Lithium ion battery monomer / module capacity estimation method based on deep migration

Abstract

The invention relates to a lithium ion battery monomer / module capacity estimation method based on a deep migration model, which comprises the following steps: collecting charge-discharge cycle data in a lithium ion battery aging test process, and taking the maximum capacity in the whole discharge process in the lithium ion battery cycle process as a capacity label; according to the relation between the battery polarization reaction intensity and the battery capacity, by means of depolarization characteristic data of constant-voltage charging current and relaxation voltage, multi-dimensional practical characteristics which are not affected by a charging starting point are extracted, the Spearman correlation coefficient between the selected multi-dimensional practical characteristics and the capacity is calculated, the characteristics with the correlation coefficient larger than 0.8 are reserved, and the battery capacity is calculated. Carrying out fusion processing on the features by using PCA, and taking the previous four principal components as dimension reduction features; and constructing a hybrid neural network DBN-SE-LSTM, and carrying out transfer learning by taking the dimension reduction features as input for online estimation of the capacity. The method can indirectly evaluate the capacity of the lithium ion battery monomer or module according to part of battery data, and has high estimation precision and good applicability.

Description

Technical Field

[0001] The present invention relates to the field of lithium-ion battery key state estimation, and in particular to a method for estimating the capacity of a lithium-ion battery cell / module based on a deep migration model. Background Art

[0002] Lithium-ion batteries are widely used in key areas such as portable electronic devices, new energy electric vehicles, and large-scale energy storage power stations due to their high energy density, long cycle life, and pollution-free. However, during long-term storage or frequent charging and discharging, the internal structure of the battery will change, such as loss of cyclable lithium and collector corrosion, which will lead to gradual degradation of battery performance. Capacity, as a key parameter to measure battery performance, can not only reflect the battery's state of health (SOH) and remaining useful life (RUL), but is also a key factor in judging the inconsistency of cells in the battery pack and the overall charging strategy. Therefore, accurate prediction of battery capacity is of great significance to the overall monitoring and safety management of the battery.

[0003] At present, battery capacity estimation methods are mainly divided into two categories: model-based methods and data-driven methods. Model-based methods reflect the capacity decay process of lithium-ion batteries by constructing mathematical relationship expressions. However, the construction of battery models is challenging and requires a comprehensive and in-depth understanding of the electrochemical characteristics of batteries. In addition, the accurate identification of model parameters is difficult and usually depends on the convergence and convergence speed of the filtering algorithm. The identification results are closely related to ambient temperature, battery materials, etc. Data-driven methods do not need to consider the complex electrochemical reaction process inside the battery, but directly use historical data to estimate the battery capacity, which has high transferability and generalization. Generally, the determination of data-driven models and the selection of model input features play a key role in the estimation accuracy, but there are still the following problems. First, in terms of data-driven models, traditional models are difficult to apply to the nonlinear capacity decay process of lithium-ion batteries under different conditions, and the model training samples are strict, often requiring a large amount of original data in the target field. Second, the currently selected features are based on complete charge and discharge curves, which are easily affected by the random charging starting point of the battery, and are difficult to apply online.

[0004] Therefore, the present invention proposes a lithium-ion battery cell-module capacity estimation method based on deep migration. On the one hand, by studying the typical charging and discharging process, multi-dimensional practical features that can reflect the polarization attenuation characteristics of the battery are extracted. On the other hand, by integrating the deep belief network, the long short-term memory network and the squeeze-incentive mechanism, a hybrid neural network with adaptive channel attention capability is constructed. Based on this network, a deep migration model combining semi-supervised learning and reconstruction fine-tuning migration strategy is proposed to improve the accuracy of lithium-ion battery capacity estimation. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention aims to provide a capacity estimation method for lithium-ion battery cells / modules based on deep migration. The capacity parameters of battery cells or modules are accurately estimated through the actual usable data of lithium-ion batteries and the deep migration model, thereby providing an important basis for the overall monitoring and safety management of batteries.

[0006] The technical solution of the present invention to solve the technical problem is: A method for estimating the capacity of a lithium-ion battery cell / module based on a deep migration model, the method comprising the following steps: Step S1: Collecting charge and discharge cycle data during the lithium-ion battery aging test process, and taking the maximum capacity of the entire discharge process during the lithium-ion battery cycle as the capacity label; Step S2: Feature selection 1) Based on the relationship between the intensity of battery polarization reaction and battery capacity, the depolarization characteristic data of constant voltage charging current and relaxation voltage are used to extract multi-dimensional practical features that are not affected by the charging starting point, including: Select the initial current value, average current value and charging capacity in the last 10 minutes before the end of constant voltage charging, and record them as I t ,I m , Q t ; Select the time and charge capacity from the constant voltage charge current of 0.15C to the end of charge, and record them as T and Q respectively. i ; The exponential function is used to fit the battery relaxation voltage curve, and the fitting curve is expressed as:

[0007] In the formula, V represents the fitting voltage, S represents the intercept, a1-a2 is the amplitude of the exponential decay link, τ1-τ2 is the time constant, and t is the time; The fitting coefficients S, a1, a2, τ1, τ2 are used as features; 2) Calculate the Spearman correlation coefficient between the selected multi-dimensional features and the capacity, discard the features with a correlation coefficient less than 0.7, and linearly combine the two features a1 and a2 as new fusion features, retain the features with a correlation coefficient greater than 0.8, and preliminarily screen and determine the constant pressure link I t ,I m , Q t , T and Q i and relaxation voltage fitting coefficients a1+a2, S and τ1 for subsequent processing; 3) All the preliminarily screened features are combined into a matrix A, and the preliminarily screened features are fused using PCA technology, with the first four principal components PC1, PC2, PC3, and PC4 used as dimensionality reduction features; The principal components PC1, PC2, PC3, and PC4 are used as the input of the hybrid neural network DBN-SE-LSTM; Step S3: Construct a hybrid neural network DBN-SE-LSTM: add an SE adaptive channel attention module to each RBM layer of the DBN network, use an LSTM network in the linear output layer of the DBN network, use the high-order information learned by the DBN network on the input data as the input of the LSTM network, and connect the output of the LSTM network to the fully connected layer to obtain the estimated value of the capacity; Step S4: Transfer learning: (1) The PC1, PC2, PC3, PC4 and capacity labels after feature selection are divided into the source domain and the target domain. The source domain data all have capacity labels, while the target domain has only a small amount of capacity-labeled data and a large amount of uncapacity-labeled data. (2) Unlabeled data pre-training and domain adaptation: The hybrid neural network DBN-SE-LSTM is pre-trained using data without capacity labels in the target and source domains. The maximum mean difference (MMD) method is used to reduce the feature distribution difference between the source and target domains, and obtain a feature-sharing pre-training model, which enables the pre-training model to learn high-level abstract information shared by the source and target domains. (3) Supervised training of source domain data: Use labeled data from the source domain to conduct supervised training of the entire network and build the transfer learning basic model TL1. In this process, the back propagation mechanism is used to update the parameters of each hidden layer to capture the mapping relationship between the battery capacity and the deep feature information learned by the hybrid neural network DBN-SE-LSTM. (4) Training the adaptation layer of the TL1 model: Freeze all hidden layer parameters of the trained basic model TL1 except the adaptation layer; train the adaptation layer using labeled data from the target domain to quickly capture the mapping relationship between the target domain features and the battery capacity, and obtain the migration model TL2; (5) Fine-tuning the entire network: Based on the TL2 model, unfreeze all hidden layers and fine-tune the parameters of the entire network using a learning rate of 0.00001 and a training number of 40-60 times to obtain the trained deep migration model TL-DBN-SE-LSTM; Step S5: Input the features of the lithium-ion battery cell / module to be estimated into the trained deep migration model TL-DBN-SE-LSTM after feature selection for online estimation of capacity.

[0008] Furthermore, the number of RBM layers is 4, and the number of RBM hidden layer units is [4, 130, 80, 40, 30]; the number of LSTM network layers is 3, and the number of units is [30, 40, 20]; the number of fully connected layers is 3, and the number of units is [15, 10, 1].

[0009] Furthermore, in the battery capacity estimation at the module level, the depolarization characteristic data of the constant voltage charging current and the relaxation voltage in step S2 are the total current in the constant voltage charging stage and the relaxation voltage of the worst performing series cell.

[0010] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a capacity estimation method for lithium-ion battery cells / modules based on deep migration. The depolarization characteristics of constant voltage current and relaxation voltage decay are used to extract multidimensional practical features that are not affected by the charging starting point. At the same time, the correlation coefficient method and principal component analysis method are introduced to pre-process the features for online capacity estimation. By integrating the deep belief network (DBN), the long short-term memory network (LSTM) and the squeeze-excitation (SE) mechanism, a hybrid neural network with adaptive channel attention capability is constructed, and a migration strategy combining semi-supervised learning and reconstruction fine-tuning is proposed to build a deep migration model to improve the capacity estimation accuracy of lithium-ion batteries. The features extracted by the present invention can adapt to the random (different charging starting points) charging application scenarios of lithium-ion batteries, and with the constructed deep migration model, the capacity of lithium-ion battery cells or modules can be indirectly evaluated based on partial battery data, with high estimation accuracy and good applicability.

[0011] In the present invention, lithium-ion battery data of various working conditions are used to train a hybrid neural network, and the trained deep migration model is used for lithium-ion battery capacity estimation.

[0012] 1. The present invention utilizes the depolarization characteristics of constant voltage current and relaxation voltage decay to extract polarization decay characteristics that are not affected by the charging starting point, specifically the current characteristics ten minutes after charging and the voltage characteristics at the end of charging. Compared with traditional health characteristics such as capacity increment, constant voltage rise time, constant voltage drop time and average voltage, the characteristics proposed in this article can eliminate the influence of the randomness of the charging starting point, ensure a strong correlation with the battery capacity, reduce the amount of calculation and data collection, and can be well matched to lithium-ion batteries in actual use.

[0013] 2. By embedding the SE model into each unit RBM of the DBN network, the focusing ability of key information when it is transmitted between hidden layers is ensured, and the feature extraction ability of the DBN model is also improved. At the same time, a migration strategy combining semi-supervised learning and reconstruction fine-tuning is proposed to effectively solve the problem of low model estimation accuracy under small samples. The results of the embodiment show that this method can improve the model performance by nearly 40%, and can fit the battery capacity well under the same sample size. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Schematic diagram of the overall process of lithium-ion battery capacity estimation.

[0015] Figure 2 Schematic diagram of the training method that combines semi-supervised learning with reconstruction fine-tuning transfer strategy.

[0016] Figure 3 This is the capacity decay curve of lithium-ion battery.

[0017] Figure 4 It is the dual polarization equivalent circuit model of lithium-ion battery.

[0018] Figure 5 The following is a typical charge and discharge process diagram, where (a) shows the change of current 1000s before the end of constant voltage charging under different cycles of the battery. Figure 5 (b) The change of relaxation voltage under different cycles of the battery.

[0019] Figure 6 This is a graph showing the change trend between the normalized eight features of the NCM-CY-35-05_1-#2 battery and the capacity.

[0020] Figure 7 Schematic diagram of the structure of the hybrid neural network.

[0021] Figure 8 Figure 2 is the capacity estimation results and error distribution diagram of Experiment A, where (a) is the curve diagram of the capacity estimation results of the Experiment A model, and (b) is the error distribution diagram of the capacity estimation of the Experiment A model.

[0022] Fig. 9Figure 2 is the capacity estimation results and error distribution diagram of Experiment B, where (a) is the curve diagram of the capacity estimation results of the Experiment B model, and (b) is the error distribution diagram of the capacity estimation of the Experiment B model. DETAILED DESCRIPTION

[0023] Specific embodiments are given below in conjunction with the accompanying drawings. The specific embodiments are only used to illustrate the technical solutions of the present invention in detail and are not intended to limit the protection scope of the present application.

[0024] The present invention provides a capacity estimation method for lithium-ion battery cells / modules based on depth migration, such as Figure 1 , Figure 2 As shown, the specific steps include: Step S1: collecting charge and discharge cycle data of different types of lithium-ion battery aging test processes; This example selects 2 NCA batteries and 1 NCM battery aging test data sets from the Tongji University aging data set. The nominal capacity of the batteries is 3.5Ah. The battery cycle process includes CC-CV charging, 30 minutes of rest, CC discharge, and 30 minutes of rest. Among them, the NCA type battery has an operating temperature of 25℃, a charging rate of 0.25C, a discharge rate of 1C, an upper and lower limit voltage of 4.2V and 2.65V, and a constant voltage charging current of 0.05C. The NCM type battery has an operating temperature of 35℃, a charging rate of 0.5C, a discharge rate of 1C, an upper and lower limit voltage of 4.2V and 2.5V, and a constant voltage charging current of 0.05C.

[0025] The capacity to be estimated and the capacity label used in the present invention both adopt the maximum capacity of the entire discharge process during the lithium ion battery cycle.

[0026] The original charge and discharge data are processed, including the removal of abnormal values ​​and the extraction of cycle data. The capacity decay curves of the three batteries are as follows: Figure 3 As shown, the three batteries finally retained 340, 348, and 1177 charge and discharge cycle data respectively.

[0027] Step S2: polarization attenuation feature extraction; During the charging process of the battery, the electrode potential will deviate from the equilibrium potential. This phenomenon is called the polarization effect of the battery. According to the cause of polarization, it can be divided into ohmic polarization, activation polarization and concentration polarization. The corresponding internal resistances are ohmic polarization internal resistance, activation polarization internal resistance and concentration polarization internal resistance. Figure 4 R o , R a , R cAs shown. During the constant voltage charging process, the charging current gradually decreases, resulting in a decrease in polarization voltage and a decrease in polarization degree. For the voltage relaxation link, the battery is in an open circuit state, so only activation polarization and concentration polarization occur inside the battery. Due to the existence of polarization capacitance, the relaxation voltage will decay exponentially until it reaches the equilibrium potential.

[0028] In summary, the voltage and current changes in the constant voltage charging and voltage relaxation stages can reflect the degree of battery polarization, so the features extracted based on these two stages are called polarization attenuation features. In addition, these two stages are in the late stage of charging and are very little affected by the random charging starting point of the battery, which can adapt to the actual use scenario of the battery.

[0029] Depend on Figure 5 As shown in Figure (a), as the number of cycles increases, the maximum current shows a decreasing trend, and the current curve gradually becomes smooth. Therefore, the initial current value, average current value, and charging capacity in the last 10 minutes before the end of constant voltage charging are selected as three features, which are denoted as I t ,I m and Q t ; In reality, the charging end time cannot be determined in advance, so the time from constant voltage charging current from 0.15C to charging end and the charging capacity are selected as the fourth and fifth characteristics, denoted as T and Q respectively i By combining the data, it is found that when the constant voltage charging current reaches 0.15C, the time from the end of charging is often more than 10 minutes, so the battery management system only needs to start collecting data when the charging current reaches 0.15C.

[0030] The charge capacity calculation expression in the above characteristics is as follows:

[0031] Where I is the battery charging current, t is the duration, t1 is the start time of charging capacity recording, and t2 is the end time of charging capacity recording.

[0032] from Figure 5 As can be seen from Figure (b), with the increase in the number of cycles, the battery relaxation voltage curve shows a clear downward trend, and with the increase in time, the voltage difference gradually increases, and under a single cycle, the battery voltage decays exponentially, so an exponential function is used to fit the battery relaxation voltage curve, and the fitting coefficient is used as a feature.

[0033]

[0034] In the formula, V represents the fitting voltage, S represents the intercept, a1-a2 is the amplitude of the exponential decay link, and τ1-τ2 is the time constant. The fitting coefficients S, a1, a2, τ1, τ2 are taken as the sixth to tenth features.

[0035] The above ten characteristics are all irrelevant to the charging start time and are more in line with the actual situation.

[0036] Step S3: feature processing; The following formula is used to verify the correlation between the feature and the battery capacity. The Spearman correlation coefficient between the feature and the capacity is shown in Table 1. It can be seen that the τ2 feature correlation is less than 0.7, so the feature is discarded. The a1 and a2 feature correlations are different for battery conditions and materials, but both show negative correlations. Therefore, a linear combination is used to fuse the two features. It can be seen from the table that the correlation coefficients of the fused features are all greater than 0.9, which is a significant improvement compared to before fusion. According to the correlation analysis results, the constant voltage link I was initially screened and determined. t , I m , Q t , T, Q i And eight characteristics of relaxation voltage fitting coefficient a1+a2, S, τ1.

[0037] I t is the initial current value in the last 10 minutes before the end of constant voltage charging, I m The current average value of the last 10 minutes before the end of constant voltage charging, Q t The charging capacity in the last 10 minutes before the end of constant voltage charging, the time T from constant voltage charging current from 0.15C to the end of charging, Q i The constant voltage charging current is the charging capacity from 0.15C to the end of charging, a1+a2 is the sum of the fitting coefficients a1 and a2, S is the intercept, and τ1 is the time constant;

[0038]

[0039] The closer the absolute value of the number in this table is to 1, the better the effect is.

[0040] from Figure 6 It can be seen that some of the normalized features have some curve overlap, indicating that there may be redundant information between the features. The PCA method is used to further fuse the features. The main steps of PCA are as follows: 1. Normalize the matrix A composed of 8 features to obtain , the formula is as follows:

[0041] In the formula and are the mean and variance of the i-th column respectively.

[0042] 2. Calculate the feature matrix The covariance matrix D of is as follows:

[0043] Where a is the sample size.

[0044] 3. Calculate the eigenvalue and eigenvector using the following formula:

[0045] In the formula is the eigenvalue, and E is the unit matrix. Arrange the eigenvalues ​​in descending order, that is, , and the corresponding eigenvector is , the matrix composed of these vectors is the indirect feature matrix obtained through principal component analysis.

[0046] 4. Calculate the variance contribution rate C of each principal component i , the formula is as follows:

[0047] Where the numerator is the i-th eigenvalue and the denominator is the sum of the k eigenvalues.

[0048] Table 2 shows the variance contribution of each principal component. It can be seen that for different types of batteries, the variance contribution of each principal component is somewhat different, but the cumulative contribution of the first four principal components is more than 99%, indicating that these four principal components can represent the overall change of the data, so PC1, PC2, PC3, and PC4 are selected as dimensionality reduction features.

[0049]

[0050] PC1, PC2, PC3, and PC4 are fused with I t ,I m , Q t , T and Q i , a1+a2, S and τ1.

[0051] Step S4: Construct hybrid neural network DBN-SE-LSTM (see Figure 7 ): The SE adaptive channel attention module is embedded in each unit RBM of the DBN network. Through the squeeze-excitation function of SE, the weights of each unit in the hidden layer are automatically calculated, giving important neurons a larger contribution ratio, thereby improving the feature extraction ability and high-order information mining ability of DBN. At the same time, the time series prediction ability of the LSTM network is used to learn the relationship between capacity and input, and a hybrid neural network DBN-SE-LSTM is constructed.

[0052] 1. Related basic algorithms: 1.1 The deep belief network (DBN) is composed of a stack of restricted Boltzmann machines (RBMs). The joint energy function E and joint probability distribution P of the visible vector v and the hidden vector h can be expressed as:

[0053] Where m v 、m h Represents the number of visible neurons and hidden neurons; W ij Represents the weight from the i-th visible neuron to the j-th hidden neuron; α j , β i represents the bias of the jth hidden neuron and the i-th visible neuron, and A(W,α,β) is the allocation function.

[0054] When the visible layer (hidden layer) neuron state is given, the hidden layer (visible layer) neurons will be independent of the RBM, so the hidden layer (visible layer) activation probability function can be expressed as:

[0055] In the formula Represents the activation function sigmoid.

[0056] DBN training includes two processes: unsupervised learning and supervised fine-tuning. In the unsupervised learning process, no labels for training data are provided, and parameter updates are derived based on the contrastive divergence (CD) algorithm, which can be described as:

[0057] In the formula represents the RBM training learning rate, represents the expected value of the data distribution, Represents the expected value after model reconstruction.

[0058] After the unsupervised training, labeled data is used for supervised fine-tuning, which uses the back-propagation (BP) algorithm to update the model parameters.

[0059] 1.2 Long Short-Term Memory Network The long short-term memory network (LSTM) effectively solves the problems of gradient vanishing and gradient exploding that may occur in the training process of traditional recurrent neural networks (RNNs) by introducing three gating mechanisms (input gate, forget gate, and output gate). The basic principle formula is as follows:

[0060] In the formula, i t 、f t and t Represents the activation vectors of the input gate, forget gate, and output gate respectively; x tand h t Represent the input and hidden state at the current moment respectively; W and b represent the weight matrix and bias vector that need to be learned in the network respectively; and is the intermediate state and output state of the unit; is the sigmoid activation function; tanh is the hyperbolic tangent function.

[0061] 1.3 Squeeze-motivate attention The squeeze-excitation (SE) channel attention consists of three parts: squeeze, excitation, and scaling. The principle is as follows: Squeeze block: Use the global average pooling layer to compress the channel global information into the channel descriptor, the expression is as follows:

[0062] Where c represents the cth channel, z is the channel descriptor, u represents the feature map, and u(i) represents the feature value of the i-th position in the channel.

[0063] Excitation block: Use the sigmoid activation function to calculate the weight of each channel. The expression is as follows:

[0064] Where W1 and W2 represent linear layers, is the ReLU activation function.

[0065] Scaling block: Apply the learned channel attention weight s to the original channel, expressed as follows:

[0066] In the formula Represents the Hadamard operation.

[0067] In the present invention, an LSTM network suitable for processing time series problems is used to replace the linear output layer of DBN to enhance the network's learning ability for the nonlinear attenuation process of lithium-ion battery capacity. The SE adaptive channel attention module is integrated into the RBM layer of the DBN network, so that the network can dynamically adjust the weights of each neuron in each hidden layer, ensure that important information is effectively transmitted between layers, and enhance the high-order feature information mining ability of the DBN network.

[0068] Step S5: Migration strategy and model training: Considering that the battery does not perform a complete charge and discharge process every time in actual application, it is impossible to obtain the current capacity parameters of the battery, but when the battery is fully charged, the features selected in this article can be obtained, and this situation is very common, which means that there is a large amount of unlabeled data in the target domain. As for labeled data, it can be obtained through regular battery capacity correction technology. Of course, this state is rare, which means that there is a small amount of unlabeled data in the target domain. Under such small sample conditions, traditional training methods will not be able to accurately estimate the state of the target domain. The present invention implements a migration strategy combining semi-supervised learning and reconstruction fine-tuning on the basis of the above-mentioned DBN-SE-LSTM to improve the battery capacity estimation accuracy under small samples. The present invention proposes a combination of semi-supervised learning and reconstruction fine-tuning to reduce the dependence of labeled samples in the target domain. The specific implementation steps are as follows: (1) Battery data partitioning: All battery data (a large amount of corresponding PC1, PC2, PC3, PC4 and capacity labels) are divided into source domain and target domain. The source domain data are all labeled with capacity, while the target domain contains only a small amount of data with capacity labels and a large amount of data without capacity labels.

[0069] (2) Unlabeled data pre-training and domain adaptation: The DBN-SE-LSTM network is pre-trained using a large amount of data without capacity labels in the target domain and the source domain. At the same time, the maximum mean difference (MMD) method is used to narrow the feature distribution difference between the source domain and the target domain, and obtain a pre-training model based on feature sharing. This enables the pre-training model to learn the high-level abstract information shared by the source domain and the target domain, thereby improving the generalization ability of the model and enhancing the migration ability.

[0070] (3) Supervised training of source domain data: Based on the feature-sharing pre-training model, a large amount of labeled data from the source domain is used to conduct supervised training of the entire model network to build the transfer learning basic model TL1. In this process, the back-propagation mechanism is used to update the parameters of each hidden layer to capture the mapping relationship between the battery capacity and the deep feature information learned by DBN-SE-LSTM.

[0071] (4) Training of the adaptation layer of the TL1 model: Freeze all hidden layer parameters of the trained basic model TL1 except the adaptation layer; train the adaptation layer using a small amount of labeled data from the target domain to quickly capture the mapping relationship between the target domain features and the battery capacity, and obtain the migration model TL2.

[0072] (5) Fine-tuning the entire network: Based on the TL2 model, all hidden layers are unfrozen, and the parameters of the entire network are fine-tuned using an extremely small learning rate (the learning rate in this embodiment is 0.00001) and a small number of training times (the number of training times in this embodiment is 50 times) to obtain the final deep migration model TL-DBN-SE-LSTM, that is, the capacity estimation model with optimal weights and bias terms.

[0073] Step S6: Use the prediction set data and the trained capacity estimation model to estimate the capacity.

[0074] Relative error (RE), mean absolute percentage error (MAPE), root mean square percentage error (RMSPE) and determination coefficient R2 were used as evaluation indicators to verify the accuracy and effectiveness of the model and method.

[0075]

[0076] The training data of experiment A is NCA-CY25-0.25_1-#4 battery, and the prediction data is NCA-CY25-0.25_1-#5. The models used are DBN-SE-LSTM, DBN-LSTM and DBN. The prediction performance of the hybrid neural network is verified. The prediction results and errors are shown in Figure 8 The training data of experiment B are NCA-CY25-0.25_1-#4, NCA-CY25-0.25_1-#5, and 20% of NCM-CY35-05_1-#2 batteries, and the prediction data are the remaining 80% of NCM-CY35-05_1-#2. The models used are TL-DBN-SE-LSTM and DBN-SE-LSTM to verify the effectiveness of the migration strategy. The prediction results and errors are shown in Fig. 9 (a) Fig. 9 (b) The evaluation indicators of experiments A and B are shown in Table 3. The parameters of the model in the experiment are shown in Table 4.

[0077]

[0078] The larger the MAPE(%) and RMSPE(%), the lower the estimation accuracy of capacity. 2 The closer it is to 1, the higher the estimation accuracy. In Experiment B, for different types of batteries, the accuracy of the deep transfer model after transfer learning is higher.

[0079]

[0080] From Table 3 and Figure 8 , Fig. 9It can be seen that the DBN-SE-LSTM network of the present invention has the best estimation effect, and by integrating the deep neural network, the estimation performance of the model can be improved by nearly 40%. The migration strategy of the present invention can accurately estimate the battery capacity of the target domain under the sample, with an estimation error within 0.6%, which significantly improves the estimation effect compared with the original model without migration capability.

[0081] The above results of the present invention are all based on the perspective of battery cells, but in actual applications, batteries are deployed in the form of modules, and in order to reduce the difficulty of battery PACK and the data burden of the battery management system, the voltage and current data of the single cell will not be available, but the total current of the module and the voltage of the series-connected single cell will be directly obtained through the voltage and current sensor. Therefore, in the battery capacity estimation at the module level, the depolarization characteristic data of the constant voltage charging current and the relaxation relaxation voltage are the total current in the constant voltage charging stage and the relaxation voltage of the worst-performing series-connected single cell. In addition, in view of the similarity between single cell and module battery data, based on the hybrid neural network and migration strategy proposed in the present invention, and using a large amount of single cell battery data and a small amount of module battery data, it is possible to achieve accurate estimation of the capacity of lithium-ion battery packs under small sample conditions.

[0082] The method of the present invention can provide capacity-level parameters for the overall monitoring and safety management of lithium-ion batteries, thereby ensuring the safety and reliability of lithium-ion batteries in practical applications.

[0083] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A method for estimating the capacity of a lithium-ion battery cell / module based on a deep migration model, characterized in that: The estimation method comprises the following steps: Step S1: Collecting charge and discharge cycle data during the lithium-ion battery aging test process, and taking the maximum capacity of the entire discharge process during the lithium-ion battery cycle as the capacity label; Step S2: Feature selection 1) Based on the relationship between the intensity of battery polarization reaction and battery capacity, the depolarization characteristic data of constant voltage charging current and relaxation voltage are used to extract multi-dimensional practical features that are not affected by the charging starting point, including: Select the initial current value, average current value and charging capacity in the last 10 minutes before the end of constant voltage charging, and record them as I t ,I m , Q t ; Select the time and charge capacity from the constant voltage charge current of 0.15C to the end of charge, and record them as T and Q respectively. i ; The exponential function is used to fit the battery relaxation voltage curve, and the fitting curve is expressed as: In the formula, V represents the fitting voltage, S represents the intercept, a1-a2 is the amplitude of the exponential decay link, τ1-τ2 is the time constant, and t is the time; The fitting coefficients S, a1, a2, τ1, τ2 are used as features; 2) Calculate the Spearman correlation coefficient between the selected multi-dimensional features and the capacity, discard the features with a correlation coefficient less than 0.7, and linearly combine the two features a1 and a2 as new fusion features, retain the features with a correlation coefficient greater than 0.8, and preliminarily screen and determine the constant pressure link I t ,I m , Q t , T and Q i and relaxation voltage fitting coefficients a1+a2, S and τ1 for subsequent processing; 3) All the preliminarily screened features are combined into a matrix A, and the preliminarily screened features are fused using PCA technology, with the first four principal components PC1, PC2, PC3, and PC4 used as dimensionality reduction features; The principal components PC1, PC2, PC3, and PC4 are used as the input of the hybrid neural network DBN-SE-LSTM; Step S3: Construct a hybrid neural network DBN-SE-LSTM: add an SE adaptive channel attention module to each RBM layer of the DBN network, use an LSTM network in the linear output layer of the DBN network, use the high-order information learned by the DBN network on the input data as the input of the LSTM network, and connect the output of the LSTM network to the fully connected layer to obtain the estimated value of the capacity; Step S4: Transfer learning: (1) The PC1, PC2, PC3, PC4 and capacity labels after feature selection are divided into the source domain and the target domain. The source domain data all have capacity labels, while the target domain has only a small amount of capacity-labeled data and a large amount of uncapacity-labeled data. (2) Unlabeled data pre-training and domain adaptation: The hybrid neural network DBN-SE-LSTM is pre-trained using data without capacity labels in the target and source domains. The maximum mean difference (MMD) method is used to reduce the feature distribution difference between the source and target domains, and obtain a feature-sharing pre-training model, which enables the pre-training model to learn high-level abstract information shared by the source and target domains. (3) Supervised training of source domain data: Use labeled data from the source domain to conduct supervised training of the entire network and build the transfer learning basic model TL1. In this process, the back propagation mechanism is used to update the parameters of each hidden layer to capture the mapping relationship between the battery capacity and the deep feature information learned by the hybrid neural network DBN-SE-LSTM. (4) Training the adaptation layer of the TL1 model: Freeze all hidden layer parameters of the trained basic model TL1 except the adaptation layer; train the adaptation layer using labeled data from the target domain to quickly capture the mapping relationship between the target domain features and the battery capacity, and obtain the migration model TL2; (5) Fine-tuning the entire network: Based on the TL2 model, unfreeze all hidden layers and fine-tune the parameters of the entire network using a learning rate of 0.00001 and a training number of 40-60 times to obtain the trained deep migration model TL-DBN-SE-LSTM; Step S5: Input the features of the lithium-ion battery cell / module to be estimated into the trained deep migration model TL-DBN-SE-LSTM after feature selection for online estimation of capacity.

2. The method according to claim 1, characterized in that The number of RBM layers is 4, and the number of RBM hidden layer units is [4, 130, 80, 40, 30]; the number of LSTM network layers is 3, and the number of units is [30, 40, 20]; the number of fully connected layers is 3, and the number of units is [15, 10, 1].

3. The method according to claim 1, characterized in that In the battery capacity estimation at the module level, the depolarization characteristic data of the constant voltage charging current and the relaxation voltage in step S2 are the total current in the constant voltage charging stage and the relaxation voltage of the worst performing series cell.

Images