Lithium battery health state monitoring method based on electroacoustic data fusion

Abstract

The invention discloses a lithium battery health state monitoring method based on electroacoustic data fusion, and the method comprises the steps: 1), collecting original data, 2), extracting and screening guided wave signal features from the collected original signal data, and carrying out the time scale transformation of the guided wave features, and obtaining a guided wave feature subset which is high in correlation with SOH, is high in sensitivity, and is low in redundancy; 3) building a data driving model; and 4) verifying the effectiveness of the electroacoustic fusion method. According to the invention, by fusing the electrical data and the acoustic guided wave characteristics, the internal and external degradation characteristics of the battery can be captured more comprehensively; the electrical data reflects the state change of the active material of the battery, and the acoustic guided wave characteristics reveal the change of the internal structure of the battery, so that the internal and external degradation characteristics of the battery are represented more comprehensively, the accuracy of SOH estimation of the data-driven model is improved, and the SOH estimation precision is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of manufacturing, and particularly relates to a method for monitoring the state of health of a lithium battery based on electroacoustic data fusion. Background Art

[0002] Lithium-ion batteries have become an indispensable energy solution in modern life due to their high energy density and long lifespan. However, with the widespread application of lithium batteries, their safety issues (such as fires and explosions) have also attracted extensive attention. The battery management system is a key component to ensure the safe and efficient operation of the battery pack, and the state of health (SOH) of the battery is an important parameter for evaluating battery aging. SOH is usually defined as the ratio of the current maximum available capacity of the battery to its nominal capacity. As the battery continues to degrade, when the capacity retention rate drops to 70-80%, the safety risk of the lithium battery increases significantly. Therefore, accurate SOH estimation is crucial for ensuring the safety and performance of lithium batteries.

[0003] Existing SOH estimation methods can be mainly divided into traditional model-based methods and data-driven methods. Model-based methods include electrochemical models and equivalent circuit models. Electrochemical models describe the distributed behavior of battery states by quantifying internal physical and chemical phenomena, while equivalent circuit models represent the dynamic behavior of the battery using various electrical components and configurations. These traditional model-based SOH estimation methods face challenges in computational complexity and accuracy in practical applications. Data-driven methods have received much attention due to their high precision, flexibility, and independence from physical and chemical mechanisms. For example, methods such as long short-term memory recurrent neural networks (LSTM) and deep convolutional neural networks (DCNN) perform well in capturing long-term data dependencies and capacity degradation patterns during battery aging. However, these data-driven methods that only rely on electrical data (such as current and voltage) may not fully characterize the degradation mechanisms inside the battery, and this single technical means limits the accuracy of SOH estimation.

[0004] In recent years, ultrasonic technology has received extensive attention in the field of battery state estimation, mainly due to its low cost, real-time performance, and non-destructive characteristics. By monitoring the real-time ultrasonic propagation characteristics inside the battery, physical and chemical state information inside the battery can be effectively obtained, thus providing new method support for battery SOH estimation. Although ultrasonic technology shows certain potential in battery SOH estimation, the research in this field is still in its infancy, and no relatively systematic SOH estimation method has been formed. Aiming at the limitation of only relying on electrical data for SOH estimation in the prior art, this case mainly solves the following problems:

[0005] 1) Existing methods cannot fully characterize the internal degradation mechanism of the battery: Traditional data-driven SOH estimation methods based on electrical data cannot fully capture the complex internal degradation mechanism of the battery, resulting in limited estimation accuracy.

[0006] 2) The research and application of ultrasonic technology in SOH estimation are insufficient: Although ultrasonic technology shows potential in battery state monitoring, its application in SOH estimation in existing research is still insufficient, and a systematic SOH estimation model has not been formed. Summary of the Invention

[0007] Object of the Invention: To overcome the above deficiencies, the object of the present invention is to provide a lithium battery health state monitoring method based on electro-acoustic data fusion. By fusing electrical data and acoustic wave guide characteristics, it can more comprehensively capture the internal and external degradation characteristics of the battery; electrical data reflects the state changes of battery active materials, while acoustic wave guide characteristics reveal the changes in the internal structure of the battery, aiming to more comprehensively characterize the internal and external degradation characteristics of the battery, thereby improving the accuracy of data-driven model SOH estimation and significantly enhancing the SOH estimation accuracy.

[0008] Technical Solution: To achieve the above object, the present invention provides a lithium battery health state monitoring method based on electro-acoustic data fusion, including:

[0009] S1): First, collect the original signal data through a digital oscilloscope, and at the same time collect the electrical data of the battery through a battery test system;

[0010] S2): Extract and screen the wave guide signal characteristics from the original signal data collected in S1), and perform time-scale transformation processing on the wave guide characteristics to obtain a wave guide feature subset that meets the requirements of high correlation, high sensitivity, and low redundancy with SOH;

[0011] S3): Build a data-driven model. Suppose the wave guide feature subset contains p features, then 2 p -1 different combinations of wave guide acoustic characteristics can be constructed; these combinations are respectively fused with electrical parameters (current, voltage), and the case of using only electrical parameters is used as a reference control group. Finally, 2 p data fusion schemes can be obtained. These synchronously collected electro-acoustic data are all from the time-series change curves during the complete charging process, including the dynamic characteristic information of the entire battery charging process;

[0012] S4): Verify the effectiveness of the electro-acoustic fusion method.

[0013] In the method for monitoring the health state of a lithium battery based on electroacoustic data fusion described in the present invention, the specific process of extracting and screening guided wave signal features from the original signal data collected in S1) in S2) is as follows: S21): First, preprocess the original signal data collected in S1): First, centralize the original waveform, subtract the mean value of the original signal to remove the DC component of the signal, make the average value of the signal zero, and thus remove the baseline shift in the signal;

[0014] S22): Take the average of every 10 adjacent guided wave signals to achieve preliminary filtering;

[0015] S23): Finally, use the Butterworth filtering algorithm to further perform noise reduction processing, and improve the signal-to-noise ratio by optimizing the filtering window length and polynomial order;

[0016] S24): Then analyze the filtered signal. Let s(t) represent the signal amplitude of the guided wave in the time domain, and S(f) represent the spectrum of the guided wave signal after Fourier transform. Its expression is: S(f) = ∫s(t)e -j2πft dt

[0017] where t represents the time variable and f represents the frequency variable;

[0018] By analyzing the time-domain and frequency-domain characteristics of the guided wave, extract the waveform amplitude, flight time, time-domain centroid, energy integral, peak-to-peak value, skewness coefficient, kurtosis coefficient, waveform index, frequency-domain peak value, frequency-domain centroid, and frequency-domain weighted standard deviation, and finally obtain the guided wave feature set shown in Table 1;

[0019] In the table, H(·) represents the function used to calculate the Hilbert envelope of the signal, T s represents the total duration of the signal, and the frequency mean μ f is expressed as:

[0020]

[0021] In the method for monitoring the health state of a lithium battery based on electroacoustic data fusion described in the present invention, the specific process of performing time-scale transformation processing on the guided wave features in S2) is as follows:

[0022] S201): In the battery aging experiment, collect the guided wave signal every 30 s and perform data preprocessing to obtain the signal time-scale waveform;

[0023] S202): Extract features from the signal time-scale waveform to form a continuous and periodic time evolution curve. To analyze the feature changes, divide and superimpose the feature curve by cycle to obtain the evolution trend of the feature curve with the number of cycles in step ③;

[0024] S203): In practical applications, the battery discharge process is complex and affected by the charge-discharge rate and dynamic load. However, the charging data is stable and highly consistent. During the battery aging process, the characteristic curve mainly shows a vertical shift (see Figure 2 Step ④), which is due to internal changes such as the degradation of electrode materials, the change of electrolyte composition, and the accumulation of mechanical stress, affecting the propagation characteristics of guided waves. To quantify this phenomenon, the evolution of the characteristic curve with the battery SOH is characterized by tracking the change in the mean value of the characteristic curve during the charging process. The mean value of the characteristic curve during the charging process is expressed as:

[0025]

[0026] where F i represents the characteristic; n represents the cycle number, with a value range from 1 to N, and N represents the total number of cycles;

[0027] T represents the duration of each cycle of the charging process;

[0028] For the convenience of analysis, the SOH value is expressed as A0. To directly compare characteristics with different units and scales, the minimum-maximum method is used to normalize the characteristic mean and SOH, expressed as:

[0029]

[0030] In the formula, (·) * represents the normalization function.

[0031] In the method for monitoring the health state of a lithium battery based on electroacoustic data fusion described in the present invention, since not all the characteristics extracted from the guided wave signals can effectively characterize the battery SOH, it is necessary to perform systematic feature screening to identify those features with the highest correlation and sensitivity. The specific screening algorithm is as follows:

[0032] First, the Pearson correlation coefficient analysis method is used to evaluate the correlation coefficient ρ ij between the mean value of each characteristic curve and the battery SOH, and the feature selection is further optimized through sensitivity analysis. A linear regression model between the eigenvalue and SOH is established, and the feature sensitivity index β i (i.e., the regression coefficient) is calculated. The expressions are respectively represented as:

[0033]

[0034] In the formula, γ i represents the bias term of the linear regression model corresponding to the feature F i ;

[0035] Secondly, a correlation coefficient threshold |ρ 0i |> 0.9 and a sensitivity threshold |β i|>0.8, filter feature subsets with high correlation and high sensitivity to SOH And calculate the weighted score corresponding to each feature in the feature subset. The weighted score can be expressed as:

[0036]

[0037] From these weighted scores, the feature with the highest score is selected and recorded as F best ;

[0038] Finally, to reduce feature redundancy, the subset is calculated The inter-correlation matrix between features, setting the correlation coefficient threshold between features (|ρ ij |<0.95), if the correlation coefficient between a feature and other features is greater than the threshold, the feature is removed to obtain a feature subset

[0039] To avoid F best The optimal feature subset F is eliminated. s Should be expressed as:

[0040]

[0041] Thus, a subset of waveguide features that meets the requirements of high correlation, high sensitivity and low redundancy is obtained.

[0042] The present invention also includes a battery testing platform for implementing the lithium battery health status monitoring method based on electroacoustic data fusion, including a battery testing system, an ultrasonic testing system and a temperature box, wherein the battery testing system and the ultrasonic testing system are both connected to the temperature box;

[0043] The battery testing system includes a tester for battery charge and discharge detection and a mid-position computer for electrical data acquisition, and the tester and the mid-position computer are both connected to the temperature box;

[0044] The ultrasonic testing system includes a signal generator for generating a five-peak Hanning window tone pulse with a center frequency of 60 kHz and a digital oscilloscope for collecting waveguide signal data. Both the signal generator and the digital oscilloscope are connected to a temperature box.

[0045] In the battery testing platform described in the present invention, the temperature range of the incubator is 0-60° C., ensuring that the battery is at a constant ambient temperature.

[0046] The present invention also includes a data-driven SOH estimation framework for verifying the effectiveness of the electroacoustic fusion method in Claim 1. The framework includes three groups of hybrid neural network models, namely, the fusion model of one-dimensional CNN and bidirectional LSTM with attention mechanism (CNN-BiLSTM-AM), the fusion model of one-dimensional CNN and active state tracking LSTM (CNN-ASTLSTM), and the fusion model of one-dimensional CNN and standard LSTM (CNN-LSTM).

[0047] And a set of independent standard LSTM models are set as benchmark models. These models can more comprehensively capture the complex dynamic characteristics during the battery charging and discharging process by fusing electrical and acoustic features, thereby improving the accuracy of SOH estimation.

[0048] The SOH estimation method of the neural network model based on electroacoustic data fusion according to the present invention is specifically as follows:

[0049] 801): First, fuse the electrical parameter curve and the acoustic guided wave feature curve of the complete charging process to form the multi-dimensional input of the model;

[0050] 802): Use different electroacoustic data input combinations of Battery #1 and Battery #2 (including individual electrical parameters, individual guided wave acoustic features, and electroacoustic fusion features) to train four neural network models (CNN-BiLSTM-AM, CNN-ASTLSTM, CNN-LSTM, and the standard LSTM benchmark model) respectively; each data combination is used as the input of the model to comprehensively evaluate the influence of different features on the SOH estimation performance;

[0051] 803): Use the corresponding data combination of Battery #3 as the input to test the trained model to verify the effectiveness and accuracy of the proposed method.

[0052] From the above technical solutions, the present invention has the following beneficial effects:

[0053] 1. The lithium battery health state monitoring method based on electroacoustic data fusion according to the present invention can more comprehensively capture the internal and external degradation characteristics of the battery by fusing electrical data and acoustic guided wave features; the electrical data reflects the state change of the battery active material, while the acoustic guided wave feature reveals the change of the battery internal structure, aiming to more comprehensively characterize the internal and external degradation characteristics of the battery, thereby improving the accuracy of the data-driven model SOH estimation and significantly enhancing the SOH estimation accuracy; and it has strong adaptability and can still maintain a high estimation accuracy under partial charging data (SOC range 20-100%).

[0054] 2. By introducing acoustic guided wave features, the present invention enhances the stability of model prediction. Acoustic features are highly sensitive to changes in the internal structure of the battery and can effectively handle noise and uncertainties in electrical data. Experimental results show that the electro-acoustic data fusion method performs excellently in uncertainty analysis. The 95% confidence intervals of the CNN-ASTLSTM and CNN-BiLSTM-AM models are narrow and stable, especially in the rapid degradation stage of the battery. In contrast, the prediction stability of the LSTM model is slightly insufficient compared to other models, with a wider confidence interval and obvious fluctuations, but overall, the model still maintains good performance.

[0055] 3. The present invention proposes a systematic feature screening algorithm. By performing correlation and sensitivity analysis, features highly correlated with SOH are selected to avoid feature redundancy and improve the performance and efficiency of data-driven models. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic structural diagram of the experimental platform in the present invention;

[0057] Figure 2 It is a flow chart of the time-scale transformation of guided wave data features in the present invention;

[0058] Figure 3 It is a schematic diagram of the neural network architecture and method framework based on electro-acoustic data fusion in the present invention;

[0059] Figure 4 It is a schematic diagram of the SOH estimation result of the data-driven model in the present invention;

[0060] Figure 5 It is a schematic diagram of the SOH estimation result of partial charge data in the present invention;

[0061] Figure 6 It is a schematic diagram of the stability analysis results of four data-driven models in the present invention; among them, (a) is the schematic diagram of the stability analysis result of the CNN-ASTLSTM model, (b) is the schematic diagram of the stability analysis result of the CNN-BiLSTM-AM model, (c) is the schematic diagram of the stability analysis result of the CNN-LSTM model, and (d) is the schematic diagram of the stability analysis result of the LSTM model. DETAILED DESCRIPTION OF THE INVENTION

[0062] The present invention will be further clarified below with reference to the accompanying drawings and specific embodiments.

[0063] Embodiment 1

[0064] As Figure 1 shown in the figure, a lithium battery health state monitoring method based on electro-acoustic data fusion, characterized in that it includes:

[0065] S1): First, collect the original signal data through a digital oscilloscope, and at the same time collect the electrical data of the battery through a battery test system.

[0066] It should be noted that the data collection method can be as follows: The battery is charged from 3.0V to 4.45V at a current of 1C (4150mA) at 45°C, followed by constant voltage charging until the current drops to C / 20 (207.5mA). After being fully charged, it is left standing for 5 minutes, and then discharged at a current of 1C to 2.8V. There is a 5-minute rest between each charge-discharge cycle. A total of 90 charge-discharge cycles are carried out. During this period, the current, voltage, and guided wave signal data are synchronously collected at 30-second intervals, and the guided wave signal is continuously collected 10 times every 30s. It should be noted that this is just one of the methods, and other appropriate collection methods can be selected according to actual needs as long as they can meet the actual needs.

[0067] S2): Extract and screen the guided wave signal features from the original signal data collected in S1), and perform time-scale transformation processing on the guided wave features to obtain a guided wave feature subset that meets the requirements of high correlation, high sensitivity, and low redundancy with SOH.

[0068] S3): Build a data-driven model. Suppose the guided wave feature subset contains p features, then 2 p -1 different combinations of guided wave acoustic features can be constructed; these combinations are respectively fused with electrical parameters (current, voltage), and at the same time, the case of using only electrical parameters alone is used as a reference control group. Finally, 2 p data fusion schemes can be obtained. These synchronously collected electro-acoustic data all come from the time-series change curve during the complete charging process and contain the dynamic characteristic information of the entire battery charging process.

[0069] It should be noted that during the data fusion process, the electrical and acoustic guided wave data in each cycle are respectively realized based on the independent acquisition mechanisms of the battery test system and the ultrasonic test system. In view of the inherent difference in the sampling frequency between the two acquisition systems, in order to ensure the precise correspondence of the number of data points within the cycle, the present invention uses an interpolation method to perform time-series alignment processing on the non-synchronously collected data sequences.

[0070] For example, in cycle 1, assume that the number of data points acquired by the electrical signal acquisition system is 320, while the number of data points acquired by the acoustic signal acquisition system is 340. Due to the periodic characteristics, the number of electrical and acoustic signal data points in cycle 10 also remains above 300. To eliminate the problem of data point mismatch caused by different sampling frequencies of different acquisition systems, the present invention uses an interpolation method to uniformly adjust the number of data points of electrical and acoustic signals to 400, thereby ensuring the precise correspondence of the number of data points within each cycle. During the interpolation process, linear interpolation is used to ensure the integrity of the data curve while reasonably increasing the number of data points. Among them, the core of the interpolation algorithm lies in maintaining the change trend of the data curve, avoiding introducing additional noise or distortion due to interpolation operations, thereby ensuring the accuracy and reliability of subsequent data fusion.

[0071] S4): Verify the effectiveness of the electro-acoustic fusion method.

[0072] In this embodiment, the specific process of extracting and screening guided wave signal features from the original signal data collected in S1) in S2) is as follows:

[0073] S21): First, preprocess the original signal data collected in S1): First, centralize the original waveform by subtracting the mean value of the original signal to remove the DC component of the signal, making the average value of the signal zero, thereby removing the baseline shift in the signal;

[0074] S22): Take the average of every 10 adjacent guided wave signals to achieve preliminary filtering;

[0075] S23): Finally, use the Butterworth filtering algorithm to further perform noise reduction processing, and improve the signal-to-noise ratio by optimizing the filtering window length and polynomial order;

[0076] S24): Then analyze the filtered signal. Let s(t) represent the signal amplitude of the guided wave in the time domain, and S(f) represent the spectrum of the guided wave signal after Fourier transform. Its expression is:

[0077] s(f) = ∫s(t)e -j2πft dt

[0078] where t represents the time variable and f represents the frequency variable;

[0079]

[0080] By analyzing the time-domain and frequency-domain characteristics of the guided wave, extract the waveform amplitude, flight time, time-domain centroid, energy integral, peak-to-peak value, skewness coefficient, kurtosis coefficient, waveform index, frequency-domain peak value, frequency-domain centroid, and frequency-domain weighted standard deviation, and finally obtain the guided wave feature set as shown in the above table;

[0081] In the table, H(·) represents the function used to calculate the Hilbert envelope of the signal, and T s represents the total duration of the signal, and the frequency mean μ f is expressed as:

[0082]

[0083] For the method for monitoring the state of health of a lithium battery based on electroacoustic data fusion, the specific process of performing time-scale transformation processing on the guided wave characteristics in S2) is as follows:

[0084] S201): In the battery aging experiment, collect the guided wave signal every 30 s and perform data preprocessing to obtain the signal time-scale waveform;

[0085] S202): Extract features from the signal time-scale waveform to form a continuous and periodic time evolution curve. To analyze the feature changes, divide and superimpose the feature curve by cycle to obtain the evolution trend of the feature curve with the number of cycles in step ③;

[0086] S203): In practical applications, the battery discharge process is complex and affected by the charge and discharge rate and dynamic load, while the charging data is stable and highly consistent. During the battery aging process, the feature curve mainly shows a vertical shift (such as Figure 2 step ④), which is due to internal changes such as the degradation of the electrode material, the change of the electrolyte composition, and the accumulation of mechanical stress, affecting the guided wave propagation characteristics. To quantify this phenomenon, by such as Figure 2 tracking the change of the mean value of the feature curve during the charging process in step ⑤ to characterize the evolution of the feature curve with the battery SOH, the mean value of the feature curve during the charging process is expressed as:

[0087]

[0088] where, F i represents the feature; n represents the number of cycles, and the value range is from 1 to N, and N represents the total number of cycles;

[0089] T represents the duration of each cycle of the charging process;

[0090] For the convenience of analysis, the SOH value is expressed as A0. In order to directly compare features with different units and scales, the minimum-maximum method is used to normalize the feature mean and SOH, which is expressed as:

[0091]

[0092] In the formula, (·)* represents the normalization function.

[0093] For the method for monitoring the health state of a lithium battery based on electroacoustic data fusion, since not all the characteristics of the guided wave signals extracted can effectively characterize the battery SOH, it is necessary to perform systematic feature screening to identify those features with the highest correlation and the highest sensitivity. The specific screening algorithm is as follows:

[0094] First, use the Pearson correlation coefficient analysis method to evaluate the correlation coefficient ρij between the mean value of each characteristic curve and the battery SOH, and further optimize the feature selection through sensitivity analysis. Establish a linear regression model between the feature value and the SOH, and calculate the feature sensitivity index β i (i.e., the regression coefficient), and the expressions are respectively represented as:

[0095]

[0096] In the formula, γ i represents the bias term of the linear regression model corresponding to the feature F i ;

[0097] Secondly, set the correlation coefficient threshold |ρ 0i |>0.9 and the sensitivity threshold |β i |>0.8, and screen the feature subset that is highly correlated and highly sensitive to the SOH and calculate the weighted score corresponding to each feature in the feature subset. The weighted score can be expressed as:

[0098]

[0099] Select the feature with the highest score from these weighted scores and denote it as F best ;

[0100] Finally, to reduce the feature redundancy, calculate the mutual correlation matrix between the features in the subset Set the correlation coefficient threshold between features (|ρ ij |<0.95). If the correlation coefficient between a feature and other features is greater than this threshold, then eliminate this feature to obtain the feature subset To avoid F best from being eliminated, the optimal feature subset F s should be expressed as:

[0101]

[0102] Thus, a guided wave feature subset that meets the requirements of high correlation, high sensitivity, and low redundancy is obtained.

[0103] Example 2

[0104] A battery test platform includes a battery test system, an ultrasonic test system, and an incubator. The battery test system and the ultrasonic test system are both connected to the incubator;

[0105] The battery test system includes a tester for battery charge and discharge detection and a middle computer for electrical data acquisition, and both the tester and the middle computer are connected to an incubator;

[0106] It should be noted that the charge and discharge tester uses a CT-4008-5V12A-204n tester for battery charge and discharge testing, and other suitable testers can also be selected according to actual needs;

[0107] The middle computer uses a CT-ZWJ-4’S-T-1U middle computer for electrical data acquisition of the battery. It should be noted that other suitable middle computers can be selected according to actual needs;

[0108] The temperature range of the incubator is 0 - 60 °C to ensure that the battery is in a constant ambient temperature.

[0109] The ultrasonic test system includes a signal generator for generating a five-peak Hanning window tone pulse with a center frequency of 60 kHz and a digital oscilloscope for collecting guided wave signal data, and both the signal generator and the digital oscilloscope are connected to the incubator.

[0110] It should be noted that the battery is placed in the incubator. During operation, the battery test system is connected to the positive and negative electrodes of the battery, and the signal generator and the digital oscilloscope in the ultrasonic test system are respectively connected to two piezoelectric transducers bonded to the surface of the battery, so that the corresponding electrical and acoustic guided wave data can be detected and collected. Other battery test systems and ultrasonic test systems can also be used according to actual needs, and the corresponding data can be collected according to actual needs.

[0111] In this embodiment, the signal generator uses a Tektronix AFG31021 signal generator, and the digital oscilloscope uses a Tektronix TBS2104B digital oscilloscope. It should be noted that suitable signal generators and digital oscilloscopes can be selected according to actual needs as long as they can meet actual needs.

[0112] Figure 1 The T and R markings represent piezoelectric transducers with a specification of circular PZT-5A with a diameter of 12 mm and a thickness of 1.95 mm, which are respectively used for the transmission and reception of guided wave signals. The distance between the two piezoelectric transducers is 50 mm; the battery is a lithium-ion soft-pack battery with a capacity of 4150 mAh, with dimensions of 110×44×4.8 mm, and uses a graphite / nickel cobalt manganese oxide (NMC) chemical system. In order to study the influence of battery aging on guided wave signals, an accelerated aging experiment is carried out.

[0113] The specific process is as follows: The battery is charged from 3.0V to 4.45V at a current of 1C (4150mA) at 45°C, followed by constant voltage charging until the current drops to C / 20 (207.5mA). After being fully charged, it is left standing for 5 minutes, and then discharged at a current of 1C to 2.8V. It is left standing for 5 minutes between each charge-discharge cycle, and a total of 90 charge-discharge cycles are carried out. During this period, current, voltage, and guided wave signal data are synchronously collected at 30-second intervals, and the guided wave signal is continuously collected 10 times every 30s.

[0114] Example 3

[0115] A data-driven SOH estimation framework for verifying the effectiveness of the electro-acoustic fusion method. The framework includes three groups of hybrid neural network models, namely, the fusion model of one-dimensional CNN and attention mechanism bidirectional LSTM, the fusion model of one-dimensional CNN and active state tracking LSTM, and the fusion model of one-dimensional CNN and standard LSTM;

[0116] And a set of independent standard LSTM models are set as benchmark models. These models can more comprehensively capture the complex dynamic characteristics during the battery charge-discharge process by fusing electrical and acoustic features, thereby improving the accuracy of SOH estimation.

[0117] It should be noted that the above models are existing, so they will not be elaborated here. Specifically as follows: AST-LSTM: CN110824364 B, November 19, 2021

[0118] CNN-LSTM, CNN-ASTLSTM: Penghua Li et al. "An end-to-end neural network framework for state-of-health estimation and remaining useful life prediction of electric vehicle lithium batteries".

[0119] CNN-BiLSTM-AM: Zhenyu Zhu et al. "Attention-based CNN-BiLSTM for SOH and RUL estimation of lithium-ion batteries".

[0120] LSTM: CN 111983457A. November 24, 2020

[0121] In this embodiment, a method for estimating SOH of a neural network model based on electro-acoustic data fusion is as follows:

[0122] 801): First, fuse the electrical parameter curve and the acoustic guided wave characteristic curve of the complete charging process to form the multi-dimensional input of the model;

[0123] 802): Use different combinations of electro-acoustic data inputs of Battery #1 and Battery #2 (including individual electrical parameters, individual guided wave acoustic characteristics, and electro-acoustic fusion characteristics) to train four neural network models (CNN-BiLSTM-AM, CNN-ASTLSTM, CNN-LSTM, and an independent standard LSTM benchmark model) respectively; each data combination is used as the input of the model to comprehensively evaluate the influence of different characteristics on the SOH estimation performance;

[0124] 803): Use the corresponding data combination of Battery #3 as the input to test the trained model to verify the effectiveness and accuracy of the proposed method.

[0125] The electro-acoustic data fusion method proposed in this embodiment innovatively combines electrical parameters (including current and voltage) with the screened acoustic guided wave characteristics (energy integral E, time domain centroid T c , time of flight TOF), significantly enhancing the accuracy of SOH estimation. The technical advantages of the electro-acoustic data fusion method of the present invention are described in detail below with reference to the accompanying drawings:

[0126] (1) Improve the accuracy of SOH estimation

[0127] The electro-acoustic data fusion method can more comprehensively capture the internal and external degradation characteristics of the battery by combining electrical data and acoustic guided wave characteristics. Electrical data mainly reflects the state change of the battery active material, while the acoustic guided wave characteristics reveal the changes in the internal structure of the battery (such as mechanical deformation and material change). This complementarity enables the data-driven model to more accurately estimate the SOH of the battery. The SOH estimation results under different data fusion schemes are shown. As Figure 4 shown in the SOH estimation results of the data-driven model: (a)(i) control group, only using electrical data, guided wave characteristics as input, (b)-(h) represent the fusion of three randomly combined guided wave characteristics with electrical data as input); among them, Figure 4 (a)(i) respectively represent the results of using electrical data (I, U) and acoustic guided wave characteristics (TOF, Tc, E) alone. It can be seen from the figure that when only electrical data is available, the prediction curve shows large fluctuations, while after fusing guided wave characteristics such as (Tc, E) or (TOF, Tc, E), the prediction effect of the battery SOH is significantly improved ( Figure 4(g), (h)). Compared with the electro-acoustic fusion data input, the SOH prediction effect of the single data source of (TOF, Tc, E) is slightly worse, specifically manifested in the fact that the prediction curves of the four data-driven models show slight jitters in the range of 1 - 30 cycles.

[0128] To comprehensively evaluate the estimation effect of SOH, the root mean square error RMSE, the mean absolute error MAE, and the coefficient of determination R 2 are introduced as evaluation indicators. Among them, RMSE reflects the average deviation degree between the predicted value and the true value, and the smaller the value, the higher the prediction accuracy; MAE measures the average absolute error of the prediction result and intuitively shows the actual size of the prediction deviation; R 2 characterizes the goodness of fit of the model, and its value range is between 0 and 1. The closer it is to 1, the stronger the correlation between the prediction result of the model and the true value, and the better the fitting effect.

[0129] The comprehensive evaluation of the above three indicators can comprehensively reflect the prediction performance of the model. The SOH prediction evaluation results of the representative combinations are summarized in Table 2. From the data in Table 2, it can be seen that under the (I, U+TOF, Tc, E) combination, the SOH estimation accuracy of the four models reaches the optimal level. Specifically, compared with the feature combination that only uses electrical data, the R2 index of the LSTM model is significantly improved by 22% under the (I, U+TOF, Tc, E) combination; the other models also achieve a 6 - 8% performance improvement on the basis of the original high SOH estimation accuracy. At the same time, the R 2 evaluation index of only using the acoustic guided wave features (TOF, Tc, E) decreases slightly compared with the (I, U+E, Tc, TOF) combination, and this result further confirms the effectiveness of the electro-acoustic data fusion method in improving the accuracy of the model's SOH estimation. It should be noted that this method shows good model applicability and has achieved significant performance improvements in various data-driven models, indicating that the electro-acoustic data fusion strategy has wide applicability.

[0130]

[0131] (2) Applicable to partial charging data

[0132] In practical applications, the charging data of the battery may be incomplete or only cover part of the charging interval. The electro-acoustic data fusion method can still maintain a high estimation accuracy in the scenario of partial charging data (SOC range 20 - 100%), demonstrating its adaptability in practical applications. Figure 5Table 3 shows the SOH estimation and evaluation results under partial charge data respectively. The estimation accuracy of most models decreases significantly when only electrical data (I, U) are used. While the SOH estimation accuracy of the combination using only acoustic features (TOF, Tc, E) and the electro-acoustic data fusion combination (I, U+E, Tc, TOF) still remains at a high level. This indicates that the electro-acoustic data fusion method can still maintain high performance in the case of incomplete data and is applicable to actual application scenarios.

[0133]

[0134] (3) Model prediction stability

[0135] The electro-acoustic data fusion method not only improves the estimation accuracy but also enables the model to have better SOH estimation stability. By introducing acoustic features, the model can better cope with the noise and uncertainty in electrical data, thus maintaining stable performance under complex working conditions. Figure 6 The uncertainty analysis results of four models when using the electro-acoustic data fusion method are shown. The CNN-ASTLSTM and CNN-BiLSTM-AM models show extremely high stability under electro-acoustic data fusion, with a narrow and stable 95% confidence interval, especially in the rapid battery degradation stage. In contrast, the prediction stability of the LSTM model ( Figure 6 (d)) is slightly insufficient compared with other models, with a wider confidence interval range and obvious fluctuations. However, from the overall prediction effect, this model still maintains good performance and can meet the actual application requirements.

[0136] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements can still be made, and these improvements should also be regarded as the protection scope of the present invention.

Claims

1. A method for monitoring the health state of a lithium battery based on electroacoustic data fusion, characterized in that: Including: S1): Original data acquisition. First, the original guided wave signal data is acquired through a digital oscilloscope, and at the same time, the electrical data of the battery is acquired through a battery test system. S2): Extract and screen the guided wave signal features from the original guided wave signal data collected in S1), and perform time-scale transformation processing on the guided wave features to obtain a guided wave feature subset that meets the requirements of high correlation, high sensitivity, and low redundancy with SOH. S3): Build a data-driven model. Suppose the guided wave feature subset contains p features, then 2 p -1 different combinations of guided wave acoustic features can be constructed; fuse these combinations of guided wave acoustic features with electrical parameters respectively, and take the case of using electrical parameters alone as the benchmark control group. Finally, 2 p data fusion schemes can be obtained. These synchronously collected electro-acoustic data are all from the time-series change curves during the complete charging process, and contain the dynamic characteristic information of the whole battery charging process; S4): Verify the effectiveness of the electro-acoustic fusion method.

2. The method for monitoring the health state of a lithium battery based on electroacoustic data fusion according to claim 1, characterized in that: The specific process of extracting and screening the guided wave signal features from the original signal data collected in S1) in S2) is as follows: S21): First, preprocess the original signal data collected in S1): First, centralize the original waveform, subtract the mean value of the original signal to remove the DC component of the signal, make the average value of the signal zero, so as to remove the baseline shift in the signal. S22): Take the average of every 10 adjacent guided wave signals to achieve preliminary filtering. S23): Finally, use the Butterworth filtering algorithm to further perform noise reduction processing, and improve the signal-to-noise ratio by optimizing the filtering window length and polynomial order. S24): Then analyze the filtered signal. Let s(t) represent the signal amplitude of the guided wave in the time domain, and S(f) represent the spectrum of the guided wave signal after Fourier transform. Its expression is: S(f) = ∫ s(t)e -j2πft dt where t represents the time variable and f represents the frequency variable. By analyzing the time-domain and frequency-domain characteristics of the guided wave, a guided wave feature set is finally obtained. In the table, H(·) represents the function used to calculate the Hilbert envelope of the signal, and T s represents the total duration of the signal, and the frequency mean μ f is expressed as:

3. The method for monitoring the health state of a lithium battery based on electroacoustic data fusion according to claim 2, wherein: The specific process of performing time-scale transformation processing on the guided wave features in S2) is as follows: S201): In the battery aging experiment, collect the guided wave signal every 30 s and perform data preprocessing to obtain the signal time-scale waveform. S202): Extract features from the signal time-scale waveform to form a continuous and periodic time evolution curve. To analyze the feature changes, divide and superimpose the feature curve by cycle to obtain the evolution trend of the feature curve with the number of cycles. S203): Characterize the evolution of the feature curve with the battery SOH by tracking the change of the mean value of the feature curve during the charging process. The mean value of the feature curve during the charging process is expressed as: Among them, F i represents a feature; n represents the number of cycles, with a value range from 1 to N, and N represents the total number of cycles; T represents the duration of each cycle of the charging process. The SOH value is expressed as A0. In order to be able to directly compare features with different units and scales, the minimum-maximum method is used to normalize the feature mean and SOH, which is expressed as: where (·) * denotes a normalization function.

4. The lithium battery health state monitoring method based on electro-acoustic data fusion according to claim 2 It is characterized in that: Since not all the extracted guided wave signal features can effectively characterize the battery SOH, it is necessary to perform systematic feature screening to identify those features with the highest correlation and sensitivity. The specific screening algorithm is as follows: First, the Pearson correlation coefficient analysis method is used to evaluate the correlation coefficient ρ between the mean values of each characteristic curve and the battery SOH ij , and the feature selection is further optimized through sensitivity analysis, a linear regression model between the eigenvalue and the SOH is established, and the feature sensitivity index β i (i.e., the regression coefficient) is calculated, and the expressions are respectively represented as: where γ i represents the feature F i corresponding to the bias term of the linear regression model; Secondly, set the correlation coefficient threshold |ρ 0i | > 0.9 and the sensitivity threshold |β i | > 0.8 to screen the feature subsets that are highly correlated with SOH and have high sensitivity And calculate the weighted scores corresponding to each feature in the feature subset. The weighted score can be expressed as: Select the feature with the highest score from these weighted scores and denote it as F best ; Finally, to reduce the feature redundancy, calculate the subset inter-correlation matrix between features, set the correlation coefficient threshold between features. If the correlation coefficient of a feature with other features is greater than the threshold, then remove the feature Obtain a feature subset To avoid F best from being eliminated, the optimal feature subset F s should be expressed as: Thus, a guided wave feature subset that meets the requirements of high correlation, high sensitivity, and low redundancy is obtained.

5. A battery test platform for implementing the method for monitoring the state of health of a lithium battery based on electroacoustic data fusion according to any one of claims 1 to 4, characterized in that: Including a battery test system, an ultrasonic test system, and an incubator. The battery test system and the ultrasonic test system are both connected to the incubator. The battery test system includes a tester for battery charge and discharge and a middle computer for electrical data acquisition. The charge and discharge tester and the middle computer are both connected to the incubator. The ultrasonic testing system includes a signal generator for generating a five-peak Hanning window tone pulse with a center frequency of 60 kHz and a digital oscilloscope for collecting guided wave signal data. Both the signal generator and the digital oscilloscope are connected to an incubator.

6. The battery test platform according to claim 5, characterized in that: The temperature range of the incubator is 0 - 60 °C to ensure that the battery is in a constant ambient temperature.

7. A data-driven SOH estimation framework for verifying the effectiveness of the electroacoustic fusion method in claim 1, characterized in that: The framework contains three groups of hybrid neural network models, namely, the fusion model of one-dimensional CNN and attention mechanism bidirectional LSTM, the fusion model of one-dimensional CNN and active state tracking LSTM, and the fusion model of one-dimensional CNN and standard LSTM. And a set of independent standard LSTM models are set as benchmark models. These models can capture the complex dynamic characteristics during the battery charging and discharging process more comprehensively by fusing electrical and acoustic features, thereby improving the accuracy of SOH estimation.

8. A method for estimating the SOH of a neural network model based on electroacoustic data fusion according to claim 7, characterized in that: The specific estimation method is as follows: 801): First, fuse the electrical parameter curve and the acoustic guided wave feature curve of the complete charging process to form the multi-dimensional input of the neural network model. 802): Use different combinations of electro-acoustic data inputs of Battery #1 and Battery #2 to train the four neural network models respectively; each data combination is used as the input of the model to comprehensively evaluate the influence of different features on the SOH estimation performance. 803): Use the corresponding data combination of Battery #3 as the input to test the trained model to verify the effectiveness and accuracy of the proposed method.

Images