A Smart Online Estimation Method for SOC of Lithium-ion Batteries Based on Mechanical Stress Characteristics

By considering the interaction between mechanical stress and other physical quantities, the BiMamba-X neural network model is used to perform SOC estimation of lithium batteries, which solves the problem that the impact of mechanical stress in the prior art is not considered, and improves the accuracy and stability of SOC estimation of lithium-ion batteries. sex.

CN118858992BActive Publication Date: 2024-11-29HENAN INST OF SCI & TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411351875.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-26
Publication Date
2024-11-29
Estimated Expiration
2044-09-26

AI Technical Summary

Technical Problem

The existing SOC estimation method of lithium batteries fails to effectively consider the impact of mechanical stress on the state of charge of lithium-ion batteries, especially under the actual driving conditions of soft-packed lithium-ion batteries, resulting in insufficient estimation accuracy.

Method used

Using the intelligent online estimation method of lithium battery SOC based on mechanical stress characteristics, the mechanical stress, temperature, current and voltage data are obtained, and the BiMamba-X neural network model is used for processing, combined with the parallel structure of the Mamba-X module, the temperature changes are captured. Mechanical stress and current-voltage characteristics improve estimation accuracy.

Benefits of technology

The accuracy and calculation stability of SOC estimation of lithium-ion batteries are improved, ensuring mutual relationship processing in different time scales and spatial modes, and are suitable for critical state estimation within the battery operating range.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118858992B_ABST
    Figure CN118858992B_ABST
Patent Text Reader

Abstract

The invention discloses an intelligent online estimation method for SOC of a lithium battery based on mechanical stress characteristics, comprising the following steps: obtaining data of mechanical stress, temperature, current and voltage of a soft-pack battery during charging and discharging, and recording them as raw data; sending the raw data to a first linear layer for processing, thereby converting the input dimension into a hidden layer dimension and obtaining conversion data; sending the conversion data to a residual layer, and obtaining a battery SoC estimation after the output of the residual layer is sequentially subjected to a root mean square normalization layer, a second linear layer and a Sigmoid activation function; taking temperature, current, voltage and mechanical stress measurement values ​​as input, and capturing the temperature change characteristics, mechanical stress characteristics and current-voltage characteristics of a soft-pack lithium-ion battery during charging and discharging by using a BiMamba‑X neural network model, thereby avoiding information loss and excessive compression, so that the model can retain more original information, focus on key states within the working range of the battery, and improve estimation accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of battery management systems, and in particular to a lithium battery SOC intelligent online estimation method based on mechanical stress characteristics. Background Art

[0002] Lithium batteries have attracted widespread attention due to their excellent energy storage density, power tolerance and low cost. They are now widely used in consumer electronics, power battery systems and large-scale energy storage systems. Battery charge capacity (SoC) is an important parameter in the lithium battery cycle, which characterizes the remaining capacity of the battery in the charge and discharge cycle. Accurate estimation of SoC is crucial and has important significance for the reliable driving of electric vehicles and the state estimation and decision-making of energy storage systems.

[0003] SoC value changes with the charge and discharge cycle, and is affected by temperature, mechanical stress and battery aging. However, SoC cannot be measured directly, and must be estimated approximately by measuring some other physical quantities and using certain mathematical models or algorithms. At present, the commonly used SoC estimation methods are roughly as follows: ampere-hour integration method, open circuit voltage method, Kalman filter method and neural network method. The ampere-hour integration method is simple and easy to use, but due to the error in current measurement, the error gradually increases with time accumulation, and the method cannot give the initial SoC value; the open circuit voltage method establishes the mapping relationship between OCV and SoC, but OCV can only be obtained when the battery is left to stand for a certain period of time to reach a balanced state, which hinders the application of the above methods in electric vehicles; the Kalman filter method relies on the accuracy of the established circuit equivalent model, and has weak adaptability to various battery types and working conditions in the real world; the neural network method can automatically summarize, learn and organize the collected data to obtain the internal relationship of these data. In the SoC estimation method, the data that can be easily collected mainly include current, voltage and temperature. At present, most battery SoC estimation methods combine these three data that can be directly obtained for estimation, while some other important data, such as mechanical stress, can only be obtained under specific conditions. This ubiquitous stress mainly comes from the volume change, external pressure and material structure change during the battery charging and discharging process. The existing battery SoC estimation method ignores the impact of mechanical stress on the estimation of lithium-ion battery state of charge (SoC). In addition, compared with lithium-ion batteries with metal shells, the performance of soft-pack lithium-ion batteries is more seriously affected by mechanical stress under actual driving conditions. Therefore, the stress changes caused by external mechanical loads that affect battery operation should also be considered in SoC estimation. Summary of the invention

[0004] In view of the problems in the prior art, the present invention provides an intelligent online estimation method for lithium battery SOC based on mechanical stress characteristics, aiming to comprehensively consider the interaction of parameters such as mechanical stress, temperature, current and voltage, and improve the accuracy of SoC estimation.

[0005] The intelligent online estimation method of lithium battery SOC based on mechanical stress characteristics includes the following steps:

[0006] Step 1: Obtain the data of mechanical stress, temperature, current and voltage of the soft-pack battery during the charging and discharging process, and record them as raw data; specifically:

[0007] Step 1.1: clamp the soft-pack battery by a mechanical stress testing device, and obtain uniform mechanical stress between the mechanical stress testing device and the outer surface of the soft-pack battery;

[0008] Step 1.2: Place the mechanical stress test device and the soft-pack battery in a constant temperature box, and then collect the temperature of the negative electrode ear of the soft-pack battery through a temperature sensor;

[0009] Step 1.3: Test the soft pack battery through the battery testing equipment and obtain the current and voltage data of the soft pack battery;

[0010] Step 2: Send the original data to the first linear layer for processing, thereby converting the input dimension to the hidden layer dimension and obtaining the converted data;

[0011] Step 3: Send the converted data to the residual layer. The residual layer includes a first residual module and a second residual module. The processing of the converted data in the residual layer includes the following steps:

[0012] Step 3.1: After the converted data passes through the RMS normalization layer in the first residual module and the first Mamba-X module, the converted data is directly added to the output of the first Mamba-X module through a residual connection to obtain the first superimposed data;

[0013] Step 3.2: After the first superimposed data passes through the RMS normalization layer in the second residual module and the second Mamba-X module, the output of the second Mamba-X module is superimposed on the first superimposed data through a residual connection to obtain second superimposed data;

[0014] Step 4: The second superimposed data is sequentially passed through the RMS normalization layer, the second linear layer, and the Sigmoid activation function to obtain the SoC estimation of the battery.

[0015] Further, the first Mamba-X module and the second Mamba-X module have the same structure and both include a first linear projection layer, a second linear projection layer and a third linear projection layer; the data processing process in the first Mamba-X module and the second Mamba-X module is to send the input data to the first linear projection layer and the second linear projection layer respectively, and the output of the first linear projection layer is sent to the first activation function SiLU; the output of the second linear projection layer is successively sent to the second activation function SiLU after convolution Conv and normalization Norm, the output of the first activation function SiLU and the output of the second activation function SiLU are added and processed by the SMM layer and the third activation function SiLU respectively, the output of the SMM layer and the output of the third activation function SiLU are multiplied and combined and sent to the third linear projection layer, and the output of the third linear projection layer is the output of the first Mamba-X module or the second Mamba-X module.

[0016] Further: in step 1, the initial preload force between the mechanical stress testing device and the soft-pack battery is set to 400N, and several soft-pack batteries are charged to 3.65 V at different rates with constant current and constant voltage, and cut off at 0.05 C, then left for 2 h, discharged at constant current to a cut-off voltage of 2.0 V, and then left for 2 h; wherein, the soft-pack battery is charged and discharged at rates of 0.5 C, 1 C, 1.5 C and 2 C at constant temperatures of 10 ℃, 25 ℃ and 40 ℃, respectively.

[0017] Further: the mechanical stress testing device includes two fixed steel plates and two clamping plates made of semi-rigid materials, the side dimensions of the clamping plates are the same as the side dimensions of the soft-pack battery, the two clamping plates are clamped on the soft-pack battery and overlap with its side surfaces; the two clamping plates are clamped between the two fixed steel plates, a pressure sensor is arranged between one of the clamping plates and the fixed steel plates, the clamping plate is fitted with the pressure sensor, the pressure sensor is fixed on the corresponding fixed steel plate, and the other clamping plate is fitted with the corresponding fixed steel plate; the spacing between the two fixed steel plates remains unchanged.

[0018] Further, the pressure sensor is embedded in the corresponding fixed steel plate, and another clamping plate is fitted with the corresponding fixed steel plate; after the two clamping plates clamp the soft-pack battery and the pressure sensor, the two fixed steel plates are fixed by four bolts and nuts around them, so that the distance between the two fixed steel plates remains unchanged.

[0019] Further: The model of the pressure sensor is DYMH-102, and it is made of high-precision strain gauge and 42CrMoA alloy steel.

[0020] Further, the surface of the pressure sensor is nickel-plated.

[0021] The beneficial effects of the present invention are as follows: taking temperature, current, voltage and mechanical stress measurement values ​​as input, by utilizing the BiMamba-X neural network to capture the temperature change characteristics, mechanical stress characteristics and current-voltage characteristics of soft-pack lithium-ion batteries during the charging and discharging process, and by using the parallel structure of the Mamba-X module, avoiding information loss and excessive compression, so that the model can retain more original information, while processing the relationship between different time scales, spatial patterns or characteristics, focusing on the key states within the working range of the battery, not only improving the estimation accuracy but also ensuring the calculation stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The process of the present invention;

[0023] Figure 2 It is a structural block diagram of the present invention;

[0024] Figure 3 The mechanical stress testing device of the present invention;

[0025] Figure 4 SoC estimation and error results of the BiMamba-X model at 25°C and without mechanical stress, (a) 0.5°C, (b) 1°C, (c) 1.5°C, (d) 2°C;

[0026] Figure 5 SoC estimation and error results of the BiMamba-X model at 25°C and mechanical stress conditions, (a) 0.5°C, (b) 1°C, (c) 1.5°C, (d) 2°C

[0027] Figure 6 SoC estimation and error results of the BiMamba-X model at 10°C and mechanical stress conditions, (a) 0.5°C, (b) 1°C, (c) 1.5°C, (d) 2°C;

[0028] Figure 7 SoC estimation and error results of the BiMamba-X model at 40°C and mechanical stress conditions, (a) 0.5°C, (b) 1°C, (c) 1.5°C, (d) 2°C;

[0029] Figure 8 It is the structural diagram of the RNN model;

[0030] Fig. 9 This is the structural diagram of the LSTM model;

[0031] Fig.10SoC estimation and error results of the LSTM model at 25°C, 1s data sampling interval and mechanical stress conditions, (a) 0.5°C, (b) 1°C, (c) 1.5°C, (d) 2°C;

[0032] Fig.11 SoC estimation and error results of the RNN model at 25°C, 1s data sampling interval and mechanical stress conditions, (a) 0.5°C, (b) 1°C, (c) 1.5°C, (d) 2°C;

[0033] Fig.12 SoC estimation and error results of the BiMamba-X model at 25°C, 1s data sampling interval and mechanical stress conditions, (a) 0.5°C, (b) 1°C, (c) 1.5°C, (d) 2°C;

[0034] Fig.13 SoC estimation and error results of the LSTM model at 25°C, 10s data sampling interval and mechanical stress conditions, (a) 0.5°C, (b) 1°C, (c) 1.5°C, (d) 2°C;

[0035] Fig.14 SoC estimation and error results of the RNN model at 25°C, 10s data sampling interval and mechanical stress conditions, (a) 0.5°C, (b) 1°C, (c) 1.5°C, (d) 2°C;

[0036] Fig.15 SoC estimation and error results of the BiMamba-X model at 25°C, 10s data sampling interval, and mechanical stress conditions, (a) 0.5°C, (b) 1°C, (c) 1.5°C, and (d) 2°C. DETAILED DESCRIPTION

[0037] The present invention is described in detail below in conjunction with the accompanying drawings. The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention. The directional terms such as left, middle, right, top, and bottom in the embodiments of the present invention are only relative concepts or are based on the normal use state of the product, and should not be considered as restrictive.

[0038] The intelligent online estimation method of lithium battery SOC based on mechanical stress characteristics includes the following steps:

[0039] Step 1: Obtain the data of mechanical stress, temperature, current and voltage of the soft-pack battery during the charge and discharge process, and record them as raw data. By capturing multiple parameters, a comprehensive data set reflecting the working state of the battery can be obtained, thereby achieving a more accurate SoC estimation; the initial preload force between the mechanical stress test device and the soft-pack battery is set to 400 N, and several soft-pack batteries are charged to 3.65 V at different rates with constant current and constant voltage, and cut off at 0.05 C, then left for 2 h, and discharged at constant current to a cut-off voltage of 2.0 V, and then left for 2 h; Among them, the soft-pack battery is charged and discharged at rates of 0.5 C, 1 C, 1.5 C and 2 C at constant temperatures of 10 ℃, 25 ℃ and 40 ℃, respectively; specifically:

[0040] Step 1.1: clamp the soft-pack battery by a mechanical stress testing device, and obtain uniform mechanical stress between the mechanical stress testing device and the outer surface of the soft-pack battery;

[0041] Step 1.2: Place the mechanical stress test device and the soft-pack battery in a constant temperature box, and then collect the temperature of the negative electrode ear of the soft-pack battery through a temperature sensor;

[0042] Step 1.3: Test the soft-pack battery through the battery testing equipment and obtain the current and voltage data of the soft-pack battery; divide the original data into a training set, a validation set and a test set;

[0043] Six batteries of the same model were tested simultaneously to reduce the uncertainty of the test. The detailed parameters of the batteries are shown in Table 1. The charging and discharging experiments are designed to obtain data on mechanical stress, temperature, current, and voltage of lithium batteries during the charging and discharging process;

[0044] Table 1 Battery Detailed Parameters

[0045]

[0046] Step 2: Send the original data to the first linear layer for processing, so as to convert the input dimension into the hidden layer dimension and obtain the converted data; the process of converting the input data into the hidden layer dimension can not only reduce the redundancy of the data, but also fuse the feature information, ensuring that the subsequent network can better learn and generalize complex patterns;

[0047] Step 3: Send the converted data to the residual layer. The residual layer includes a first residual module and a second residual module. The processing of the converted data in the residual layer includes the following steps:

[0048] Step 3.1: After the converted data passes through the RMS normalization layer in the first residual module and the first Mamba-X module, the converted data is directly added to the output of the first Mamba-X module through a residual connection to obtain the first superimposed data;

[0049] Step 3.2: After the first superimposed data passes through the RMS error normalization layer in the second residual module and the second Mamba-X module, the output of the second Mamba-X module is superimposed on the first superimposed data through a residual connection to obtain the second superimposed data; similar to the first residual module, multiple residual modules are superimposed to enhance the gradient flow, ensure stable training, and capture more complex features. The second residual module ensures that the output of the previous layer in the deep network is not lost, enhances the generalization ability of the model, and prevents the gradient from disappearing;

[0050] Step 4: The second superimposed data is passed through the RMS normalization layer, the second linear layer, and the Sigmoid activation function in sequence to obtain the SoC estimation of the battery; the RMS normalization layer ensures that the data is in a good state before the final estimation, the second linear layer helps to adjust the output to a dimension suitable for SoC estimation, and the Sigmoid activation function maps the output to a probability distribution (between 0 and 1), which is suitable for SoC estimation.

[0051] Among them, the mechanical stress testing device includes two fixed steel plates and two clamping plates made of semi-rigid materials. The side dimensions of the clamping plates are the same as the side dimensions of the soft-pack battery. The two clamping plates are clamped on the soft-pack battery and overlap with its side. The two clamping plates are clamped between the two fixed steel plates. A pressure sensor is arranged between one of the clamping plates and the fixed steel plate. The clamping plate is fitted with the pressure sensor. The pressure sensor is fixed on the corresponding fixed steel plate. For easy installation, the pressure sensor is embedded in the corresponding fixed steel plate, and the other clamping plate is fitted with the corresponding fixed steel plate. The spacing between the two fixed steel plates remains unchanged. After clamping the soft-pack battery and the pressure sensor, the fixed steel plate is fixed by four bolts and nuts around it, so that the spacing between the two fixed steel plates remains unchanged. The pressure sensor is DYMH-102, which is made of high-precision strain gauges and 42CrMoA alloy steel. It has good heat resistance, small creep, small thermal hysteresis, good long-term stability, and surface nickel plating for corrosion resistance, which is suitable for a variety of working conditions.

[0052] The first Mamba-X module and the second Mamba-X module have the same structure and both include a first linear projection layer, a second linear projection layer and a third linear projection layer; the data processing process in the first Mamba-X module and the second Mamba-X module is as follows: the input data is sent to the first linear projection layer and the second linear projection layer respectively, and the output of the first linear projection layer is sent to the first activation function SiLU; the output of the second linear projection layer is successively sent to the second activation function SiLU after convolution Conv and normalization Norm, the output of the first activation function SiLU and the output of the second activation function SiLU are added and processed by the SMM layer and the third activation function SiLU respectively, the output of the SMM layer and the output of the third activation function SiLU are multiplied and combined and sent to the third linear projection layer, and the output of the third linear projection layer is the third activation function SiLU. The output of the first Mamba-X module or the second Mamba-X module; the normalization layer standardizes the input data to keep its mean and variance consistent, ensuring that the subsequent layers are stable when processing input data of different scales, preventing gradient explosion or disappearance, and enhancing the training stability and convergence speed of the model; the linear layer projection layer is used to map multi-dimensional data to a low-dimensional space, compress information for further processing, reduce dimensions, and extract important features for output prediction; the activation function SiLU combines the linearity of ReLU and the smoothness of Sigmoid to provide nonlinear transformation, which helps the model better capture complex patterns while maintaining a smooth transition of values ​​and avoiding gradient disappearance; the convolution layer can extract features with local dependencies in space, which is particularly suitable for processing time series or two-dimensional data; the normalization layer further ensures the convergence of training. The SSM layer captures patterns and trends over long time spans or large ranges, which helps improve long-term predictions of battery behavior. The residual connection allows model training to be more efficient by adding the input directly to the output, and makes the optimization of deep networks smoother while maintaining the transmission of original input features. This architectural design solves the shortcomings of the model in terms of nonlinear feature expression (mechanical stress), time-dependent feature capture, gradient vanishing, training stability, information loss, etc. By introducing multiple uses of the SiLU activation function, long-term dependency processing of the SMM layer, and dual-path information flow, the performance of the model is enhanced, which is especially suitable for processing complex, multi-dimensional, and multi-time-scale tasks such as battery SoC estimation.

[0053] The present invention is experimented and tested, comprising the following steps:

[0054] Step 5: Configure the initial parameters of the BiMamba-X model, including the iteration round, learning rate, weight, number of hidden layer neurons, etc. These parameters can be set through command line parameters or configuration files;

[0055] Step 6: Use the Adam optimization algorithm. The key formula of the Adam optimizer is as follows:

[0056] Compute the exponential moving average of the gradient:

[0057]

[0058] in It's time The first moment estimate of (i.e., the mean of the gradient), It's time The gradient of Usually close to 1 (e.g. 0.9);

[0059] Compute the exponential moving average of the squared gradients:

[0060]

[0061] in It's time The second moment estimate of (i.e., the uncentered variance of the gradient), Also usually close to 1 (e.g. 0.999)

[0062] right and To perform bias correction:

[0063]

[0064]

[0065] here Indicates the number of iterations. This correction can be reduced in the early stage and The deviation makes the estimation more accurate in the early stage of training;

[0066] Update parameters:

[0067]

[0068] in is the parameter at time The value of is the learning rate, is a very small number (e.g. 1e-8) to prevent the denominator from being zero;

[0069] Select and define an appropriate loss function. Here, we use the L1 loss function, which calculates the absolute error between the estimated value and the true value. The calculation formula is shown in (6). The choice of loss function will directly affect the convergence speed and final effect of the model.

[0070]

[0071] in yes The true value of the moment, yes The estimated value of the time, Indicates the total number of data;

[0072] Step 7: Perform forward propagation of the network. Forward propagation refers to the process of starting from the input layer, passing through a series of hidden layers, and outputting the results to the output layer. In each layer, the input is linearly combined with the weight of the corresponding layer, and then nonlinearly transformed by the activation function to obtain the output and pass it to the next layer. This process can be expressed as:

[0073]

[0074] in, Indicates The output of the layer, is the activation function of this layer, is the weight matrix of this layer, is the output of the previous layer, is the bias of the layer;

[0075] Step 8: Calculate the error (loss) of the output layer, and then perform back propagation to derive the partial derivative of the error. The back propagation formula is shown in (8). The loss value is calculated using the defined loss function (L1). The Adam optimization method is used to update the network weights and biases. The optimizer updates the parameters of each layer based on the calculated gradient information, gradually reducing the loss value.

[0076]

[0077] in, is the loss function, and are any two vectors, and They are about The gradient and about The gradient is:

[0078] Step 9: Use different evaluation indicators (such as MSE, RMSE, MAE, R2) to evaluate the performance of the model and measure the generalization ability and estimation accuracy of the model by its performance on the test set;

[0079] The performance of the BiMamba-X model is verified mainly through four important evaluation indicators, namely, mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2 ), their formulas are as follows:

[0080]

[0081]

[0082]

[0083]

[0084] in, Indicates the total number of data. and They are The true value and estimated value of SoC at a moment. MSE evaluates the overall error of the model through squared error. RMSE is used to measure the discreteness of the model error. Both are sensitive to large errors. The smaller the value, the higher the model accuracy. MAE is used to evaluate the average magnitude of the model error. The smaller the value, the higher the model accuracy. R² reflects the degree of explanation of the model to the data. The closer the value is to 1, the stronger the model's ability to explain the data and the higher the accuracy.

[0085] in, Figure 4 The estimation of battery SoC and the corresponding error results are shown without mechanical stress input. Figure 5 The estimates and errors are shown when the model includes mechanical stress data. The data in Table 2 further confirm that the model considering mechanical stress is superior to the model without mechanical stress in all evaluation indicators. For example, at a 1C discharge rate, the MAE with mechanical stress input is 0.30%, compared with 0.73% without mechanical stress input, and the error is reduced by 58.9%. In addition, the improvement in the coefficient of determination (R²) also shows that the accuracy and reliability of the prediction of the model including mechanical stress data have been significantly enhanced, and the R² value has been improved to 0.9998. These results highlight the importance of considering battery mechanical stress in battery management systems.

[0086] Table 2 Comparison of experimental results with and without mechanical stress conditions

[0087]

[0088] In order to verify the generalization performance of the BiMamba-X network model, this study estimated the SoC of soft-pack lithium batteries discharged at two temperature conditions of 10°C and 40°C, and evaluated the model performance at different discharge rates (0.5C, 1C, 1.5C, 2C). Figure 6 and Figure 7The estimation results and errors at 10°C and 40°C are shown respectively. Table 3 lists the specific data at each temperature in detail. The experimental results reveal the significant impact of temperature on the accuracy of SoC estimation. Compared with the normal temperature environment of 25°C, the evaluation indicators of the model in the low temperature environment of 10°C and the high temperature environment of 40°C are slightly increased. However, even under different temperature conditions, the maximum MSE of the model is 0.0049%, the maximum RMSE is 0.54%, the maximum MAE is 0.49%, and the minimum R² is 0.9996. Specifically, at 10°C, the average MSE of the model at each discharge rate is 0.0029%, the RMSE is 0.41%, the MAE is 0.37%, and the R² is 0.9996. At 40°C, the average MSE of the model at each discharge rate is 0.0027%, the RMSE is 0.52%, the MAE is 0.39%, and the R² is 0.9996. These results show that the BiMamba-X model can maintain high accuracy at different temperatures and discharge rates, although temperature changes will have a certain impact on the accuracy of SoC estimation. The average MSE, RMSE, MAE and R² of the model are 0.0028%, 0.47%, 0.33% and 0.9996, respectively, demonstrating its excellent prediction performance and robustness under different environments and discharge rates. This further demonstrates the wide applicability and reliability of the BiMamba-X model under various conditions. Through these evaluations, it can be seen that the BiMamba-X network model exhibits stable performance at different temperatures and can effectively cope with the challenges of temperature changes to lithium battery SoC estimation. This provides strong support for its widespread promotion in practical applications.

[0089] Table 3 Comparison of experimental results under different temperature conditions

[0090]

[0091] Under 25°C and different discharge rates, this paper compares the performance of three different network models (LSTM, RNN and BiMamba-X) for lithium battery SoC estimation at different data sampling intervals. LSTM and RNN are classic deep learning models commonly used to process time series data. Since SOC estimation involves battery time series data (such as voltage, current, etc.), they are suitable comparison objects. Among them, RNN is a basic recurrent neural network that can capture the dependencies in sequence data, but there is a long-term dependency problem, that is, it cannot handle the gradient vanishing of long sequence data well; LSTM, as an improved version of RNN, solves the long-term dependency problem of RNN by introducing a gating mechanism (input gate, forget gate, output gate), so it performs better when processing long time series data, especially for complex sequence dependencies in battery SOC estimation. By selecting these two classic models, the advantages of the improved BiMamba-X model over these traditional methods in processing battery SOC estimation tasks can be effectively demonstrated, especially in processing long sequences and complex data patterns.

[0092] The RNN and LSTM unit structures are as follows Figure 8 and Fig. 9 As shown. The basic structure of RNN includes the input sequence and hidden state , respectively represent the time step The input and through the recursive formula, RNN in time step The hidden state at a certain moment is determined by the hidden state at the previous moment and the current input. RNN updates the hidden state using formula (13):

[0093]

[0094] in, is the weight matrix input to the hidden layer, is the weight matrix from hidden layer to hidden layer, is the bias term, tanh is the activation function, and tanh or ReLu function is usually used.

[0095] Output It is usually determined by the following formula (14):

[0096]

[0097] in, is the weight matrix from the hidden layer to the output layer, is the bias term.

[0098] LSTM introduces three key gates: forget gate, input gate, output gate, and a cell state for memory. Through the gate mechanism, LSTM can decide which information to keep, update, and ignore at each time step. The following is the working principle of LSTM:

[0099] Forget gate: controls which information will be discarded from the cell state, as shown in formula 15.

[0100]

[0101] in, Control the proportion of forgetting, and are weights and biases, is the sigmoid activation function.

[0102] Input gate: controls what new information will be added to the cell state, as shown in formula 16.

[0103]

[0104] in, Control the input ratio, new candidate memory information:

[0105]

[0106] Update the status unit:

[0107]

[0108] Unit Status The unit state of the previous step It is obtained after adjustment by the forget gate and weighted combination of the new input information.

[0109] Output gate: controls the hidden state output of the current time step. The formula is:

[0110]

[0111] The final hidden state:

[0112]

[0113] Figure 10-12The estimation and error results of the three models under the conditions of 25°C and 1 s data sampling interval are shown, and the specific evaluation index data are shown in Table 5. Under the data sampling interval of 1 s, the LSTM, RNN and BiMamba-X models all performed well, among which the BiMamba-X model performed particularly well. The maximum MSE, RMSE and MAE indicators of the BiMamba-X model at different rates were 0.0023%, 0.48% and 0.30%, respectively, while the maximum MSE, RMSE and MAE indicators of the LSTM model were 0.0094%, 0.97% and 0.86%, respectively, and the maximum MSE, RMSE and MAE indicators of the RNN model were 0.0104%, 1.02% and 0.94%, respectively, and the R² value of the BiMamba-X model was closer to 1, indicating that the BiMamba-X model had the best prediction accuracy and stability under different discharge rates. In contrast, the error indicators of LSTM and RNN were slightly higher, but they still maintained a high prediction accuracy, which was suitable for SOC estimation tasks with high-frequency data sampling.

[0114] Table 4 Comparison results of different models at 1s data sampling interval

[0115]

[0116] Figure 13-15 The estimation and error results of the three models under the conditions of 25℃ and 10 s data sampling interval are shown, and the specific evaluation index data are shown in Table 5. Under the data sampling interval of 10 s, the performance of each model has declined, but the BiMamba-X model still maintains the lowest error and higher R² value. The maximum MSE, RMSE and MAE indicators of the BiMamba-X model at different magnifications are 0.0028%, 0.53% and 0.39%, respectively, which are 18%, 9% and 23% higher than the results under the data sampling interval of 1 s, but overall it shows good adaptability and robustness to lower frequency data sampling. LSTM and RNN have increased in MSE and RMSE indicators, especially at higher magnifications, the errors are more obvious, which shows that they are more sensitive to the data sampling interval.

[0117] In summary, the BiMamba-X model shows excellent prediction accuracy and stability under both high-frequency and low-frequency data sampling conditions. Its powerful data processing and generalization capabilities make it the preferred model for SoC estimation under different sampling conditions. In contrast, although the LSTM and RNN models perform well under high-frequency sampling, the significant increase in errors at low-frequency sampling limits their application scope. Therefore, in actual application scenarios that pursue high-precision SoC estimation and need to cope with variable sampling conditions, the BiMamba-X model undoubtedly has higher application value.

[0118] Table 5 Comparison results of different models under 10s data sampling interval

[0119]

[0120] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An intelligent online estimation method for lithium battery SOC based on mechanical stress characteristics, characterized by: The following steps are involved: Step 1: Obtain the data of mechanical stress, temperature, current and voltage of the soft-pack battery during the charging and discharging process, and record them as raw data; Specifically: Step 1.1: clamp the soft-pack battery by a mechanical stress testing device, and obtain uniform mechanical stress between the mechanical stress testing device and the outer surface of the soft-pack battery; Step 1.2: Place the mechanical stress test device and the soft-pack battery in a constant temperature box, and then collect the temperature of the negative electrode ear of the soft-pack battery through a temperature sensor; Step 1.3: Test the soft pack battery through the battery testing equipment and obtain the current and voltage data of the soft pack battery; Step 2: Send the original data to the first linear layer for processing, thereby converting the input dimension to the hidden layer dimension and obtaining the converted data; Step 3: Send the converted data to the residual layer. The residual layer includes a first residual module and a second residual module. The processing of the converted data in the residual layer includes the following steps: Step 3.1: After the converted data passes through the RMS normalization layer in the first residual module and the first Mamba-X module, the converted data is directly added to the output of the first Mamba-X module through a residual connection to obtain the first superimposed data; Step 3.2: After the first superimposed data passes through the RMS normalization layer in the second residual module and the second Mamba-X module, the output of the second Mamba-X module is superimposed on the first superimposed data through a residual connection to obtain second superimposed data; Step 4: The second superposition data is sequentially passed through a root mean square normalization layer, a second linear layer, and a sigmoid activation function to obtain a SoC estimate of the battery; Among them, the structures of the first Mamba-X module and the second Mamba-X module are the same and both include a first linear projection layer, a second linear projection layer and a third linear projection layer; the data processing process in the first Mamba-X module and the second Mamba-X module is as follows: the input data is sent to the first linear projection layer and the second linear projection layer respectively, and the output of the first linear projection layer is sent to the first activation function SiLU; the output of the second linear projection layer is successively sent to the second activation function SiLU after convolution Conv and normalization Norm, the output of the first activation function SiLU and the output of the second activation function SiLU are added and processed by the SMM layer and the third activation function SiLU respectively, the output of the SMM layer and the output of the third activation function SiLU are multiplied and combined and sent to the third linear projection layer, and the output of the third linear projection layer is the output of the first Mamba-X module or the second Mamba-X module.

2. The method for intelligent online estimation of lithium battery SOC based on mechanical stress characteristics according to claim 1, characterized in that: In step 1, the initial preload force between the mechanical stress test device and the soft-pack battery is set to 400 N, and several soft-pack batteries are charged to 3.65 V at different rates with constant current and constant voltage, and cut off at 0.05 C, then left for 2 h, discharged at constant current to a cut-off voltage of 2.0 V, and then left for 2 h; among them, the soft-pack batteries are charged and discharged at rates of 0.5 C, 1 C, 1.5 C and 2 C at constant temperatures of 10 ℃, 25 ℃ and 40 ℃, respectively.

3. The method for intelligent online estimation of lithium battery SOC based on mechanical stress characteristics according to claim 1, characterized in that: The mechanical stress testing device includes two fixed steel plates and two clamping plates made of semi-rigid materials. The side dimensions of the clamping plates are the same as the side dimensions of the soft-pack battery. The two clamping plates are clamped on the soft-pack battery and overlap with its side surfaces. The two clamping plates are clamped between the two fixed steel plates. A pressure sensor is arranged between one of the clamping plates and the fixed steel plates. The clamping plate is fitted with the pressure sensor. The pressure sensor is fixed on the corresponding fixed steel plate. The other clamping plate is fitted with the corresponding fixed steel plate. The spacing between the two fixed steel plates remains unchanged.

4. The method for intelligent online estimation of lithium battery SOC based on mechanical stress characteristics according to claim 3 is characterized in that: The pressure sensor is embedded in the corresponding fixed steel plate, and another clamping plate is fitted with the corresponding fixed steel plate; after the two clamping plates clamp the soft-pack battery and the pressure sensor, the two fixed steel plates are fixed by four bolts and nuts around them, so that the distance between the two fixed steel plates remains unchanged.

5. The method for intelligent online estimation of lithium battery SOC based on mechanical stress characteristics according to claim 3 is characterized in that: The model of the pressure sensor is DYMH-102, and it is made of high-precision strain gauge and 42CrMoA alloy steel.

6. The method for intelligent online estimation of lithium battery SOC based on mechanical stress characteristics according to claim 5, characterized in that: The surface of the pressure sensor is nickel plated.

Citation Information

Patent Citations

  • Soft package lithium ion battery SOC estimation method and system considering stress influence

    CN115079008A

  • Lithium ion battery SOC estimation method based on neural network

    CN117313532A