status assessment modeling method based on transfer learning.

[0097] In a third aspect, the present invention provides a computer-readable storage medium, which stores computer instructions. When the computer instructions are executed on a computer, the computer executes the above-mentioned lithium-ion battery health status assessment modeling method based on transfer learning.

[0098] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in this field.

Claims

1. A lithium-ion battery health status assessment modeling method based on transfer learning, characterized in that include: Determine the performance parameters of lithium-ion battery energy storage systems that have changing trends during their life cycle; After each discharge cycle is completed, the discharge capacity of the lithium-ion battery energy storage system is used as the health status parameter, which is divided into four levels, and a corresponding label is set for each level; Obtain the performance parameter data and label data of the lithium-ion battery energy storage system during the development and testing phase, sort them in time sequence, and construct the source domain dataset F S , including the input set X S and the output set Y S ; Using the source domain dataset F S The mapping relationship model between the six performance parameters and state labels established by the stacked autoencoder algorithm is trained to obtain the source domain F S The model parameter set Construct the target domain dataset F T , and divide it into a target domain normal state dataset and a target domain abnormal state dataset, and use the target domain normal state dataset to adjust the model parameter set Optimize and obtain the target domain F T The model parameter set Then use the target domain abnormal state dataset to calibrate the model parameter set Conduct tests to verify the accuracy of the mapping relationship model; The mapping relationship model after verification is a lithium-ion battery health status assessment model, which can perform health assessment on the lithium-ion battery energy storage system.

2. The lithium-ion battery health status assessment modeling method based on transfer learning according to claim 1 is characterized in that: Performance parameters include direct and indirect measurements, specifically: Direct measurements include: battery internal temperature, charge and discharge internal resistance, and maximum voltage difference of single cells; Indirect measurement values ​​include: discharge median voltage, constant current charging time, constant voltage charging voltage value; The direct measurement value records the maximum value within a discharge cycle in real time as a data sample, and the discharge capacity and indirect measurement value are recorded directly after a discharge cycle is completed.

3. The lithium-ion battery health status assessment modeling method based on transfer learning according to claim 1 is characterized in that: The discharge capacity is divided into four levels, which are defined as: If the discharge capacity is less than 85%, it is "abnormal" and the label is set to "3"; When the discharge capacity is between 85% and 90%, it is "accelerated decay", and the label is set to "2"; When the discharge capacity is between 90% and 95%, it is "fading", and the label is set to "1"; A discharge capacity greater than or equal to 95% is defined as good and the label is set to "0".

4. The lithium-ion battery health status assessment modeling method based on transfer learning according to claim 3 is characterized in that: Source domain dataset F S The construction process is as follows: Construct source domain dataset F S The input set X S When: Get six performance parameters of the lithium-ion battery energy storage system with decay characteristics at different times, namely: battery internal temperature x1, charge and discharge internal resistance x2, single cell maximum voltage difference x3, discharge median voltage x4, constant current charging time x5, constant voltage charging voltage value x6; express the input set X in matrix form S : Where k is the discharge cycle, and t represents the time when the parameter is recorded in the first discharge cycle; Construct source domain dataset F S The output set Y S When: During the development and testing of the lithium-ion battery energy storage system, the system status label is set according to the discharge capacity each time; the output set Y is represented in matrix form S : Among them, Z1, Z2, Z3, and Z4 are set to labels corresponding to the ranges of the four discharge capacities: when the battery discharge capacity is labeled "0", Z1 is set to 1, and Z2, Z3, and Z4 are set to 0; similarly, when the battery discharge capacity labels are "1", "2", and "3" respectively, Z2, Z3, and Z4 are set to 1 respectively, and the rest are set to 0.

5. The lithium-ion battery health status assessment modeling method based on transfer learning according to claim 1, characterized in that: Using the source domain dataset F S Training the mapping relationship model between the six performance parameters and the state labels established by the stacked autoencoder algorithm includes the following steps: S5-1 Source domain dataset F S Input set X S The data is normalized so that the value range of the sample data is between [0,1]; S5-2 Using the normalized input set X S The mapping relationship model between performance parameters and state labels is trained through two processes: forward unsupervised learning and backward supervised fine-tuning to obtain the source domain F S The model parameter set The model parameter set includes the weights and biases of the stacked autoencoders.

6. The lithium-ion battery health status assessment modeling method based on transfer learning according to claim 5 is characterized in that: In forward unsupervised learning, we use a stacked autoencoder network to obtain the input set X S The reconstruction value z is based on the input set X S The mean square error between the reconstructed value z and the original value z is used as the objective function J. The minimum value of J is used as the target for training layer by layer to obtain the optimal value of weights and biases. The model parameter set is denoted as θ S ; The objective function J is specifically: Where k is the discharge period, z i is the performance parameter reconstruction value obtained by the self-stacked encoder, x i is the performance parameter in the input set.

7. The lithium-ion battery health status assessment modeling method based on transfer learning according to claim 6 is characterized in that: Use the backpropagation algorithm for backward supervised fine-tuning, specifically: State classification results using stacked autoencoder outputs during feed-forward unsupervised learning and the source domain output set Y S The corresponding state label y establishes the minimum cost function J B : By minimizing the cost function J B To optimize and update the model parameter set θ S The weights and biases in are as follows: Among them, L is the number of stacked autoencoder stacking layers, Represents the model parameter set θ S The weight between the i-th input layer neuron and the j-th hidden layer neuron in the n-th stacked autoencoder; Represents the model parameter set θ S In the n-th layer stacked autoencoder, the bias of the i-th input layer neuron, α is the learning rate; After completing the supervised fine-tuning, we obtain the source domain F S The model parameter set 8. The lithium-ion battery health status assessment modeling method based on transfer learning according to claim 4 is characterized in that: Based on constructing the source domain F S The target domain dataset F is constructed by using the dataset method T , when dividing it into the target domain normal state data set and the target domain abnormal state data set, the data labels are divided according to the battery capacity: The performance parameters corresponding to the discharge capacity labels "0" and "1" are used as the target domain normal state input set; The performance parameters corresponding to the discharge capacity labels "2" and "3" are used as the target domain abnormal state input set.

9. A terminal device, characterized in that: include: a memory for storing instructions executed by at least one processor; A processor, configured to execute instructions stored in a memory to implement a method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and when the computer instructions are executed on a computer, the computer is enabled to execute the method according to any one of claims 1 to 8.

Images