A Joint Estimation Method for SOH-RUL of Lithium-ion Batteries Based on the Dozerformer Model

Abstract

This invention belongs to the field of battery technology based on specific computational models, and discloses a joint estimation method for SOH-RUL of lithium-ion batteries based on the Dozerformer model, which can realize and effectively improve the performance of the Dozerformer model for joint SOH-RUL estimation. The method includes: First, using the Dozerformer model to learn the mapping relationship between battery aging characteristics and SOH and RUL, decomposing battery degradation data into periodic components and trend components, introducing a sparse attention mechanism composed of local sparsity, step size sparsity, and variation sparsity to learn the long-term and short-term features of the periodic components, and using a linear layer to learn the feature information of the trend component. The estimation results of the two are combined to obtain the joint SOH-RUL estimation result. Second, a smooth quadratic loss function is used to guide the Dozerformer model to reduce the model's sensitivity to outliers during training. Finally, based on the optimal weights and biases of the Dozerformer model, the SOH and RUL of the lithium-ion battery are estimated, and the SOH-RUL estimation result is output.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of battery technology based on specific computational models, and particularly to a joint estimation method for SOH-RUL of lithium-ion batteries based on the Dozerformer model. Background Technology

[0002] Currently, lithium-ion batteries are widely used in various fields, such as electric vehicles, energy storage devices, and mobile electronic devices, due to their advantages such as convenient charge-discharge cycles, high energy density, and long cycle life. However, as the number of charge-discharge cycles accumulates, the battery's lifespan decreases, affecting the reliability and safety of devices equipped with lithium-ion batteries and posing potential hazards to users. Battery State of Health (SOH) and Remaining Useful Life (RUL) are two important battery parameters that describe the degree of battery aging from two different perspectives. Therefore, accurately estimating the battery's SOH and RUL is crucial for improving overall battery performance. It is generally recommended that users replace the battery when the SOH value drops to 80% or the RUL value approaches 0. The formulas for calculating battery SOH and RUL are as follows: RUL t =rr(2) where SOHt is the SOH value at the tth charge-discharge cycle, Ct is the maximum capacity at the tth charge-discharge cycle, CN is the nominal capacity of the battery, RULt is the RUL value at the tth charge-discharge cycle, r is the total number of cycles when the battery reaches the end of its life, and rt is the number of t cycles used.

[0003] Many researchers have conducted extensive research on battery SOH and RUL estimation, and current estimation methods are mainly divided into two categories: model-based and data-driven methods. Model-based state estimation methods establish a battery physical model and use mathematical formulas to describe the operation and degradation characteristics of lithium batteries, such as equivalent circuits, electrochemical models, and empirical models. Data-driven estimation methods do not require establishing a battery physical model; instead, they train the model using observable battery external parameters (usually voltage, current, temperature, etc.) to map the nonlinear relationship between external parameters and SOH and RUL. Since machine learning models are relatively simple and cannot effectively fit the nonlinear relationships of lithium batteries from large amounts of data, deep learning has gradually gained attention in the field of battery SOH and RUL estimation. Existing researchers have designed a model with a multi-channel parallel structure of CNN, LSTM, and gated recurrent units (GRU) with attention mechanisms for SOH and RUL estimation, and validated the model's performance on different battery data.

[0004] CNNs, limited by the size of convolutional kernels, lack the ability to capture global sequence information, resulting in poor ability to capture long-term dependencies. While LSTM or its variants have stronger long-term dependency capabilities than CNNs, their sequential structure leads to information loss even in very long sequences. Transformer models, with their unique self-attention mechanism, can directly establish global dependencies, allowing each time step to interact with any other time step, easily handling long-term dependencies. A lithium-ion battery RUL estimation was implemented using a Transformer model, but SOH prediction was not achieved. Variational mode decomposition was used to reduce data noise, followed by feature extraction using a Transformer model to achieve SOH estimation, but the performance requirements for SOH estimation were not met. While the above Transformer-based estimation methods achieved good prediction results, they did not meet the requirement of simultaneously estimating battery SOH and RUL, and did not consider the periodicity and trends in battery degradation data. Due to the sensor's own measurement accuracy or other factors, outliers exist in the actual collected data. The models in the above methods are all trained using the mean squared error (MSE) loss function. The MSE gradient increases linearly with the prediction error, making the model very sensitive to outliers. Therefore, even a small number of outliers in the data may cause the overall performance of the model to drop sharply. Summary of the Invention

[0005] To overcome the aforementioned technical deficiencies, this invention provides a joint estimation method for SOH-RUL of lithium-ion batteries based on the Dozerformer model, the main technical solution of which is as follows:

[0006] S1. The Dozerformer model is used to learn the mapping relationship between battery aging features and SOH and RUL. The battery aging features are decomposed into periodic components and trend components. A sparse attention mechanism consisting of local sparsity, step size sparsity and change sparsity is introduced to learn the long-term and short-term features in the periodic components. A linear layer is used to learn the feature information of the trend components.

[0007] S2. During model training, the weights and biases of the Dozerformer model are optimized using a smooth quadratic loss function.

[0008] S3. Use the Dozerformer model with optimal weights and biases to estimate the SOH and RUL of the lithium-ion battery and output the SOH-RUL estimation results.

[0009] Furthermore, for an input sequence X, where L is the sequence length and N is the number of features, the main calculation process of the Dozerformer model is as follows:

[0010] S1.1 Decompose the input sequence X into trend components and periodic components;

[0011] S1.2 Calculate the prediction results on the trend components;

[0012] S1.3. Embed the periodic components into c feature maps with dimension invariance and divide them into p patches;

[0013] S1.4 Utilize the Transformer encoder to learn the sequence feature information in the periodic component, and introduce local sparsity and stride sparsity strategies in the self-attention mechanism to sparsify the attention weights, thereby obtaining the sparsified attention mechanism result. Subsequently, through the feedforward network layer and residual connection, the final output of the encoder is generated.

[0014] S1.5. Using the Transformer decoder, the information output by the encoder is decoded through a combination of self-attention and cross-attention mechanisms to finally obtain the decoder output.

[0015] S1.6. Use 1×1 convolution to obtain the prediction results on the periodic components;

[0016] S1.7. Add the prediction results on the trend component and the cycle component to obtain the final result.

[0017] The embodiments of the present invention, by employing the above-described technical solutions, can achieve the following technical effects:

[0018] 1. The Dozerformer model adopted introduces three attention components to improve the joint estimation performance of SOH-RUL, namely local attention to capture short-term dependencies, step-size attention to learn the periodicity of SOH and RUL, and dynamic expansion attention to review historical data points to capture the long-term trend of battery SOH and RUL.

[0019] 2. A smooth quadratic loss function is used to guide the optimization of the weights and biases of the Dozerformer model. Compared to the traditional MSE loss function, the smooth quadratic loss function can effectively reduce the model's sensitivity to outliers and effectively improve the Dozerformer model's performance in joint estimation of battery SOH-RUL. Attached Figure Description

[0020] The accompanying drawings exemplify embodiments and form part of the specification, serving together with the textual description to explain exemplary implementations of the embodiments. The illustrated embodiments are for illustrative purposes only and do not limit the scope of the claims. Throughout the drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0021] Figure 1This is a diagram illustrating the overall framework for the joint estimation of SOH-RUL in lithium-ion batteries in this invention.

[0022] Figure 2 This is a flowchart illustrating the joint estimation method of SOH-RUL for lithium-ion batteries based on the Dozerformer model in this invention.

[0023] Figure 3 The MSE and SQL loss function curves and their derivative curves are shown in this invention;

[0024] Figure 4 Characteristic CH10-CH15 and curves showing the changes in SOH and RUL with the number of cycles;

[0025] Figure 5 Estimated SOH and RUL curves for different models on the CH10-CH15 test set;

[0026] Figure 6 SOH and RUL estimation curves for different loss functions on the CH10-CH15 test set. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0028] It should be noted that the descriptions involving "first," "second," etc., in the embodiments of the present invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0029] In the description of this invention, it should be understood that the numerical labels before the steps do not indicate the order in which the steps are performed, but are only used to facilitate the description of this invention and to distinguish each step, and therefore should not be construed as a limitation of this invention.

[0030] To address the aforementioned issues, this invention investigates the use of the Dozerformer model for joint estimation of SOH and RUL in lithium-ion batteries. Furthermore, we employ a novel smoothed quadratic loss function to reduce the model's sensitivity to outliers and guide the adjustment of model weights.

[0031] The overall framework for joint estimation of SOH-RUL in lithium-ion batteries is as follows: Figure 1 As shown, the Dozerformer model is used to learn the mapping relationship between battery aging characteristics and SOH and RUL, and a smooth quadratic loss function is used to guide the optimization of the Dozerformer model's weights and biases. Finally, the battery SOH-RUL estimation results are visualized. The Dozerformer model and smooth quadratic loss function are described in detail below.

[0032] To facilitate understanding of the technical solutions provided in the embodiments of the present invention by those skilled in the art, the relevant technologies are described below:

[0033] See Figure 2 The SOH-RUL joint estimation method for lithium-ion batteries based on the Dozerformer model in this embodiment of the invention includes:

[0034] S1. The Dozerformer model is used to learn the mapping relationship between battery aging features and SOH and RUL. The battery aging features are decomposed into periodic components and trend components. A sparse attention mechanism consisting of local sparsity, step size sparsity and change sparsity is introduced to learn the long-term and short-term features in the periodic components. A linear layer is used to learn the feature information of the trend components.

[0035] For SOH and RUL estimation tasks, many prediction methods based on LSTM and its variants have been established. However, as the sequence length increases, LSTM and its variants may experience gradient explosion or vanishing, and computational efficiency decreases accordingly. In contrast, the Transformer model avoids traditional recursive computation by employing a self-attention mechanism to compute battery degradation features in parallel over the entire historical sequence, thus effectively capturing global information. However, the traditional Transformer model ignores the impact of periodicity and trend in time series data on the battery degradation process. Therefore, this paper proposes the Dozerformer model for joint SOH-RUL estimation of lithium batteries, which can fully utilize the periodicity and trend in the battery degradation process to improve estimation performance.

[0036] For an input sequence X, where L is the sequence length and N is the number of features, the main computation process of the Dozerformer model is as follows:

[0037] S1.1 Decompose the input sequence X into trend components and periodic components, and the calculation formula is as follows:

[0038]

[0039] Xs=X-Xt∈RL×N(5),

[0040] Where Xt is the trend component, Xs is the period component, Avgpool is the average pooling operation, and Padding is the padding operation.

[0041] S1.2 Calculate the prediction results on the trend component, and the calculation formula is as follows:

[0042]

[0043] Where T is the predicted sequence length. For a trainable weight matrix, is a trainable bias vector.

[0044] S1.3, the periodic component X s The dimensionality-invariant embedding is used to embed into c feature maps and divided into p patches. The calculation formula is as follows:

[0045]

[0046] Among them, Reshape changes the data dimension so that the data can be embedded into the feature map, Conv2D is a two-dimensional convolution operation with a 3×1 kernel, and Patch is a patching operation. Processing each channel independently allows the model to learn the features and patterns of each channel better, which can improve the model's expressive power

[14] . Here, the channel independence method is also used to input each variable into the subsequent model separately. The patch sequence obtained after applying channel independence to each feature is shown in the following formula:

[0047]

[0048] S1.4. The Transformer encoder is used to learn the sequence feature information in the periodic components. Local sparsity and stride sparsity strategies are introduced into the self-attention mechanism to sparsify the attention weights, thereby obtaining the sparsified attention mechanism result. Then, the final output of the encoder is generated through the feedforward network layer and residual connection.

[0049] The self-attention mechanism in a typical Transformer encoder is a multi-head self-attention mechanism, and its calculation process is as follows:

[0050]

[0051] Where H represents the number of long positions, Wq W k and W v Let b be a trainable weight matrix. q b k and b v Here, is the trainable bias vector. Traditional Transformer models employ a full attention mechanism, learning all information from the entire sequence. However, as sequence length increases, information redundancy leads to overfitting and a rapid decline in computational efficiency. Sparse attention mechanisms can reduce computation and lower information redundancy. In self-attention mechanisms, two different methods—local sparsity and stride sparsity—are used to sparsify the obtained attention weights.

[0052] Specifically, local sparsity refers to querying weights on a fixed window, thus ignoring the remainder. For a local window of size w, the local sparse mask matrix can be represented as:

[0053] Where ij represents the element in the i-th row and j-th column of the matrix.

[0054] Sparse stride means querying weights at a fixed stride, ignoring and setting other weights to zero. A sparse stride mask matrix with stride size 's' can be represented as:

[0055]

[0056] The sparse mask matrix obtained by merging the local sparse mask matrix and the stride sparse mask matrix is ​​shown in the following equation:

[0057]

[0058] Using equation (17), the sparsified attention mechanism can be obtained as follows:

[0059]

[0060]

[0061] Finally, the encoder output is obtained through the feedforward network layer and residual connection:

[0062]

[0063] Where ReLU is the activation function and LayerNorm is the layer normalization process.

[0064] S1.5. Using the Transformer decoder, the information output by the encoder is decoded through a combination of self-attention and cross-attention mechanisms to finally obtain the decoder output.

[0065] Unlike the encoder, the decoder employs not only self-attention but also cross-attention, where the key and value are derived from the encoder's output. The decoder's computation process is as follows:

[0066] The input to the decoder is processed using the same multi-head self-attention mechanism as the Transformer encoder (i.e., local sparsity and stride sparsity strategies are introduced into the self-attention mechanism to sparsify the attention weights). The output of the multi-head self-attention mechanism is obtained through the calculation process of equations (11)-(20). The cross-attention mechanism not only adopts local sparsity and stride sparsity, but also a variable sparsity method. The local sparsity matrix, stride sparsity matrix, and variable sparsity matrix in the cross-attention mechanism can be expressed as follows:

[0067]

[0068] Among them, len p To predict the length, len v Used to specify the length of a variable-length sequence.

[0069] The final sparse matrix obtained by merging the three sparse matrices for the cross-attention mechanism is:

[0070]

[0071] Then, the encoder's output is used as the key and value of the cross-attention mechanism, and the decoder's input is used as the query. After sparsification, the cross-attention mechanism result is obtained as follows:

[0072]

[0073] Crossattn dec The final output of the decoder can be obtained using equations (21)-(22). That is, the final prediction result.

[0074] S1.6. The prediction results for the periodic components are obtained using 1×1 convolution, and the calculation formula is as follows:

[0075]

[0076] S1.7. The prediction results on the trend component and the cycle component are added together to obtain the final result, Y:

[0077] S2. During model training, the weights and biases of the Dozerformer model are optimized using a smooth quadratic loss function.

[0078] Currently, mean squared error (MSE) and L2 loss functions are commonly used in time series prediction. However, these are highly sensitive to outliers during model training because the gradients of MSE and L2 increase with the prediction error. Furthermore, the gradient of the L1 loss function is constant, which, while insensitive to outliers, slows down model convergence. To reduce the model's sensitivity to outliers during training while maintaining a certain convergence speed, this paper employs the SQL function to dynamically adjust the gradient to guide model training. A rational quadratic function is introduced into the MES loss function to form the SQL loss function. The curves of MSE and the SQL loss function with different α values ​​at c=1, along with their derivative curves, are shown below. Figure 3 As shown. The SQL expression is as follows:

[0079] Where α∈[0,1] is the smoothing coefficient, c∈(0,+∞), and y a For the true value of the sequence, These are the predicted values ​​for the sequence.

[0080] Figure 3 The figures show the MSE and SQL loss function curves and their derivative curves in this invention. From... Figure 3 As can be seen from the function curve in part a, the SQL loss function is smoother than the MSE loss function, and its smoothness increases with the increase of the α value. Figure 3 Part b shows that the derivative of the SQL loss function does not increase linearly with increasing error, but tends to a relatively small value. Therefore, it does not have a significant cumulative error effect like the MSE loss function. Furthermore, as the value of α increases, the derivative of SQL approaches 0 more effectively, thus reducing the impact of outliers on the model.

[0081] S3. Use the Dozerformer model with optimal weights and biases to estimate the SOH and RUL of the lithium-ion battery and output the SOH-RUL estimation results.

[0082] The estimated data used are lithium-ion battery aging data, such as battery internal resistance, average temperature, charging capacity, maximum temperature, discharging capacity, minimum temperature, and charging time. The maximum and minimum temperatures are both charging temperatures. After standardizing the above aging characteristic data, it is input into a Dozerformer model with optimal weights and biases.

[0083] In this embodiment of the invention, the Dozerformer model employs three attention components to improve the joint estimation performance of SOH-RUL. Local attention captures short-term dependencies, step-size attention learns the periodicity of SOH and RUL, and dynamically expanded attention reviews historical data points to capture long-term trends in battery SOH and RUL. Simultaneously, a smooth quadratic loss function is used to guide the optimization of the Dozerformer model's weights and biases. Compared to the traditional MSE loss function, the smooth quadratic loss function effectively reduces the model's sensitivity to outliers, significantly improving the Dozerformer model's performance in joint battery SOH-RUL estimation.

[0084] Finally, to verify the technical effectiveness of the present invention, comparisons were made with other advanced Transformer-based time series models and ablation experiments on public datasets, verifying the effectiveness of the Dozerformer model and the smooth quadratic loss function in the joint estimation of battery SOH-RUL. Specifically:

[0085] This invention utilizes a dataset of 124 commercial LiFePO4 graphite battery degradation data collected from the Toyota Research Center at Stanford University, MIT. The basic parameters of the LiFePO4 graphite batteries used are shown in Table 1. To investigate the impact of different fast-charging conditions on lifespan, all batteries were charged to 80% at 72 current groups ranging from 3.6C to 6C, and then charged at the remaining constant current and constant voltage stages with the same current. Unlike the diversity of charging, each discharge current was 4C. This paper selects battery aging data from channels 10-15 of the third batch of data for testing, naming it CH10-CH15. Furthermore, we use the internal resistance, charging time, charging capacity, discharging capacity, minimum temperature, maximum temperature, and average temperature from the data as input features for model training. The curves showing the changes of these seven features, as well as SOH and RUL, with the number of cycles for CH10-CH15 are shown below. Figure 4 As shown above, the different features exhibit different orders of magnitude. To prevent features with larger numerical values ​​from weakening the influence of features with smaller numerical values ​​in the model, we perform standardization on the dataset. The standardization formula is as follows: Where μ is the data mean and σ is the data standard deviation.

[0086] Table 1: Basic Battery Parameters

[0087]

[0088] The model was built and tested using Python 3.10 and PyTorch 2.0 frameworks, with the Adam algorithm used as the adaptive optimizer. The specific parameter settings for the Dozerformer model are shown in Table 2. To evaluate the accuracy of the SOH and RUL estimates, we used three evaluation metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²). The expressions for these three metrics are shown below:

[0089]

[0090] Where T is the number of samples in the sequence, a t For the t-th measurement value, The t-th estimated value, This is the average value.

[0091] Table 2: Dozerformer Model Parameters

[0092]

[0093]

[0094] The performance of the proposed Dozerformer model in joint SOH-RUL estimation on CH10-CH15 was tested. The ratio of the model's training, validation, and test sets was 0.15:0.15:0.7 for each battery degradation dataset. To demonstrate the superiority of the Dozerformer model in joint SOH-RUL estimation, we compared it with state-of-the-art prediction models Transformer, Informer, Autoformer, and Fedformer. The parameters of these four comparison models are shown in Table 3. All models in this section were trained using the MSE loss function. The SOH and RUL estimation curves of the Dozerformer and comparison models on the CH10-CH15 test set are shown in Table 3. Figure 5 As shown in Table 4, the evaluation indexes for SOH estimation are shown in Table 5, and the evaluation indexes for RUL estimation are shown in Table 5.

[0095] Table 3: Comparison of Model Parameter Settings

[0096]

[0097] Table 4: Evaluation Indicators for SOH Estimation on CH10-CH15 by Different Models

[0098]

[0099]

[0100] Table 5: Evaluation Indicators of RUL Estimation for Different Models on CH10-CH15

[0101]

[0102]

[0103] from Figure 5 As can be seen from the SOH and RUL estimation results in Table 4-5, the Transformer and Informer models show a near-linear relationship and high evaluation metrics for both SOH and RUL estimations. This is mainly because the traditional multi-head self-attention mechanism in these models cannot effectively learn the trends in the battery degradation path. Compared to these two models, the Autoformer and Fedformer models, by introducing a sequence decomposition component, make the estimated curves closer to the true values. However, they exhibit significant deviations from the true values ​​and large fluctuations in predicted values ​​at the end of the battery life. The Dozerformer model used in this paper performs best in the joint SOH-RUL estimation. This model decomposes the data into trend and periodic components and then effectively learns the lithium battery degradation characteristics through three attention components, thereby significantly improving the performance of the joint SOH-RUL estimation.

[0104] Comparison of experimental results with different loss functions:

[0105] The impact of the SQL loss function on the joint SOH-RUL estimation of the Dozerformer model was tested on CH10-CH15. The ratio of training, validation, and test sets for each battery degradation dataset remained at 0.15:0.15:0.7. Furthermore, the proposed SQL loss function was compared with common L1 and MSE loss functions. It is worth noting that all parameters are consistent in this section except for the loss function; the parameter settings for the Dozerformer model are shown in Table 2. The SOH and RUL estimation curves of the three different loss functions on the CH10-CH15 test sets are shown below. Figure 6 As shown in Table 6, the evaluation indexes for SOH estimation are shown in Table 7, and the evaluation indexes for RUL estimation are shown in Table 7.

[0106] Table 6: Evaluation Indicators for SOH Estimation on CH10-CH15 under Different Loss Functions

[0107]

[0108]

[0109] Table 7: Evaluation metrics for RUL estimation on CH10-CH15 with different loss functions

[0110]

[0111] from Figure 6 As can be seen from the SOH and RUL estimation results in Tables 6-7: (1) In the SOH estimation results, the MSE loss function showed the worst estimation performance, mainly because the MSE is very sensitive to outliers, which leads to overfitting of the model training and the worst prediction performance. The average values ​​of MAE (%), RMSE (%) and R2 of the L1 loss function on the six cells were 0.2042%, 3.098% and 0.9925, respectively; the average values ​​of MAE (%), RMSE (%) and R2 of the SQL loss function on the six cells were 0.204%, 0.392% and 0.9927, respectively. Although the L1 loss function is better than the SQL loss function in terms of evaluation metrics on CH10, CH13 and CH15, its average evaluation metrics on all cells are worse than the SQL loss function. Therefore, the SQL loss function has the best overall performance. In summary, the SQL loss function can maintain a certain convergence speed while reducing the sensitivity to outliers during model training.

[0112] Conclusion: Accurate estimation of battery state of harm (SOH) and relative uptime (RUL) is crucial for improving the energy management capabilities of electric vehicles. This paper proposes a Dozerformer model with an SQL loss function for the joint estimation of SOH and RUL of lithium-ion batteries. To fully learn the overall trend and degradation patterns of the battery degradation process, the Dozerformer model decomposes the data into trend and periodic components and introduces a sparse attention mechanism consisting of a local sparse matrix, a step-size sparse matrix, and a variable sparse matrix to improve the accuracy of the SOH-RUL joint estimation. To reduce the model's sensitivity to outliers during training, an SQL function is used to dynamically adjust the gradient to guide the Dozerformer model training. This improves model performance and effectively enhances the accuracy of SOH-RUL estimation.

[0113] Simulation experiments were conducted on six different batteries using different models and loss functions. Compared to other advanced prediction models based on Transformer, the proposed Dozerformer model performs better in both SOH estimation and RUL estimation, thus enabling it to better learn the lithium battery degradation process. Furthermore, the SQL loss function used in this paper achieved average values ​​of 0.204%, 0.392%, and 0.9927 for SOH estimation across the six batteries, and average values ​​of 6.7912%, 6.7918, and 0.998 for RUL estimation, making it the best performing loss function among all listed loss functions.

[0114] Obviously, those skilled in the art should understand that the modules or steps of the above-described embodiments of the present invention can be implemented using general-purpose computer devices. They can be centralized on a single computer device or distributed across a network of multiple computer devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computer device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular combination of hardware and software.

[0115] It should be noted that the above are merely preferred embodiments of the present invention and do not limit the scope of patent protection of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A joint estimation method for SOH-RUL of lithium-ion batteries based on the Dozerformer model, characterized in that, include: S1. The Dozerformer model is used to learn the mapping relationship between battery aging features and SOH and RUL. The battery aging features are decomposed into periodic components and trend components. A sparse attention mechanism consisting of local sparsity, step size sparsity and change sparsity is introduced to learn the long-term and short-term features in the periodic components. A linear layer is used to learn the feature information of the trend components. For an input sequence X, where L is the sequence length and N is the number of features, the main calculation process of the Dozerformer model is as follows: S1.1 Decompose the input sequence X into trend components and periodic components; S1.2 Calculate the prediction results on the trend components; S1.

3. Embed the periodic components into c feature maps with dimension invariance and divide them into p patches; S1.4 Utilize the Transformer encoder to learn the sequence feature information in the periodic component, and introduce local sparsity and stride sparsity strategies in the self-attention mechanism to sparsify the attention weights, thereby obtaining the sparsified attention mechanism result. Subsequently, through the feedforward network layer and residual connection, the final output of the encoder is generated. S1.

5. Using the Transformer decoder, the information output by the encoder is decoded through a combination of self-attention and cross-attention mechanisms to obtain the final decoder output; specifically, this includes the following steps: S1.5.

1. The input of the decoder is processed using the same masked multi-head self-attention mechanism as the Transformer encoder to obtain the multi-head self-attention output result; S1.5.

2. The output of the Transformer encoder is used as the key and value of the cross-attention mechanism, and the input of the decoder is used as the query. Local sparse matrix, step sparse matrix and change sparse matrix are introduced into the cross-attention mechanism to decode the encoder information and obtain the output result of the cross-attention mechanism. S1.5.

3. The output of the multi-head self-attention mechanism and the output of the cross-attention mechanism are fused through the feedforward network layer and residual connection to generate the final output of the decoder, i.e. the final prediction result. S1.

6. Use 1×1 convolution to obtain the prediction results on the periodic components; S1.

7. Add the prediction results on the trend component and the cycle component to obtain the final result; S2. During model training, the weights and biases of the Dozerformer model are optimized using a smooth quadratic loss function. S3. Use the Dozerformer model with optimal weights and biases to estimate the SOH and RUL of the lithium-ion battery and output the SOH-RUL estimation results.

2. The lithium-ion battery SOH-RUL joint estimation method according to claim 1, characterized in that, Local sparsity refers to the weights within a fixed query window, thus ignoring the remainder. For a local window of size w, the local sparse mask matrix can be represented as: Step sparsity means querying weights at a fixed step size, ignoring and setting other weights to zero; a step sparsity mask matrix with a step size of s can be represented as: At this point, merging the local sparse mask matrix and the stride sparse mask matrix yields the sparse mask matrix for the self-attention mechanism, as shown in the following equation: Using equation (17), the sparsified attention mechanism is obtained as follows: The encoder output, obtained through the feedforward network layer and residual connections, is: Where ReLU is the activation function and LayerNorm is the layer normalization process.

3. The lithium-ion battery SOH-RUL joint estimation method according to claim 1, characterized in that, The local sparse matrix, step-size sparse matrix, and variation sparse matrix in the cross-attention mechanism can be represented as follows: The local sparse matrix, the step-size sparse matrix, and the variational sparse matrix are combined to obtain the final sparse matrix for the cross-attention mechanism: The result of the cross-attention mechanism after sparsification is as follows:

4. The method according to claim 1, characterized in that, The expression for the smooth quadratic loss function is: Where α∈[0,1] is the smoothing coefficient, c∈(0,+∞), and y a For the true value of the sequence, These are the predicted values ​​for the sequence.

5. The lithium-ion battery SOH-RUL joint estimation method according to claim 1, characterized in that, Battery aging characteristics are the features corresponding to lithium-ion battery aging data, which include: battery internal resistance, average temperature, charging capacity, charging temperature, discharging capacity, and charging time.

Images