Lithium battery state of health estimation method based on hybrid neural network and transfer learning

Abstract

The present application relates to a kind of lithium battery health state estimation method based on hybrid neural network and transfer learning, belong to battery storage technology field field.The method is: extracting battery cycle data and constitutes characteristic curve, carries out data preprocessing to it, and carries out outlier screening and replacement, constitutes source domain data set;Extract the characteristic curve data of lithium battery data of different types or different working conditions, and constitute target domain data set;Pre-training model is obtained using source domain data set to CNN-GRU hybrid neural network;The pre-training model is used as base learner estimation model, and is trained in target domain data set using Tradaboost.R2 algorithm, to update sample weight and base learner weight;Lithium battery SOH estimation value output, while calculating corresponding evaluation index value.The present application can improve the accuracy of lithium battery health state estimation.

Description

Technical Field

[0001] This invention belongs to the field of battery storage technology and relates to the estimation of the state of health of lithium batteries, specifically to a method for estimating the state of health of lithium batteries based on hybrid neural networks and transfer learning. Background Technology

[0002] A Battery Management System (BMS) is a crucial link between the onboard lithium-ion battery and the electric vehicle. It integrates the monitoring and management of lithium-ion batteries or battery packs, ensuring their safety and reliability and maximizing power output. As a vital component of the BMS, the lithium-ion battery health assessment can predict the current discharge capacity and remaining cycle life of the battery. This not only effectively guarantees battery safety but also improves the driving range of the electric vehicle, significantly contributing to its safe and reliable operation.

[0003] However, during use, lithium batteries gradually degrade in performance and capacity due to factors such as ambient temperature and increased charge-discharge cycles, leading to a reduction in the driving range of electric vehicles. Furthermore, lithium battery performance degradation is usually accompanied by increased internal resistance and significantly increased heat generation, accelerating internal side reactions and further speeding up degradation. This results in a rapid decline in battery life and may even trigger thermal runaway, causing safety incidents. Therefore, it is necessary to monitor the current health status of lithium batteries and assess their current degree of degradation. Estimating the battery health status allows the battery management system to take timely and effective management and control measures, thereby improving the overall safety and reliability of the battery and ensuring its continued operation in good condition for a longer period.

[0004] With the rise of artificial intelligence, lithium battery health assessment methods based on deep neural networks have become increasingly crucial. Currently, data-driven SOH estimation mainly relies on indirect health factors and the number of training samples. However, the correlation between indirect health factors and SOH on different batteries can change due to battery inconsistency and degradation, thus affecting the accuracy of lithium battery SOH estimation. In addition, for small sample scenarios, i.e., when the number of current battery training samples is small, the prediction accuracy of models built using data-driven methods is low. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a lithium battery health state estimation method based on hybrid neural networks and transfer learning, which solves the defects of manually extracting indirect health factors and improves the SOH estimation accuracy under small sample conditions.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for estimating the health status of lithium batteries based on hybrid neural networks and transfer learning includes the following steps:

[0008] S1. Extract battery cycle data and construct feature curves, perform data preprocessing, and filter and replace outliers to form the source domain dataset.

[0009] Regarding step S1, this invention mainly analyzes the impact of external factors on battery capacity, including:

[0010] a. Discharge rate: Under the same conditions, the larger the discharge current, the faster the battery capacity decreases. The discharge current is one of the factors affecting the degradation of lithium batteries.

[0011] b. Cycle Count: As the number of charge-discharge cycles of lithium batteries increases, their capacity generally tends to degrade. However, there are instances of sudden capacity changes during this degradation process. This is because there is a resting phase in the experiment, and the battery degradation trajectory shows capacity recovery at certain monitoring points, leading to sudden capacity changes. As the number of charge-discharge cycles increases, the battery capacity will gradually decrease.

[0012] c. Battery capacity degrades rapidly in low-temperature environments, resulting in faster battery degradation.

[0013] For step S1, several common health factors during the lithium battery charging stage are extracted, including:

[0014] 1. Time difference ΔT between equal voltage intervals during constant current charging phase: in, This is expressed as the voltage reaching V. low Time, Indicates that the voltage reaches V high Time;

[0015] 2. Voltage charging integral A1 during constant current charging stage; Current charging integral A2 during constant voltage charging stage: Where t0 represents the constant current charging start time. t represents the time it takes for the voltage to reach 4.2V. end It indicates the time it takes for the current to reach the cutoff current; V represents the voltage at time t, and I represents the current at time t.

[0016] 3. Constant current charging stage: constant current charging time T1; constant voltage charging stage: constant voltage charging time T2: T1 = T cc T2 = T cv Among them, T cc T represents the constant current charging time. cv Indicates the constant voltage charging time;

[0017] 4. Temperature integral TA1 during constant current charging stage; Temperature integral TA2 during constant voltage charging stage: Where T represents the temperature value at time t;

[0018] 5. The maximum slope K of the voltage curve during the constant current charging stage cc The maximum slope K of the current curve during the constant current charging stage cv :K cc =max(k vcc ), K cv =max(k icv ), where k vcc k represents the slope of the voltage transformation during the constant current charging phase. icv This represents the slope of the current transformation during the constant current charging phase.

[0019] Based on the above, the main part of step S1 is as follows:

[0020] (1) Feature curve selection: Select the feature curve of lithium battery in charging state as data input. Among them, the voltage range curve of lithium battery constant current charging stage is used as current feature curve; the current range curve of lithium battery constant voltage charging stage is used as voltage feature curve; the curve from the voltage start time of constant current charging stage to the current drop time of constant voltage charging stage is used as temperature feature curve.

[0021] (2) Noise Reduction: Due to interference from equipment or the external environment, transient failures, etc., sensors may introduce errors, and some data may show fluctuating values, thereby reducing accuracy and leading to incorrect estimation results, resulting in inaccurate data detection and recording, or data loss. Therefore, it is necessary to denoise the data. This invention uses the moving average formula method for noise reduction.

[0022] (3) Normalization: The feature curves after denoising are normalized by the min-max method.

[0023] (4) Resampling: The extracted feature curves are fitted using a cubic polynomial. Each feature curve after fitting is resampled at the same sampling interval to obtain the target features. The target features constitute the source domain dataset.

[0024] S2. Extract characteristic curve data from lithium battery data of different types or under different operating conditions to form the target domain dataset.

[0025] S3. Use the source domain dataset to pre-train the CNN-GRU hybrid neural network to obtain a pre-trained model.

[0026] In step S3, the pre-training steps include:

[0027] 1. Utilize CNN to automatically extract degradation information from battery cycling data. Process the nonlinear relationship between input data and SOH through convolution, pooling, and activation operations to extract hidden feature information in the cycling data;

[0028] 2. Employ multi-layer CNN for feature extraction to improve the correlation between extracted features and SOH;

[0029] 3. Use GRU to extract features from the feature extraction model and build a model. Extract hidden temporal features from the degenerate features and establish the mapping relationship between features and SOH to obtain the SOH estimation model.

[0030] Furthermore, the CNN-based feature extraction model is constructed as follows: the features reconstructed through a sliding window are passed through the input layer, then through a convolutional layer (Conv1), followed by pooling and activation functions for shallow feature extraction. Then, through a convolutional layer (Conv2), further pooling and activation functions are used for deep feature extraction. The CNN feature extraction model processes the input data through n convolution and pooling operations, enabling deep feature extraction of the current cycle's features. Finally, the Flatten layer flattens the extracted deep features into a one-dimensional vector for output.

[0031] The CNN-GRU hybrid neural network is constructed by sequentially connecting a CNN convolutional module and a GRU convolutional module. The CNN convolutional module consists of two consecutive stacked 4-layer convolutional layers and a max-pooling layer, followed by a fully connected layer. The GRU convolutional module consists of a GRU layer and a fully connected layer, where the fully connected layer is used for estimating the state of harmonics (SOH) of the lithium battery.

[0032] Furthermore, the CNN-GRU hybrid neural network adopts a combined model for temporal input, namely, different weight combinations of single loop and sliding window, which combines the advantages of both to achieve a more accurate SOH estimation of the capacity regeneration point and effectively learn temporal features.

[0033] The SOH estimate of the combined model is calculated as follows:

[0034]

[0035] In the formula, f1(·) represents a single-cycle CNN-GRU model, and f2(·) represents a sliding-window CNN-GRU model; w it This represents the weight of the i-th model at time t;

[0036]

[0037] In the formula, e ij This represents the prediction error value at time point j:

[0038]

[0039] in, Let y represent the predicted value of the i-th model at time point j. j This represents the true SOH value at time point j. The model weights at time point t should also satisfy the following condition:

[0040] w it ≥0

[0041] And w 1t +w 2t =1

[0042] When using a CNN-GRU hybrid neural network to estimate the state of harmonics (SOH) of lithium-ion batteries, the mean squared error is used as the loss function to evaluate the prediction error. The loss function is as follows:

[0043]

[0044] In the formula, N represents the number of lithium battery data cycles, and SOH i Let SOH represent the predicted SOH value for the i-th data sample. i ture This represents the true SOH value corresponding to the i-th data sample.

[0045] S4. Use the pre-trained model as the base learner to estimate the model, and train it on the target domain dataset using the Tradaboost.R2 algorithm to update the sample weights and base learner weights.

[0046] For the Tradaboost.R2 algorithm, the SOH estimate is a weighted output of the base learner predictions. Its main calculation formula is as follows:

[0047]

[0048] Where, f(x,θ) t ) represents the base learner, which in this invention is the CNN-GRU model, θ t w represents the optimal parameters of the base learner. i This represents the weights of the base learners in a strong learner composed of base learners.

[0049] In step S4, the pre-trained model is transferred and its parameters are tuned using the constructed target domain dataset, including the following steps:

[0050] 1. Setting up the base learner:

[0051] First, a CNN-GRU hybrid neural network is used to pre-train the source domain dataset to learn the parameters of each part, and then the parameters are transferred to the parameters of the new base learner network.

[0052] Then, when retraining using the target domain dataset, it is necessary to freeze the weights of the shallow convolutional network and the GRU layer network, and then train the deep network.

[0053] A backpropagation suppression algorithm is introduced for specific network layers. The specific formula for the backpropagation suppression algorithm is as follows:

[0054]

[0055] In the formula, Let represent the gradient value calculated in the previous step, λ represent the suppression parameter with a value range of (0,1), l represent the learning rate, b represent the bias parameter, and L represent the learning rate. g Let L represent the g-th loss function value. t Indicates a specific network layer.

[0056] 2. Initial weight settings for Tradaboost.R2:

[0057] Using the KMM algorithm to determine the initial weights of the source and target domains is equivalent to obtaining a relatively accurate initial weight before the Tradaboost.R2 algorithm starts iterating, thereby reducing the risk of model overfitting.

[0058] The KMM algorithm is shown below:

[0059]

[0060] In the formula, β i This represents the weighting factor for the source domain sample data. This represents a sample of source domain data. Let H represent the target domain data sample, φ(·) be the mapping function from the original space to RKHS, and H represent the regenerated Hilbert space-RKHS space with feature kernel k.

[0061] Derivation of KMM:

[0062]

[0063] In the formula, c represents a constant, and It can be simplified:

[0064]

[0065]

[0066] in, By simplifying the above equation, the final objective function takes the form of:

[0067]

[0068] 3. Model framework construction:

[0069] Step 1: Determine the source domain dataset Ds = {Xs, Ys}, where (x i ,y i Given that )∈Ds and i=1,2,3,…,n; and the target domain dataset Dt={Xt,Yt} is determined, then (x j ,y j )∈Dt and j=1,2,3,…,m; merge the source domain dataset and the target domain dataset T=Ds∪Dt;

[0070] Step 2: Set the initial weights for dataset T and the maximum number of iterations N, i.e., the number of base learners; where the weight coefficients for samples in the source domain dataset are... The sample weight coefficients of the target domain dataset are

[0071] Step 3: Obtain the source domain weights = KMM(X) S ,X T ), where X S X represents the source domain data sample. T This represents a sample of data from the target domain.

[0072] Step 4: Initialize the weight vector:

[0073]

[0074] Step 5: Update source domain sample weight parameters:

[0075]

[0076] Step 6: Normalize the weight vector:

[0077]

[0078] Step 7: Obtain the base learner h through model fine-tuning. t The dataset T and the weight distribution p on T will be merged. t By training with an existing base learner model, we obtain the regressor: f t :X→Y; X represents the source domain space, and Y represents the target domain space;

[0079] Step 8: Calculate f t In T target Error rate on:

[0080]

[0081] The overall regression error of Dt is:

[0082]

[0083] Step 9: Set the target domain weight update parameters:

[0084] β t =ε t / (1-ε t )

[0085] Step 10: Update the weight vectors of the training samples in the source and target domains:

[0086]

[0087] Step 11: Determine if the number of iterations is less than N. If it is less, return to step 6; if it is greater, proceed to step 12.

[0088] Step 12: Obtain the lithium battery SOH estimation model by weighting the last N / 2 base learners:

[0089]

[0090] In the formula, θ t Let α represent the optimal parameters of the base learner. i This represents the learning rate of the base learner.

[0091] S5 outputs the estimated SOH value of the lithium battery and calculates the corresponding evaluation index value.

[0092] In step S5, three regression evaluation metrics are used: Mean Square Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE), and their calculation formulas are as follows:

[0093]

[0094]

[0095]

[0096] The beneficial effects of this invention are as follows: This invention adopts a hybrid neural network model of CNN-GRU, and at the same time, it can improve the correlation between indirect health factors and SOH through transfer learning; in addition, even when the number of battery training samples is small, the accuracy of the constructed prediction model can be improved by collecting different working conditions or different types of data.

[0097] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0098] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0099] Figure 1 A diagram illustrating the overall evaluation process for lithium batteries;

[0100] Figure 2 This is a schematic diagram illustrating the range selection and resampling of input battery data;

[0101] Figure 3 For CNN feature extraction model;

[0102] Figure 4 This is a schematic diagram of a CNN-GRU-based SOH estimation model;

[0103] Figure 5 The estimation results and weight distribution of SOH for different battery combination models;

[0104] Figure 6 These are the initial weights of the source domain during the transfer learning process. Figure 6 (a) represents the initial weights of the source domain in Experiment 1. Figure 6 (b) represents the initial weights of the source domain in Experiment 2;

[0105] Figure 7 A schematic diagram of the RMSE distribution during the training process of the base learner. Figure 7 (a) shows the RMSE distribution during the training process of the base learner in Experiment 1. Figure 7 (b) shows the RMSE distribution during the training process of the base learner in Experiment 2;

[0106] Figure 8 The SOH estimation results for experiments 1–4 are as follows. Figure 8 (a) shows the estimated SOH results for Experiment 1. Figure 8 (b) shows the estimated SOH results for Experiment 2. Figure 8 (c) shows the estimated SOH results for Experiment 3. Figure 8 (d) shows the estimated SOH results for Experiment 4. Detailed Implementation

[0107] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0108] Please see Figures 1 to 8 This describes a method for estimating the health status of lithium batteries based on hybrid neural networks and transfer learning. The method includes the following steps:

[0109] S1. Extract battery cycle data and construct feature curves, perform data preprocessing, and filter and replace outliers to form the source domain dataset.

[0110] S2. Extract characteristic curve data from lithium battery data of different types or under different operating conditions to form the target domain dataset;

[0111] S3. Use the source domain dataset to pre-train the CNN-GRU hybrid neural network to obtain a pre-trained model;

[0112] S4. Using the pre-trained model as a base learner, the Tradaboost.R2 algorithm is used to perform transfer learning to build a transfer learning-based SOH estimation model. At the same time, the target domain dataset is used for training, thereby updating the sample weights and base learner weights.

[0113] S5 outputs the estimated SOH value of the lithium battery and calculates the corresponding evaluation index value.

[0114] This embodiment uses a publicly available lithium battery dataset from the National Aeronautics and Space Administration (NASA) as the raw data for a study on the health status assessment of lithium batteries. This dataset uses commercially available lithium batteries (lithium cobalt oxide positive electrode, graphite negative electrode, rated capacity 2Ah) cycled at a range of ambient temperatures (4℃, 24℃, 43℃), charged using a common constant-current-constant voltage (CC-CV) protocol and different discharge schemes. The dataset includes intra-cycle measurements of terminal current, voltage, and battery temperature, as well as inter-cycle measurements of discharge capacity and EIS impedance readings. The dataset is provided in MAT file format and contains a `cycle` structure, which records data for three experimental types: charging, discharging, and impedance. It also includes room temperature, experimental time, and a `data` structure. The `data` structure records relevant data for the current experimental type, mainly including charge / discharge voltage, charge / discharge current, battery temperature, and test time. The failure threshold for this experimental lifespan is defined as a decrease in capacity to approximately 70% of the rated capacity, indicating the end of the battery's lifespan. In this experiment, the battery failure threshold was defined as a decrease in capacity to approximately 70% of the rated capacity, indicating that the lithium battery life had reached its end. Batteries B0005, B0006, and B0034 from the NASA dataset were used to assess battery health, while batteries CX2-33, CX2-38, CS2-33, and CS2-38 from the CALCE battery dataset were selected as transfer learning samples for evaluating the model.

[0115] The data preprocessing in step S1 includes four steps. The first step is to obtain current characteristic curve data, voltage characteristic curve data, and temperature characteristic curve data from the NASA lithium battery dataset.

[0116] The second step is to denoise the battery dataset using the moving average method. The core idea is to take the arithmetic mean of the observations in a certain n-region surrounding the current observation, and then use this arithmetic mean to replace the current observation. Here, n is the size of the moving window, which is set to 3.

[0117]

[0118] The third step is data standardization, which corrects data of different scales and specifications according to a unified standard, thereby eliminating the differences in results caused by different units of measurement. The specific formula is:

[0119]

[0120] By readjusting the values ​​of each dimension of the data, the final data vector falls within the [0,1] interval.

[0121] The fourth step is to fit the extracted feature curves, using a cubic polynomial as the fit.

[0122] y = at 3 +bt 2 +ct+d

[0123] In step S2, a target domain dataset is constructed in the CALCE battery dataset, which will be used later to fine-tune the pre-trained battery health status assessment model.

[0124] Step S3, constructing the CNN-GRU pre-trained hybrid neural network model mainly includes the following steps:

[0125] 1. Feature extraction design based on CNN

[0126] The input battery data is range-selected and resampled, such as... Figure 2 As shown in the figure. The battery time-series features are constructed by combining a weighted approach with a single-loop method and a sliding window method on the input data, as shown in the following equation:

[0127]

[0128] The weights are typically determined based on the inverse of the mean squared prediction error of a single model at the most recently observed time point.

[0129]

[0130] The processed feature data is input into a multi-layer CNN for feature extraction. The feature extraction model is as follows: Figure 3 As shown.

[0131] The features extracted from the CNN model after splitting are recombined into a sequence and output to the GRU model. Finally, the GRU model is used to extract the temporal feature information present in the cyclic features to complete the SOH estimation of the lithium battery. This realizes an end-to-end CNN-GRU SOH estimation structure model, such as... Figure 4 As shown.

[0132] Let g(·) denote the feature extraction model of the CNN, and let the sequence input of the model be [t1, t2, ..., tm], where:

[0133]

[0134] In the formula, m represents the size of the sliding window, and ti represents the data at the i-th time point of the window. The reconstructed sequence output is denoted as [t1′,t2′,…,tm′], and for ti′, ti′=g(ti), i.e., the sequence output is [g(t1),g(t2),…,g(tm)]. Let GRU be denoted as f(·), then the final SOH estimate is:

[0135] SOH=f(g(t1),g(t2),…,g(tm))

[0136] Finally, the model was applied to the test set and the corresponding evaluation indicators were obtained. The SOH estimation results and weight distribution of batteries B0005, B0006, and B0007 are as follows: Figure 5 As shown.

[0137] In step S4, for the case of small samples, the constructed target domain dataset is used to perform transfer learning and parameter tuning on the pre-trained model, as described below:

[0138] Step 1: Determine the source domain dataset Ds = {Xs, Ys}, where (x i ,y i Given that )∈Ds and i=1,2,3,…,n; and the target domain dataset Dt={Xt,Yt} is determined, then (x j ,y j )∈Dt and j=1,2,3,…,m; merge the source domain dataset and the target domain dataset T=Ds∪Dt;

[0139] Eight batteries were selected from the NASA lithium battery dataset and the CALCE dataset, and four sets of experiments were designed, as shown in Table 1. Let K be the denoted K. train (D) represents the first 20% of the data in dataset D, K test (D) represents the last 80% of the data in dataset D, Ds represents the source domain, Dt represents the target domain, and Te represents the test set.

[0140] Table 1 4 groups of experiments

[0141]

[0142] Transfer learning between similar batteries was analyzed using Experiments 1 and 2. In Experiment 1, B0034 and B0005 were charged under the same conditions and at the same ambient temperature. B0005 had a discharge current of 2A and a discharge cutoff voltage of 2.7V, while B0034 had a discharge current of 4A and a discharge cutoff voltage of 2.2V. In Experiment 2, CS2-33 and CS2-38 were charged under the same conditions, at the same nominal capacity, and at the same ambient temperature. CS2-33 had a discharge current of 0.55A, while CS2-38 had a discharge current of 1.1A.

[0143] Experiments 3 and 4 were used to analyze the transfer learning between dissimilar batteries. In Experiments 3 and 4, dissimilar batteries (NASA and CALCE batteries) were selected to perform knowledge transfer between each other to verify the effectiveness of the algorithm. Batteries B0006 and B0007 have a nominal capacity of 2Ah, a discharge current of 2A, and discharge cutoff voltages of 2.5V and 2.2V, respectively. Batteries CX2-33 and CX2-38 have a nominal capacity of 1350mAh, a discharge current of 0.67A, and a discharge cutoff voltage of 2.7V for both.

[0144] Experiment 1 uses battery B0034 as the source domain data, the first 20% of the data from battery B0005 as the target domain data, and the remaining data as test data. Experiment 2 uses battery CS2-38 as the source domain data, the first 20% of the data from battery CS2-33 as the target domain data, and the remaining data as test data. Experiment 3 uses battery B0007 as the source domain data, the first 20% of the data from battery CX2-38 as the target domain data, and the remaining data as test data. Experiment 4 uses battery CS2-33 as the source domain data, the first 20% of the data from battery B0006 as the target domain data, and the remaining data as test data.

[0145] Step 2: Initialize parameters. Set the initial weights for dataset T and the maximum number of iterations N, i.e., the number of base learners; where the weight coefficients for samples in the source domain dataset are... The sample weight coefficients of the target domain dataset are

[0146] Step 3: Obtain the source domain weights β = KMM(X) S ,X T ), where X S X represents the source domain data sample. T This represents a sample of data from the target domain.

[0147] Step 4: Initialize the weight vector:

[0148]

[0149] Step 5: Update source domain sample weight parameters:

[0150]

[0151] Step 6: Normalize the weight vector:

[0152]

[0153] For Experiments 1 and 2, the initial weights of the source domain are as follows: Figure 6 As shown.

[0154] Step 7: Obtain the base learner h through model fine-tuning.t The datasets T and the weight distribution p on T will be merged. t By training with an existing base learner model, we obtain the regressor: f t :X→Y; X represents the source domain space, and Y represents the target domain space;

[0155] Step 8: Calculate f t In T target Error rate on:

[0156]

[0157] The overall regression error of Dt is:

[0158]

[0159] Step 9: Set the target domain weight update parameters:

[0160] β t =ε t / (1-ε t )

[0161] Step 10: Update the weight vectors of the training samples in the source and target domains:

[0162]

[0163] Step 11: Determine if the number of iterations is less than N. If it is less, return to step 6; if it is greater, proceed to step 12.

[0164] For Experiments 1 and 2, the RMSE of the base learner training process is as follows: Figure 7 As shown, Figure 7 As can be seen, during the training of the base learner, the RMSE distributions of Experiment 1 and Experiment 2 remained at relatively small values ​​and were relatively stable.

[0165] Step 12: Obtain the lithium battery SOH estimation model by weighting the last N / 2 base learners:

[0166]

[0167] In the formula, θ t Let α represent the optimal parameters of the base learner. i This represents the learning rate of the base learner.

[0168] The final experimental results are as follows Figure 8 And as shown in Table 2.

[0169] Table 2 Experimental Results

[0170]

[0171]

[0172]

[0173] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for estimating the state of health of lithium batteries based on hybrid neural networks and transfer learning, characterized in that: The method includes the following steps: S1. Extract battery cycle data and construct feature curves, perform data preprocessing, and filter and replace outliers to form the source domain dataset. S2. Extract characteristic curve data from lithium battery data of different types or under different operating conditions to form the target domain dataset; S3. Use the source domain dataset to pre-train the CNN-GRU hybrid neural network to obtain a pre-trained model; S4. Use the pre-trained model as the base learner to estimate the model, and train it on the target domain dataset using the Tradaboost.R2 algorithm to update the sample weights and base learner weights. S5. Output the estimated SOH value of the lithium battery, and calculate the corresponding evaluation index value at the same time; The CNN-GRU hybrid neural network adopts a combined model for temporal input, namely, different weight combinations of single loop and sliding window, which combines the advantages of both to achieve a more accurate SOH estimation of the capacity regeneration point and effectively learns temporal features. The SOH estimate of the combined model is calculated as follows: In the formula, This represents a single-cycle CNN-GRU model. This represents a sliding window CNN-GRU model; Indicates time t Upper i The weights of each model; In the formula, Indicates a point in time j Prediction error value: in, Indicates the first i The model at time point j The predicted value at that location, Indicates the first j The actual SOH values ​​at each time point.

2. The lithium battery health status estimation method according to claim 1, characterized in that: Step S1 is as follows: The voltage range curve during the constant current charging stage of the lithium battery is obtained as the current characteristic curve; the current range curve during the constant voltage charging stage of the lithium battery is obtained as the voltage characteristic curve; the curve from the voltage start time during the constant current charging stage to the current decrease time during the constant voltage charging stage is obtained as the temperature characteristic curve. The extracted feature curves are denoised using the moving average formula method. The feature curves after denoising are normalized using the min-max method; The extracted feature curves are fitted using a cubic polynomial. Each feature curve after fitting is then resampled at the same sampling interval to obtain the target features, which constitute the source domain dataset.

3. The lithium battery health status estimation method according to claim 1, characterized in that: In step S2, a target dataset is constructed on a battery dataset of the same or similar type, which is used during model fine-tuning.

4. The lithium battery health status estimation method according to claim 1, characterized in that: In step S3, the CNN-GRU hybrid neural network includes a CNN convolutional module and a GRU convolutional module connected in sequence; the CNN convolutional module includes two consecutive 4-layer convolutional layers and a max pooling layer stacked together, and finally a fully connected layer stacked together; the GRU convolutional module includes a GRU layer and a fully connected layer, wherein the fully connected layer is used for the estimation of SOH of the lithium battery.

5. The lithium battery health status estimation method according to claim 1, characterized in that: Step S3, pre-training the CNN-GRU hybrid neural network includes: CNN is used to automatically extract degradation information from battery cycling data. Convolution, pooling and activation operations are used to process the nonlinear relationship between input data and SOH, and to extract hidden feature information in cycling data. Multi-layer CNN is used for feature extraction to improve the correlation between extracted features and SOH; The GRU is used to extract features from the feature extraction model to build the model. Hidden temporal features are extracted from the degenerate features to establish the mapping relationship between features and SOH, thus obtaining the SOH estimation model.

6. The lithium battery health status estimation method according to claim 1, 4, or 5, characterized in that: When using a CNN-GRU hybrid neural network to estimate the state of harmonics (SOH) of lithium-ion batteries, the mean squared error is used as the loss function to evaluate the prediction error. The loss function is as follows: In the formula, N This indicates the number of battery data cycles. Indicates the first i The predicted SOH value for each data sample Indicates the first i The actual SOH value corresponding to each data sample.

7. The lithium battery health status estimation method according to claim 1, characterized in that: In step S4, the Tradaboost.R2 algorithm is used to perform transfer learning and parameter tuning on the pre-trained model on the target domain dataset, including the following steps: 1) Set up the base learner: First, a CNN-GRU hybrid neural network is used to pre-train the source domain dataset to learn the parameters of each part, and then the parameters are transferred to the parameters of the new base learner network. Then, when retraining using the target domain dataset, it is necessary to freeze the weights of the shallow convolutional network and the GRU layer network, and then train the deep network. 2) Use the KMM algorithm to determine the initial weights for the source and target domains; 3) Constructing the model framework: Step 1: Determine the source domain dataset ,have and i =1, 2, 3, ... n Determine the target domain dataset. ,have and j =1, 2, 3, ... m Merge the source domain dataset and the target domain dataset. ; Step 2: Set up the dataset T Initial weights, set maximum number of iterations N That is, the number of base learners; where, for the source domain dataset, the sample weight coefficient is... The sample weight coefficients of the target domain dataset are ; Step 3: Obtain source domain weights ,in This represents a sample of source domain data. This represents a sample of data from the target domain. Step 4: Initialize the weight vector: Step 5: Update source domain sample weight parameters: Step 6: Normalize the weight vector: Step 7: Obtain the base learner through model fine-tuning. The dataset will be merged. T and T Weight distribution on Train the model using the existing base learner to obtain the regressor: ; X Represents the source domain space. Y Represents the target domain space; Step 8, Calculation exist Error rate on: but Dt The overall regression error is: Step 9: Set the target domain weight update parameters: Step 10: Update the weight vectors of the training samples in the source and target domains: Step 11: Determine if the number of iterations is less than [a certain number]. N If it is less than, return to step 6; if it is greater than, proceed to step 12. Step 12, through the final N The lithium battery SOH estimation model is obtained by weighting the two base learners: In the formula, This represents the optimal parameters of the base learner. This represents the learning rate of the base learner.

8. The lithium battery health status estimation method according to claim 7, characterized in that: The calculation process of the KMM algorithm is as follows: In the formula, This represents the weighting factor for the source domain sample data. This represents a sample of source domain data. This represents a sample of data from the target domain. This represents the mapping function from the original space to RKHS. H Indicates having a characteristic kernel The regenerated Hilbert space - RKHS space.

9. The lithium battery health status estimation method according to claim 1, characterized in that: In step S5, the evaluation indicators include mean square error, root mean square error, and mean absolute error. The mean square error is as follows: The root mean square error is as follows: The mean absolute error is as follows: In the formula, This represents the estimated SOH value for lithium batteries. This represents the average value of the SOH estimates.

Images