Lithium battery life prediction method

Abstract

The invention relates to the technical field of lithium battery prediction, in particular to a lithium battery life prediction method, which comprises the following steps of: automatically adjusting a sampling range of a convolution kernel and a critical stage of focusing capacity from linear attenuation to nonlinear abrupt drop by a model according to battery degradation characteristics through dynamic adaptive Bayesian convolution, capturing capacity inflection points in real time, and predicting the service life of a lithium battery. The triple sparse attention mechanisms cooperate to compress attention and reduce global calculation complexity, attention dynamic weighting key mutation nodes are selected, sliding window attention strengthens local fluctuation features, and high-sensitivity detection of sudden drop of electric quantity is ensured. According to the dynamic feature pruning, redundant interference is eliminated through multi-scale attention weights, voltage / capacity curve features strongly associated with capacity attenuation are reserved, multiple types of batteries are adapted, failure early warning can still be stably triggered under complex working conditions, equipment shutdown or potential safety hazards caused by sudden reduction of electric quantity are effectively prevented, and the reliability of the system is improved. And high-reliability guarantee is provided for battery health management.

Description

Technical Field

[0001] This invention relates to the field of lithium battery prediction technology, and in particular to a method for predicting lithium battery life. Background Technology

[0002] The background technology for lithium battery prediction involves multiple aspects, including battery state of charge estimation, remaining life (RUL) prediction, and battery performance monitoring. Lithium-ion batteries are widely used in many fields due to their high energy density and long cycle life. However, the SOC (State of Charge) of lithium-ion batteries is affected by various factors, making accurate estimation difficult. Therefore, accurate SOC estimation plays a crucial role in preventing overcharging and over-discharging and extending battery cycle life.

[0003] Furthermore, battery performance deteriorates as capacity decreases and impedance increases, making it prone to problems such as internal short circuits and thermal runaway. Accurately estimating battery health (SOH) and predicting remaining life (RUL) to improve battery reliability is therefore crucial. The performance degradation of lithium-ion batteries severely impacts the stable operation of electric vehicles, large-scale energy storage, and other devices. Monitoring the performance degradation process of lithium-ion batteries and achieving accurate prediction of cycle life has become a current research hotspot. Real-world application scenarios are complex, and usage strategies and environmental factors are highly random and variable. How do these factors affect the performance degradation of lithium-ion batteries, and can a simple method be used to simplify these complex, randomly generated influences?

[0004] Lithium-ion battery energy storage systems are widely used in deep peak shaving applications after new energy sources are integrated into the power system due to their high energy density and long cycle life. However, because lithium-ion batteries pose certain safety risks at the end of their lifespan, timely and reliable lifespan prediction methods are particularly important for early warning. Summary of the Invention

[0005] To achieve the above objectives, this invention proposes a method for predicting the lifespan of lithium batteries, comprising the following steps:

[0006] Step 1: Data feature construction. Collect the charging capacity Q and charging voltage V during the lithium battery charging process as input features, and use the actual capacitance of the corresponding charging cycle as the output label to construct the QV capacity mapping relationship.

[0007] Step 2: Data preprocessing, including outlier cleaning of the dataset, including deleting missing values ​​and infinity values, and using the interquartile range method to correct outlier extreme values;

[0008] Step 3: Data standardization and completion. The capacity characteristics of different battery models are independently normalized. After normalization, the missing time series data are completed by linear interpolation.

[0009] Step 4: Temporal feature extraction. Perform sliding window processing on the QV sequence of each charging segment, use the current capacity as the prediction label, and generate three-dimensional input data.

[0010] Step 5: Model building and training. Build a prediction model based on the U-net architecture, including an encoder, decoder, and skip connection module. The encoder uses a Bayesian regularized one-dimensional adaptive convolution module, the decoder restores resolution through deconvolution, and the skip connections fuse features through the Attention-Cut module. The model is trained using a hybrid loss function.

[0011] In one example, the outlier correction specifically involves calculating Q1 and Q3 of the dataset, defining the outlier range as LowerBound = Q1 - 1.5IQR and UpperBound = Q3 + 1.5IQR, and replacing outliers outside the range with the median of adjacent valid data.

[0012] In one example, the sliding window processing is specifically implemented by setting the sliding window length N and the step size to 1, and converting the QV sequence into a three-dimensional data structure of (B,2,N), where B is the number of samples, 2 is the Q / V dual channel, and N is the time series length.

[0013] In one example, the encoder's Bayesian convolution module includes a dynamic adaptive unit that automatically adjusts the number and spatial location of convolution kernel sampling points according to the input distribution, and a Bayesian regularization layer that imposes probability distribution constraints on the network weights.

[0014] In one example, the deconvolution operation of the decoder is implemented using transposed convolution, the deconvolution kernel parameters are initialized by bilinear interpolation, and the upsampling rate is strictly matched with the encoder downsampling rate layer by layer.

[0015] In one example, the Attention-Cut module includes a multimodal attention unit and a feature pruning fusion unit. The multimodal attention unit simultaneously performs compressed attention, selective attention, and sliding window attention operations. The feature pruning fusion unit aligns the feature dimensions through adaptive average pooling and linear interpolation, and then performs a Hadamard product operation with the decoder features to achieve dynamic and adaptive cross-level feature modulation.

[0016] In one example, the compressed attention pairs perform block compression on the time-series data and compute a global compressed representation.

[0017] In one example, the selected attention is guided by the global importance score generated by compressed attention, dynamically filtering key time blocks and performing fine-grained calculations only on high-weight blocks.

[0018] In one example, the sliding window attention captures local window dependencies and dynamically adjusts the receptive field size.

[0019] The lithium battery life prediction method proposed in this invention can bring the following beneficial effects:

[0020] 1. This invention employs a phased sparse attention gating mechanism (compression, selection, and sliding window attention) to accurately capture key abrupt changes in battery capacity degradation within a time series. Compressed attention reduces global computational complexity through a block aggregation strategy while preserving the long-term degradation trend of the capacity curve. Selection attention, based on the dynamic saliency score of compressed attention, focuses on the nonlinear inflection point where capacity decreases from the rated value to around 80%, suppressing redundant interference in the stable interval. Selection attention perceives details of capacity fluctuations, avoiding the loss of fine-grained features. Combined with the dynamic sampling mechanism of one-dimensional adaptive Bayesian convolution, the model can adaptively adjust the spatiotemporal coverage of the convolution kernel according to the distribution characteristics of the input data, focusing on enhancing the feature extraction capability for the critical capacity degradation stage (e.g., 75%–80%). This combination of techniques effectively solves the problem of inflection point misjudgment caused by fixed convolution kernels or global attention redundancy in traditional methods, significantly improving the judgment accuracy when the capacity approaches the failure threshold, and ensuring reliable early warning is triggered when the actual battery capacity approaches 80%.

[0021] 2. This invention employs Bayesian regularization and dynamic feature pruning techniques to significantly improve the model's adaptability to noisy data, cross-model differences, and partial data loss. Bayesian convolution constrains network parameters through probabilistic weight distribution, suppressing interference from sensor noise or abnormal fluctuations on capacity prediction and avoiding false failures due to local data distortion. The dynamic feature pruning module adaptively prunes cross-level features through multi-scale attention weights, removing redundant information while retaining key temporal patterns strongly correlated with capacity decay. Furthermore, the data preprocessing stage uses outlier correction based on interquartile range and Lagrange interpolation to complete the data, which can repair temporal breaks caused by data acquisition interruptions or equipment failures, ensuring the integrity and smoothness of the input data. This technical design allows the model to stably output highly reliable capacity prediction results under complex operating conditions without relying on specific battery models or idealized data conditions, providing a unified and robust solution for battery health management in multiple scenarios. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0023] Figure 1 This is a schematic diagram of the overall model architecture of a lithium battery life prediction method;

[0024] Figure 2 This is a schematic diagram of the BayeARconv1d module architecture of a lithium battery lifetime prediction method.

[0025] Figure 3 This is a schematic diagram of the MultiAttention module architecture of a lithium battery life prediction method.

[0026] Figure 4 This is a schematic diagram of the Cut module architecture for a lithium battery life prediction method. Detailed Implementation

[0027] To more clearly illustrate the overall concept of the present invention, a detailed description will be provided below with reference to the accompanying drawings and examples.

[0028] In the description of this invention, it should be understood that the terms "center," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0029] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0030] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a communication connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0031] In this invention, unless otherwise expressly specified and limited, the first feature "on" or "below" the second feature may be in direct contact with the first and second features, or indirect contact through an intermediate medium. In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0032] like Figures 1 to 4 As shown, this invention proposes a method for predicting the lifespan of lithium batteries, comprising the following steps:

[0033] Step 1: Measure the experimental data, using the voltage V and capacitance Q of the charging section as the data set input;

[0034] Step 2: Using 80% of the data as the training set and 20% as the test set, when handling outliers, especially extreme outliers, the Interquartile Range (IQR) method is used to identify and correct them. This method is a robust statistic that does not depend on the data distribution and is particularly suitable for non-normally distributed datasets. First, sort the data from smallest to largest, then divide it into four equal parts, each containing 25% of the data. The first quartile (Q1) is the value at the 25th percentile, meaning 25% of the values ​​in the entire dataset are below this number. The second quartile (Q2) is the value at the 50th percentile, i.e., the median, meaning 50% of the values ​​in the entire dataset are below this number. The third quartile (Q3) is the value at the 75th percentile, meaning 75% of the values ​​in the entire dataset are below this number. The IQR is the difference between the third quartile (Q3) and the first quartile (Q1), i.e., IQR = Q3 - Q1.

[0035] The upper and lower ranges of reasonable data are:

[0036] LowerBound = Q1 - 1.5IQR (2)

[0037] UpperBound = Q3 + 1.5IQR (3)

[0038] Step 3: For each column of data, x = (x1 x2 x3 … x n The normalization process is performed using min-max normalization, and the formula is as follows:

[0039]

[0040] The purpose of normalization is to unify the dimensions of various features, preventing different dimensions from having an unbalanced impact on the weight updates of deep neural networks. Normalization ensures that each feature is on the same scale, thereby improving the training efficiency and convergence speed of the model.

[0041] In this patent, to improve data quality, interpolation was performed on the processed data using the Lagrange interpolation method. The core of this method lies in constructing a polynomial to smooth the experimental data and effectively eliminate noise. Compared to linear interpolation, Lagrange interpolation demonstrates superior adaptability when dealing with a large number of data points and more complex data structures. Battery datasets, with their numerous sample points, perfectly align with the application scenario of Lagrange interpolation.

[0042] The essence of Lagrange interpolation lies in fitting a set of data point samples to obtain a polynomial P. Taking a battery dataset as an example, we take the time step t as the independent variable and the charging feature that needs to be interpolated as the dependent variable y. i The following formula can be obtained:

[0043]

[0044] P(t) is the expression obtained from n given sample points (t0, y0), (t1, y1), (t2, y2)...(t... n ,y n The resulting Lagrange polynomial.

[0045]

[0046] L i (t) is a Lagrange polynomial. Due to the existence of this term, when t = t i At this time, it means that L i (t)=y i This means that the value of a point is determined only by that point, and is less affected by surrounding points, resulting in higher accuracy.

[0047] Step 4: Apply sliding window processing to the dataset after the above processing, extract the QV value of the charging segment as a feature, and then calculate the actual battery capacity Y at this point. i As a tag, as follows:

[0048]

[0049] Step 5: Construct the encoder and decoder; specific module implementation:

[0050] BayeARconv1d Module Design

[0051] The Bayesian Convolution method is employed, replacing the deterministic parameters in traditional convolution with probability weights, and calculating the posterior distribution p(f|X,Y) of a given dataset based on Bayes' rule. For a new input x... * The prediction of its output value can be achieved by integrating over all possible functions f, as shown in the following formula:

[0052] q(y * |x * ,X,Y)=∫p(y * |f * )p(f * |x * ,X,Y)df * (8)

[0053] However, in the above formula, q(y) * |x * Solving the problem in practice presents challenges because it requires enumerating an infinite number of possible f(x,y). * Integrating the integral makes it difficult to calculate directly. To address this issue, Bayesian convolution employs variational inference, fixing f... * To simplify the calculation process, we can use the distribution form of f. * The distribution follows a normal distribution. The entire problem can be transformed into how to make the assumed distribution p(w) closer to the true distribution. The loss function that needs to be optimized in this process is:

[0054]

[0055] KL divergence: used to describe the degree of approximation between two distributions. This is the approximate posterior distribution of parameter w given a dataset D (usually obtained through variational inference, where θ is the variational parameter). p(w): This is the prior distribution of parameter w, representing the hypothesis before the observed data. The parameter optimization process is as follows.

[0056]

[0057] This is the log-likelihood function. The expectation under the approximate posterior distribution q(w|θ) measures how well the model fits the data (the larger the expectation, the better the fit).

[0058] After obtaining the loss function, backpropagation is required. To make backpropagation faster and more convenient, this paper adopts local reparameterization. Since the convolution kernel parameter W is located in f... * The distribution follows a normal distribution. Therefore, the output value Y after linear operation by the convolution kernel j Alternatively, the parameters μ and σ of the normal distribution can be used. 2 To represent, as follows:

[0059]

[0060] Then the backpropagation is from the original W to μ,σ 2 The differentiation process is converted to Y j For μ,σ 2 During the differentiation process, since Y j The number of elements is much smaller than W, thus reducing a large amount of computation. The Bayesian method is used to make BayeARconv1d effectively resist overfitting.

[0061] Implementing BayeARconv1d adaptive convolution:

[0062] 1. Dynamically predict the length of the convolution kernel for each spatial location using a regular convolution kernel. Input feature map Predict the length feature map l=f using subnetworks θ (X): where l is between (0,1) by the Sigmoid activation constraint, and l represents the expected relative sampling length at each position in the input graph. The value is mapped to the actual physical range through a linear transformation: L = a1·l + b1, which is transformed into the actual sampling length required at each position, where a1 and b1 are adjustable parameters.

[0063] 2. Adaptively determine the number of sampling points K_L based on the size of the convolution kernel, and globally average the output feature map L: L mean =mean(L). Then, using the modulation coefficient n and the parity correction function φ:

[0064]

[0065] The obtained K_L is the globally optimal odd number of sampling points.

[0066] 3. Construct an adaptive sampling grid. Since the sampling grid of a typical one-dimensional convolution is G: [1,2,3,…,N], the index of the position needs to be offset. Therefore, the offset factor is calculated for the position:

[0067]

[0068] k l L is the scaling factor. i The elements of the feature map L obtained above are adjusted using the Hadamard product to create a standard grid, and then the adaptive offset matrix R is calculated:

[0069] R = Z⊙G,

[0070] 4. Interpolation and Convolution Calculation: The coordinate values ​​in the sampling grid R obtained in step 3 are still decimals. Therefore, bilinear interpolation is used to transform the coordinates of the sampling grid to integers. Finally, based on the spatiotemporal information reflected in the obtained feature map L, K_L sampling center points are taken on the sampling map R (each sampling point contains its own offset matrix). The resulting matrix is ​​H, which contains the coordinate index of each sampling point on the input feature map. The sampling map is then reconstructed based on the coordinate index, and a Bayesian convolution kernel of size K_L is selected to convolve the reconstructed feature map.

[0071] Therefore, the above operations demonstrate how BayeARconv1d performs accurate and efficient adaptive convolution.

[0072] MultiAttention module design

[0073] MuliAttention utilizes gating units to fuse a triple attention mechanism for output. The three attention mechanisms are: global compressed attention (low-complexity coarse-grained modeling), selective attention (focusing on important blocks, balancing efficiency and accuracy), and local window attention (capturing proximity relationships).

[0074] First, the input sequence x is transformed linearly to generate Query(Q), Key(K), and Value(V), where W... Q W K W V This is the learnable parameter matrix.

[0075] Global compressed attention: K is processed in blocks, K blocks =reshape(K,[B,T / C,C·d) model ]), using MLP to process the compressed and segmented K blocks The following is its compression formula:

[0076]

[0077] in

[0078] The goal is to compute a global representation of the segmented time series without having to calculate the correlation at each time step individually, thus significantly reducing computational complexity. This representation is then multiplied by Q to generate the importance of each time step to the overall time series, which is the compressed attention score attn. comp The formula is as follows:

[0079]

[0080] Selective attention aims to reduce computational complexity while preserving key information by dynamically sparsifying and computing only the subset of key-value pairs most relevant to the current query. This is achieved through coarse-grained attention (attn). comp The selection process guides the choice and avoids randomness, resulting in the selected block selectedK.

[0081] Sliding window attention: Preserves the local window (size: win_size) at the end of the sequence, captures local dependencies, and selects the window.

[0082]

[0083] They obtained compressedK, selectedK, and window respectively. K After dividing the data into three blocks, the attention calculation begins, using the following formula:

[0084]

[0085] Finally, the three attention branches are dynamically weighted and fused using a gating network to produce the output. The parameters of the gating unit are determined by the global matrix Q. The formula is as follows:

[0086]

[0087] Attention works through the following steps:

[0088] Input data preprocessing:

[0089] The input data is first transformed in dimension, converting the original convolution output (shape ) into a format suitable for subsequent attention calculations.

[0090] Multimodal attention unit:

[0091] The MultiAttention module processes temporal features in parallel using three attention mechanisms.

[0092] Compressing attention:

[0093] The input sequence is divided into blocks along the time dimension. Each block is compressed into a representation vector using a multilayer perceptron (MLP). This block-based strategy reduces the computational complexity from 0 to 1. By calculating the similarity between each time step and the block, a matrix A is obtained, representing the importance of each time step to the block.

[0094] Select attention:

[0095] By using the importance matrix obtained from compressed attention, the top n most important blocks are selected and concatenated to form the Key and Value for attention selection. This process reduces computation through dynamic sparsity while preserving key information.

[0096] Sliding window attention:

[0097] By preserving the local window at the end of the sequence and capturing local dependencies, we can better capture the local patterns of the time series.

[0098] Gated network convergence:

[0099] The outputs of the multimodal attention units are fused through a gating network. The gating network dynamically weights the outputs of the three attention branches according to the global matrix, ultimately generating the fused attention result.

[0100] (3) Cut module

[0101] The Cut module integrates semantic information from different levels through scale-adaptive transformation and feature interaction. First, there's the adaptive transformation; assuming the feature map of the i-th layer is: Through scale alignment transformation:

[0102]

[0103] The goal is to ensure that the feature map size of other layers is the same as the feature map size of this layer before performing feature mapping on this layer. Figure 1 This process involves performing a Hadamard product with other layers (including the current layer) and then adding the results together. This integrates information from all layers to output richer semantic information. The calculation formula is as follows:

[0104]

[0105] Belongs to the process The processed feature maps of each layer.

[0106] In the data preprocessing stage, this method improves data quality through operations such as finely partitioning the dataset, handling outliers and missing values, and normalization, providing a reliable foundation for model training. The application of the sliding window technique constructs an effective feature set, enabling the model to better learn the battery capacity variation patterns. In terms of model architecture, the ByesARCnn-Unet model based on the U-net architecture innovatively combines a one-dimensional adaptive Bayesian convolution module and an Attention-Cut fusion module. The former dynamically optimizes feature extraction through an adaptive mechanism and a Bayesian framework, avoiding overconfidence learning; the latter utilizes a multimodal attention mechanism and dynamic feature pruning to efficiently fuse temporal features, enabling the model to accurately capture the spatiotemporal characteristics of battery data. In addition, the introduction of the Bayesian framework enhances the model's stability and generalization ability, allowing it to make robust predictions even under high uncertainty. These innovative designs and optimization strategies enable this method to achieve a balance between efficiency and accuracy in lithium battery life prediction tasks, significantly improving prediction accuracy and generalization performance.

[0107] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0108] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for predicting the lifespan of a lithium battery, characterized in that: Includes the following steps: Step 1: Data feature construction. Collect the charging capacity Q and charging voltage V during the lithium battery charging process as input features, and use the actual capacitance of the corresponding charging cycle as the output label to construct the QV capacity mapping relationship. Step 2: Data preprocessing, including outlier cleaning of the dataset, including deleting missing values ​​and infinity values, and using the interquartile range method to correct outlier extreme values; Step 3: Data standardization and completion. The capacity characteristics of different battery models are independently normalized. After normalization, the missing time series data are completed by linear interpolation. Step 4: Temporal feature extraction. Perform sliding window processing on the QV sequence of each charging segment, use the current capacity as the prediction label, and generate three-dimensional input data. Step 5: Model building and training. Build a prediction model based on the U-net architecture, including an encoder, decoder, and skip connection module. The encoder uses a Bayesian regularized one-dimensional adaptive convolution module, the decoder restores resolution through deconvolution, and the skip connections fuse features through the Attention-Cut module. The model is trained using a hybrid loss function.

2. The lithium battery life prediction method according to claim 1, characterized in that... The outlier correction specifically involves calculating Q1 and Q3 of the dataset, defining the outlier range as LowerBound = Q1 - 1.5IQR and UpperBound = Q3 + 1.5IQR, and replacing extreme values ​​outside the range with the median of adjacent valid data.

3. The lithium battery life prediction method according to claim 1, characterized in that: The specific method of the sliding window processing is as follows: set the sliding window length N, the step size is 1, and convert the QV sequence into a three-dimensional data structure of (B,2,N), where B is the number of samples, 2 is the Q / V dual channel, and N is the time sequence length.

4. The lithium battery life prediction method according to claim 1, characterized in that: The encoder's Bayesian convolution module includes a dynamic adaptive unit that automatically adjusts the number and spatial position of convolution kernel sampling points according to the input distribution, and a Bayesian regularization layer that applies probability distribution constraints to the network weights.

5. The lithium battery life prediction method according to claim 1, characterized in that: The deconvolution operation of the decoder is implemented using transposed convolution, and the deconvolution kernel parameters are initialized through bilinear interpolation. The upsampling rate and the downsampling rate of the encoder are strictly matched layer by layer.

6. The lithium battery life prediction method according to claim 1, characterized in that: The Attention-Cut module includes a multimodal attention unit and a feature pruning fusion unit. The multimodal attention unit simultaneously performs compressed attention, selective attention, and sliding window attention operations. The feature pruning fusion unit aligns the feature dimensions through adaptive average pooling and linear interpolation, and then performs a Hadamard product operation with the decoder features to achieve dynamic and adaptive cross-level feature modulation.

7. The lithium battery life prediction method according to claim 6, characterized in that: The compressed attention pair performs block compression on the time-series data and calculates a global compressed representation.

8. The lithium battery life prediction method according to claim 6, characterized in that: The selected attention is guided by the global importance score generated by compressed attention, dynamically filtering key time blocks, and performing fine-grained calculations only on high-weight blocks.

9. The lithium battery life prediction method according to claim 6, characterized in that: The sliding window attention captures local window dependencies and dynamically adjusts the receptive field size.

Images