A lithium battery SOH estimation method based on adaptive fusion of multi-level sequence information

Abstract

This invention proposes a lithium-ion battery SOH estimation method based on adaptive fusion of multi-level sequence information. The invention constructs a novel deep learning-based sequence model, consisting of two cascaded multi-level fusion modules and a bidirectional LSTM layer. Leveraging the proposed model's ability to adaptively extract and fuse multi-level sequence information, this model addresses the problem of insufficient battery data and inadequate extraction, thus achieving relatively accurate online SOH estimation of lithium-ion batteries. Furthermore, the proposed model possesses long-term memory, further improving online estimation accuracy. Experiments using battery degradation data from the NASA lithium-ion dataset validate the proposed network model through simulation. The results demonstrate that the model can successfully complete the online SOH estimation task of lithium-ion batteries while maintaining high robustness and accuracy.

Description

Technical Field

[0001] This invention belongs to the field of battery management technology, and relates to lithium battery management system technology, specifically to a lithium battery SOH estimation method based on multi-level sequence information adaptive fusion. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles (EVs) due to their long lifespan, high capacity, and wide operating temperature range. Timely and accurate monitoring of the EV battery status is crucial to ensuring the safety, reliability, and durability of EVs. However, prolonged and frequent battery use will inevitably shorten its lifespan. Furthermore, improper charging and use can accelerate battery aging and even lead to safety issues. Therefore, accurately estimating the state of health (SOH) of lithium-ion batteries has become a key element in ensuring the safe operation of electric vehicles.

[0003] Traditional methods for online estimation of state-of-the-art (SOH) in lithium-ion batteries mainly include direct measurement methods and model-based methods. Direct measurement methods involve offline measurement of aging indicators such as capacity and internal resistance of the lithium-ion battery, and then calculating the estimated SOH value using a formula. However, this type of method is open-loop, easily affected by uncertainties, and its estimation accuracy is highly dependent on the measuring instruments and techniques. Model-based methods mainly include electrochemical modeling, equivalent circuit methods, and empirical modeling. These methods establish models by considering the material properties and degradation mechanisms of the lithium-ion battery, thereby predicting the battery's SOH. However, currently established battery degradation models struggle to accurately reflect the battery's degradation process, resulting in poor robustness and accuracy in prediction.

[0004] With the rapid development of deep learning and the powerful data processing capabilities of deep neural networks, numerous online SOH estimation methods for lithium batteries based on deep neural networks have emerged. Unlike traditional methods, neural network-based methods treat the battery data modeling process as a black box, directly learning its internal dynamic characteristics through a large amount of charge and discharge data, and establishing a nonlinear relationship between SOH and raw data such as voltage, current, and temperature. Among them, convolutional neural networks (CNNs) are widely used in online SOH estimation of lithium batteries due to their advantages in time series forecasting. However, their single, sequential network structure cannot extract rich sequential information, resulting in poor model generalization and inaccurate estimation of lithium battery SOH. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by proposing a lithium-ion battery SOH estimation method based on multi-level sequence information adaptive fusion. A prediction model based on a multi-level fusion module and a bidirectional LSTM layer is constructed to complete the online SOH estimation task for lithium-ion batteries. Extensive experiments demonstrate that this method can achieve online SOH estimation of lithium-ion batteries while maintaining high robustness and estimation accuracy.

[0006] The online SOH estimation method for lithium batteries based on multi-level sequence information adaptive fusion specifically includes the following steps:

[0007] Step 1: The collected raw data is preprocessed using the min-max normalization method to map the data uniformly to the range [0,1], thus summarizing and unifying the statistical distribution of the lithium battery sample data and accelerating the convergence speed of the network model. The normalization formula is as follows:

[0008]

[0009] Where x′ represents the normalized data, x represents the original data in the dataset, and x' represents the normalized data. min x is the maximum value in the original data sample. max It is the minimum value in the original data sample.

[0010] Step 2: Construct a novel network model based on a multi-level fusion module and a bidirectional LSTM layer, which includes two cascaded multi-level fusion modules with identical structures, a bidirectional LSTM layer, a Flatten layer, and two fully connected layers. The multi-level fusion module includes a deep convolutional layer and a shallow convolutional layer, each containing network layers of different depths.

[0011] Step 3: Set the learning rate, training batch size, and training period. Input the data samples from the training set after normalization in Step 1 into the network model in Step 2. Save the network weight parameters at the end of training. Input the test set data samples after normalization in Step 1 into the proposed network model to verify the online estimation effect of SOH of lithium battery.

[0012] Preferably, the two cascaded multi-level fusion modules are used for feature extraction. Within each multi-level fusion module, two parallel convolutional layers are designed for extracting information at different levels. The deep convolutional layer comprises three one-dimensional convolutional modules, each consisting of a convolutional layer, an activation layer, and a max-pooling layer. The shallow convolutional layer contains only one convolutional layer, which facilitates the model's extraction of sequential information at different levels. The specific steps for feature extraction and data fusion by the multi-level fusion modules are as follows:

[0013] (a) Deep convolutional layers and shallow convolutional layers obtain initial extracted information through their respective branches; among them, deep convolutional layers obtain richer semantic information, while shallow convolutional layers obtain more raw data information;

[0014] (b) The outputs of deep and shallow convolutional layers adaptively fuse these two pieces of information to fully reflect the internal correlation state of the battery data, thereby solving to some extent the problem of inaccurate estimation results caused by insufficient battery data and inadequate extraction;

[0015] (c) The operation process of the designed multi-level information fusion module H is as follows:

[0016] H(X input )=F1(X input )+F2(X input (2)

[0017] Among them, X input The original input information for this network model is F1, which represents the feature extraction operator of the deep convolutional layer, and F2 represents the feature extraction operator of the shallow convolutional layer.

[0018] Preferably, the output formula of the bidirectional LSTM network is as follows:

[0019] h t =f(w1X) t+ w2h t-1 (3)

[0020] h t '=f(w3X t +w2h' t+1 (4)

[0021] O t =g(w4h4+w6h') t (5)

[0022] Among them, h t h is the output of the feedforward layer at time t. t-1 h' is the output of the feedforward layer at time t-1. t h' is the output of the reverse layer at time t. t+1 X is the output of the reverse layer at time t+1. t For the data input at time t, w i The weight matrix is ​​denoted by i, which takes values ​​from 1 to 6, and O is the weight matrix. t Let f(·) and g(·) be the output of the bidirectional LSTM network layer at time t, and let f(·) and g(·) denote the MLP operators.

[0023] Preferably, the battery data used in step 1 is the NASA lithium battery random usage dataset, which includes aging and degradation data samples of batteries of models B0005, B0006, and B0018.

[0024] Preferably, the learning rate of the network model in step 3 is set to 0.001, the training batch size is set to 32, and the training period is set to 200.

[0025] The present invention has the following effects:

[0026] Using a multi-level fusion module can acquire data features with different semantic information, thus solving to some extent the problem of inaccurate estimation results caused by insufficient battery data extraction. Using a bidirectional LSTM can model serialized information and effectively learn more time-series sample data information. This method combines a multi-level fusion module with a bidirectional LSTM to acquire rich semantic information while possessing the advantage of long-term memory, further improving prediction accuracy. Extensive simulation experiments demonstrate that this method can achieve end-to-end online estimation tasks and can be applied to serialized information prediction scenarios such as electric vehicle battery capacity estimation. Attached Figure Description

[0027] Figure 1 The flowchart shows the online SOH estimation method for lithium batteries based on adaptive fusion of multi-level sequence information.

[0028] Figure 2 This is a schematic diagram of a network model structure based on a combination of multi-level fusion modules and bidirectional LSTM.

[0029] Figure 3 Battery health status estimation results in the examples. Detailed Implementation

[0030] The present invention will be further explained and described below with reference to the accompanying drawings.

[0031] The experimental environment used in this embodiment is as follows: CPU Intel(R) Core(TM) i5-10600KF CPU@4.10Ghz, GPU RTX 3070, GPU memory 8GB, Python version 3.7, CUDA version 11.1, the deep learning framework used is TensorFlow-GPU 2.3.0, and the data used comes from the battery prediction dataset of NASA's Center of Excellence for Prediction.

[0032] like Figure 1 As shown, the online SOH estimation method for lithium batteries based on multi-level sequence information adaptive fusion specifically includes the following steps:

[0033] Step 1: To determine the degradation trend of lithium battery health under different operating conditions, this embodiment selects datasets B0005, B0006, B0018, and B0029 as training sets, and B0005 as the test set. The battery model used in this dataset is the LG Chem 18650 lithium battery, which can operate normally in a voltage range of 3.2-4.2V and a rated capacity of 2.1Ah. Related documentation indicates that the battery reaches the end of its lifespan when its capacitance drops to 70% of its nominal capacitance (2Ah) (1.4Ah).

[0034] Step 2: Construct a network model based on multi-layered fusion modules and bidirectional LSTM layers. The model structure is as follows: Figure 2 As shown, specifically, the model includes two multi-level sequence information fusion modules and a bidirectional LSTM layer. Within each multi-level sequence information fusion module, the invention meticulously designs two parallel convolutional layers for extracting information at different levels. The deep convolutional layer comprises three one-dimensional convolutional modules, each consisting of a convolutional layer, an activation layer, and a max-pooling layer; in contrast, the shallow convolutional layer contains only one convolutional layer, which facilitates the model's extraction of sequenced information at different levels. In this module, the deep convolutional layer acquires richer semantic information, while the shallow convolutional layer acquires more raw data information. These two types of information are then adaptively fused to fully reflect the internal correlation state of the battery data, thereby addressing to some extent the problem of inaccurate estimation results caused by insufficient battery data and inadequate extraction. The operation process of the designed multi-level information fusion module H is as follows:

[0035] H(X input )=F1(X input )+F2(X input (2)

[0036] Among them, X input The original input information for this network model is F1, which represents the feature extraction operator of the deep convolutional layer, and F2 represents the feature extraction operator of the shallow convolutional layer.

[0037] Then, the outputs of the deep and shallow convolutional layers are adaptively integrated into a bidirectional LSTM layer with 3 hidden layers. The bidirectional LSTM layer consists of two ordinary LSTM layers (forward and backward), and its output is determined by both LSTMs, thus achieving long-term memory while obtaining more accurate estimation results. The network output formula of the bidirectional LSTM at time t is as follows:

[0038] h t =f(w1X) t+ w2h t-1 (3)

[0039] ht '=f(w3X t +w2h' t+1 (4)

[0040] O t =g(w4h4+w6h') t (5)

[0041] Among them, h t h is the output of the feedforward layer at time t. t-1 h' is the output of the feedforward layer at time t-1. t h' is the output of the reverse layer at time t. t+1 X is the output of the reverse layer at time t+1. t For the data input at time t, w i (i from 1 to 6) is the weight matrix, O t Let f(·) and g(·) be the output of the bidirectional LSTM network layer at time t, and let f(·) and g(·) denote the MLP operators.

[0042] Finally, the data is processed through a Flatten layer and two fully connected layers to output the online estimation results of the SOH of the lithium battery.

[0043] Step 3: To verify the effectiveness of the proposed network model, the lithium battery dataset from Step 1 was divided into training and test sets in a 3:1 ratio. The raw battery data from the training set was input into the proposed network model based on a multi-level fusion module and a bidirectional LSTM layer. The online estimation performance of the lithium battery SOH of this model was then verified. The entire network model was trained using the Adam optimizer with a fixed learning rate of 0.001 and a training period of 200. To fully verify the effectiveness of the proposed model, two different loss functions were used as training objectives in the experiment. The calculation formulas are as follows:

[0044]

[0045]

[0046] Where N is the total number of original lithium battery samples. and Q i These are the model output predicted value and the actual labeled value for the i-th data point, respectively.

[0047] When the training cycle reaches the set 200, the training is completed, the network weight parameters at this time are saved, and the battery health estimation model is obtained.

[0048] Step 4: Input the test set data into the network model trained in Step 3. The test results are shown in the table below compared with the CNN+LSTM network model:

[0049] Estimation Model MAE MSE Error CNN-LSTM 0.0790 0.0158 0.07 This method 0.0231 0.0137 0.025

[0050] Table 1

[0051] Where MAE and MSE are the mean absolute error and mean squared error loss functions, respectively, and Error is the error in the online estimation of SOH of lithium batteries. Table 1 shows that the online SOH estimation method for lithium batteries based on multi-level sequence information adaptive fusion proposed in this invention can effectively achieve the task of online SOH estimation of lithium batteries and ensure high accuracy. The test results in the dataset are as follows: Figure 3 As shown.

Claims

1. A lithium battery SOH estimation method based on multi-level sequence information adaptive fusion, characterized in that, The method specifically includes the following steps: Step 1: Normalize the collected battery data; Step 2: Construct a novel network model based on a multi-layered fusion module and a bidirectional LSTM layer, which includes two cascaded multi-layered fusion modules with identical structures, a bidirectional LSTM layer, a Flatten layer, and two fully connected layers. Each multi-layered fusion module includes a deep convolutional layer and a shallow convolutional layer, each containing network layers of different depths. The deep convolutional layer comprises three one-dimensional convolutional modules, each consisting of a convolutional layer, an activation layer, and a max-pooling layer. The shallow convolutional layer contains only one convolutional layer. Step 3: Set the learning rate, training batch size, and training period. Input the data samples from the training set after normalization in Step 1 into the network model in Step 2. Save the network weight parameters at the end of training. Input the test set data samples after normalization in Step 1 into the network model to verify the online estimation effect of SOH of lithium battery.

2. The lithium battery SOH estimation method based on multi-level sequence information adaptive fusion as described in claim 1, characterized in that: In step 1, the collected raw data are preprocessed using the min-max normalization method to map the data uniformly to the range [0,1], thus summarizing and unifying the statistical distribution of the lithium battery sample data; the normalization formula is as follows: Where x′ represents the normalized data, x represents the original data from the NASA Lithium-ion Random Use Dataset, and x min x is the minimum value in the original data sample. max This represents the maximum value in the original data sample.

3. The lithium battery SOH estimation method based on multi-level sequence information adaptive fusion as described in claim 1, characterized in that: The cascaded two multi-level fusion modules perform feature extraction; within each multi-level fusion module, two parallel convolutional layers for extracting information at different levels are designed; the specific steps for feature extraction and data fusion in the multi-level fusion modules are as follows: (a) Deep convolutional layers and shallow convolutional layers obtain initial extracted information through their respective branches; (b) The outputs of deep and shallow convolutional layers fully reflect the internal correlation state of battery data by adaptively fusing the information from both. (c) The operation process of the designed multi-level information fusion module H is as follows: H(X input )=F1(X input )+F2(X input ) (2) Among them, X input The original input information for this network model is F1, which represents the feature extraction operator of the deep convolutional layer, and F2 represents the feature extraction operator of the shallow convolutional layer.

4. The lithium battery SOH estimation method based on multi-level sequence information adaptive fusion as described in claim 1, characterized in that: The output formula of the bidirectional LSTM layer is as follows: h t =f(w1X t +w2h t-1 ) (3) h t '=f(w3X t +w2h' t+1 ) (4) O t =g(w4h4+w6h' t ) (5) Among them, h t h is the output of the feedforward layer at time t. t-1 h' is the output of the feedforward layer at time t-1. t h' is the output of the reverse layer at time t. t+1 X is the output of the reverse layer at time t+1. t For the data input at time t, w i The weight matrix is ​​denoted by i, which takes values ​​from 1 to 6, and O is the weight matrix. t Let f(·) and g(·) be the bidirectional LSTM layer output at time t, and let f(·) and g(·) denote the MLP operators.

5. The lithium battery SOH estimation method based on multi-level sequence information adaptive fusion as described in claim 1, characterized in that: The battery data used in step 1 is the NASA lithium battery random usage dataset, which includes aging and degradation data samples of batteries of models B0005, B0006, and B0018.

6. The online SOH estimation method for lithium batteries based on multi-level sequence information adaptive fusion as described in claim 1, characterized in that: In step 3, the learning rate of the network model is set to 0.001, the training batch size is set to 32, and the training period is set to 200.

Images