A lithium ion battery state of charge estimation method based on multi-modal data fusion

Abstract

The application belongs to the technical field of battery state of charge estimation, and discloses a lithium ion battery state of charge estimation method based on multi-modal data fusion. The method comprises the following steps: constructing a multi-scale convolutional neural network; performing bidirectional information interaction between acoustic signals and electrical signals to obtain fusion features; constructing a time series prediction model based on sequence decomposition; inputting the fusion features into the time series prediction module for modeling; and generating a final SOC estimation result based on the prediction output features. The application constructs a complete multi-modal SOC estimation framework by constructing a multi-scale convolutional neural network, a bidirectional cross-attention fusion module, and a sequence decomposition prediction model, realizes the collaborative utilization of acoustic information and electrical information, and improves the accuracy and robustness of SOC estimation.

Description

Technical Field

[0001] This invention relates to the field of battery state of charge assessment technology, and in particular to a method for estimating the state of charge of lithium-ion batteries based on multimodal data fusion. Background Technology

[0002] Currently, State of Charge (SOC) estimation for lithium-ion batteries is one of the key functions of a Battery Management System (BMS). SOC estimation characterizes the battery's current remaining capacity, and its accuracy directly affects the battery system's safety, energy management efficiency, and lifespan. Therefore, various technical solutions have been proposed to address the SOC estimation problem.

[0003] In traditional methods, State of Charge (SOC) estimation is typically based on the battery's electrical signals. For example, coulometric methods estimate battery capacity changes by integrating the charging and discharging current over time; open-circuit voltage methods estimate SOC by establishing a mapping relationship between SOC and the battery's open-circuit voltage; additionally, there are methods based on equivalent circuit models or electrochemical models, which construct internal dynamic models of the battery and combine them with algorithms such as Kalman filtering for real-time SOC estimation. These methods usually rely on external measurable signals such as voltage, current, and temperature, and achieve SOC extrapolation through model calculations or data fitting.

[0004] With the development of machine learning technology, data-driven SOC estimation methods have emerged in recent years. These methods typically collect time-series data such as voltage, current, and temperature during battery operation and use deep learning models (e.g., convolutional neural networks, recurrent neural networks, or long short-term memory networks) to build an SOC prediction model. By training the model with a large amount of experimental data, it learns the nonlinear relationship between battery operating parameters and SOC, thereby achieving SOC estimation.

[0005] However, relying solely on electrical signals is insufficient to directly reflect changes in the internal structure of a battery. Therefore, some studies have begun to introduce ultrasonic detection technology. In this approach, ultrasonic transducers are typically placed on the battery surface. By exciting ultrasonic waves and receiving the echo signals, the acoustic response information inside the battery is obtained. Since the insertion and extraction of lithium ions in the positive and negative electrode materials causes changes in the electrode material density, elastic modulus, and internal structural state, it alters the propagation characteristics of ultrasonic waves within the battery. Existing technologies typically extract parameters such as the time of flight (TOF), signal amplitude (SA), and frequency domain characteristics of the ultrasonic signal to characterize changes in the battery's internal state and further establish estimation models for State of Charge (SOC) or State of Health (SOH).

[0006] Building upon this foundation, some studies have proposed SOC estimation schemes based on multimodal data fusion. These schemes typically collect electrical data such as battery voltage and current, as well as acoustic data such as ultrasonic echo signals, simultaneously extracting feature information from different modes. Then, an SOC estimation model is constructed through feature fusion. For example, in practical implementation, voltage and current data during battery charging and discharging are first collected, while echo signals are simultaneously acquired using an ultrasonic transducer. Subsequently, feature extraction is performed on both the electrical and acoustic signals to obtain feature parameters reflecting the battery's operating state. Next, the features from different modes are fused and input into a machine learning model for training or prediction; finally, the SOC estimation result is output.

[0007] However, the aforementioned existing technologies typically only utilize electrical data such as voltage, current, and temperature as model inputs. While these methods can achieve high accuracy within the training data range, they suffer from insufficient generalization ability under battery aging, temperature variations, or complex operating conditions due to the lack of information characterizing changes in the battery's internal structure. Furthermore, existing data-driven methods often employ a single-scale feature learning approach during feature extraction, making it difficult to simultaneously capture both local dynamic features and overall trends in battery operating data. Summary of the Invention

[0008] The main objective of this invention is to provide a method for estimating the state of charge of lithium-ion batteries based on multimodal data fusion, so as to solve the technical problems in the prior art.

[0009] To achieve the above objectives, this invention provides a method for estimating the state of charge of lithium-ion batteries based on multimodal data fusion, the method comprising the following steps: S10, construct a multi-scale convolutional neural network; S20, performs bidirectional information interaction between acoustic and electrical signals to obtain fusion features; S30, Construct a time series prediction model based on sequence decomposition; S40, The fused features are input into the time series prediction module for modeling to obtain the predicted output features; S50, Based on the predicted output features, generate the final SOC estimation result.

[0010] Optionally, step S10 includes the following steps: S110, construct the initial multi-scale convolutional neural network; S120, based on the initial multi-scale convolutional neural network, the input sequence is processed through three parallel branches to output three features; S130, the three features are spliced ​​together to form the spliced ​​features; S140 processes the concatenated features through a convolutional layer with a kernel size of 1. S150, the processed features are sequentially passed through an activation function layer, a max pooling layer, and a Dropout layer to construct the final multi-scale convolutional neural network.

[0011] Optionally, step S20 includes the following steps: S210, Construct a bidirectional cross-attention module, which includes two symmetrical information interaction directions; S220, in the first direction, acoustic data is used as a query vector, and electrical data is used as a key vector and a value vector, respectively. Attention calculation is performed on the electrical data to obtain the output of the first direction; S230, in the second direction, electrical data is used as a query vector, and acoustic data is used as a key vector and a value vector, respectively. Attention calculation is performed on the acoustic data to obtain the output of the second direction. S240, the outputs of the first direction and the outputs of the second direction are added element by element to obtain the fusion feature.

[0012] Optionally, the attention calculation of the electrical data includes the following steps: S2210, by multiplying with the weight matrix, maps the input data into the three vectors Q, K, and V required by the attention mechanism, as shown in the following formula:

[0013]

[0014]

[0015] In the formula: X ultraX represents the input ultrasound feature sequence tensor. cv Represents the input electrical data sequence tensor; , , These represent the learnable weight matrices for Query, Key, and Value, respectively. S2220 calculates the attention score and substitutes Q, K, and V into the corresponding formula to obtain the output for the first direction. The calculation formula is as follows:

[0016] In the formula: Q ultra This represents the query vector in the first direction; K ultra This represents the key vector in the first direction; V ultra Represents the value vector in the first direction; d k Indicates the dimension of the key vector; softmax This represents the normalization function, used to convert attention scores into attention weights; Output ultra This indicates the output in the first direction.

[0017] Optionally, the attention calculation on the acoustic data includes the following steps: S2310, by multiplying with the weight matrix, maps the input data into the three vectors Q, K, and V required by the attention mechanism, as shown in the following formula:

[0018]

[0019]

[0020] In the formula: X cv This represents the input electrical data sequence tensor. X ultra W represents the input ultrasound feature sequence tensor; q cv and W v cv These represent the learnable weight matrices used to generate the query vector, key vector, and value vector in the second direction, respectively. Q cv , K cv These represent the query vector, key vector, and value vector in the second direction, respectively. S2320 calculates the attention score and substitutes Q, K, and V into the corresponding formula to obtain the output for the first direction. The calculation formula is as follows:

[0021] In the formula: Q cv , K cv and V cv These represent the query vector, key vector, and value vector in the second direction, respectively. d k Indicates the dimension of the key vector; softmax This represents a function that normalizes the attention score. Output cv This represents the output obtained by attention calculation in the second direction.

[0022] Optionally, in step S240, the fusion features are obtained based on the following formula: Fused=Norm(Output ultra +Output cv ) In the formula: Output ultra Indicates the output in the first direction. Output cv This indicates the output in the second direction; Norm represents the normalization function; Fused represents the fused feature.

[0023] Optionally, step S30 includes the following steps: S310, based on the moving average operation, extracts the long-term trend of the data, as shown in the following formula:

[0024] In the formula: Trend t Indicates time t The long-term trend term extracted from the data; W This represents the length of the moving average window, preferably a positive odd number; i This represents the relative position index within the moving average window; X t+i Represents the time in the input sequence t + i Data at the location; t Indicates the current time; S320, subtracts the trend component from the original input to obtain the seasonal term containing local details:

[0025] In the formula: Seasonal t Indicates time t The seasonal term at the input data is used to characterize local detail changes and dynamic responses in the input data; X t Represents the time in the input sequence t The original data at the location; Trend t Indicates time t The long-term trend item at the location; S330, inputs the separated seasonal terms into a multilayer perceptron for nonlinear mapping; S340, based on the following formula, adds the processed seasonal term to the original trend term to reconstruct a new sequence:

[0026] In the formula: Output t Indicates time t The output sequence obtained by reconstruction at the point; Seasonal p This represents the seasonal term after nonlinear mapping via a multilayer perceptron. Trend t Indicates time t The long-term trend item at the location; S350 introduces residual connections and layer normalization to accelerate model convergence and stabilize gradient propagation to obtain the final output.

[0027] Optionally, step S40 includes the following steps: S410 inputs the preprocessed and feature-engineered ultrasonic feature sequences and electrical data in parallel into the two branch entrances of the network; S420 extracts features from the input data based on two MSCNNs with the same and independent structures, mapping the original input data sequence into a deep feature representation. S430 inputs the fused features into the time series prediction module for modeling.

[0028] Optionally, step S50 includes the following steps: S510, input the predicted output features of the time series prediction module into the output layer for feature mapping to obtain the corresponding SOC estimate; S520, The SOC estimate is processed so that the processed SOC estimate corresponds to the range of the actual state of charge of the battery. S530 outputs the processed result as the final SOC estimate of the lithium-ion battery at the current moment.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: The lithium-ion battery state-of-charge estimation method based on multimodal data fusion provided in this application obtains fused features through bidirectional information interaction between acoustic and electrical signals. By constructing a multi-scale convolutional neural network, a bidirectional cross-attention fusion module, and a sequence decomposition prediction model, a complete multimodal SOC estimation framework is built, realizing the synergistic utilization of acoustic and electrical information, thereby improving the accuracy and robustness of SOC estimation. Attached Figure Description

[0030] Figure 1 A flowchart illustrating the lithium-ion battery state-of-charge estimation method based on multimodal data fusion provided in this application embodiment; Figure 2 A schematic diagram of the MSCNN structure provided in the embodiments of this application; Figure 3 A schematic diagram of the cross-attention structure provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of the time series prediction model provided in the embodiments of this application; Figure 5 The SOC estimation results are under standard conditions; Figure 6 The SOC estimation results under different temperature conditions are provided for the embodiments of this application.

[0031] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0032] It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the application. Rather, these embodiments are provided to make the disclosure more thorough and complete, and to fully convey the scope of the disclosure to those skilled in the art.

[0033] To address the aforementioned technical problems, embodiments of this application provide a method for estimating the state of charge (SOC) of lithium-ion batteries based on multimodal data fusion, such as... Figure 1 As shown, the method may include the following steps: S10, construct a multi-scale convolutional neural network.

[0034] In an exemplary embodiment, step S10 includes the following steps: S110, construct the initial multi-scale convolutional neural network; S120, based on the initial multi-scale convolutional neural network, the input sequence is processed through three parallel branches to output three features; S130, the three features are spliced ​​together to form the spliced ​​features; S140 processes the concatenated features through a convolutional layer with a kernel size of 1. S150, the processed features are sequentially passed through an activation function layer, a max pooling layer, and a Dropout layer to construct the final multi-scale convolutional neural network.

[0035] Convolutional Neural Networks (CNNs) are often used to process one-dimensional time series data due to their strong ability to extract local features. However, because the kernel size is fixed, standard one-dimensional CNNs often struggle to simultaneously extract both local detail features and long-term dependencies. This limitation is particularly evident in SOC estimation tasks based on ultrasound signals, as electrochemical mechanisms manifest at different signal scales.

[0036] Essentially, changes in State of Charge (SOC) involve the insertion and extraction of lithium ions between the positive and negative electrodes. This process alters the macroscopic physical parameters of the battery, particularly Young's modulus and density. Consequently, these changes cause variations in the time of flight (TOF) and signal amplitude (SA). Simultaneously, microscopic interactions at the battery's internal interfaces induce fine-grained waveform distortions. Single-scale convolutions struggle to capture these dual characteristics simultaneously: small kernels lack sufficient receptive fields to track macroscopic trends, while large kernels often overlook subtle local features. Therefore, this invention introduces a Multi-Scale Convolutional Neural Network (MSCNN) that adaptively extracts features at different scales through a parallel convolutional structure.

[0037] Figure 2 The structure of MSCNN is demonstrated, its core being the processing of the input sequence through three parallel branches. Each branch employs a one-dimensional CNN of different scales, with kernel sizes of 1, 3, and 5, respectively. Subsequently, the features output from the three branches are concatenated to form a hybrid feature tensor containing rich information.

[0038] Since direct concatenation would double the feature dimension, to reduce computational complexity, the concatenated features are processed again through a convolutional layer with a kernel size of 1. This operation not only makes the model lightweight but also promotes the fusion of information between macro-trend features and micro-detail features.

[0039] Finally, to further optimize the feature representation, the compressed features are sequentially passed through an activation function layer, a max pooling layer, and a dropout layer. This structural design not only ensures the lightweight nature of the model but also provides a more comprehensive feature representation capability, thereby significantly improving the model's ability to map complex battery signals into accurate SOC values.

[0040] S20 performs bidirectional information interaction between acoustic and electrical signals to obtain fusion features.

[0041] In an exemplary embodiment, step S20 includes the following steps: S210, Construct a bidirectional cross-attention module, which includes two symmetrical information interaction directions; S220, in the first direction, acoustic data is used as a query vector, and electrical data is used as a key vector and a value vector, respectively. Attention calculation is performed on the electrical data to obtain the output of the first direction; S230, in the second direction, electrical data is used as a query vector, and acoustic data is used as a key vector and a value vector, respectively. Attention calculation is performed on the acoustic data to obtain the output of the second direction. S240, the outputs of the first direction and the outputs of the second direction are added element by element to obtain the fusion feature.

[0042] In the characterization of the internal state of a battery, acoustic signals can reflect changes in the battery's physical structural characteristics, specifically the changes in electrode lattice structure and material density caused by lithium-ion migration. Electrical signals, on the other hand, directly reflect the battery's electrochemical reaction rate and energy state. These two modes of data reflect the state changes of lithium-ion batteries from different perspectives and are highly complementary in terms of information.

[0043] However, existing single signal processing techniques are insufficient to address the problem of inconsistent distribution of key information among features of different modalities, and simple feature splicing is insufficient to effectively capture the correlation between different modalities.

[0044] To achieve effective fusion of features from two modalities, this embodiment provides a feature fusion method based on bidirectional cross-attention. This method enables bidirectional information interaction between acoustic and electrical signals, thereby generating a fused feature representation containing multi-physics interaction information. The specific scheme is as follows: The bidirectional cross-attention module comprises two symmetrical directions of information interaction. In the first direction, acoustic data serves as the query vector, used to perform attention calculations on electrical data and extract key information, while the electrical data serves as the key vector and value vector, respectively. Figure 3A cross-attention structure was demonstrated, which enhances feature fusion by establishing a bidirectional information exchange mechanism between different modalities to model the correlation between acoustic and electrical features. In an exemplary embodiment, the attention calculation of the electrical data includes the following steps: S2210, by multiplying with the weight matrix, maps the input data into the three vectors Q, K, and V required by the attention mechanism, as shown in the following formula:

[0045]

[0046]

[0047] In the formula: X ultra X represents the input ultrasound feature sequence tensor. cv Represents the input electrical data sequence tensor; , , These represent the learnable weight matrices for Query, Key, and Value, respectively. S2220 calculates the attention score and substitutes Q, K, and V into the corresponding formula to obtain the output for the first direction. The calculation formula is as follows:

[0048] In the formula: Q ultra This represents the query vector in the first direction; K ultra This represents the key vector in the first direction; V ultra Represents the value vector in the first direction; d k Indicates the dimension of the key vector; softmax This represents the normalization function, used to convert attention scores into attention weights; Output ultra This indicates the output in the first direction.

[0049] In an exemplary embodiment, the attention calculation of the acoustic data includes the following steps: S2310, by multiplying with the weight matrix, maps the input data into the three vectors Q, K, and V required by the attention mechanism, as shown in the following formula:

[0050]

[0051]

[0052] In the formula: X cv This represents the input electrical data sequence tensor. X ultra W represents the input ultrasound feature sequence tensor; q cv and W v cv These represent the learnable weight matrices used to generate the query vector, key vector, and value vector in the second direction, respectively. Q cv , K cv These represent the query vector, key vector, and value vector in the second direction, respectively. S2320 calculates the attention score and substitutes Q, K, and V into the corresponding formula to obtain the output for the first direction. The calculation formula is as follows:

[0053] In the formula: Q cv K cv and V cv These represent the query vector, key vector, and value vector in the second direction, respectively; d k The key vector dimension is represented by `softmax`, which represents the function used to normalize the attention scores. Output cv This represents the output obtained by attention calculation in the second direction.

[0054] In an exemplary embodiment, in step S240, the fusion features are obtained based on the following formula: Fused=Norm(Output ultra +Output cv ) In the formula: Output ultra Indicates the output in the first direction. Output cv This indicates the output in the second direction; Norm represents the normalization function; Fused represents the fused feature.

[0055] S30, Construct a time series prediction model based on sequence decomposition.

[0056] Data from lithium-ion batteries during charging and discharging exhibits complex time-varying characteristics. Specifically, the signal can be viewed as a superposition of two different physical processes: one is a long-term trend term reflecting changes in SOC, and the other is a short-term seasonal term affected by instantaneous current changes or other transient factors.

[0057] Traditional recurrent neural networks or fully connected networks typically process input data uniformly without distinguishing between different components. This not only increases the optimization difficulty of the model but also easily leads to overfitting and reduced prediction performance. Therefore, this invention constructs a time series prediction model based on sequence decomposition, utilizing the decomposition mechanism of the Autoformer model. This model first extracts the long-term trend term from the input sequence through a moving average operation. Then, it subtracts the trend term from the original input to obtain a seasonal term containing local details. Subsequently, it performs a nonlinear mapping on the seasonal term and reconstructs it with the trend term. Finally, it obtains the model output through residual connections and layer normalization. Figure 4 The structure of this time series forecasting model is demonstrated, which mainly includes three steps: moving average decomposition, seasonal term modeling, and residual connection. This model effectively predicts the trend of battery SOC changes by decomposing the input sequence into trend and seasonal components and modeling and reconstructing them separately.

[0058] In an exemplary embodiment, step S30 includes the following steps: S310, based on the moving average operation, extracts the long-term trend of the data, as shown in the following formula:

[0059] In the formula: Trend t Indicates time t The long-term trend term extracted from the data; W This represents the length of the moving average window, preferably a positive odd number; i This represents the relative position index within the moving average window; X t+i Represents the time in the input sequence t + i Data at the location; t Indicates the current time; S320, subtracts the trend component from the original input to obtain the seasonal term containing local details:

[0060] In the formula: Seasonal t Indicates time t The seasonal term at the input data is used to characterize local detail changes and dynamic responses in the input data; X t Represents the time in the input sequence t The original data at the location; Trend t Indicates time t The long-term trend item at the location; S330, inputs the separated seasonal terms into a multilayer perceptron for nonlinear mapping; S340, based on the following formula, adds the processed seasonal term to the original trend term to reconstruct a new sequence:

[0061] In the formula: Output t Indicates time t The output sequence obtained by reconstruction at the point; Seasonal p This represents the seasonal term after nonlinear mapping via a multilayer perceptron. Trend t Indicates time t The long-term trend item at the location; S350 introduces residual connections and layer normalization to accelerate model convergence and stabilize gradient propagation to obtain the final output.

[0062] S40, The fused features are input into the time series prediction module for modeling.

[0063] In an exemplary embodiment, step S40 includes the following steps: S410 inputs the preprocessed and feature-engineered ultrasonic feature sequences and electrical data in parallel into the two branch entrances of the network; S420 extracts features from the input data based on two MSCNNs with the same and independent structures, mapping the original input data sequence into a deep feature representation. S430 inputs the fused features into the time series prediction module for modeling.

[0064] Specifically, to separate the trend term and seasonality, this invention constructs a time series prediction module based on the sequence decomposition method proposed in the Autoformer model. First, the long-term trend component is extracted from the input features through a moving average operation. Second, the seasonal component, containing local details, is separated by subtracting the trend component from the input sequence. Then, the seasonal component is input into a multilayer perceptron for nonlinear feature transformation. Finally, the transformed seasonal component and the trend component are reconstructed, and the module output is obtained through residual connections and layer normalization.

[0065] Subtracting the trend component from the original input yields the seasonal term, which includes local details, as shown in the following formula:

[0066] Here, "Seasonal" represents the dynamic response, such as instantaneous voltage fluctuations and microscopic acoustic distortions. This decomposition strategy allows the model to focus more on learning non-stationary fluctuation characteristics. Subsequently, the separated seasonal term is input into a Multi-Layer Perceptron (MLP) for nonlinear mapping. This MLP contains two fully connected layers and a Gaussian Error Linear Unit (GELU) activation function, and its operation can be represented as follows:

[0067] In the formula: Seasonal p This represents the seasonal term after nonlinear mapping via a multilayer perceptron. Seasonal This represents the separated seasonal term, used to characterize local detail changes and dynamic responses in the input data; W 1 and W 2 represents the weight matrices of the first fully connected layer and the second fully connected layer in the multilayer perceptron, respectively; b 1 and b 2 represents the bias terms of the first fully connected layer and the second fully connected layer, respectively; GELU This represents the Gaussian error linear unit activation function, used to perform a nonlinear transformation on the seasonal term.

[0068] Furthermore, since trend terms are generally quite stable, this invention retains the original trend terms without processing them. Subsequently, the processed seasonal term is added to the original trend term to reconstruct a new sequence:

[0069] In the formula: Output t Indicates time t The output sequence obtained by reconstruction at the point; Seasonal p This represents the seasonal term after nonlinear mapping via a multilayer perceptron. Trend t Indicates time t The long-term trend item at the location.

[0070] Finally, residual connections and layer normalization are introduced to accelerate model convergence and stabilize gradient propagation. The final output is:

[0071] In the formula: Output final This represents the final output feature after residual connection and layer normalization. Output t Indicates time tThe reconstructed features output by the time series prediction module; X represents the input features of this module; LayerNorm The layer normalization function is used to normalize the features after the residuals are summed.

[0072] S50, Based on the predicted output features, generate the final SOC estimation result.

[0073] In an exemplary embodiment, step S50 includes the following steps: S510, input the predicted output features of the time series prediction module into the output layer for feature mapping to obtain the corresponding SOC estimate; S520, The SOC estimate is processed so that the processed SOC estimate corresponds to the range of the actual state of charge of the battery. S530 outputs the processed result as the final SOC estimate of the lithium-ion battery at the current moment.

[0074] Specifically, after multi-scale convolutional feature extraction, bidirectional cross-attention fusion, and time-series prediction based on sequence decomposition, the predicted output features already contain the joint representation of acoustic and electrical information, as well as long-term trend information and local dynamic information reflecting changes in SOC. To convert these predicted output features into a target variable that can directly characterize the remaining battery power state, this embodiment inputs the predicted output features into the output layer, preferably using a fully connected layer for linear mapping, thereby generating an estimated SOC value.

[0075] In some implementations, when the SOC label is normalized during model training, the SOC estimate can be further denormalized during the output stage to obtain an SOC estimate consistent with the actual battery state of charge range. The final output SOC estimate is used to characterize the remaining state of charge of the lithium-ion battery at the current moment and can provide a basis for energy management, charge / discharge control, and safety warnings in the battery management system.

[0076] In the embodiments provided in this application, fusion features are obtained by performing bidirectional information interaction between acoustic and electrical signals; a complete multimodal SOC estimation framework is constructed by building a multi-scale convolutional neural network, a bidirectional cross-attention fusion module, and a sequence decomposition prediction model, thereby realizing the synergistic utilization of acoustic and electrical information and improving the accuracy and robustness of SOC estimation.

[0077] To verify the effectiveness of the above embodiments, an experiment was conducted using lithium-ion battery data collected under strict laboratory conditions. The dataset used to train the lithium-ion battery state-of-charge estimation model contained 28 complete charge-discharge cycles.

[0078] Figure 5 The SOC estimation results for six consecutive test cycles under a standard temperature of 25℃ are presented. Table 1 lists the evaluation metrics for each cycle. By fusing multimodal data, this method achieves high-precision SOC estimation with a mean absolute error (MAE) of 1.0733%, a root mean square error (RMSE) of 1.3537%, and a coefficient of determination R0. 2 The value is 0.9978. Furthermore, during long-cycle SOC estimation, the model's error fluctuations are small, consistently remaining at a low level.

[0079] Table 1

[0080] The model achieves a low SOC estimation error, mainly due to the following two reasons.

[0081] First, the information is richer. Traditional single-modal models typically rely solely on ultrasonic signals, primarily reflecting changes in the battery's internal mechanical modulus and structure. While ultrasonic signals are highly sensitive to State of Charge (SOC), in some cases, changes in mechanical state may lag behind changes in electrochemical state. Therefore, multimodal models further incorporate voltage and current data, as electrical signals can reflect changes in electrochemical state more promptly. Furthermore, the cross-attention feature fusion module can fuse electrical and acoustic information, thereby compensating for the insufficient feature information in single-modal models and significantly improving the accuracy of SOC estimation.

[0082] Secondly, the decomposition strategy in the time series forecasting model separates the long-term trend term, which varies with SOC, from the short-term seasonal term. This decoupling design makes the model more robust when dealing with noisy data.

[0083] Furthermore, ambient temperature is a significant external factor affecting the accuracy of lithium-ion battery SOC estimation. To evaluate the model's adaptability under different temperature conditions, this experiment used data collected at 15℃, 20℃, 25℃, 30℃, and 35℃ as the test set.

[0084] Figure 6The SOC estimation results of the multimodal model under different temperature conditions are presented, and Table 2 provides the corresponding evaluation metrics. Specifically, by analyzing the results and comparing the mean absolute error (MAE) and root mean square error (RMSE) under different temperature conditions, it can be seen that when the temperature decreases to 15℃, the MAE is 1.43%. When the temperature increases to 35℃, the MAE rises to 1.4776%, and the RMSE increases to 1.9031%. The errors remain within a small range, indicating that the model still has good SOC estimation capabilities.

[0085] Table 2

[0086] At lower ambient temperatures, electrolyte viscosity increases, thus affecting lithium-ion migration. Simultaneously, the electrochemical reaction rate at the electrode surface slows down, and the internal resistance of the battery increases significantly. Furthermore, low temperatures also affect the ultrasonic propagation characteristics within the battery. The increased Young's modulus of the electrodes leads to a decrease in ultrasonic propagation speed.

[0087] At higher temperatures, the internal activity of the battery increases, and the electrochemical reaction rate rises. Simultaneously, increased temperature causes the internal layered structure of the battery to expand, thereby reducing the speed of ultrasonic propagation. Furthermore, high temperatures may trigger side reactions and generate trace amounts of gas. All of these factors can affect the accuracy of SOC estimation.

[0088] Despite these influencing factors, the model's maximum MAE remained below 1.5% over a wide temperature range. This is mainly due to the information complementarity mechanism between the two modal data, thus ensuring the model's high robustness over a wide temperature range.

[0089] To verify the applicability of the multimodal model to batteries with different aging levels, the SOC estimation performance of the model was tested on batteries with SOH ranging from 100% to 90%. As shown in Table 3, it can be clearly observed that the estimation error of the model gradually increases as the SOH decreases. In Table 3, SOH (State of Health) represents the current health status of the battery, expressed as a percentage, with 100% indicating a brand-new state. The SOH value gradually decreases as the battery ages. MAE (Mean Absolute Error) represents the mean absolute error, the average of the absolute differences between the predicted and actual values ​​of the target variable. It measures the average error between the model's predicted and actual values; a smaller value indicates a more accurate prediction. RMSE (Root Mean Squared Error) represents the root mean square error, the square root of the average of the squares of the differences between the predicted and actual values. It gives greater weight to larger errors and is also used to evaluate the model's prediction accuracy; a smaller value is better. R² (R-squared) represents the coefficient of determination, reflecting the proportion of the variance in the dependent variable that can be explained by the independent variable. Its value ranges from 0 to 1; the closer to 1, the better the model fits the data, meaning the stronger the model's ability to explain the changes in the variable. MaxError represents the maximum error, the maximum value of the error between the predicted and actual values, reflecting the most extreme error situation in the model's predictions.

[0090] Table 3

[0091] When the SOH decreases to 90%, the estimation error shows a gradual upward trend. When the SOH decreases to 90%, the MAE increases to 1.5183%, the RMSE increases to 1.8289%, and the R... 2 The error decreased from 0.9978 to 0.9961. Although the error increased with increasing aging, the MAE of the model remained below 1.6% even under 90% SOH conditions. This indicates that the proposed multimodal fusion framework has good robustness.

[0092] The model maintains a low error rate within a certain range, primarily due to its sequence decomposition mechanism. This mechanism effectively filters out transient fluctuations caused by aging, allowing the model to focus more on the main trends in SOC changes. Furthermore, when data in one modality fluctuates, bidirectional cross-attention enables the model to adaptively focus on data from another modality, further enhancing its robustness.

[0093] To systematically verify the advantages of the core modules in the proposed framework and further explore their performance advantages in the SOC estimation task, this invention designed a comparative experiment. The experimental results are shown in Table 4 below.

[0094] Table 4

[0095] Comparing the model of this invention with iTransformer, although iTransformer slightly outperforms in the MAE metric, the model of this invention performs better in the RMSE metric, and RMSE better reflects the robustness of the model. A lower RMSE indicates that the model of this invention produces fewer large biases in the SOC estimation task, thus demonstrating stronger stability. This is mainly due to the multi-scale feature extraction capability of MSCNN, the efficient information fusion capability of the bidirectional cross-attention mechanism, and the sequence decomposition mechanism, which enables the model to simultaneously focus on overall trend features and local detailed features.

[0096] The overall experimental results show that the proposed multimodal feature fusion framework can not only achieve high SOC estimation accuracy, but also has stronger estimation robustness.

[0097] In the description of this application, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0098] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0099] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0100] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0101] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0102] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

[0104] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

Claims

1. A method for estimating the state of charge of a lithium-ion battery based on multimodal data fusion, characterized in that, The method includes the following steps: S10, construct a multi-scale convolutional neural network; S20, performs bidirectional information interaction between acoustic and electrical signals to obtain fusion features; S30, Construct a time series prediction model based on sequence decomposition; S40, The fused features are input into the time series prediction module for modeling to obtain the predicted output features; S50, Based on the predicted output features, generate the final SOC estimation result.

2. The method for estimating the state of charge of a lithium-ion battery based on multimodal data fusion according to claim 1, characterized in that, Step S10 includes the following steps: S110, construct the initial multi-scale convolutional neural network; S120, based on the initial multi-scale convolutional neural network, the input sequence is processed through three parallel branches to output three features; S130, the three features are spliced ​​together to form the spliced ​​features; S140 processes the concatenated features through a convolutional layer with a kernel size of 1. S150, the processed features are sequentially passed through an activation function layer, a max pooling layer, and a Dropout layer to construct the final multi-scale convolutional neural network.

3. The method for estimating the state of charge of a lithium-ion battery based on multimodal data fusion according to claim 1, characterized in that, Step S20 includes the following steps: S210, Construct a bidirectional cross-attention module, which includes two symmetrical information interaction directions; S220, in the first direction, acoustic data is used as a query vector, and electrical data is used as a key vector and a value vector, respectively. Attention calculation is performed on the electrical data to obtain the output of the first direction; S230, in the second direction, electrical data is used as a query vector, and acoustic data is used as a key vector and a value vector, respectively. Attention calculation is performed on the acoustic data to obtain the output of the second direction. S240, the outputs of the first direction and the outputs of the second direction are added element by element to obtain the fusion feature.

4. The method for estimating the state of charge of a lithium-ion battery based on multimodal data fusion according to claim 3, characterized in that, The attention calculation of the electrical data includes the following steps: S2210, by multiplying with the weight matrix, maps the input data into the three vectors Q, K, and V required by the attention mechanism, as shown in the following formula: In the formula: X ultra X represents the input ultrasound feature sequence tensor. cv Represents the input electrical data sequence tensor; , , These represent the learnable weight matrices for Query, Key, and Value, respectively. S2220 calculates the attention score and substitutes Q, K, and V into the corresponding formula to obtain the output for the first direction. The calculation formula is as follows: In the formula: Q ultra This represents the query vector in the first direction; K ultra This represents the key vector in the first direction; V ultra Represents the value vector in the first direction; d k Indicates the dimension of the key vector; softmax This represents the normalization function, used to convert attention scores into attention weights; Output ultra This indicates the output in the first direction.

5. The method for estimating the state of charge of a lithium-ion battery based on multimodal data fusion according to claim 1, characterized in that, The attention calculation of the acoustic data includes the following steps: S2310, by multiplying with the weight matrix, maps the input data into the three vectors Q, K, and V required by the attention mechanism, as shown in the following formula: In the formula: X cv This represents the input electrical data sequence tensor. X ultra W represents the input ultrasound feature sequence tensor; q cv and W v cv These represent the learnable weight matrices used to generate the query vector, key vector, and value vector in the second direction, respectively. Q cv , K cv These represent the query vector, key vector, and value vector in the second direction, respectively. S2320 calculates the attention score and substitutes Q, K, and V into the corresponding formula to obtain the output for the first direction. The calculation formula is as follows: In the formula: Q cv , K cv and V cv These represent the query vector, key vector, and value vector in the second direction, respectively. d k Indicates the dimension of the key vector; softmax This represents a function that normalizes the attention score. Output cv This represents the output obtained by attention calculation in the second direction.

6. The method for estimating the state of charge of a lithium-ion battery based on multimodal data fusion according to claim 3, characterized in that, In step S240, the fusion features are obtained based on the following formula: Fused=Norm(Output ultra +Output cv ) Where: Output ultra Output represents the output in the first direction. cv This indicates the output in the second direction; Norm represents the normalization function; Fused represents the fused feature.

7. The method for estimating the state of charge of a lithium-ion battery based on multimodal data fusion according to claim 1, characterized in that, Step S30 includes the following steps: S310, based on the moving average operation, extracts the long-term trend of the data, as shown in the following formula: In the formula: Trend t Indicates time t The long-term trend term extracted from the data; W This represents the length of the moving average window, preferably a positive odd number; i This represents the relative position index within the moving average window; X t+i Represents the time in the input sequence t + i Data at the location; t Indicates the current time; S320, subtracts the trend component from the original input to obtain the seasonal term containing local details: In the formula: Seasonal t Indicates time t The seasonal term at the input data is used to characterize local detail changes and dynamic responses in the input data; X t Represents the time in the input sequence t The original data at the location; Trend t Indicates time t The long-term trend item at the location; S330, inputs the separated seasonal terms into a multilayer perceptron for nonlinear mapping; S340, based on the following formula, adds the processed seasonal term to the original trend term to reconstruct a new sequence: In the formula: Output t Indicates time t The output sequence obtained by reconstruction at the point; Seasonal p This represents the seasonal term after nonlinear mapping via a multilayer perceptron. Trend t Indicates time t The long-term trend item at the location; S350 introduces residual connections and layer normalization to accelerate model convergence and stabilize gradient propagation to obtain the final output.

8. The method for estimating the state of charge of a lithium-ion battery based on multimodal data fusion according to claim 1, characterized in that, Step S40 includes the following steps: S410 inputs the preprocessed and feature-engineered ultrasonic feature sequences and electrical data in parallel into the two branch entrances of the network; S420 extracts features from the input data based on two MSCNNs with the same and independent structures, mapping the original input data sequence into a deep feature representation. S430, the fused features are input into the time series prediction module for modeling. The fused features are input into the time series prediction module based on sequence decomposition for trend term extraction, seasonal term separation, nonlinear mapping, and sequence reconstruction to obtain the prediction output features.

9. The method for estimating the state of charge of a lithium-ion battery based on multimodal data fusion according to claim 1, characterized in that, Step S50 includes the following steps: S510, input the predicted output features of the time series prediction module into the output layer for feature mapping to obtain the corresponding SOC estimate; S520, The SOC estimate is processed so that the processed SOC estimate corresponds to the range of the actual state of charge of the battery. S530 outputs the processed result as the final SOC estimate of the lithium-ion battery at the current moment.

Images