Lithium battery life prediction method based on EMD framework

Abstract

The invention relates to a lithium ion battery life prediction method, and belongs to the field of battery life prediction and intelligent maintenance. The method comprises the steps that S1, a battery capacity degradation sequence is collected, and integrity is checked and normalized; s2, decomposing the sequence by using an improved complete set empirical mode decomposition algorithm, and dividing the sequence into a high-frequency component and a low-frequency component according to a zero-crossing rate; s3, modeling the high-frequency component: fusing multi-scale channel interactive attention, a time sequence convolutional network and a hybrid expert model, and extracting short-term fluctuation and capacity recovery features; s4, modeling a low-frequency component: introducing a two-way gating circulation unit network constrained by a double-index degradation model, and simulating a long-term trend; and S5, constructing a high-frequency migration module through tensor decomposition, improving cross-battery generalization, and fusing high and low frequency results to output a residual life prediction value. According to the method, a dual-channel framework combining signal decomposition, deep learning and physical modeling is combined, the prediction precision and adaptability under complex degradation are improved, and the method is suitable for various battery systems.

Description

Technical Field

[0001] This invention relates to the field of battery health prediction technology, and specifically to a lithium battery life prediction method based on the EMD framework. Background Technology

[0002] Lithium-ion batteries (LIBs) are widely used in electric vehicles, renewable energy storage, and portable electronic devices due to their high energy density, long lifespan, and low self-discharge rate. However, during long-term charge-discharge cycles, the performance of lithium-ion batteries gradually degrades due to internal chemical and physical changes, leading to a decrease in capacity and reduced driving range. Typically, when the battery capacity drops to 70% to 80% of its initial capacity, it is considered to have reached the end of life (EOL) and needs to be replaced promptly to avoid malfunctions and safety hazards. Therefore, accurately predicting the remaining useful life (RUL) of lithium-ion batteries is crucial for ensuring safe system operation and optimizing battery management.

[0003] The capacity degradation process of lithium-ion batteries exhibits complex characteristics such as nonlinearity, non-stationarity, and capacity regeneration (CR). Capacity regeneration refers to the temporary recovery of battery capacity during non-operating cycles due to the partial dissipation of unstable byproducts between the electrodes and electrolyte, significantly increasing the difficulty of predicting rated capacity up to zero (RUL). Existing RUL prediction methods mainly fall into two categories: physical model-based methods, data-driven methods, and hybrid methods.

[0004] Physical model-based methods predict battery degradation by constructing mathematical models (such as empirical models, equivalent circuit models, and electrochemical models) that reflect the battery degradation mechanism. These methods rely on accurate prior knowledge, involve complex model construction, incur high computational costs, and have limited generalization ability under complex operating conditions. For example, patent CN111426952A, entitled "A Method for Predicting the Lifetime of a Lithium-ion Battery," employs a prediction method based on a capacity decay model. This method relies on a deep understanding of the complex electrochemical reaction mechanisms within the battery, involving relatively complex chemical and physical processes, thus requiring stringent conditions in practical applications. Furthermore, errors in the model construction process can significantly deviate from the lifetime prediction results, thereby reducing its practical value and reference significance in battery lifetime prediction.

[0005] Data-driven methods, which construct statistical or machine learning models by analyzing historical data, have gained widespread attention in recent years with the development of deep learning technology. While data-driven methods offer flexibility and ease of implementation, their performance is highly dependent on large amounts of high-quality data. Furthermore, their long-term prediction stability remains insufficient when facing complex characteristics such as non-stationarity and capacity regeneration. For example, patent number CN202110474679.8, entitled "A Method and System for Predicting the Lifespan of Power Batteries for New Energy Vehicles," employs a data-driven approach to predict battery lifespan. However, this method is highly dependent on the model structure and data quality used; any errors in the model or data will significantly affect the accuracy of the prediction results.

[0006] To address the aforementioned issues, this invention proposes a lithium-ion battery RUL prediction method and system based on improved complete set empirical mode decomposition (ICEEMDAN), differential modeling, and transfer learning, in order to improve prediction accuracy, robustness, and cross-operating condition adaptability. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for predicting the remaining life of lithium-ion batteries based on the EMD framework and transfer learning, which solves the problems of insufficient prediction accuracy and limited generalization ability of existing methods when dealing with nonlinearity, non-stationarity and capacity regeneration phenomena of battery capacity sequences.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] Obtain capacity sequence data of lithium-ion batteries and normalize the data;

[0010] The battery capacity sequence data is decomposed using an improved complete set empirical mode decomposition and adaptive noise algorithm (ICEEMDAN) to obtain multiple intrinsic mode function (IMF) components and a residual component.

[0011] Calculate the zero-crossing rate of the IMF component and the residual component. If the zero-crossing rate is greater than 0.01, it is classified as a high-frequency component; otherwise, it is classified as a low-frequency component.

[0012] For the high-frequency components, an ensemble model integrating multi-scale channel interactive attention mechanism (MSICA), modern temporal convolutional network (ModernTCN), and hybrid expert (MoE) is used for prediction. Transfer learning based on adaptive tensor decomposition is designed and introduced to obtain the predicted values ​​of the high-frequency components.

[0013] For the low-frequency components, a bidirectional gated cyclic unit (PC-BiGRU) model with a double exponential degradation model as a physical constraint is used for prediction to obtain the predicted values ​​of the low-frequency components.

[0014] The predicted values ​​of the high-frequency components and the low-frequency components are added together to obtain the predicted capacity value of the lithium-ion battery.

[0015] Preferably, the battery capacity sequence is decomposed using an improved complete set empirical mode decomposition and adaptive noise algorithm, including:

[0016] Add adaptive Gaussian white noise to the original capacity sequence to generate a perturbed sequence:

[0017] C (i) (t)=C(t)+β0·E1(ξ (i) (t)), i = 1, 2, ..., K

[0018] Where β0 is the initial noise intensity, ξ (i) (t) represents the Gaussian white noise generated independently for the i-th time.

[0019] The first-order residual is calculated by integrating the local means of the average perturbation sequence, and the first-order IMF component is extracted:

[0020] IMF1(t) = C(t) - R1(t)

[0021] Iteratively decompose the higher-order IMF components until the residuals cannot be further decomposed. The original capacity can be expressed as:

[0022]

[0023] Where K is the total number of IMF components, R k (t) represents the final residual.

[0024] Preferably, the zero-crossing rate is used to distinguish between high-frequency and low-frequency components. Specifically, the zero-crossing rate of each IMF component is calculated. If the zero-crossing rate is greater than a preset threshold of 0.01, it is determined to be a high-frequency component; otherwise, it is a low-frequency component. The zero-crossing rate calculation rules are as follows:

[0025]

[0026] Where n zero The value represents the number of zero crossings, where N is the length of the battery capacity sequence and P is the zero-crossing rate.

[0027] Preferably, the ensemble model for high-frequency component prediction includes:

[0028] The MSICA mechanism is used to fuse multi-head self-attention and channel attention mechanisms, and local features at different time scales are extracted through multi-scale one-dimensional convolution (with kernel sizes of 1, 3, and 5).

[0029] F k =Conv1D k (X), k∈{1,3,5}

[0030]

[0031] Where F k ∈R B×T×C Let B represent the feature map at the k-th scale, where B, T, and C represent the batch size, time step, and number of channels, respectively.

[0032] And channel dependency modeling is performed using a multi-head attention module:

[0033] F attn =MutilHead(Q,K,V)=Concat[head1,…,head k W O

[0034] α=σ(MLP(AvgPool(F attn ))+MLP(MaxPool(F attn )))

[0035] The final output is the weighted features:

[0036] X′=X·α

[0037] Where σ represents the Sigmoid activation function.

[0038] Using the ModernTCN module, depthwise separable convolutions with large kernel functions are employed to expand the effective receptive field and capture dependencies over long time spans:

[0039] Z conv =DepthwiseConv(X′)+PointwiseConv(X′)

[0040] Z = Z conv +X′

[0041] in The output of the multi-scale channel attention mechanism module is used as the input to ModernTCN. This is the output obtained through residual connection.

[0042] Using the MoE model, a precise characterization of the degradation modes of high-frequency components is achieved through multiple expert networks and sparse gating networks:

[0043]

[0044] Where M is the number of experts, E i (z) represents the output of the i-th expert network, G(z) i This represents the weight vector of the gating network.

[0045] Preferably, the PC-BiGRU model incorporates a bi-exponential degradation model into the prediction process by designing a composite loss function, ensuring that the low-frequency prediction results conform to the physical laws of battery degradation, specifically including:

[0046] A BiGRU model is constructed, and information from both the forward and backward propagation of the capacity sequence is utilized to further enhance the modeling of low-frequency component trends:

[0047]

[0048] in and W represents the hidden state output of the BiGRU in the forward and backward directions, respectively. t and V t b represents the weight. t This represents the bias vector.

[0049] A double-exponential model is constructed, which is a classic physical model capable of accurately describing the capacity degradation process of lithium batteries. Its degradation model can be expressed as:

[0050]

[0051] Where a, b, c, and d are parameters, and τ k It is noise.

[0052] The joint loss function is defined as the weighted sum of the prediction error of the BiGRU model and the low-frequency capacity of the double exponential model:

[0053]

[0054]

[0055]

[0056] Preferably, the method further includes a transfer learning method based on adaptive tensor decomposition, which improves the robustness of high-frequency component prediction in small-sample scenarios by sharing the high-order feature subspaces of the source and target domains, specifically including:

[0057] The high-frequency components of the source and target domains are represented as tensors, and then decomposed into a sequence of low-order core tensors using Tensor-Train decomposition. The decomposed tensors... Represented as:

[0058]

[0059] Statistical features are extracted from high-frequency components in the target domain, and the optimal rank of tensor decomposition is predicted through a fully connected network.

[0060] Optimize the combined loss of reconstruction error, transfer consistency, and rank regularization term, and update the core tensor and rank prediction parameters. The optimization problem is formalized as:

[0061]

[0062] λ1 and λ2 are hyperparameters that adjust the intensity of feature alignment and rank regularization, respectively.

[0063] Preferably, the low-frequency prediction result is added to the high-frequency prediction result to obtain the final prediction result. Attached Figure Description

[0064] Figure 1 The flowchart below shows the battery life prediction method of the present invention.

[0065] Figure 2 This is a schematic diagram of the high-frequency prediction model structure of the present invention;

[0066] Figure 3 This is a schematic diagram of the low-frequency prediction model structure of the present invention; Detailed Implementation

[0067] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments to further understand the purpose, solution and effect of the present invention, but this should not be construed as a limitation on the scope of protection of the present invention.

[0068] Specific implementation method one: A lithium battery lifetime prediction method based on the EMD framework, implemented through the following steps:

[0069] S1. Data Acquisition and Preprocessing:

[0070] To obtain the capacity sequence data of lithium-ion batteries, and to eliminate dimensional differences, the capacity data is normalized and mapped to the [0,1] interval. The formula is as follows:

[0071]

[0072] where X norm This is the normalized capacity data, where X is the original capacity data. max and X min These are the maximum and minimum values ​​of the original capacity data, respectively.

[0073] S2. Decompose the capacity sequence using ICEEMDAN:

[0074] The normalized capacity sequence was decomposed using ICEEMDAN to obtain multiple intrinsic mode function (IMF) components and one residual component. The specific process is as follows:

[0075] Add adaptive Gaussian white noise to the original capacity sequence C(t) to generate a perturbed sequence:

[0076] C (i) (t)=C(t)+β0·E1(ξ (i) (t)), i = 1, 2, ..., K

[0077] Where β0 is the initial noise intensity, ξ (i) (t) represents the Gaussian white noise generated independently for the i-th time.

[0078] By integrating the local means of the averaged perturbation sequences, the first-order residuals are calculated and the first-order IMF components are extracted:

[0079] IMF1(t) = C(t) - R1(t)

[0080] Iteratively decompose the higher-order IMF components until the residuals cannot be further decomposed. The original capacity can be expressed as:

[0081]

[0082] Where K is the total number of IMF components, R k (t) represents the final residual.

[0083] S3, Frequency Allocation:

[0084] The IMF components are divided into high-frequency and low-frequency components based on the zero-crossing rate. The zero-crossing rate is defined as the number of alternating positive and negative signs per unit length of signal, and is calculated using the following formula:

[0085]

[0086] If the zero-crossing rate P > 0.01, it is determined to be a high-frequency component; otherwise, it is a low-frequency component.

[0087] S4. High-frequency component prediction:

[0088] For high-frequency components, an ensemble model integrating Multi-Scale Channel Interactive Attention (MSICA), Modernized Temporal Convolutional Network (ModernTCN), and Hybrid Expert (MoE) is used for prediction. The specific process is as follows:

[0089] S41. Construct the MSCIA module, extract local features at different time scales through multi-scale one-dimensional convolution (kernel sizes of 1, 3, and 5), and fuse them into a unified tensor:

[0090] F k =Conv1D k (X), k∈{1,3,5}

[0091] S42. Subsequently, inter-channel dependencies are modeled using a multi-head attention module to generate attention weights:

[0092] Fattn =MutilHead(Q,K,V)=Concat[head1,…,head k W O

[0093] α=σ(MLP(AvgPool(F attn ))+MLP(MaxPool(F attn )))

[0094] S43, The final output is:

[0095] X′=X·α

[0096] S44. Construct the ModernTCN module, which uses depthwise separable convolution (DWConv) with a large kernel function to expand the receptive field, capture long-term dependencies, and enhances multivariate collaborative modeling capabilities through the decoupled ConvFFN module:

[0097] Z conv =DepthwiseConv(X′)+PointwiseConv(X′)

[0098] Z = Z conv +X′

[0099] S45. Construct the MoE module, perform predictions using multiple expert networks and sparse gating networks, and output:

[0100]

[0101] S5, Low-frequency component prediction:

[0102] For low-frequency components, a PC-BiGRU model incorporating a double exponential degradation model is used for prediction.

[0103] S51 and BiGRU capture long-term trends in time series through forward and backward GRU layers:

[0104]

[0105] S52. Constructing a double-exponential model:

[0106]

[0107] S53. Define the joint loss function as the weighted sum of the prediction error of the BiGRU model and the low-frequency capacity of the double exponential model:

[0108]

[0109]

[0110]

[0111] in The loss function for BiGRU is... Let λ be the physical constraint loss function, and λ be the tradeoff coefficient.

[0112] S6. The high-frequency prediction part is designed and incorporates transfer learning based on adaptive tensor decomposition:

[0113] S61. Represent the high-frequency components of the source and target domains as tensors, and decompose them into a low-order core tensor sequence through Tensor Train decomposition.

[0114] S62. Extract statistical features from high-frequency components in the target domain and predict the decomposition rank using a fully connected network;

[0115] S63, Optimize overall loss:

[0116]

[0117] S7. Add the high-frequency prediction results to the low-frequency prediction results to obtain the final prediction result. The above description is merely an exemplary embodiment of the present invention and is not intended to limit the scope of protection of the present invention, which is determined by the appended claims.

Claims

1. A lithium battery lifetime prediction method based on the EMD framework, characterized in that, This includes the following steps: S1. Obtain the battery capacity degradation sequence as data and perform normalization processing; S2. The data is decomposed using an improved complete set empirical mode decomposition and adaptive noise algorithm to obtain multiple IMF components and a residual component. S3. Calculate the zero-crossing rate of the multiple components obtained from the decomposition. If the zero-crossing rate is greater than 0.01, it is classified as a high-frequency component; otherwise, it is classified as a low-frequency component. S4. Construct an integrated model of multi-scale channel interactive attention mechanism (MSICA), modern temporal convolutional network (ModernTCN) and hybrid expert (MoE) for prediction, and design a transfer learning method based on adaptive tensor decomposition to improve the robustness of high-frequency component prediction in small sample scenarios by sharing the high-order feature subspaces of the source and target domains. S5. Construct a bidirectional gated cyclic unit model (PC-BiGRU) with a double exponential degradation model as a physical constraint to make predictions and obtain the predicted values ​​of low-frequency components. S6. Add the obtained high-frequency prediction results and low-frequency prediction results to obtain the final prediction result.

2. The method for predicting the lifespan of a lithium battery as described in claim 1, characterized in that, Step S1 includes: S21. Add adaptive Gaussian white noise to the original capacity sequence to generate a perturbation sequence: C (i) (t)=C(t)+β0·E1(ξ (i) (t),i=1,2,…,K Where β0 is the initial noise intensity, ξ (i) (t) represents the Gaussian white noise generated independently for the i-th time. S22. Calculate the first-order residual by integrating the local mean of the average perturbation sequence, and extract the first-order IMF component: IMF1(t) = C(t) - R1(t) S23. Iteratively decompose the higher-order IMF components until the residuals cannot be further decomposed. The original capacity can be expressed as: Where K is the total number of IMF components, R k (t) represents the final residual.

3. The lithium battery life prediction method according to claim 1, characterized in that, The zero-crossing rate described in S3 is used to distinguish between high-frequency and low-frequency components. By calculating the zero-crossing rate of each IMF component, if the zero-crossing rate is greater than 0.01, it is determined to be a high-frequency component; otherwise, it is a low-frequency component. The zero-crossing rate calculation rules are as follows: Where n zero The value represents the number of zero crossings, where N is the length of the battery capacity sequence and P is the zero-crossing rate.

4. The lithium battery life prediction method according to claim 1, characterized in that, S3 includes: S31. Design the MSCIA mechanism, which integrates multi-head self-attention and channel attention mechanisms. It extracts local features at different time scales through multi-scale one-dimensional convolutions with kernel sizes of 1, 3, and 5. F k =Conv1D k (X),k∈{1,3,5} Where F k ∈R B×T×C Let B represent the feature map at the k-th scale, where B, T, and C represent the batch size, time step, and number of channels, respectively. S32, and channel dependency modeling is performed using a multi-head attention module: F attn =MutilHead(Q,K,V)=Concat[head1,…,head k ]W O α=σ(MLP(AvgPool(F attn ))+MLP(MaxPool(F attn ))) S33. The final output is the weighted feature: X'=X·α Where σ represents the Sigmoid activation function. S34. To further enhance the modeling capability for high-frequency features, a ModernTCN module is introduced after the multi-scale interactive attention module to further process the extracted features. The specific algorithm is as follows: Z conv =DepthwiseConv(X′)+PointwiseConv(X′) Z=Z conv +X' in The output of the multi-scale channel attention mechanism module is used as the input to ModernTCN. This is the output obtained through residual connection. S35. A hybrid expert (MoE) model is introduced to predict the processed features, improving the representation ability of high-frequency components. The MoE consists of multiple expert networks and a gated network, where each expert network is a fully connected layer. The network computation process is as follows: Where M is the number of experts, E i (z) represents the output of the i-th expert network, G(z) i This represents the weight vector of the gating network. S36. Design a transfer learning approach based on adaptive tensor decomposition, representing the high-frequency components of the source and target domains as tensors, and then decomposing them into a sequence of low-order core tensors using TensorTrain. The decomposed tensors... Represented as: S37. Extract statistical features from the high-frequency components of the target domain and predict the optimal rank of the tensor decomposition using a fully connected network; S38. Optimize the combined loss of reconstruction error, transfer consistency, and rank regularization term, and update the core tensor and rank prediction parameters. The optimization problem is formalized as: λ1 and λ2 are hyperparameters that adjust the intensity of feature alignment and rank regularization, respectively.

5. The lithium battery life prediction method according to claim 1, characterized in that, The PC-BiGRU model in S5 incorporates a double-exponential degradation model into the prediction process by designing a composite loss function, thereby achieving physical constraints on the degradation trend of low-frequency components. Step S5 is as follows: S51. Construct a BiGRU model. BiGRU is essentially a two-layer GRU network, consisting of forward and backward GRU layers. Can this network simultaneously utilize information from both the forward and backward directions to further enhance the modeling of low-frequency component trends? in and W represents the hidden state output of the BiGRU in the forward and backward directions, respectively. t and V t b represents the weight. t This represents the bias vector. S52. Construct a double-exponential model. The double-exponential model is a classic physical model that can accurately describe the capacity degradation process of lithium batteries. Its degradation model can be expressed as: Where a, b, c, and d are parameters, and τ k It is noise. S53. Define the joint loss function as the weighted sum of the prediction error of the BiGRU model and the low-frequency capacity of the double exponential model: Where λ is the tradeoff coefficient, used to balance the weights of the two losses.

6. The lithium battery life prediction method according to claim 1, characterized in that, S6 describes adding the high-frequency prediction results and the low-frequency prediction results to obtain the final prediction result.

Images