A lithium battery health status estimation method based on time-frequency dual-flow feature representation

Abstract

A lithium battery health state estimation method based on time-frequency dual-stream feature representation obtains a battery data set and extracts data related to the battery health capacity; constructs an SFF Block module, including a time domain component, a frequency domain component, and a fusion component, and inputs the extracted data into the time domain component, the frequency domain component, and the residual connection respectively; in the time domain component, a CNN-BGRU network is used to preliminarily extract local information; in the frequency domain component, fast discrete Fourier transform and convolution are used to extract global frequency information; in the fusion component, the outputs of the time domain component and the frequency domain component are fused in the time and frequency domains to obtain a fused feature; residual connections are arranged at the output position of the fused feature; multiple SFF Block modules are stacked to construct SFFNet, and an evidence deep learning module is added after SFFNet to construct SFFNet Uncertainty, which outputs the prediction interval and confidence interval of the health state respectively.

Description

Technical Field

[0001] The present invention belongs to the field of battery technology, and in particular relates to a lithium battery state of health (SOH) estimation method based on time-frequency dual-flow feature characterization. Background Art

[0002] Currently, battery safety solutions are generally categorized into three types: intrinsic safety, passive safety, and active safety. Active safety utilizes intelligent algorithms to provide early warnings and lifespan assessments. One of the functions of a battery management system (BMS) is to monitor and manage batteries. As batteries age, they are affected by factors like temperature and voltage, resulting in a series of complex chemical changes that can present hidden safety risks. State of health (SOH) is closely linked to battery safety, making it crucial and meaningful to use a BMS to quickly and accurately estimate battery SOH.

[0003] Currently, methods for estimating the healthy capacity of batteries are divided into experience-based methods and performance-based methods, including open-circuit voltage method, ampere-hour integration method, Kalman filter, and neural network method. Different methods have different advantages and limitations. Among them, the open-circuit voltage method has a simple algorithm and low hardware cost, but it is affected by the operating environment, has large errors at the end of discharge, and is not accurate enough. The ampere-hour integration method has a simple algorithm and low hardware cost, but has weak anti-interference ability and average accuracy. The Kalman filter has high accuracy and can achieve online dynamic estimation with low computational complexity, but it has high requirements for the equivalent circuit model it uses, making it difficult to fully match the actual situation. The neural network method has strong fitting ability and high accuracy, but it incurs high computational overhead.

[0004] After searching, the following SOH prediction technology literature is relatively close:

[0005] Document 1: A hybrid neural network model with improved input for state ofcharge estimation of lithium-ion battery at low temperatures

[0006] Reference 2: Deep Evidential Regression

[0007] Reference 1 proposes a hybrid algorithm based on a CNN-BWGRU network. This method optimizes the influence of battery information on the results through a "multi-moment input" structure and a bidirectional network. A convolutional neural network (CNN) is used to learn the characteristic parameters of the input, while a bidirectional weighted gated recurrent unit (BWGRU) can improve the network's fitting performance at low temperatures by changing the weights. This network exhibits strong generalization, estimation accuracy, and robustness.

[0008] Reference 2 proposes a new method for estimating continuous targets and corresponding evidence using non-Bayesian neural networks, and quantifies epistemic and arbitrary uncertainty.

[0009] As represented by Reference 1, the current method of estimating battery SOH based on raw time domain signals has been widely explored by scholars in recent years and has achieved relatively ideal performance, which has verified the rationality and effectiveness of predicting battery SOH based on time domain signals.

[0010] Reference 2 points out that neural networks are increasingly being deployed in safety-critical areas, where calibrated, robust, and effective uncertainty measurements are crucial. In the battery field, confidence interval estimation has significant practical applications, such as battery safety early warning and abnormal data detection.

[0011] Combining reflections on References 1 and 2, we conclude that current data-driven algorithms have the following problems: First, most methods using battery data consider only the time domain signal and fail to effectively utilize its frequency domain information. However, the voltage and current in battery data have periodic characteristics, meaning that there is information to be mined in the frequency domain. Therefore, current methods do not fully consider this information. Second, in battery SOH estimation, if a direct prediction is performed, in fact (from a probabilistic perspective), it is a direct regression of its mean, without considering the uncertainty inherent in the problem. It is impossible to obtain confidence in the model's predicted value or a confidence interval for the SOH. Using only a single predicted value as the prediction result of the lithium-ion battery's health status is inaccurate and unscientific. Summary of the Invention

[0012] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a lithium battery health state estimation method based on time-frequency dual-stream feature characterization, which mines both local and global information of battery data from the time domain and frequency domain, and then uses the time and frequency domain feature fusion module to fuse the local information and global information, ultimately more accurately predicting the SOH structure and introducing the uncertainty of the data and model into the SOH estimation, thereby increasing the credibility of the lithium battery health state prediction results. The technical solution proposed by the present invention can extract features more comprehensively and improve the accuracy of the predicted SOH results.

[0013] In order to achieve the above object, the technical solution adopted by the present invention is:

[0014] A method for estimating the health status of a lithium battery based on time-frequency dual-flow feature characterization includes the following steps:

[0015] Step 1: Obtain a battery data set and preprocess the battery data, i.e., remove outliers and perform standardization.

[0016] Step 2: extracting data related to battery health capacity from the pre-processed battery data;

[0017] Step 3: Construct an SFF Block module. The SFF Block module includes a time domain component, a frequency domain component, and a fusion component. The extracted data are input into the time domain component, the frequency domain component, and the residual connection respectively. In the time domain component, the CNN-BGRU network is used to preliminarily extract local information. In the frequency domain component, the amplitude and phase angle components are obtained by fast discrete Fourier transform, and then convolution is performed to extract global frequency information. In the fusion component, the outputs of the time domain component and the frequency domain component are fused in the time domain and frequency domain to obtain a fused feature. The residual connection is arranged at the output position of the fused feature.

[0018] Step 4: Stack multiple SFF Block modules to construct SFFNet, add an evidence deep learning module after SFFNet, and construct SFFNet Uncertainty; train SFFNet and SFFNet Uncertainty separately, the output of SFFNet is the prediction interval of the health status, and the output of SFFNet Uncertainty is the confidence interval of the health status.

[0019] In one embodiment, in step 1, the battery data includes time, current, voltage, impedance, temperature, and capacity; in step 2, the extracted data related to the battery health capacity includes the battery current, voltage, temperature, and capacity during battery charging and discharging.

[0020] In one embodiment, the method for removing outliers is:

[0021] For data x1, x2, ..., x n , calculate its arithmetic mean And the residual error of each data, and calculate the standard deviation σ, if a data x i The residual error v i Satisfy the following formula:

[0022]

[0023] Then we think x iis an outlier containing gross errors, which is removed; where n is the number of data in the battery data set, 1≤i≤n;

[0024] The standardization is to perform minimum-maximum standardization on the data and map all values to the range of [0.0, 1.0]. The conversion formula is as follows:

[0025]

[0026] Among them, x and x * Respectively, they represent the values of the data before and after minimum-maximum normalization, and max and min represent the maximum and minimum values of all battery data.

[0027] In one embodiment, in step 3, in the time domain component, CNN is first used to slide along the time dimension to process the time domain features, and then the processed data is input into the bidirectional GRU to learn the bidirectional correlation, thereby completing the preliminary extraction of local information; in the frequency domain component, the data is first subjected to a fast discrete Fourier transform to obtain a frequency domain signal corresponding to the original data, each frequency domain signal represents the complex weight of its corresponding complex exponential harmonic component, that is, the original data is decomposed into a weighted sum of several complex exponential harmonic components to reflect the multi-scale periodicity of the original data; then CNN is used to perform a convolution operation on the amplitude and phase angle of the obtained frequency domain signal, and after multi-layer convolution, the obtained data is subjected to a fast inverse Fourier transform to obtain a time domain signal; in the fusion component, CBAM is used to fuse the outputs of the time domain component and the frequency domain component, and the CBAM includes a spatial attention mechanism and a channel attention mechanism, and the global information G and the fused information I are obtained by the following formula:

[0028] G = SpatialAttention (TF)

[0029] I=ChannelAttention(G+T)

[0030] Where T is the feature extracted by the time domain component, F is the feature extracted by the frequency domain component, SpatialAttention and ChannelAttention represent the spatial attention mechanism and channel attention mechanism respectively.

[0031] In one embodiment, the time domain component, the convolution template in the CNN has multiple; in the bidirectional GRU, the control coefficient z of the update gate t and reset gate control coefficient r t It is expressed as follows

[0032] z t =Sigmoid(W (z) x t +U (z) ht-1 )

[0033] r t =Sigmoid(W (r) x t +U (r) h t-1 )

[0034] Among them, x t is the input at the current moment, h t-1 is the hidden state of the previous moment, W (z) 、U (z) is the weight matrix of the linear mapping in the update gate, W (r) 、U (r) To reset the weight matrix of the linear mapping in the gate

[0035] The functions of the update gate and reset gate are expressed as:

[0036] h′ t =tanh(Wx t +r t ⊙Uh t-1 )

[0037] h t =(1-z t )⊙h t-1 +z t ⊙h′ t

[0038] Where h′ t is the new hidden state after input adjustment, h t It is the weighted sum of the hidden state at the previous moment and the new hidden state at the current moment. ⊙ represents the item-by-item product, that is, the corresponding elements of the two matrices are multiplied. t It represents the control coefficient. The closer it is to 1, the more data is remembered; the closer it is to 0, the more data is forgotten. W and U are the weight matrices of the candidate hidden states.

[0039] In one embodiment, the frequency domain component, for a signal sequence s with a length of T t , t=0,...,T-1, the fast discrete Fourier transform is as follows:

[0040]

[0041] The above formula is used to convert the time domain signal into a frequency domain signal, and then machine learning methods are used to extract features from the frequency domain signal. For a complex number a+j·b, its amplitude A and phase angle θ are calculated using the following formula:

[0042]

[0043]

[0044] Where j refers to the imaginary unit;

[0045] The obtained amplitude sequence and phase angle sequence are convolved separately. After passing through multiple layers of convolution, the transformed amplitude sequence and phase angle sequence are recombined into an imaginary number sequence.

[0046] Perform a fast inverse Fourier transform on the obtained imaginary number sequence to convert it into a time domain signal. The fast inverse Fourier transform is as follows:

[0047]

[0048] In one embodiment, the channel attention mechanism is expressed as follows:

[0049] I=Sigmoid(W C (AvgPooling(TF))+W C (MaxPooling(TF)))

[0050] Among them, W C It is the linear mapping weight matrix in the channel attention mechanism, which is used to represent the correlation between modeling channels; AvgPooling and MaxPooling refer to the average pooling layer and the maximum pooling layer respectively. In the channel attention mechanism, pooling is performed in the time dimension;

[0051] The spatial attention mechanism is expressed as follows:

[0052] G=Sigmoid(CNN([AvgPooling(TF),MaxPooling(TF)]))

[0053] In the spatial attention mechanism, pooling is performed in the channel dimension.

[0054] In one embodiment, during the SFFNet training process, the mean square error is used as the loss function to describe the difference between the predicted health state and the actual health state; the mean absolute error and the root mean square error are used to describe the accuracy of the predicted health state; during the SFFNet Uncertainty training process, evidence deep learning is used, and the loss function consists of two parts: the likelihood term and the evidence term. The smaller the likelihood term, the more it fits the distribution of the data, and the smaller the evidence term, the greater the uncertainty of the distribution corresponding to the error; the predicted health state is described using four distribution parameters, that is, a distribution interval; the mean absolute error and the root mean square error are used to describe the accuracy of the predicted health state.

[0055] In one embodiment, the SFFNet Uncertainty uses the evidential deep learning method to explicitly model the uncertainty of the prediction results using the output value after SFFNet, thereby realizing the uncertainty expression of the prediction results.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] (1) The present invention no longer uses the time domain information in the battery data alone, but proposes an SFFNet module that includes time domain, frequency domain processing and fusion parts. Among them, the proposed SFFNet module includes three components: time domain component, frequency domain component and time domain and frequency domain fusion part. First, the time domain and frequency domain information of the battery data are extracted, and then the extracted information is input into the fusion module. The fusion module can fuse the global information and local information learned from the two branches, and then add a residual structure from the perspective of gradient retention during training and preventing data from being lost during network processing. Finally, after being processed by multiple SFFNet modules, the data achieves the purpose of utilizing global information and local information.

[0058] (2) The present invention no longer uses a single predicted output value as the prediction result. Instead, it uses the evidence deep learning method to place the prior on the model output, explicitly model the uncertainty of the prediction result, and directly obtain the predicted distribution of SOH. Ultimately, the structure of SOH can be predicted more accurately, increasing the credibility of the prediction result. The technical solution proposed by the present invention can extract features more comprehensively and improve the accuracy of the predicted SOH results. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flow chart of the overall framework of the present invention.

[0060] Figure 2 This is the overall network structure diagram of the present invention.

[0061] Figure 3 This is a flow chart of battery data preprocessing according to the present invention.

[0062] Figure 4 This is a specific structural diagram of the SFF Block of the present invention.

[0063] Figure 5 This is a flow chart of network training of the present invention.

[0064] Figure 6 This is the battery SOH prediction flow chart of the present invention. DETAILED DESCRIPTION

[0065] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0066] like Figure 1 As shown, the present invention provides a lithium battery health status estimation method based on time-frequency dual-flow feature characterization, comprising the following steps:

[0067] Step 1: Obtain a battery data set and preprocess the battery data, that is, remove outliers and standardize them.

[0068] In this step, battery data can be collected from various experiments. This primarily includes time, current, voltage, impedance, temperature, and capacity. Since battery charging and discharging are cyclical, this periodicity often results in distinct patterns in the frequency domain. Therefore, this frequency domain of battery data holds a wealth of information to be explored, allowing for further data mining.

[0069] Step 2: Extract data related to battery health capacity from the preprocessed battery data.

[0070] In accordance with the purpose of the present invention, the extracted data related to the battery health capacity includes the battery current, voltage, temperature and capacity during battery charging and discharging.

[0071] Step 3: Construct an SFF Block module. The SFF Block module includes a time domain component, a frequency domain component, and a fusion component. The extracted data is input into the time domain component, the frequency domain component, and the residual connection respectively. In the time domain component, the CNN-BGRU network is used to preliminarily extract local information. In the frequency domain component, the amplitude and phase angle components are obtained using the fast discrete Fourier transform, and then convolution is performed to extract global frequency information. In the fusion component, the outputs of the time domain component and the frequency domain component are fused in the time and frequency domain features to obtain a fused feature. The residual connection is arranged at the output position of the fused feature.

[0072] The SFF Block module extracts local and global information from battery data in both the time and frequency domains, and further processes and fuses these extracted features to produce the final fused features. The SFF Block module has three input branches, all of which are data related to battery health capacity extracted from the battery data. Two of the input branches are time and frequency domain components, and the third branch is a constructed residual connection.

[0073] In this step, we make full use of the different modal information (time domain and frequency domain information) in the battery data and integrate them reasonably. From the perspective of feature extraction, in the time domain, we use convolution or recurrent neural networks to extract the original signal. The feature extraction performed at this time corresponds to the local information of the original signal. In the frequency domain, we perform a convolution operation on the amplitude and phase angle obtained from the original data through fast discrete Fourier transform. The feature extraction performed at this time corresponds to the global information of the original signal, thereby achieving the purpose of supplementing the local features extracted in the time domain with global features.

[0074] In step 4, multiple SFF Block modules are stacked to construct SFFNet. The evidence deep learning module is added to SFFNet to construct SFFNet Uncertainty. SFFNet and SFFNet Uncertainty are trained separately. The output of SFFNet is the prediction interval of the health state, and the output of SFFNet Uncertainty is the confidence interval of the health state, achieving the final SOH estimation.

[0075] In this step, uncertainty modeling is performed to consider the uncertainty of the data and the model, and the confidence interval of SOH is obtained to improve the characterization ability of the model, so as to more accurately predict the battery SOH. Furthermore, through SFFNetUncertainty, the output value after SFFNet is used to use the evidence deep learning method to explicitly model the uncertainty of the prediction result, thereby realizing the uncertainty expression of the prediction result. This part uses the evidence deep learning method to better consider the uncertainty caused by incomplete information in actual production or the test encountering data outside the distribution of the training set, and explicitly model the uncertainty. The output result is the confidence interval of SOH, which realizes the uncertainty expression of the prediction result of the battery health status, improves the characterization ability of the model, and increases the credibility of the prediction result. The overall network structure of the present invention is as follows Figure 2 shown.

[0076] In the embodiment of the present invention, reference Figure 3 In step 1, the collected data is preprocessed to check whether there are any outliers. If so, the outlier data is processed. After processing the outliers, the data is standardized.

[0077] This embodiment uses the Laida criterion to remove outliers from the data, which can be described as follows:

[0078] For data x1, x2, ..., x n , calculate its arithmetic mean And the residual error of each data, and calculate the standard deviation σ according to the Bessel formula, etc., if a data x iThe residual error v i Satisfy the following formula:

[0079]

[0080] Then we think x i It is an outlier containing gross errors and is removed. Where n is the number of data in the battery dataset, 1≤i≤n.

[0081] The standardization is to perform minimum-maximum normalization on the data and map all values to the range of [0.0, 1.0]. The minimum-maximum normalization is performed on the data obtained after removing outliers. The conversion formula is as follows:

[0082]

[0083] Among them, x and x * Respectively represent the values of the data before and after minimum-maximum normalization, max and min represent the maximum and minimum values of all battery data

[0084] In the embodiment of the present invention, reference Figure 4 In step 3, in the time domain component, a CNN and bidirectional GRU model are used to extract local information from the battery data. First, a CNN is used to process the time domain features by sliding along the time dimension. The sliding CNN can more comprehensively learn the continuous patterns in the input data. The processed data is then fed into a bidirectional GRU to learn bidirectional correlations, completing the initial extraction of local information.

[0085] This paper chooses GRU over the more complex Long Short-Term Memory (LSTM) neural network for two reasons. First, compared to LSTM, GRU has a simpler structure, making it better suited to less powerful hardware devices such as mobile computing. Second, with the same settings, GRU has fewer parameters than LSTM. This setting reduces the number of parameters in the time domain feature model, thus leaving sufficient parameter space for feature extraction in the frequency domain while minimizing overfitting.

[0086] Furthermore, in the time domain component, the specific details of CNN convolution are:

[0087] Specifically, for a one-dimensional convolution template g with a length of N i , i = 0, ..., N-1, a convolution operation in CNN can be expressed by the following formula,

[0088]

[0089] Depending on the settings, a CNN can have multiple convolutional templates, the number of which is called the number of channels. Theoretically, each template can model a specific pattern in the data. The advantage of multiple templates is that they can model multiple mixed patterns in the data. After convolution and dimensionality increase using a multi-channel CNN, the processed data is fed into a GRU to extract time-domain features from the original data.

[0090] The specific details of the bidirectional GRU network are:

[0091] GRU is a recurrent neural network based on the hidden Markov hypothesis, which includes two gate control links: update gate and reset gate. The control coefficient z of the update gate is t and reset gate control coefficient r t It is expressed as follows:

[0092] z t =Sigmoid(W (z) x t +U (z) h t-1 )

[0093] r t =Sigmoid(W (r) x t +U (r) h t-1 )

[0094] Among them, x t is the input at the current moment, h t-1 is the hidden state of the previous moment, W (z) 、U (z) is the weight matrix of the linear mapping in the update gate, W (r) 、U (r) is the weight matrix for the linear mapping in the reset gate. The update gate helps the model decide how much past information to carry forward to the future. On the other hand, using the update gate, the model can decide to copy all information from the past to reduce the risk of vanishing gradients. The reset gate primarily determines how much past information to forget, which can be calculated using the following expression. Specifically, the effects of the two gates can be expressed as:

[0095] h′ t =tanh(Wx t +r t ⊙Uh t-1 )

[0096] h t =(1-z t )⊙h t-1 +z t ⊙h′ t

[0097] Where h′ t is the new hidden state after input adjustment, h t It is the weighted sum of the hidden state at the previous moment and the new hidden state at the current moment. ⊙ represents the item-by-item product, that is, the corresponding elements of the two matrices are multiplied. t represents the control coefficient. The closer it is to 1, the more data is memorized. The closer it is to 0, the more data is forgotten. W and U are the weight matrices for the candidate hidden states. Therefore, the update gate determines the influence of the old and new hidden states on the current hidden state, while the reset gate determines the influence of the past hidden state on the new hidden state.

[0098] In the frequency domain, the data is first subjected to a fast discrete Fourier transform (FFT) to obtain the frequency domain signal corresponding to the original data. Each frequency domain signal represents the complex weight of its corresponding complex exponential harmonic component, decomposing the original data into the weighted sum of several complex exponential harmonic components, thereby reflecting the multi-scale periodicity of the original data. A CNN is then used to perform convolution operations on the amplitude and phase angle of the obtained frequency domain signal. After multiple layers of convolution, the obtained data is subjected to a fast inverse Fourier transform (FIFT) to obtain the time domain signal, facilitating the subsequent fusion of information from the time and frequency domains using the time-frequency fusion structure.

[0099] Furthermore, in the frequency domain component, for a signal sequence s with a length of T t , t=0,…,T-1, the fast discrete Fourier transform is as follows:

[0100]

[0101] The above formula converts the time domain signal into a frequency domain signal, and then uses machine learning methods to extract features from the frequency domain signal. Since the frequency domain signal is a complex signal, it is not convenient to process it directly. After obtaining the frequency domain signal, its corresponding amplitude and phase angle are first calculated. For a complex number a + j·b, its amplitude A and phase angle θ are calculated using the following formula:

[0102]

[0103]

[0104] Where j refers to the imaginary unit.

[0105] The obtained amplitude sequence and phase angle sequence are convolved separately. After passing through multiple layers of convolution, the transformed amplitude sequence and phase angle sequence are recombined into an imaginary number sequence.

[0106] In order to facilitate the subsequent fusion module to fuse the extracted features, it is necessary to perform a fast inverse Fourier transform on the obtained imaginary number sequence to convert it into a time domain signal. The fast inverse Fourier transform is as follows:

[0107]

[0108] The transformed data is used as the frequency domain features extracted by the model. It is worth mentioning that in many cases, convolution in the frequency domain is equivalent to multiplication in the time domain. This theory supports the rationale of using convolution in the frequency domain to extract global features to supplement local features extracted in the time domain.

[0109] After obtaining the features of the time and frequency domains, directly adding or subtracting the two is obviously too rough. Although it is expected that the two extract different features (decoupling the two features), in fact, the features extracted by the two are likely to contain aliased redundant information. Therefore, the present invention adopts a more advanced method. That is, in the fusion component, CBAM is used to fuse the outputs of the time domain component and the frequency domain component. CBAM includes two modules: Spatial Attention Mechanism and Channel Attention Mechanism. The global information G (Global) and the fused information I (Information) are obtained by the following formula:

[0110] G = SpatialAttention (TF)

[0111] I=ChannelAttention(G+T)

[0112] Where T is the feature extracted from the time domain component (Time), F is the feature extracted from the frequency domain component (Frequency), and SpatialAttention and ChannelAttention represent the spatial attention mechanism and channel attention mechanism, respectively. This paper uses TF to remove redundant information from the frequency domain features and then extracts the global feature G through the spatial attention mechanism. The global feature G is then added to the time domain feature T and adjusted using the channel attention mechanism to obtain a more representative comprehensive feature that combines global and local features.

[0113] Furthermore, in the fusion component, the channel attention mechanism, after obtaining the weight of each feature channel, the weight can be applied to the original channels, so that the importance of different channels can be learned. The channel attention mechanism can be expressed as follows:

[0114] I=Sigmoid(W C (AvgPooling(TF))+W C (MaxPooling(TF)))

[0115] Among them, WC It is the linear mapping weight matrix in the channel attention mechanism, used to represent the correlation between modeled channels. AvgPooling and MaxPooling refer to the average pooling layer and the maximum pooling layer, respectively. First, the global spatial features of each channel are used as the representation of that channel, that is, the input features are subjected to the average pooling layer and the maximum pooling layer. The pooling results are then fed into the multi-layer perceptron (MLP) to learn the degree of dependence of each channel and adjust different features accordingly. The two output branches are then added and mapped using the sigmoid function to finally obtain the channel attention value. In the channel attention mechanism, pooling is performed in the time dimension.

[0116] The spatial attention mechanism can capture the most important feature areas, which can be expressed as follows:

[0117] G=Sigmoid(CNN([AvgPooling(TF),MaxPooling(TF)]))

[0118] First, the input features are compressed to the channel size in order to learn spatial features. This involves performing an average pooling layer and a maximum pooling layer on the input features. The two pooling results are then concatenated according to the channel dimension and convolution is performed on the concatenated result. Unlike the channel-wise attention mechanism, in the spatial attention mechanism, pooling is performed in the channel dimension.

[0119] In an embodiment of the present invention, step 4 adopts the SFF Block stacking method to construct SFFNet as a whole, and from the perspective of retaining gradients during training and preventing data information loss during network processing, residual connections are added between each SFF Block.

[0120] The training process of SFFNet and SFFNet Uncertainty networks includes the following steps:

[0121] Step 1: Training data preparation

[0122] Because different battery datasets contain different data related to battery health and capacity, we first preprocess the dataset to check for outliers. If so, we address these outliers. We then filter data related to battery health and capacity (typically temperature, circuit, and voltage) and normalize the selected data to the range [0.0, 1.0]. During training, the input variables are divided into three parts: the time domain component, the frequency domain component, and the residual connection used to establish the error.

[0123] Step 2: Loss function setting (loss function, parameter update rules)

[0124] During SFFNet training, the mean squared error (MSE) is used as the loss function to describe the difference between the predicted health status and the true health status. The mean absolute error (MAE) and root mean squared error (RMSE) are used to describe the accuracy of the predicted health status.

[0125] During SFFNet Uncertainty training, evidence-based deep learning is used. The loss function consists of two parts: a likelihood term and an evidence term. A smaller likelihood term represents a closer fit to the data distribution, while a smaller evidence term indicates greater uncertainty in the distribution of the error. The predicted health status is described using four distribution parameters, i.e., a distribution interval. The accuracy of the predicted health status is described using mean absolute error and root mean square error.

[0126] The loss function MSE describes the average value of the square of the distance between the model prediction value and the true value of the sample. The specific calculation formula is as follows:

[0127]

[0128] Among them, f(x) is the SOH predicted by SFFNet, y i is the actual SOH size, and n is the number of samples.

[0129] The SFFNet Uncertainty loss function consists of two parts: the likelihood term and the evidence term. The specific formula is as follows:

[0130] L=L NLL +λL R

[0131] The mean absolute error (MAE) describes the average distance between the model prediction value and the true value of the sample. The specific calculation formula is as follows:

[0132]

[0133] Among them, f(x) is the SOH predicted by SFFNet, y i is the actual SOH size, and n is the number of samples.

[0134] The root mean square error (RMSE) describes the average value of the distance between the model prediction value and the true value of the sample. The specific calculation formula is as follows:

[0135]

[0136] Step 3: Network parameter settings

[0137] During the training process, SFFNet uses a stack of 3 SFF Block modules, each of which contains three components: time domain component, frequency domain component, and time-frequency domain fusion. In the time domain component, CNN contains 6 channels, the convolution kernel size is set to 7, and the number of neurons in BGRU is 32; in the frequency domain component, CNN contains 6 channels, the convolution kernel size is set to 7. The optimization algorithm used in network training is very important for improving the training speed and convergence of deep learning networks. Therefore, the present invention uses the Adam optimizer in training, the initial learning rate is set to 1e-5, and the batch_size of the training data used in backpropagation is set to 64.

[0138] A further improvement of the present invention is that the explicit modeling of uncertainty in the prediction results specifically includes the following steps:

[0139] This invention considers both the uncertainty of the data itself and the uncertainty of the model (from a Bayesian perspective). Data uncertainty aims to model the distribution of the battery SOH itself, while model uncertainty aims to model the confidence level of the model prediction. From a Bayesian perspective, data uncertainty is equivalent to estimating the likelihood function, while model uncertainty is equivalent to estimating the prior. The modeling approach is described in detail below.

[0140] According to the law of large numbers, without loss of generality, assume that the SOH of the battery is Y and obeys a Gaussian distribution N(·|μ,σ 2 ). And use the full probability to expand P(μ,σ 2 )=P(μ|σ 2 )P(σ 2 ), assuming that its standard deviation σ and mean μ follow an inverse gamma distribution and another Gaussian distribution respectively:

[0141] σ 2 ~Γ -1 (·|α,β)

[0142] μ~N(·|γ,σ 2 v -1 )

[0143] Among them, α, β and γ, σ 2 v -1 are the parameters of the inverse gamma distribution and the Gaussian distribution respectively. The reason for this assumption is that the prior probability and the corresponding posterior probability are conjugate distributions, which will greatly simplify the difficulty of derivation and make them have the same distribution type. Finally, P(μ,σ 2 ) is written as the standard inverse gamma distribution (NIG: Normal Inverse-Gamma distribution)

[0144]

[0145] Due to the nature of the distribution variance, we assume that the evidence is φ = 2ν + α. This evidence term will be used in the subsequent loss function design. Further, based on the above joint distribution, the distribution of the data and the distribution of its distribution parameters m = [α, β, γ, v] can be characterized, and the following conclusions can be obtained:

[0146] E[μ]=γ

[0147]

[0148]

[0149] The uncertainty of the data can be expressed as E[σ 2 ], the uncertainty of the model can be measured by Var[μ]. In summary, the last fully connected layer in SFFNet can be modified from predicting a mean μ to predicting four distribution parameters m, and the loss function can be replaced by the likelihood term L NLL and evidence item L R The weighted form, where the likelihood term refers to the negative log-likelihood of the data, can be directly obtained by changing the logarithm of the likelihood function. The evidence term and the final loss are shown below.

[0150] L R =\y-γ\φ

[0151] L=L NLL +λL R

[0152] Specifically, minimizing the negative log-likelihood allows the model to fit the data distribution, while minimizing the evidence term increases the uncertainty of the error distribution. In summary, by making assumptions about the battery SOH distribution and designing a loss function, we incorporate data and model uncertainty into the SOH estimation. This better accounts for the uncertainty caused by incomplete information or testing against data outside the training set distribution in real-world production, thus increasing practical value.

[0153] Before testing, the present invention first preprocesses the battery data in the test set, similar to the training set, by handling outliers and normalizing them. Variables related to battery health are then fed into the trained SFFNet network for prediction, yielding the predicted SOH results for each moment. In uncertainty modeling, the SFFNet Uncertainty network generates confidence intervals for the SOH.

[0154] In one embodiment of the present invention, the battery data selected was from the CS2 battery cycle test dataset published by the Computer-Assisted Product Lifecycle Engineering (CALCE) project at the University of Maryland. This battery type is a lithium cobalt oxide battery with a rated capacity of 1100 mAh. Data from CS2_35, CS2_36, CS2_37, and CS2_38, cycled at room temperature (23°C) and a constant current of 1C, were used for training and testing. Three datasets were selected for training and one for testing.

[0155] The data preprocessing process is as follows:

[0156] All CS2 cells in the CALCE dataset underwent the same charging profile, following a standard constant current / constant voltage protocol with a constant current rate of 0.5C until the voltage reached 4.2 V, and then maintained at 4.2 V until the charge current dropped below 0.05 A. The cutoff voltage during discharge was 2.7 V, and the cells were randomly named, with the nth numbered CS2 cell being named “CS2_n”.

[0157] Each CS2 battery was cycled multiple times under the conditions adjacent to its name in the table. The data file for each cell contains an Excel file of the recorded data generated by the test. The data used, CS2_35, CS2_36, CS2_37, and CS2_38, were all tested using an Arbin battery tester. The data information included includes 17 variables such as time, current, voltage, discharge capacity, and charge capacity. In this dataset, two variables, current and current, were selected as variables for predicting SOH. A total of four experiments were conducted, and in each experiment, three datasets were selected as training sets and one dataset as a test set. (For example, CS2_35, CS2_36, and CS2_37 were used as training sets, and CS2_38 data was used as a test set).

[0158] The SFFNet network construction and training process is as follows Figure 5 It uses a stack of three SFF Block modules. Each SFFBlock module contains a time domain component, a frequency domain component, and a fusion component. In the time domain component, the CNN has 6 channels, the convolution kernel size is set to 7, and the number of neurons in the BGRU is 32. In the frequency domain component, the CNN has 6 channels, the convolution kernel size is set to 7, and the number of neurons in the BGRU is 32.

[0159] The network uses the back-propagation algorithm. The optimization algorithm used in network training is very important for improving the training speed and convergence of deep learning networks. Therefore, the Adam optimizer is used in training, the initial learning rate is set to 1e-5, the batch_size of the training data used in back-propagation is set to 64, the number of training times is 300, and the model parameters with the best performance on the validation set are saved as the final training model.

[0160] Using the CS2_35, CS2_36, CS2_37, and CS2_38 data in the CALCE dataset, this time we use CS2_35, CS2_36, and CS2_38 as training sets, and CS2_37 as test sets. Figure 6 , the following are the specific steps for testing:

[0161] The test set data undergoes the same preprocessing as the training set, namely, outlier handling and standardization. The current and voltage in the test set are used as variables for predicting the battery's SOH and fed into the trained network for testing. The predicted SOH value corresponding to the current and voltage data at each moment is obtained. After obtaining the SOH values corresponding to all moments, they are compared with the actual SOH values at that moment, and the MAE and RMSE are calculated. Lower MAE and RMSE values indicate higher prediction accuracy for the model.

[0162] The model effect was evaluated by comparing with the CNN-BGRU method and the Transformer method, and the results are shown in Table 1. It can be seen from the results that the SFFNet Uncertainty model of the present invention can obtain better prediction results than other models.

[0163] Table 1

[0164]

[0165] It will be easily understood by those skilled in the art that the above description is merely an embodiment of the method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A lithium battery health status estimation method based on time-frequency dual-flow feature characterization, characterized in that: The steps include: Step 1: Obtain a battery data set and preprocess the battery data, i.e., remove outliers and perform standardization. Step 2: extracting data related to battery health capacity from the pre-processed battery data; Step 3: Construct an SFF Block module. The SFF Block module includes a time domain component, a frequency domain component, and a fusion component. The extracted data are input into the time domain component, the frequency domain component, and the residual connection respectively. In the time domain component, the CNN-BGRU network is used to preliminarily extract local information. In the frequency domain component, the amplitude and phase angle components are obtained by fast discrete Fourier transform, and then convolution is performed to extract global frequency information. In the fusion component, the outputs of the time domain component and the frequency domain component are fused in the time domain and frequency domain to obtain fused features. The residual connection is arranged at the output position of the fusion feature; Step 4: Stack multiple SFF Block modules to construct SFFNet, add an evidence deep learning module after SFFNet, and construct SFFNet Uncertainty; train SFFNet and SFFNet Uncertainty separately, the output of SFFNet is the prediction interval of the health status, and the output of SFFNet Uncertainty is the confidence interval of the health status.

2. The lithium battery health status estimation method based on time-frequency dual-flow feature characterization according to claim 1 is characterized in that: In step 1, the battery data includes time, current, voltage, impedance, temperature and capacity; in step 2, the extracted data related to the battery health capacity includes the battery current, voltage, temperature and capacity during battery charging and discharging.

3. The lithium battery health status estimation method based on time-frequency dual-flow feature characterization according to claim 1 is characterized in that: The method for removing outliers is as follows: For data x1, x2, ..., x n , calculate its arithmetic mean And the residual error of each data, and calculate the standard deviation σ, if a data x i The residual error v i Satisfy the following formula: Then we think x i is an outlier containing gross errors, which is removed; where n is the number of data in the battery data set, 1≤i≤n; The standardization is to perform minimum-maximum standardization on the data and map all values to the range of [0.0, 1.0]. The conversion formula is as follows: Among them, x and x * Respectively, they represent the values of the data before and after minimum-maximum normalization, and max and min represent the maximum and minimum values of all battery data.

4. The lithium battery health status estimation method based on time-frequency dual-flow feature characterization according to claim 1 is characterized in that: In step 3, in the time domain component, CNN is first used to slide along the time dimension to process the time domain features, and then the processed data is input into the bidirectional GRU to learn the bidirectional correlation, thereby completing the preliminary extraction of local information; in the frequency domain component, the data is first subjected to a fast discrete Fourier transform to obtain the frequency domain signal corresponding to the original data. Each frequency domain signal represents the complex weight of its corresponding complex exponential harmonic component, that is, the original data is decomposed into a weighted sum of several complex exponential harmonic components to reflect the multi-scale periodicity of the original data; then CNN is used to perform a convolution operation on the amplitude and phase angle of the obtained frequency domain signal, and after multi-layer convolution, the obtained data is subjected to a fast inverse Fourier transform to obtain the time domain signal; in the fusion component, CBAM is used to fuse the outputs of the time domain component and the frequency domain component. The CBAM includes a spatial attention mechanism and a channel attention mechanism, and the global information G and the fused information I are obtained by the following formula: G = SpatialAttention (TF) I=ChannelAttention(G+T) Where T is the feature extracted by the time domain component, F is the feature extracted by the frequency domain component, SpatialAttention and ChannelAttention represent the spatial attention mechanism and channel attention mechanism respectively.

5. The lithium battery health status estimation method based on time-frequency dual-stream feature characterization according to claim 4 is characterized in that: The time domain component, there are multiple convolution templates in CNN; in the bidirectional GRU, the control coefficient z of the update gate t and reset gate control coefficient r t It is expressed as follows z t =Sigmoid(W (z) x t +U (z) h t-1 ) r t =Sigmoid(W (r) x t +U (r) h t-1 ) Among them, x t is the input at the current moment, h t-1 is the hidden state of the previous moment, W (z) 、U (z) is the weight matrix of the linear mapping in the update gate, W (r) 、U (r) To reset the weight matrix of the linear mapping in the gate The functions of the update gate and reset gate are expressed as: h t =tanh(Wx t +r t ⊙Uh t-1 ) h t =(1-z t )⊙h t-1 +z t ⊙h′ t Where h′ t is the new hidden state after input adjustment, h t It is the weighted sum of the hidden state at the previous moment and the new hidden state at the current moment. ⊙ represents the item-by-item product, that is, the corresponding elements of the two matrices are multiplied. t It represents the control coefficient. The closer it is to 1, the more data is remembered; the closer it is to 0, the more data is forgotten. W and U are the weight matrices of the candidate hidden states.

6. The lithium battery health status estimation method based on time-frequency dual-flow feature characterization according to claim 4 is characterized in that: The frequency domain component, for a signal sequence s of length T t , t=0,...,T-1, the fast discrete Fourier transform is as follows: The above formula is used to convert the time domain signal into a frequency domain signal, and then machine learning methods are used to extract features from the frequency domain signal. For a complex number a+j·b, its amplitude A and phase angle θ are calculated using the following formula: Where j refers to the imaginary unit; The obtained amplitude sequence and phase angle sequence are convolved separately. After passing through multiple layers of convolution, the transformed amplitude sequence and phase angle sequence are recombined into an imaginary number sequence. Perform a fast inverse Fourier transform on the obtained imaginary number sequence to convert it into a time domain signal. The fast inverse Fourier transform is as follows:

7. The lithium battery health status estimation method based on time-frequency dual-flow feature characterization according to claim 4 is characterized in that: The channel attention mechanism is expressed as follows: I=Sigmoid(W C (AvgPooling(T-F))+W C (MaxPooling(T-F))) Among them, W C It is the linear mapping weight matrix in the channel attention mechanism, which is used to represent the correlation between modeling channels; AvgPooling and MaxPooling refer to the average pooling layer and the maximum pooling layer respectively. In the channel attention mechanism, pooling is performed in the time dimension; The spatial attention mechanism is expressed as follows: G=Sigmoid(CNN([AvgPooling(TF),MaxPooling(TF)])) In the spatial attention mechanism, pooling is performed in the channel dimension.

8. The lithium battery health status estimation method based on time-frequency dual-flow feature characterization according to claim 4 is characterized in that: During the SFFNet training process, the mean square error is used as the loss function to describe the difference between the predicted health status and the actual health status; the mean absolute error and root mean square error are used to describe the accuracy of the predicted health status; during the SFFNetUncertainty training process, evidence deep learning is used, and the loss function consists of two parts: the likelihood term and the evidence term. The smaller the likelihood term, the more it fits the data distribution, and the smaller the evidence term, the greater the uncertainty of the error corresponding distribution; the predicted health status is described using four distribution parameters, that is, a distribution interval; the mean absolute error and root mean square error are used to describe the accuracy of the predicted health status.

9. The lithium battery health status estimation method based on time-frequency dual-flow feature characterization according to claim 4 is characterized in that: The SFFNet Uncertainty uses the evidence deep learning method to explicitly model the uncertainty of the prediction results of the output values after SFFNet, thereby realizing the uncertainty expression of the prediction results.

Images