Lithium ion battery health state prediction method based on impedance spectrum frequency band characteristics

Abstract

The invention relates to the technical field of electrochemical energy storage, and discloses a lithium ion battery health state prediction method based on impedance spectrum frequency band characteristics, and the method comprises the following steps: applying a binary current excitation signal containing a plurality of frequency components to a to-be-detected lithium ion battery, and generating the excitation signal; when the excitation signal is applied, terminal voltage time domain response of the battery is collected at the sampling rate of 1 kHz to 100 kHz, and voltage response data is generated; by applying a short-time multi-valued current excitation signal containing a specific frequency band, ohm, charge transfer and diffusion multi-electrochemical process response of the battery can be synchronously excited within a few seconds, and traditional time-consuming capacity testing and frequency domain impedance spectroscopy scanning are replaced. Therefore, the system can rapidly evaluate and screen the health state of the battery in an industrial scene, and the detection efficiency is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrochemical energy storage, in particular to a lithium ion battery state of health prediction method based on impedance spectrum frequency band characteristics. BACKGROUND

[0002] Lithium ion batteries are widely used in new energy vehicles, energy storage power stations, data centers and mobile terminals, etc. The battery state of health (SOH) directly affects the safety, available capacity and overall life of the equipment. Accurate and rapid evaluation of battery SOH has become a key requirement of the industry.

[0003] Currently, due to the dependence on multiple technical paths in the lithium ion battery SOH prediction process, when batch of retired batteries are screened and production line consistency is rapidly detected, the existing capacity test method takes too long to complete the evaluation within the industrial time scale. When such a method is used for detection, it will result in low detection efficiency and cannot meet the needs of large-scale, high-throughput application scenarios.

[0004] Therefore, the present application provides a lithium ion battery state of health prediction method based on impedance spectrum frequency band characteristics to solve the above problems. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a lithium ion battery state of health prediction method based on impedance spectrum frequency band characteristics, which solves the problem of low detection efficiency and inability to meet the needs of large-scale, high-throughput application scenarios as proposed in the background art.

[0006] To achieve the above purpose, the present application provides the following technical solutions: a lithium ion battery state of health prediction method based on impedance spectrum frequency band characteristics, the method comprising the following steps: S1, a binary current excitation signal containing multiple frequency components is applied to the lithium ion battery to be tested, the signal is constructed by Walsh base function and its combination, the frequency range covers 1Hz to 1024Hz, the signal amplitude is switched between positive A and negative A, A is a preset current amplitude, the application time is 1 second to 4 seconds, and the excitation signal is generated; S2, while applying the excitation signal, the end voltage time domain response of the battery is collected at a sampling rate of 1kHz to 100kHz, the collection time is 1 second to 4 seconds, and the steady-state part of 1 second is intercepted, and the voltage response data is generated; S3, the voltage response data is preprocessed, including removing the direct current component and normalizing, to generate the preprocessed voltage signal; S4. Extract multiple impedance correlation features from the preprocessed voltage signal, including ohmic internal resistance features, mid-frequency peak frequency features, mid-frequency arc equivalent impedance features, peak imaginary part impedance features, peak real part impedance features, peak phase angle features, low-frequency inflection point frequency features, low-frequency inflection point real part features, low-frequency inflection point imaginary part features, phase angle minimum value features, phase angle minimum value corresponding frequency features, Warburg region slope features, Warburg fitting correlation coefficient features, and extended statistical features, and combine these features to generate a feature vector; S5. Input the feature vector into a pre-trained machine learning model, wherein the machine learning model is a multilayer perceptron, and the machine learning model outputs a predicted value of the battery health state (SOH). S6. Output the predicted state of health (SOH) value of the battery. The predicted SOH value is used for retired battery screening, production line consistency testing, online monitoring of energy storage systems, and electric vehicle battery diagnostics.

[0007] Preferably, applying a binary current excitation signal in S1 includes the following steps: S11. Excitation signal parameter configuration: Set the amplitude of the binary current excitation signal to switch between +A and -A, where the value of A ranges from 0.1 Ampere to 2 Ampere, and the signal duration is from 1 second to 4 seconds; S12, Multi-frequency component setting: Configure multiple frequency components covering the ohmic region, charge transfer region and diffusion region in the excitation signal, wherein the frequency components include at least five different frequencies selected from 2Hz, 4Hz, 8Hz, 16Hz, 32Hz, 64Hz, 128Hz, 256Hz and 512Hz. S13. Signal generation and application: A binary current waveform conforming to the set amplitude, duration and frequency component parameters is generated based on the Walsh basis function combination, and the waveform is applied to both ends of the lithium-ion battery through a constant current source circuit.

[0008] Preferably, the voltage response data acquisition in step S2 includes the following steps: S21. Voltage response acquisition: The battery terminal voltage is synchronously acquired using an analog-to-digital converter at a sampling rate of 10kHz to 100kHz. The acquisition period corresponds to the excitation application period to obtain time-domain voltage waveform data. S22. Data Extraction: Select a continuous 1-second segment from the collected complete data, remove the initial transient portion, and make the data meet the steady-state response conditions.

[0009] Preferably, the preprocessing of the voltage response data in step S3 includes the following steps: S31, DC removal: Calculate the average value of the voltage response data and subtract the average value from each data point; S32. Normalization processing: The amplitude of the signal after DC removal is normalized, and signal processing is achieved through linear scaling.

[0010] Preferably, the normalization process in S32 includes the following steps: S321. Signal feature extraction: Obtain the amplitude characteristics of the DC-de-DC signal, including the maximum value, minimum value, and dynamic range; S322, Normalization reference setting: Set the target normalization interval to linearly map the signal to a unified amplitude reference; S323. Linear scaling calculation: Calculate the scaling ratio parameter based on the amplitude characteristics of the signal and the target normalization interval; S324. Normalization execution: Perform a linear transformation on each data point in the DC-free signal according to the calculated proportional parameters to generate a normalized voltage signal.

[0011] Preferably, extracting impedance correlation features in S4 includes the following steps: S41. Overall feature extraction process: 14 impedance-related features are automatically calculated from the preprocessed voltage signal; S42. Specific methods for feature calculation: Ohmic internal resistance characteristics are obtained by the ratio of the maximum and minimum difference of the high-frequency voltage window to the excitation amplitude; intermediate frequency peak frequency characteristics are determined by the reciprocal of the signal oscillation period; intermediate frequency arc equivalent impedance characteristics are calculated by the intermediate frequency energy difference; peak imaginary and real impedance characteristics are extracted from the signal envelope; peak phase angle characteristics are estimated by Hilbert transform; phase angle minimum characteristics are obtained by full-band phase angle analysis; the frequency characteristics corresponding to the phase angle minimum are determined by frequency analysis at the moment of phase angle minimum; low-frequency inflection point characteristics are obtained by trend change rate analysis; Warburg region characteristics are obtained by linear fitting slope; Warburg fitting correlation coefficient characteristics are obtained by linear regression determination coefficient calculation; extended statistical characteristics include standard deviation and mean.

[0012] Preferably, inputting the feature vector into the machine learning model in step S5 includes the following steps: S51. Machine learning model configuration: The pre-trained model is a multilayer perceptron, which includes an input layer, 2 to 4 hidden layers and an output layer. The number of nodes in the input layer is 14, the number of neurons in the hidden layer is 16, 32 and 64, the activation function is ReLU, and the output layer is a single node. S52. Model Training and Inference: The model is trained using historical battery data. Before the feature vector is input, it is Z-score standardized. The model training process is executed, and the output SOH value is de-standardized to the actual capacity.

[0013] Preferably, the machine learning model configuration and training in S51 includes the following steps: S511. Loss Function Definition: The mean squared error (MSE) is used as the loss function for model training, and its calculation formula is as follows: ; in, For the first The true size of each sample, For the first The predicted capacity value for each sample. The number of samples; S512, Optimizer Configuration: Select the adaptive moment estimation optimizer to perform iterative updates on the model parameters and configure the learning rate scheduling strategy; S513, Training Monitoring Settings: Set the loss function convergence threshold and the maximum number of iterations as training termination conditions.

[0014] Preferably, the output of the SOH prediction value in step S6 includes the following steps: S61. Application scenario processing: For retired battery screening, the SOH prediction value is compared with the threshold; for production line inspection, the consistency evaluation is performed on the synchronous excitation and acquisition of multiple batteries in the module. S62. Online monitoring integration: In energy storage systems and electric vehicles, the excitation signal is superimposed on the normal operating current and is collected in real time by the BMS to complete the periodic update operation of SOH.

[0015] Preferably, the method further includes a model transfer learning step: S71. Implementation of transfer learning: For new battery models, collect a small amount of calibration data and fine-tune the last layer of the pre-trained model to adapt the model to different cathode material systems. S72. Model ensemble: An ensemble of the average outputs of multiple machine learning models to form the output mechanism of the prediction results.

[0016] Compared with existing technologies, this invention provides a method for predicting the health status of lithium-ion batteries based on impedance spectrum frequency band characteristics, which has the following beneficial effects: 1. In this invention, by applying a short-duration, multi-value current excitation signal containing a specific frequency band, the ohmic, charge transfer, and diffusion multiple electrochemical processes of the battery can be excited simultaneously within a few seconds, replacing the time-consuming traditional capacity testing and frequency domain impedance spectroscopy scanning. This enables the system to achieve rapid assessment and screening of battery health status in industrial scenarios, greatly improving detection efficiency.

[0017] 2. In this invention, by directly acquiring the time-domain voltage response under excitation, a set of predefined time-domain features strongly correlated with the internal physicochemical state of the battery are systematically extracted from it. This avoids the reliance on complex frequency domain transformation algorithms and equivalent circuit models in traditional methods, enabling the system to achieve highly reliable state prediction with simpler signal processing and lower hardware costs. At the same time, it reduces the reliance on historical battery operating data, enhancing the method's anti-interference ability and universality.

[0018] 3. In this invention, multiple extracted time-domain features are combined into a feature vector and input into a pre-trained dedicated machine learning model for decision-making. This model can automatically learn the complex nonlinear mapping relationship between features and battery aging state, enabling the system to comprehensively utilize information from multiple dimensions for prediction. This not only improves the robustness and generalization ability of the model and allows it to adapt to the complex situations of different batches, models, and even retired batteries, but also reduces the dependence on massive calibration data, thereby ensuring the accuracy and stability of the prediction results. Attached Figure Description

[0019] Figure 1 This is a flowchart of the lithium-ion battery health status prediction method based on impedance spectrum frequency band characteristics according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] For specific implementation examples, please refer to: Figure 1 A method for predicting the health status of lithium-ion batteries based on impedance spectrum frequency band characteristics, the method includes the following steps: S1. Apply a binary current excitation signal containing multiple frequency components to the lithium-ion battery under test. Construct the signal through Walsh basis functions and their combinations. The frequency range covers 1Hz to 1024Hz. The signal amplitude switches between positive A and negative A, where A is the preset current amplitude. The application time is 1 second to 4 seconds to generate the excitation signal. S2. While applying the excitation signal, the time domain response of the battery terminal voltage is acquired at a sampling rate of 1kHz to 100kHz, the acquisition time is 1 second to 4 seconds, and a steady state portion of 1 second is extracted to generate voltage response data. S3. Preprocess the voltage response data, including removing the DC component and normalizing the data, to generate a preprocessed voltage signal. S4. Extract multiple impedance correlation features from the preprocessed voltage signal, including ohmic internal resistance features, mid-frequency peak frequency features, mid-frequency arc equivalent impedance features, peak imaginary part impedance features, peak real part impedance features, peak phase angle features, low-frequency inflection point frequency features, low-frequency inflection point real part features, low-frequency inflection point imaginary part features, phase angle minimum value features, phase angle minimum value corresponding frequency features, Warburg region slope features, Warburg fitting correlation coefficient features, and extended statistical features, and combine these features to generate a feature vector; S5. Input the feature vector into the pre-trained machine learning model. The machine learning model is a multilayer perceptron. The machine learning model outputs the predicted value of the battery health state (SOH). S6 outputs the predicted state of battery health (SOH). The predicted SOH value is used for retired battery screening, production line consistency testing, online monitoring of energy storage systems, and electric vehicle battery diagnostics.

[0022] Applying a binary current excitation signal to S1 includes the following steps: S11. Excitation signal parameter configuration: Set the amplitude of the binary current excitation signal to switch between +A and -A, where the value of A ranges from 0.1 Ampere to 2 Ampere, and the signal duration is from 1 second to 4 seconds; S12, Multi-frequency component setting: Configure multiple frequency components covering the ohmic region, charge transfer region and diffusion region in the excitation signal. The frequency components include at least five different frequencies from 2Hz, 4Hz, 8Hz, 16Hz, 32Hz, 64Hz, 128Hz, 256Hz and 512Hz. S13. Signal generation and application: Based on the combination of Walsh basis functions, a binary current waveform that conforms to the set amplitude, duration and frequency component parameters is generated, and the waveform is applied to both ends of the lithium-ion battery through a constant current source circuit.

[0023] The steps involved in acquiring voltage response data in S2 are as follows: S21. Voltage response acquisition: The battery terminal voltage is synchronously acquired using an analog-to-digital converter at a sampling rate of 10kHz to 100kHz. The acquisition period corresponds to the excitation application period to obtain time-domain voltage waveform data. S22. Data Extraction: Select a continuous 1-second segment from the collected complete data, remove the initial transient portion, and make the data meet the steady-state response conditions.

[0024] The preprocessing of voltage response data in S3 includes the following steps: S31. DC Removal Processing: Calculate the average value of the voltage response data and subtract the average value from each data point, including the following steps: DC component calculation: The average value of the voltage response data is calculated as the DC component. The calculation formula is as follows: ; in, Represents the DC component. Indicates time Voltage value at time, Indicates the total number of sampling points; DC removal: Subtract the DC component from each voltage data point to generate the DC-removed signal. ; in, This represents the voltage signal after DC removal. Represents the original voltage signal; S32. Normalization processing: The amplitude of the signal after DC removal is normalized, and signal processing is achieved through linear scaling.

[0025] The normalization process in S32 includes the following steps: S321. Signal Feature Extraction: Obtain the amplitude characteristics of the signal after DC removal, including the maximum value, minimum value, and dynamic range; S322, Normalization reference setting: Set the target normalization interval to linearly map the signal to a unified amplitude reference; S323. Linear Scaling Calculation: Based on the signal amplitude characteristics and the target normalized interval, calculate the linear scaling ratio parameter, including the following steps: Scaling parameter calculation: based on the target normalization interval [ , ] and actual signal range[ , ] Calculate the scaling factor for linear scaling. and offset : ; in, This represents the voltage signal vector after DC removal. , These are the upper and lower limits of the target normalization interval; Linear transformation execution: Perform a linear transformation on each data point in the signal: ; in, This represents the normalized signal; S324, Normalization Execution: Perform a linear transformation on each data point in the DC-free signal according to the calculated proportional parameters to generate a normalized voltage signal.

[0026] Extracting impedance correlation features from S4 includes the following steps: S41. Overall feature extraction process: 14 impedance-related features are automatically calculated from the preprocessed voltage signal; S42. Specific methods for feature calculation: Ohmic internal resistance characteristics are obtained by the ratio of the maximum and minimum difference of the high-frequency voltage window to the excitation amplitude; intermediate frequency peak frequency characteristics are determined by the reciprocal of the signal oscillation period; intermediate frequency arc equivalent impedance characteristics are calculated by the intermediate frequency energy difference; peak imaginary and real impedance characteristics are extracted from the signal envelope; peak phase angle characteristics are estimated by Hilbert transform; phase angle minimum characteristics are obtained by full-band phase angle analysis; the frequency characteristics corresponding to the phase angle minimum are determined by frequency analysis at the moment of phase angle minimum; low-frequency inflection point characteristics are obtained by trend change rate analysis; Warburg region characteristics are obtained by linear fitting slope; Warburg fitting correlation coefficient characteristics are calculated by linear regression determination coefficient; extended statistical characteristics include standard deviation and mean. The calculation of the ohmic internal resistance characteristic includes the following steps: High-frequency signal extraction: Extracting high-frequency voltage signal sequences from preprocessed voltage signals. This corresponds to the components in the excitation signal with frequencies higher than 100Hz; Internal resistance calculation: Calculate the characteristics of ohmic internal resistance: ; in, For ohmic internal resistance, For the excitation current amplitude, It is a high-frequency voltage signal sequence; The calculation of intermediate frequency peak frequency characteristics includes the following steps: Intermediate frequency signal analysis: Extract the intermediate frequency oscillation component from the preprocessed voltage signal, which corresponds to the component in the excitation signal with a frequency range of 10Hz-100Hz; Periodic detection: The periodicity of the intermediate frequency peak is determined by measuring the time interval between consecutive peaks and troughs. ; Frequency Calculation: Calculate the intermediate frequency peak frequency: ; in, Indicates the intermediate frequency peak frequency. Indicates the period of the intermediate frequency peak; The calculation of the equivalent impedance characteristics of the mid-frequency arc includes the following steps: Real impedance acquisition: Obtain the real impedance value at the position corresponding to the mid-frequency peak frequency. ; Ohmic resistance subtraction: Subtract the ohmic resistance characteristic value from the mid-frequency real impedance to obtain the mid-frequency arc equivalent impedance. ; in, This represents the mid-frequency arc equivalent impedance. Represents the real part of the impedance; The calculation of the peak imaginary part impedance characteristics includes the following steps: Intermediate Frequency (IF) Signal Extraction: Extracting the intermediate frequency voltage signal sequence from the preprocessed voltage signal. This corresponds to the components in the excitation signal with frequencies in the range of 10Hz-100Hz; Imaginary part maximum search: Calculate the maximum value of the signal in the intermediate frequency region as the peak imaginary part impedance.

[0027] in, This represents the characteristic value of the peak imaginary part of the impedance. Represents an intermediate frequency voltage signal sequence; The calculation of the peak real part impedance characteristic includes the following steps: Peak moment location: Determine the moment corresponding to the intermediate frequency peak frequency. ; Real impedance calculation: Obtain the real part of the voltage signal at the peak moment as the peak real impedance. ; in, This represents the characteristic value of the peak real part impedance. Represents the real part of the impedance; The calculation of peak phase angle characteristics includes the following steps: Analytical signal construction: Perform Hilbert transform on the voltage signal to construct the analytic signal. ; in, Indicates an analytical signal. Indicates voltage signal, The Hilbert transform of a voltage signal The imaginary unit; Instantaneous phase angle calculation: Extracting the instantaneous phase angle from the analytic signal: ; in, Indicates the instantaneous phase angle; Peak phase angle extraction: at the moment corresponding to the mid-frequency peak. Extract phase angle value: ; in, Indicates the peak phase angle. This represents the phase angle at the peak value; The calculation of the phase angle minimum characteristic includes the following steps: Full-band phase angle calculation: Calculate the instantaneous phase angle of the signal across the entire time domain using the Hilbert transform. ; Minimum search: Finding the minimum value in phase angle data across the entire frequency band. ; in, This represents the eigenvalue of the minimum phase angle. Represents the voltage signal vector. Indicates the argument of the voltage signal; The calculation of the frequency characteristics corresponding to the minimum phase angle includes the following steps: Phase angle minimum value location: Determining the phase angle minimum value corresponding time ; Periodic component analysis: Analyzing the period corresponding to this moment. ; Frequency calculation: Calculate the frequency corresponding to the minimum phase angle. ; in, This indicates the frequency corresponding to the minimum phase angle. This indicates the period corresponding to the minimum phase angle; The calculation of low-frequency inflection point characteristics includes the following steps: Low-frequency trend analysis: Analyze the rate of change of voltage response in the frequency range below 1 Hz; Inflection point identification: The inflection point of the diffusion region is determined by second derivative detection; Inflection point frequency calculation: Identifying the period corresponding to the inflection point Calculate the inflection point frequency: ; in Indicates the low-frequency inflection point frequency. Indicates the period of the low-frequency inflection point; The calculation of the real part feature of low-frequency inflection points includes the following steps: Low-frequency signal extraction: Extracting low-frequency voltage signal sequences from preprocessed voltage signals. This corresponds to the components in the excitation signal with frequencies below 1Hz; Real part average value calculation: Calculate the average value of the signal in the low-frequency region as the real part feature of the low-frequency inflection point: ; in, This represents the real part eigenvalue of the low-frequency inflection point. This represents the average value of a low-frequency voltage signal sequence. The calculation of the imaginary part feature of low-frequency inflection points includes the following steps: Low-frequency signal extraction: Extracting low-frequency voltage signal sequences from preprocessed voltage signals. ; Calculation of the maximum value of the imaginary part: Calculate the maximum value of the signal in the low-frequency region as the imaginary part feature of the low-frequency inflection point.

[0028] in, This represents the imaginary part eigenvalue of the low-frequency inflection point; The slope calculation for the Warburg region includes the following steps: Diffusion impedance extraction: Extracting the imaginary impedance value from the low-frequency region. and its corresponding angular frequency ; Linear Fitting Preparation: Calculation As the independent variable, As a dependent variable; Slope Calculation: The Warburg slope is calculated using linear regression. ; in, For the Warburg slope, This represents the number of data points in the low-frequency region. Angular frequency, It is the reciprocal of the square root of the angular frequency. It is the average of the reciprocals of the square root of the angular frequency; The calculation of the Warburg fitting correlation coefficient features includes the following steps: Warburg Region Data Preparation: Obtaining Imaginary Impedance Values ​​in the Low-Frequency Diffusion Region and its fitted values ; Correlation coefficient calculation: Calculate the coefficient of determination for the Warburg fit:

[0029] in, This represents the characteristic value of the Warburg fit correlation coefficient. This is the imaginary part of the impedance. This represents the average value of the imaginary part of the impedance; The calculation of extended statistical features includes the following steps: Signal statistical analysis: Calculate statistical parameters for the preprocessed voltage signal; Standard deviation calculation: The standard deviation of a signal is calculated as one of the extended statistical features. ; in, Represents the eigenvalues ​​of the standard deviation. The total number of sampling points. The average value of the signal. The voltage signal's first One sampling point; Mean value calculation: Calculate the mean of the signal as another extended statistical feature; In S5, inputting feature vectors into a machine learning model includes the following steps: S51. Machine learning model configuration: The pre-trained model is a multilayer perceptron, which includes an input layer, 2 to 4 hidden layers and an output layer. The number of nodes in the input layer is 14, the number of neurons in the hidden layer is 16, 32 and 64, the activation function is ReLU, and the output layer is a single node. S52. Model Training and Inference: The model is trained using historical battery data. Z-score standardization is performed before the feature vectors are input. The model training process is executed, and the output SOH value is de-standardized to the actual capacity. This includes the following steps: Feature statistics calculation: Calculate the mean for each feature in the training set. and standard deviation : ; in, Indicates the first Feature values ​​of each sample The total number of training samples; Standardization execution: Perform a standardization transformation on each feature value: ; in, Represents the standardized eigenvalues. Represents the original eigenvalues; Destandardization: After the model prediction is completed, the output results are destandardized.

[0030] in, The capacity value predicted by the model. This is the actual capacity value after destandardization. The standard deviation of the capacity value. This represents the average of the capacity values.

[0031] The configuration and training of machine learning models in S51 includes the following steps: S511. Loss Function Definition: The mean squared error (MSE) is used as the loss function for model training, and its calculation formula is as follows: ; in, For the first The true size of each sample, For the first The predicted capacity value for each sample. The number of samples; S512. Optimizer Configuration: Select the adaptive moment estimation optimizer to perform iterative updates on the model parameters and configure the learning rate scheduling strategy, including the following steps: Initial learning rate setting: Set the initial learning rate. The value range is 0.001-0.01; Dynamic adjustment rule: The learning rate is adjusted using an exponential decay strategy. ; in, Indicates the first The learning rate for the next iteration. This represents the initial learning rate. The attenuation coefficient is... To decay step size, Indicates the number of iterations; Convergence monitoring: Dynamically adjust the learning rate update frequency based on the rate of change of the loss function; S513, Training Monitoring Settings: Set the loss function convergence threshold and the maximum number of iterations as training termination conditions.

[0032] The output of the SOH prediction value in S6 includes the following steps: S61. Application scenario processing: For retired battery screening, the SOH prediction value is compared with the threshold; for production line inspection, the consistency evaluation is performed on the synchronous excitation and acquisition of multiple batteries in the module. S62. Online monitoring integration: In energy storage systems and electric vehicles, the excitation signal is superimposed on the normal operating current and is collected in real time by the BMS to complete the periodic update operation of SOH.

[0033] The method also includes a model transfer learning step: S71. Transfer Learning Implementation: For new battery models, collect a small amount of calibration data and fine-tune the last layer of the pre-trained model to adapt the model to different cathode material systems. This includes the following steps: New data preparation: Collect voltage response data and corresponding capacity calibration values ​​of the new battery model under different health conditions to form a transfer learning training set; Model fine-tuning: Freeze all weight parameters in the pre-trained model except for the last fully connected layer, and iteratively update the weights of the last layer using the transfer learning training set; Weight update calculation: The gradient descent algorithm is used to update the weights of the last fully connected layer. The calculation formula is as follows: ; in, This represents the updated weight matrix. This represents the current weight matrix. Indicates the learning rate. This represents the gradient of the loss function with respect to the weights; Fine-tuning convergence: Training stops when any of the following conditions are met: the decrease in the loss function on the transfer learning training set is less than a preset threshold in consecutive preset rounds; or the preset number of fine-tuning iterations is reached to obtain a prediction model adapted to the new battery model. S72. Model Ensemble: This method integrates the average outputs of multiple machine learning models to form the output mechanism of the prediction results, including the following steps: Base model construction: Multiple machine learning models with the same structure but different initial weights and different structures are used as base models and trained independently using training data; Output: Input the feature vectors into all trained base models simultaneously to obtain the independent SOH prediction value for each model; Integrated Decision: Calculate the arithmetic mean of the SOH predictions from all base models, and output the final integrated prediction result. Integrated computing: based on The calculation formula is as follows: ; in, This indicates the health status predicted by the integrated system. Indicates the first Predicting health status using basic models Indicates the number of models.

[0034] The operation steps of the lithium-ion battery health status prediction method based on impedance spectrum frequency band characteristics are as follows: Step 1: Apply multi-frequency binary excitation signals: First, a specially designed binary current excitation signal is applied to the lithium-ion battery under test. This excitation signal is constructed based on Walsh basis functions and their combinations, with the signal amplitude switching only between two levels: positive A and negative A. The signal frequency components cover the key frequency band from 1Hz to 1024Hz, ensuring that the ohmic polarization, electrochemical polarization, and concentration polarization processes inside the battery can be simultaneously excited within a short time of 1 to 4 seconds. This multi-frequency composite excitation method replaces the traditional point-by-point frequency sweep method of electrochemical impedance spectroscopy, laying the foundation for rapidly obtaining the dynamic response of the battery across the entire frequency range.

[0035] Step 2: Time-domain voltage response signal acquisition: Simultaneously with the application of the excitation signal, the voltage response across the battery is acquired at a high sampling rate ranging from 1 kHz to 100 kHz. The acquisition time lasts from 1 to 4 seconds, and a continuous 1-second stable response segment is extracted as valid analysis data. This time-domain voltage signal fully preserves the multi-frequency component response characteristics contained in the excitation signal, accurately reflecting the dynamic behavior of the battery under different frequency excitations.

[0036] Step 3: Voltage response signal preprocessing: The acquired time-domain voltage response is preprocessed. First, DC bias components in the signal are eliminated through deDC processing, and then normalization is performed to standardize the signal amplitude. A linear scaling algorithm is used in the preprocessing to adjust the signal to a uniform amplitude reference range. This step enhances the stability and reliability of subsequent feature extraction, providing a high-quality input signal for accurately characterizing the battery impedance properties.

[0037] Step 4: Impedance Correlation Feature Extraction Fourteen key impedance correlation features were systematically extracted from the preprocessed voltage signal. These features were specifically designed to characterize key electrochemical parameters of the battery, including ohmic impedance, interfacial charge transfer impedance, and diffusion impedance, from the time-domain signal. Specifically, these include ohmic internal resistance features, mid-frequency peak frequency features, mid-frequency arc equivalent impedance features, peak imaginary part impedance features, peak real part impedance features, peak phase angle features, low-frequency inflection point frequency features, low-frequency inflection point real part features, low-frequency inflection point imaginary part features, phase angle minimum value features, frequency features corresponding to the phase angle minimum value, Warburg region slope features, Warburg fitting correlation coefficient features, and extended statistical features. All features were directly obtained from the time-domain signal using corresponding mathematical calculation methods.

[0038] Step 5: Predictive Analysis Using Machine Learning Models The 14 extracted features were combined into a feature vector, which was then input into a pre-trained machine learning model for health status prediction. The machine learning model used was a multilayer perceptron with 14 nodes in the input layer, perfectly corresponding to the feature vector dimension. The model was trained using historical battery data, with mean squared error as the loss function and an adaptive moment estimator (AME) optimized for parameters.

[0039] Step Six: Output and Application of Health Status Results: The final output is a predicted battery health status, which can be directly applied to various industrial scenarios such as rapid screening of retired batteries, consistency testing of power battery production lines, online health monitoring of energy storage systems, and diagnostics of electric vehicle batteries. This achieves rapid, accurate, and low-cost assessment of the health status of lithium-ion batteries, solving the technical challenges of long testing times, expensive equipment, and complex operations associated with traditional methods.

[0040] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0041] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A lithium-ion battery state-of-health prediction method based on impedance spectrum frequency band characteristics, characterized in that: The method comprises the following steps: S1, a binary current excitation signal containing multiple frequency components is applied to the lithium ion battery to be tested, the signal is constructed by Walsh basis functions and their combinations, the frequency range covers 1 Hz to 1024 Hz, the signal amplitude is switched between positive A and negative A, A is a preset current amplitude, the application time is 1 second to 4 seconds, and the excitation signal is generated; S2, while applying the excitation signal, the time domain response of the terminal voltage of the battery is collected at a sampling rate of 1 kHz to 100 kHz, the collection time is 1 second to 4 seconds, and the steady-state part of 1 second is intercepted, and the voltage response data is generated; S3, the voltage response data is preprocessed, including removing the direct current component and normalizing, and generating the preprocessed voltage signal; S4, a plurality of impedance-related features are extracted from the preprocessed voltage signal, including ohmic resistance features, medium-frequency peak frequency features, medium-frequency arc equivalent impedance features, peak imaginary impedance features, peak real impedance features, peak phase angle features, low-frequency inflection point frequency features, low-frequency inflection point real part features, low-frequency inflection point imaginary part features, phase angle minimum value features, phase angle minimum value corresponding frequency features, Warburg region slope features, Warburg fitting correlation coefficient features, and extended statistical features, and the features are combined to generate a feature vector; S5, the feature vector is input into a pre-trained machine learning model, the machine learning model is a multi-layer perception, and the machine learning model outputs a predicted value of the battery state of health SOH; S6, output the battery state of health SOH prediction value, the SOH prediction value is used for retired battery screening, production line consistency detection, energy storage system online monitoring and electric vehicle battery diagnosis scenarios.

2. The impedance spectroscopy frequency band feature based lithium-ion battery state of health prediction method of claim 1, wherein: The S1 of applying a binary current excitation signal comprises the following steps: S11, excitation signal parameter configuration: set the amplitude of the binary current excitation signal to switch between +A and -A, wherein the value of A ranges from 0.1 ampere to 2 ampere, and the signal duration is 1 second to 4 seconds; S12, multiple frequency component setting: configure multiple frequency components covering the ohmic region, charge transfer region and diffusion region in the excitation signal, the frequency components include at least five different frequencies of 2 Hz, 4 Hz, 8 Hz, 16 Hz, 32 Hz, 64 Hz, 128 Hz, 256 Hz and 512 Hz; S13, signal generation and application: generate a binary current waveform according to the set amplitude, duration and frequency component parameters based on the combination of Walsh basis functions, and apply the waveform to the lithium ion battery through a constant current source circuit.

3. The impedance spectroscopy frequency band feature based lithium-ion battery state of health prediction method of claim 1, wherein: The S2 of collecting voltage response data comprises the following steps: S21, voltage response collection: use an analog-to-digital converter to synchronously collect the battery terminal voltage at a sampling rate of 10 kHz to 100 kHz, the collection period corresponds to the excitation application period, and the time domain voltage waveform data is obtained; S22, data interception: select a continuous 1-second segment from the collected complete data, remove the initial transient part, and make the data meet the steady-state response condition.

4. The impedance spectroscopy frequency band feature based lithium-ion battery state of health prediction method of claim 1, wherein: The preprocessing of the voltage response data in S3 comprises the following steps: S31, DC removal processing: calculate the average value of the voltage response data, and subtract the average value from each data point; S32, normalization processing: amplitude normalization is performed on the DC removed signal, and signal processing is realized through linear scaling.

5. The impedance spectroscopy frequency band feature based lithium-ion battery state of health prediction method of claim 4, wherein: The normalization processing in S32 includes the following steps: S321, signal feature extraction: obtain the amplitude feature of the DC removed signal, including the maximum value, the minimum value and the dynamic range; S322, normalization reference setting: set the target normalization interval, and linearly map the signal to a unified amplitude reference; S323, linear scaling calculation: calculate the scaling parameter of linear scaling according to the amplitude feature of the signal and the target normalization interval; S324, normalization execution: linearly transform each data point in the DC removed signal according to the calculated scaling parameter to generate the normalized voltage signal.

6. The impedance spectroscopy frequency band based lithium-ion battery state of health prediction method of claim 1, wherein: The impedance correlation feature extraction in S4 includes the following steps: S41, overall flow of feature extraction: automatically calculate 14 impedance correlation features from the preprocessed voltage signal; S42, specific method of feature calculation: ohmic resistance feature is obtained by the ratio of the maximum and minimum difference of voltage high-frequency window to the excitation amplitude, medium-frequency peak frequency feature is determined by the reciprocal of signal oscillation period, medium-frequency arc equivalent impedance feature is calculated by medium-frequency energy difference, peak imaginary and real impedance features are extracted from signal envelope, peak phase angle feature is estimated by Hilbert transform, phase angle minimum value feature is obtained by full frequency band phase angle analysis, phase angle minimum value corresponding frequency feature is determined by frequency analysis at the time of phase angle minimum value, low-frequency inflection point feature is analyzed by trend change rate, Warburg region feature is obtained by linear fitting slope, Warburg fitting correlation coefficient feature is obtained by linear regression determination coefficient calculation, and extended statistical features include standard deviation and mean value.

7. The impedance spectroscopy frequency band based lithium-ion battery state of health prediction method of claim 1, wherein: The input of feature vector into machine learning model in S5 includes the following steps: S51, machine learning model configuration: the pre-trained model is a multi-layer perception, including an input layer, 2-4 hidden layers and an output layer, the number of input layer nodes is 14, the number of hidden layer neurons is 16, 32 and 64, the activation function uses ReLU, and the output layer is a single node; S52, model training and inference: the model is trained using historical battery data, the feature vector is standardized before input, the model training process is executed, and the SOH value is output and then standardized to the actual capacity.

8. The impedance spectroscopy frequency band feature based lithium-ion battery state of health prediction method of claim 7, wherein: The machine learning model configuration and training in S51 includes the following steps: S511, loss function definition: mean square error MSE is used as the loss function of model training, and its calculation formula is: ； wherein, is the true volume value for the th sample, is the predicted volume value for the th sample, is the number of samples; S512, optimizer configuration: adaptive moment estimation optimizer is selected to perform iterative update on model parameters, and learning rate scheduling strategy is configured; S513, training monitoring setting: set the loss function convergence threshold and the maximum number of iterations as the training termination condition.

9. The impedance spectroscopy frequency band based lithium-ion battery state of health prediction method of claim 1, wherein: The output of SOH prediction value in S6 includes the following steps: S61, application scenario processing: for retired battery screening, compare SOH prediction value with threshold value, for production line detection, synchronize excitation and collection of multiple batteries in module, and perform consistency evaluation; S62, online monitoring integration: in energy storage systems and electric vehicles, excitation signals are superimposed on normal working current, and SOH periodic updating operation is completed through real-time collection by BMS.

10. The impedance spectroscopy frequency band based lithium-ion battery state of health prediction method of claim 1, wherein: The method further comprises a model migration learning step: S71, migration learning implementation: for new type batteries, a small amount of calibration data is collected, the last layer of the pre-trained model is fine-tuned, and the model is adapted to different positive electrode material systems; S72, model integration: an integrated method of average value of multiple machine learning models is adopted to constitute an output mechanism of prediction results.

Images