New energy automobile secondary battery health state intelligent estimation method based on small sample transfer learning

Abstract

The invention discloses a new energy automobile secondary battery health state intelligent estimation method based on small sample transfer learning, and the method comprises the steps: firstly obtaining the voltage, current and time sequence data of a new energy automobile secondary battery through a BMS, carrying out the linear interpolation downsampling operation of the data of each charging cycle, and constructing a source domain data set and a target domain data set according to different degradation levels; carrying out normalization processing on all the input features and the SOH labels; then constructing an SOH evaluation model based on a ResNet-Transform-KAN network architecture, inputting a source domain sample and a target domain sample into the model at the same time, taking regression loss of the source domain and a tagged target domain and distribution alignment loss based on MMD as a joint objective function, and performing iterative training through an Adam optimizer; and finally, inputting target domain label-free data into the trained model to realize state estimation of the secondary new energy automobile battery SOH at different degradation levels. According to the method, the accuracy of estimating the SOH of the secondary battery of the new energy automobile at different degradation levels by the model is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery state estimation, and particularly relates to an intelligent estimation method for the state of health of secondary batteries of new energy vehicles based on small-sample transfer learning. Background Art

[0002] With the proposal of the global carbon neutrality goal, the increasing severity of the greenhouse effect and the exhaustion of fossil energy, lithium-ion batteries have been widely used in the new energy vehicle industry due to their advantages such as high energy density, long cycle life, good safety performance, and low self-discharge rate. However, affected by a variety of complex physical and chemical processes, the battery performance will gradually decline with the number of charge and discharge cycles. When the state of health (SOH) of the new energy vehicle battery drops below 80%, in order to ensure the use safety, these batteries are usually retired. It is estimated that by 2040, the total amount of retired lithium batteries globally will reach several million tons. If not properly handled, it will cause serious environmental pollution and huge economic losses ([1] Shahjalal M, Roy PK, Shams T, Fly A, Chowdhury JI, Ahmed MR, et al. A review on second-life of Li-ion batteries: prospects, challenges, and issues. Energy. 2022; 241.). Applying the retired new energy vehicle batteries to scenarios with lower performance requirements such as energy storage not only extends the service life of resources but also effectively reduces the risk of environmental pollution. However, compared with the batteries in the first life cycle, the internal chemical characteristics of the secondary batteries of new energy vehicles have deteriorated significantly due to long-term service, mainly manifested as a decrease in energy density, unstable electrochemical reactions, and an increased risk of thermal runaway. To ensure the safe and reliable operation of these secondary batteries in downstream applications, the key lies in the accurate estimation of the state of health (SOH) of the secondary batteries of new energy vehicles.

[0003] At present, most of the existing studies on SOH estimation mainly focus on batteries in the first life cycle ([2] Braco E, San Martín I, Sanchis P, et al. State of health estimation of second-life lithium-ion batteries under real profile operation[J]. Applied Energy, 2022, 326: 119992.). However, they usually assume that the degradation levels of the batteries are relatively consistent, the data is complete, and the operating conditions are ideal, making it difficult to be directly applied to secondary batteries of new energy vehicles. In actual application scenarios, due to different user operation habits, environmental conditions, and historical charge and discharge records, there are significant differences in the degradation levels among secondary batteries of different new energy vehicles. This change in data distribution caused by degradation differences, namely the "domain shift" phenomenon, severely restricts the accuracy and robustness of existing SOH estimation models. Therefore, conducting accurate SOH estimation for secondary batteries of new energy vehicles with different degradation levels has become a key challenge. Summary of the Invention

[0004] In order to overcome the above-mentioned shortcomings of the existing technologies, the purpose of the present invention is to provide an intelligent estimation method for the state of health of secondary batteries of new energy vehicles based on small-sample transfer learning. Through the small-sample transfer learning method, the joint optimization of batteries with different degradation levels is realized, the generalization ability of the model is improved, and the prediction accuracy of the SOH of batteries with different degradation levels can be effectively improved.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] An intelligent estimation method for the state of health of secondary batteries of new energy vehicles based on small-sample transfer learning first obtains the charging data of secondary batteries of new energy vehicles, performs linear interpolation downsampling to the same length, and constructs a source domain dataset and a target domain dataset; then builds an SOH evaluation model for secondary batteries of new energy vehicles, and simultaneously inputs the source domain data and the target domain data samples into the model. Combining the evaluation loss of the source domain data and the labeled target domain data and the distribution alignment loss based on the maximum mean discrepancy (MMD) for joint optimization to achieve small-sample transfer learning; finally, inputs the unlabeled data in the target domain into the trained evaluation model to obtain the final SOH estimation value of the secondary batteries of new energy vehicles.

[0007] An intelligent estimation method for the state of health of secondary batteries of new energy vehicles based on small-sample transfer learning includes the following steps:

[0008] Step 1: Obtain the charging data of secondary batteries of new energy vehicles through the BMS, including voltage data (V n,k,1 , …, V n,k,i , …, Vn,k,l ), the current data (I n,k,1 , …, I n,k,i , …, I n,k,l ) and its corresponding time series (T n,k,1 , …, T n,k,i , …, T n,k,l ), where i = 1, …, l, V n,k,i represents the i-th voltage data point of the k-th charging cycle of the n-th secondary battery;

[0009] Step 2: Resample the original sequence {V n,k,i , I n,k,i , T n,k,i} of the k-th charging cycle of the n-th secondary battery obtained in Step 1 to the same length l' by linear interpolation;

[0010] Step 3: Organize the resampling results of each cycle as samples according to O n,k = (V n,k , I n,k , T n,k ), and collect all K-cycle samples of the n-th secondary battery into the dataset O n = {O n,1 , O n,2 , O n,3 , ··· O n,k , ··· O n,K}; Divide all samples into the source domain dataset D s and the target domain dataset D T according to the degradation level. All samples in the source domain dataset D s have SOH labels, and the target domain dataset D T is further divided into a labeled subset and an unlabeled subset

[0011] Step 4: Perform min-max normalization processing on the original input features and SOH labels in both the source domain dataset D s and the target domain dataset D T in the [0, 1] interval,

[0012]

[0013] where x is the original input feature (voltage, current, time), x min and x max are the minimum and maximum values of the original input feature sequence respectively, x norm is the normalized feature value, SOH is the original label value of the new energy vehicle battery, SOH min and SOHmax are their maximum and minimum values, respectively, SOH norm is the labeled value obtained after normalization;

[0014] Step 5: Construct a health state evaluation model for the secondary battery of new energy vehicles. The health state evaluation model for the secondary battery of new energy vehicles is based on the ResNet-Transformer-KAN network architecture and consists of a feature extraction module and a predictor; among them, the feature extraction module consists of three layers of residual blocks and one layer of Transformer module, and the predictor consists of two layers of KAN linear layers; the Transformer module adopts a multi-head self-attention mechanism and positional encoding; finally, through a Kolmogorov-Arnold network predictor based on B-spline activation, the extracted high-order features are mapped to a scalar;

[0015] Step 6: Input the samples in the source domain dataset D S and the target domain dataset D T into the constructed health state evaluation model for the secondary battery of new energy vehicles for training, and respectively output the source domain prediction value and the target domain prediction value After passing through the feature extraction module, the source domain feature F S and the target domain feature F T are obtained; a distribution alignment term based on the maximum mean discrepancy (MMD) is added to the loss function, and the joint objective function used during training is:

[0016]

[0017] where, |D S | and are the numbers of samples in the source domain and the target domain respectively, λ is the proportion of the MMD loss in the overall objective function , F S and F T are the source domain feature and the target domain feature passing through the feature extraction module respectively; the joint objective function is minimized by the Adam optimizer, and the parameters of the evaluation model are continuously iteratively optimized to obtain a better evaluation effect;

[0018] Step 7: Input the unlabeled data in the target domain into the trained health state evaluation model for the secondary battery of new energy vehicles to obtain the final evaluation value which is the health state of the evaluated secondary battery.

[0019] The specific process of resampling it to the same length l' by linear interpolation in Step 2 is as follows:

[0020] 2.1) For the p-th resampled point, where p = 1, 2, …, l', calculate its corresponding position in the original sequence:

[0021]

[0022] 2.2) Determine the neighboring samples: Let δ = i p '- i, where i is the position after rounding down, i + 1 is the position after rounding up, and δ ∈ [0, 1) is the interpolation weight;

[0023] 2.3) Linear interpolation: Interpolate between two neighboring points in the original sequence according to the ratio δ to obtain the resampled voltage, current, and time:

[0024] V n,k,p = V n,k,i + δ × (V n,k,i+1 - V n,k,i )

[0025] I n,k,p = I n,k,i + δ × (I n,k,i+1 - I n,k,i )

[0026] T n,k,p = T n,k,i + δ × (T n,k,i+1 - T n,k,i )

[0027] V n,k,i , I n,k,i , T n,k,i represent the magnitudes of the i-th value of the n-th secondary battery at the k-th cycle before downsampling.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] The present invention first obtains the charging data of the secondary battery of a new energy vehicle through the BMS, and performs linear interpolation downsampling on the obtained voltage, current, and time series to make them unified to the same length, and respectively constructs the source domain dataset D s and the target domain dataset D T; and the maximum - minimum normalization method is used to normalize the input features and SOH labels to eliminate the scale difference and stabilize the gradient update. Subsequently, an SOH evaluation model for secondary batteries of new energy vehicles is built based on the ResNet–Transformer–KAN network architecture, and the source - domain data and target - domain data samples are simultaneously input into the model. Combining the evaluation losses of the source domain and the labeled target domain and the distribution alignment loss based on the maximum mean discrepancy (MMD) for joint optimization, small - sample transfer learning is achieved. Finally, the unlabeled data in the target domain is input into the trained model to obtain the final SOH estimation value of the secondary battery of the new energy vehicle, effectively solving the problem of low SOH evaluation accuracy caused by different degradation levels among secondary batteries of new energy vehicles.

[0030] The present invention introduces a small - sample transfer learning method to narrow the domain difference between the data of secondary batteries of new energy vehicles with different degradation levels, effectively solving the domain shift problem caused by different degradation levels of secondary batteries of new energy vehicles, thereby significantly improving the evaluation performance of the SOH of secondary batteries of new energy vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flowchart of an embodiment of the present invention.

[0032] Figure 2 It is a comparison chart of the results of different methods in the embodiment of the present invention, where (a) is the evaluation result of target - domain battery 1, (b) is the evaluation result of target - domain battery 2, (c) is the evaluation result of target - domain battery 3, and (d) is the comparison of the evaluation results of different methods. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] The present invention will be further described in detail below in conjunction with the embodiments and the drawings.

[0034] Referring to Figure 1 , an intelligent estimation method for the health state of secondary batteries of new energy vehicles based on small - sample transfer learning includes the following steps:

[0035] Step 1: Obtain the charging data of the secondary battery of the new energy vehicle through the BMS, including voltage data (V n,k,1 ,…,V n,k,i ,…,V n,k,l ), current data (I n,k,1 ,…,I n,k,i ,…,I n,k,l ) and their corresponding time series (T n,k,1 ,…,T n,k,i ,…,T n,k,l ), where i = 1,…,l, and V n,k,i represents the i - th voltage data point of the n - th secondary battery in the k - th charging cycle.

[0036] Step 2: For the original sequence {V n,k,i , I n,k,i , T n,k,i} of the nth secondary battery at the kth charge cycle obtained in Step 1, where i = 1, 2, …, l, resample it to the same length l' using linear interpolation. The specific process is as follows:

[0037] 2.1) For the pth point (p = 1, 2, …, l') after resampling, calculate its corresponding position in the original sequence:

[0038]

[0039] 2.2) Determine the neighboring samples: Let δ = i p ' - i, where i is the position after rounding down, i + 1 is the position after rounding up, and δ ∈ [0, 1) is the interpolation weight;

[0040] 2.3) Linear interpolation: Interpolate between two adjacent points in the original sequence according to the ratio δ to obtain the voltage, current, and time after resampling:

[0041] V n,k,p = V n,k,i + δ × (V n,k,i+1 - V n,k,i )

[0042] I n,k,p = I n,k,i + δ × (I n,k,i+1 - I n,k,i )

[0043] T n,k,p = T n,k,i + δ × (T n,k,i+1 - T n,k,i )

[0044] V n,k,i , I n,k,i , T n,k,i represent the magnitudes of the ith value of the nth secondary battery at the kth cycle before downsampling;

[0045] Step 3: Organize the resampling results of each cycle as samples according to O n,k = (V n,k , I n,k , T n,k ), and collect all K - cycle samples of the nth secondary battery into the dataset O n = {O n,1 , O n,2 , O n,3 , ··· O n,k , ··· On,K}; To make full use of battery data with different degradation levels under the transfer learning framework, all samples are divided into the source domain dataset D s (all with SOH labels) and the target domain dataset D T , and the target domain dataset D T is further divided into a labeled subset and an unlabeled subset for supervised fine-tuning and unsupervised distribution alignment learning respectively;

[0046] Step 4: To reduce the impact of gradient imbalance caused by feature scale differences on model training, both the original input features and SOH labels in the source domain dataset D s and the target domain dataset D T are processed by min-max normalization in the [0,1] interval. Through this operation, all data with different degradation levels are mapped to a unified scale, achieving stable gradient updates and significantly accelerating the convergence speed of the model;

[0047]

[0048] where x is the original input feature (voltage, current, time), x min and x max are the minimum and maximum values of the original input feature sequence respectively, x norm is the normalized feature value, SOH is the original label value of the new energy vehicle battery, SOH min and SOH max are its maximum and minimum values respectively, and SOH norm is the label value obtained after normalization;

[0049] Step 5: Build a health state assessment model for new energy vehicle secondary batteries. The health state assessment model for new energy vehicle secondary batteries is based on the ResNet-Transformer-KAN network architecture and consists of a feature extraction module and a predictor; among them, the feature extraction module is mainly composed of three layers of residual blocks and one layer of Transformer module, and the predictor is composed of two layers of KAN linear layers; the residual blocks are mainly used to capture local degradation features in the charging data; the Transformer module uses the multi-head self-attention mechanism and position encoding to capture global temporal dependencies in the charging data; finally, through the Kolmogorov-Arnold network predictor based on B-spline activation, the extracted high-order features are mapped to a scalar to output the health state (SOH) of the secondary battery;

[0050] Step 6: Combine the source domain dataset D s and the target domain dataset D TThe samples in are simultaneously input into the constructed health state assessment model of the secondary battery of new energy vehicles for training, and the source domain prediction values are output respectively and the target domain prediction values After passing through the feature extraction module, the source domain feature F s and the target domain feature F T are obtained; in order to make full use of a small number of labeled target domain samples and reduce the distribution difference between the source domain and the target domain, a distribution alignment term based on the maximum mean discrepancy (MMD) is added to the loss function, and the joint objective function used during training is:

[0051]

[0052] where, |D S | and are the numbers of samples in the source domain and the target domain respectively, λ is the proportion of the MMD loss in the overall objective function , F S and F T are the source domain feature and the target domain feature passing through the feature extraction module respectively; the joint objective function is minimized by the Adam optimizer, and the parameters of the evaluation model are continuously iteratively optimized to obtain a better evaluation effect;

[0053] Step 7: Input the unlabeled data in the target domain into the trained health state assessment model of the secondary battery of new energy vehicles to obtain the final evaluation value which is the health state of the evaluated secondary battery.

[0054] This embodiment verifies the effectiveness of the method of the present invention based on a secondary battery dataset of a retired Nissan new energy vehicle. The dataset consists of two different retired battery pack groups NP5 and NP6. Each pack group contains 16 lithium manganese oxide battery cells (nominal capacity of 56 Ah, charge and discharge voltage range of 2.8 V - 4.2 V), and the degradation levels of the two battery pack groups are significantly different: the SOH of NP5 gradually decays from about 80% to 73%, while the SOH of NP6 drops from about 78% to 66%.

[0055] Take the 16 battery cells in the NP5 pack group as the source domain dataset, 1 battery cell in the NP6 pack group as the labeled target domain dataset, and the other 15 battery cells as the unlabeled target domain dataset, and repeat 4 groups of transfer experiments to verify the feasibility of the present invention. In the experiment, the mean square error (MSE) and the mean absolute error (MAE) indicators are respectively used to evaluate the estimation performance of SOH, and three other methods are compared at the same time. The experimental results are shown in Table 1 and Figure 2 as follows:

[0056] Table 1

[0057]

[0058] Method 1: Train the model on a labeled source domain dataset and an unlabeled target domain dataset, use MMD to align feature distribution, and do not use labeled target domain information. Method 2: Train the model using only limited labeled data in the target domain dataset, without any source domain knowledge or domain adaptation mechanism. Method 3: Train the model only on sufficient labeled data in the source domain dataset, without considering transfer learning strategies at all.

[0059] The experimental results show that the proposed method performs best in SOH estimation of the target domain dataset: MSE reaches 0.173% and MAE is 1.094%; the MSE and MAE of method 1 are 0.346% and 1.599% respectively; the MSE of method 2 is 0.352% and MAE is 1.775%; method 3 is the worst, with MSE of 0.622% and MAE of 2.133%. The above results prove that the proposed method significantly reduces the error caused by domain shift by effectively utilizing the labeled and unlabeled data of the source domain dataset and the target domain dataset, and improves the accuracy and robustness of SOH estimation of new energy vehicle secondary batteries with different degradation levels.

[0060] The present invention is suitable for SOH evaluation of new energy vehicle secondary batteries with different degradation levels. By utilizing a small sample transfer learning strategy to reduce the domain offset caused by secondary batteries with different degradation levels, the accuracy of the model in estimating SOH of new energy vehicle batteries with different degradation levels can be improved to a certain extent.

Claims

1. An intelligent estimation method for the state of health of secondary batteries of new energy vehicles based on small-sample transfer learning, characterized in that: First, obtain the charging data of the secondary battery of a new energy vehicle, perform linear interpolation to downsample it to the same length, and construct the source domain dataset and the target domain dataset; then build an SOH evaluation model for the secondary battery of a new energy vehicle, and input the source domain data and the target domain data samples into the model at the same time, and jointly optimize the evaluation loss of the source domain data and the labeled target domain data and the distribution alignment loss based on the maximum mean discrepancy (MMD); finally, input the unlabeled data in the target domain into the trained evaluation model to obtain the final SOH estimate of the secondary battery of a new energy vehicle.

2. The method according to claim 1, characterized in that It includes the following steps: Step 1: Obtain the charging data of the secondary battery of the new energy vehicle through the BMS, including voltage data (V n,k,1 , …, V n,k,i , …, V n,k,l ), current data (I n,k,1 , …, I n,k,i , …, I n,k,l ) and their corresponding time series (T n,k,1 , …, T n,k,i , …, T n,k,l ), where i = 1, …, l, V n,k,i represents the i-th voltage data point of the n-th secondary battery in the k-th charging cycle; Step 2: For the original sequence {V n,k,i , I n,k,i , T n,k,i} of the n-th secondary battery at the k-th charging cycle obtained in Step 1, where i = 1, 2, …, l, resample it to the same length l' by linear interpolation; Step 3: Organize the resampling results of each loop into samples according to O n,k =(V n,k , I n,k , T n,k ), and collect all K-loop samples of the n-th secondary battery into the dataset O n ={O n,1 , O n,2 , O n,3 , ··· O n,k , ··· O n,K}; Divide all samples into the source domain dataset D s and the target domain dataset D T according to the degradation level. All samples in the source domain dataset D s have SOH labels, and the target domain dataset D T is further divided into a labeled subset and an unlabeled subset Step 4: For the source domain dataset D s and the target domain dataset D T both the original input features and the SOH labels in them are processed by min-max normalization in the interval [0, 1]. where x is the original input feature, including voltage, current or time, x min and x max are the minimum and maximum values of the original input feature sequence respectively, x norm is the normalized feature value, SOH is the original label value of the new energy vehicle battery, SOH min and SOH max are its maximum and minimum values respectively, SOH norm is the label value obtained after normalization; Step 5: Construct a health state evaluation model for the secondary battery of a new energy vehicle. The health state evaluation model for the secondary battery of a new energy vehicle is based on the ResNet-Transformer-KAN network architecture and consists of a feature extraction module and a predictor; among them, the feature extraction module consists of three layers of residual blocks and one layer of Transformer module, and the predictor consists of two layers of KAN linear layers; the Transformer module adopts a multi-head self-attention mechanism and position encoding; finally, through a Kolmogorov-Arnold network predictor based on B-spline activation, the extracted high-order features are mapped to a scalar; Step 6: Input the samples in the source domain dataset D s and the target domain dataset D T into the constructed health state assessment model of the secondary battery of new energy vehicles for training, and respectively output the source domain prediction value and the target domain prediction value After passing through the feature extraction module, the source domain feature F s and the target domain feature F T are obtained; Add a distribution alignment term based on the maximum mean discrepancy (MMD) to the loss function, and the joint objective function used during training is: Among them, |D S | and are the numbers of samples in the source domain and the target domain respectively, λ is the proportion of the MMD loss in the overall objective function , F S and F T are the source domain features and the target domain features passing through the feature extraction module respectively; the joint objective function is minimized by the Adam optimizer, and the parameters of the evaluation model are continuously iteratively optimized to obtain a better evaluation effect; Step 7: Input the unlabeled data in the target domain into the trained new energy vehicle secondary battery state of health assessment model to obtain the final evaluation value which is the state of health of the evaluated secondary battery.

3. The method according to claim 2, wherein The specific process of resampling it to the same length l' by linear interpolation in step 2 is as follows: 2.1) For the p-th point after resampling, p = 1, 2, …, l', calculate its corresponding position in the original sequence: 2.2) Determine neighboring samples: Let δ = i p '- i, where i is the position after rounding down, i + 1 is the position after rounding up, and δ ∈ [0, 1) is the interpolation weight; 2.3) Linear interpolation: Interpolate between two adjacent points in the original sequence according to the ratio δ to obtain the voltage, current, and time after resampling: V n,k,p = V n,k,i + δ × (V n,k,i+1 - V n,k,i ) I n,k,p = I n,k,i + δ × (I n,k,i+1 - I n,k,i ) T n,k,p = T n,k,i + δ × (T n,k,i+1 - T n,k,i ) V n,k,i , I n,k,i , T n,k,i represents the magnitude of the i-th value of the k-th cycle of the n-th secondary battery before downsampling.

Images