A Battery Aging Detection Method and System Based on Impedance Spectroscopy

Abstract

This invention relates to the field of energy storage system technology, specifically to a battery aging detection method and system based on impedance spectroscopy. The battery aging detection method and system based on impedance spectroscopy is characterized by comprising: an adaptive frequency band focusing excitation module; a synchronous sampling and frequency domain transformation module; a spectral morphology feature encoding module; an intra-cluster dynamic baseline module; a temperature compensation module; a normalized weighted scoring module; an aging grading determination module; an aging trajectory prediction module; and an online evolution module for detection configuration. Compared with existing technologies, by using spectral morphology feature encoding, the Nyquist curve is discretized into a sequence of morphological descriptors, retaining the non-parametric aging information lost in ECM fitting, thus improving the sensitivity of early aging detection; through the adaptive frequency band focusing mechanism, the excitation energy is redistributed to the most sensitive frequency band according to the aging feature sensitivity ranking of each frequency band in the previous round of detection, thereby improving the signal-to-noise ratio.

Description

Technical Field

[0001] This invention relates to the field of energy storage system technology, specifically to a battery aging detection method and system based on impedance spectroscopy. Background Technology

[0002] In electrochemical energy storage systems, batteries gradually age with the accumulation of charge-discharge cycles, manifesting as capacity decay, increased internal resistance, and decreased power output. Accurately identifying the aging stage of batteries is of great significance for energy storage power station operation and maintenance decisions, cascade utilization sorting, and safety management.

[0003] Electrochemical impedance spectroscopy (EIS) is an effective means of characterizing the internal electrochemical state of a battery. By applying a small-amplitude excitation signal to the battery terminals and measuring its response, EIS can obtain the battery's impedance characteristics at different frequencies, reflecting changes in internal processes such as electrolyte resistance, SEI film impedance, charge transfer impedance, and solid-state diffusion impedance. However, existing aging detection techniques based on impedance spectroscopy have the following drawbacks and limitations:

[0004] 1. Existing solutions rely on Equivalent Circuit Model (ECM) parameter fitting to extract aging features. While the physical meaning of ECM parameters (such as R0, R_ct, and Warburg coefficients) is clear, the fitting process is sensitive to initial values, has non-unique solutions, and the fitting results differ significantly under different ECM model structures. More fundamentally, ECM compresses the complex Nyquist curve into a finite number of scalar parameters, losing the rich aging information contained in the curve's shape—such as the asymmetric deformation of the mid-frequency semicircle, the morphological changes in the high-frequency-mid-frequency transition region, and the curvature of the low-frequency wake—non-parametric features that are often sensitive indicators of early aging.

[0005] 2. Existing methods use a fixed set of frequencies for excitation, with uniform excitation energy distributed across each frequency. However, different aging stages exhibit different frequency bands most sensitive to impedance spectrum changes: early aging is mainly reflected in SEI film thickening at high frequencies, mid-stage aging in increased charge transfer impedance at mid frequencies, and late aging is significantly manifested in diffusion impedance changes at low frequencies. Fixed uniform excitation cannot adaptively concentrate excitation energy on the most sensitive frequency bands according to the current aging stage of the battery, resulting in insufficient signal-to-noise ratio and wasted detection sensitivity.

[0006] 3. Existing baseline management methods employ a static approach—establishing a baseline upon initial commissioning and only performing simple updates thereafter. This "individual baseline" strategy ignores the impedance statistical distribution information of other normal cells within the same cluster. When the baseline cell itself exhibits manufacturing outliers, the aging offset based on the individual baseline will generate systematic errors. Existing solutions lack a mechanism to construct dynamic baselines using the statistical patterns of the cluster population.

[0007] 4. The existing solution only performs a single snapshot determination of the impedance spectrum at the current moment, outputting the instantaneous aging level. This method cannot utilize the impedance spectrum sequence accumulated from multiple historical tests of the battery ("spectral timeline"), nor can it make forward-looking predictions of future aging trends. Maintenance personnel can only see "currently at L2", but cannot know "when it is expected to enter L3", limiting the planning capabilities for preventive maintenance.

[0008] 5. Existing solutions treat each test as an independent event, and the test configuration (frequency set, excitation amplitude, sampling window) remains unchanged throughout the entire lifecycle. This "one-time configuration for permanent use" model cannot learn from historical test results, cannot identify which frequencies contribute the most to aging detection, and which frequencies have redundant information, causing the test efficiency and accuracy to stagnate as the battery ages.

[0009] 6. Existing online impedance detection schemes only cover a limited number of feature dimensions. Two-point impedance-based schemes extract impedance values ​​at only two frequency points for regression prediction; machine learning schemes based on relaxation voltage do not involve real-time excitation injection; chip-level EIS integration schemes focus on hardware circuit design. None of these schemes achieve an end-to-end approach encompassing "online multi-frequency excitation → impedance spectrum morphology encoding → adaptive frequency band focusing → dynamic population baseline → aging trajectory prediction". Summary of the Invention

[0010] To overcome the shortcomings of existing technologies, this invention provides a battery aging detection method and system based on impedance spectroscopy. By employing spectral morphology feature encoding, adaptive frequency band focusing mechanism, intra-cluster dynamic baseline, aging trajectory prediction engine, and online evolution mechanism for detection configuration, the system achieves improved sensitivity in early aging detection, enhanced signal-to-noise ratio, and elimination of the influence of individual manufacturing deviations.

[0011] To achieve the above objectives, a battery aging detection method and system based on impedance spectroscopy are designed, characterized in that the system includes:

[0012] Adaptive frequency band focusing excitation module: Based on the aging characteristic sensitivity ranking of each frequency band in the previous round of detection results, dynamically adjust the excitation frequency point set and the excitation energy distribution of each frequency point;

[0013] Synchronous sampling and frequency domain transformation module: The time-domain waveforms of excitation current and response voltage are synchronously acquired through high-speed ADC, and the impedance complex values ​​at each frequency point are calculated by FFT frequency domain transformation to construct a digital impedance spectrum;

[0014] Spectral morphology feature encoding module: This module directly extracts non-parametric features from the geometric morphology of the Nyquist curve; and retains the traditional ECM parametric features as a reference, which together with the morphological features constitute an extended feature vector;

[0015] Intra-cluster dynamic baseline module: Calculates the statistical distribution of the group impedance spectrum as a dynamic baseline using the most recent test results of all cells in the same cluster;

[0016] Temperature compensation module: Based on the Arrhenius temperature compensation model, it performs temperature correction on impedance characteristics, correcting the characteristic values ​​at the measurement temperature to the equivalent values ​​at the reference temperature;

[0017] Normalized weighted scoring module: Normalizes the feature vectors after temperature compensation and baseline offset, and then calculates the comprehensive aging score by weighted summation;

[0018] Aging grading module: Based on the comprehensive aging score, the battery is divided into four levels: L1, L2, L3, and L4 according to the threshold, and the grading confidence level is calculated.

[0019] Aging trajectory prediction module: This module maintains the historical detection feature sequence of each cell and uses the trend of the historical feature sequence to extrapolate and predict the future aging state;

[0020] Online detection configuration evolution module: This module is responsible for learning from historical detection results and automatically optimizing the detection configuration; Result output and linkage module: This module summarizes the current aging level, trajectory prediction results, population baseline comparison report and detection configuration evolution log to generate a comprehensive detection report;

[0021] The specific process of the battery aging test method is as follows:

[0022] S1, Adaptive frequency band configuration loading: The system loads the detection configuration parameters. During the first detection, the default configuration is used: M frequency points are covered at equal logarithmic intervals from 10mHz to 10kHz, and the excitation amplitude of each frequency point is evenly distributed. During subsequent detections, the optimized configuration output by the detection configuration evolution module is loaded: frequency point set, excitation amplitude distribution ratio of each frequency point, and frequency band sensitivity sorting.

[0023] S2, Adaptive frequency band focusing excitation signal generation: Generate a multi-frequency superimposed excitation signal according to the configuration in step S1: i_exc(t)=Σ(m=1toM)A_m·sin(2πf_m·t+φ_m); where i_exc(t) is the excitation current signal; M is the total number of frequency points; A_m is the excitation amplitude of the m-th frequency point, satisfying A_m≤A_max; f_m is the frequency of the m-th frequency point; φ_m is the initial phase of the Schroeder multiphase sequence optimization;

[0024] S3, Inject Excitation and Simultaneously Sample: The excitation signal generated in step S2 is superimposed on the battery terminal through the excitation injection interface, and the battery terminal voltage response v(t) and actual excitation current i(t) are simultaneously collected;

[0025] S4, FFT frequency domain transformation and impedance spectrum construction: Perform FFT on the time domain sequence to calculate the complex impedance values ​​at each frequency point: Z(f_m)=V(f_m) / I(f_m)=|Z(f_m)|·exp(j·θ(f_m)); where Z(f_m) is the complex impedance value at frequency f_m; V(f_m) is the voltage frequency domain component; I(f_m) is the current frequency domain component; |Z(f_m)| is the impedance magnitude; θ(f_m) is the impedance phase angle. Construct the Nyquist spectrum and perform quality verification.

[0026] S5, Spectral morphology feature encoding: Extract curvature sequence, arc length distribution vector and semicircle asymmetry from the geometric morphology of Nyquist curve, and extract traditional ECM parameter features as reference dimensions.

[0027] S6, Temperature Compensation Correction: Perform Arrhenius temperature compensation on impedance-type features: x_j^comp=x_j·exp(E_a_j / R_g·(1 / T_ref-1 / T_meas)); where x_j^comp is the compensated feature value; x_j is the measured value; E_a_j is the activation energy parameter; R_g is the ideal gas constant; T_ref is the reference temperature; and T_meas is the measurement temperature.

[0028] S7, Construction of dynamic baseline within the cluster: Collect detection feature data of all cells in the most recent round within the same cluster, and calculate the population truncation statistic as the dynamic baseline:

[0029] x_j^base_dyn=TrimMean(x_j^(1),x_j^(2),...,x_j^(N_cell),p_trim);

[0030] σ_j^base_dyn=TrimStd(x_j^(1),x_j^(2),...,x_j^(N_cell),p_trim);

[0031] Where x_jbase_dyn is the population dynamic baseline value of feature j; σ_jbase_dyn is the population baseline standard deviation; TrimMean is the truncated mean function; TrimStd is the truncated standard deviation function; N_cell is the total number of cells in the cluster; and p_trim is the truncation ratio.

[0032] S8, Extended Feature Normalization and Weighted Scoring: Combine ECM parameter offsets and morphological feature offsets into an extended feature vector, and perform weighted scoring.

[0033] S = Σ(j=1 to J_ecm)w_j·|Δx_j| + Σ(k=1 to J_morph)w_morph_k·d_morph_k; where S is the comprehensive aging score; J_ecm is the number of ECM parameter features; w_j is the ECM feature weight; J_morph is the number of morphological features; w_morph_k is the morphological feature weight; and d_morph_k is the offset distance of the k-th morphological feature.

[0034] S9, Aging Classification Determination: A four-level classification determination is performed based on the comprehensive aging score, and the classification confidence level ρ is calculated simultaneously.

[0035] S10, Aging Trajectory Prediction: The aging trajectory prediction engine extrapolates the trend based on the historical rating sequence {S_1,S_2,...,S_N_hist}, fitting a linear model to the rating time series: S(t)=S_0+k_aging·t; where S(t) is the predicted rating at time t; S_0 is the intercept; k_aging is the aging rate; t is time; the predicted time to enter the next level is: t_predict=(θ_next-S_current) / k_aging; where t_predict is the predicted remaining time; θ_next is the grading threshold for the next level; S_current is the current rating;

[0036] S11, Cluster-level lateral comparison and anomaly localization: Perform a lateral comparison on the current score and trajectory prediction of all cells in the same cluster;

[0037] S12, Online Evolution of Detection Configuration: After each round of full cluster detection, the information contribution of each frequency point is calculated: IC_m=|Corr(ΔZ_m,ΔS)|; where IC_m is the information contribution of the m-th frequency point; ΔZ_m is the impedance change of each cell at frequency point m; ΔS is the aging score of each cell; and Corr is the Pearson correlation coefficient.

[0038] S13, Results Output and Linkage: The output includes the current aging level, grade confidence level, aging trajectory prediction, population baseline comparison report, cluster-level consistency report, and detection configuration evolution log; and linkage alarms are executed according to the level.

[0039] S14, Data Recording and Baseline Update: Record the complete detection log, update the population dynamic baseline, and update the historical sequence of the aging trajectory prediction engine.

[0040] In the spectral morphology feature encoding module, non-parametric features include:

[0041] (1) Curvature sequence κ(s): The curvature is calculated along the arc length parameter s of the Nyquist curve to generate a curvature profile sequence. The mid-frequency semicircular deformation, asymmetry and low-frequency wake bending caused by aging will produce characteristic changes in the curvature sequence.

[0042] (2) Arc length distribution vector L_band: Divide the frequency range into K frequency bands and calculate the proportion of the arc length of the Nyquist curve in each frequency band to the total arc length. The local impedance change caused by aging will cause the arc length proportion of the corresponding frequency band to shift.

[0043] (3) Semicircle asymmetry γ: The degree of deviation of the mid-frequency semicircle from the ideal semicircle is quantified and defined as the ratio of the area of ​​the upper half of the semicircle to the area of ​​the lower half. For normal cells, γ is close to 1.0, while for aged cells, γ deviates from 1.0.

[0044] In the intra-cluster dynamic baseline module, the baseline construction method is as follows:

[0045] (1) Calculate the truncated mean and truncated standard deviation for each dimension of all cells in the cluster;

[0046] (2) When cell data is insufficient, temporarily revert to the traditional individual baseline strategy, and automatically switch to the group baseline after sufficient group data has been accumulated.

[0047] In the normalized weighted scoring module, the weight coefficients consist of two parts: fixed basic weights for ECM parameter features and adaptive weights for morphological features. The weights for morphological features are dynamically adjusted by the detection configuration evolution module based on historical sensitivity analysis.

[0048] In the aging trajectory prediction module, the prediction method is as follows:

[0049] (1) Fit a linear trend to the historical time series of the comprehensive aging score and calculate the time required for the score to grow from the current value to the threshold of each level; the output includes the current aging level, the expected time to enter the next level and the prediction confidence level;

[0050] (2) The prediction confidence is based on the size of the fitting residuals, that is, the smaller the residuals, the more certain the prediction;

[0051] (3) When the number of historical data points is less than the preset minimum, mark "Insufficient data, do not predict for now".

[0052] In the online evolution module of the detection configuration, after each round of full cluster detection, the online evolution module of the detection configuration calculates the information contribution of each frequency point to the change of aging score. Frequency points with information contribution below the threshold are marked as "redundant frequency points"; frequency points with high contribution are marked as "critical frequency points". In the next round of detection, critical frequency points are allocated more excitation energy, and redundant frequency points can be downgraded or merged into a wider frequency band to shorten the detection time.

[0053] In the results output and linkage module, linkage actions are executed according to the aging level: L2 generates observation alarm, L3 triggers maintenance work order, L4 generates emergency alarm, and trajectory prediction results provide forward-looking information such as "expected to enter L3 in N months" to assist in maintenance planning; among them, L1 is normal, L2 is mild aging, L3 is moderate aging, and L4 is severe aging.

[0054] In step S2, the adaptive amplitude allocation rule is: A_m = A_total·(λ_m / Σ(m=1 to M)λ_m); where A_total is the total excitation amplitude budget; λ_m is the excitation energy allocation coefficient for the m-th frequency point; for the first detection, λ_m = 1 / M; for subsequent detections, λ_m = max(λ_min, ρ_m^α_focus); where ρ_m is the historical sensitivity value for that frequency point; α_focus is the focusing index; and λ_min is the minimum allocation coefficient to prevent the excitation energy at any frequency point from dropping to zero.

[0055] In step S5,

[0056] (1) Curvature sequence extraction: The Nyquist curve is parameterized as a function Z(s)=(Re(s),-Im(s)) of arc length s, and the curvature at each point along the curve is calculated as: κ(s)=|d²Z / ds²| / |dZ / ds|³; where κ(s) is the curvature value at the arc length parameter s, in units of Ω⁻¹; d²Z / ds² is the second derivative of the curve with respect to the arc length; dZ / ds is the first derivative of the curve with respect to the arc length;

[0057] (2) Arc length distribution vector: Divide the frequency range into K_band frequency bands and calculate the proportion of the arc length of the Nyquist curve in each frequency band to the total arc length: L_band_k=L_k / L_total; where L_band_k is the proportion of the arc length of the k-th frequency band; L_k is the arc length of the Nyquist curve in that frequency band; and L_total is the total arc length.

[0058] (3) Semicircle asymmetry: Asymmetric deformation of the quantized mid-frequency semicircle: γ=S_upper / S_lower; where γ is the asymmetry; S_upper is the area enclosed by the upper half of the mid-frequency semicircle; S_lower is the area enclosed by the lower half.

[0059] Compared with existing technologies, this invention provides a battery aging detection method and system based on impedance spectroscopy. By encoding spectral morphology features, the Nyquist curve is discretized into a morphological descriptor sequence, preserving the non-parametric aging information lost in ECM fitting, thus improving the sensitivity of early aging detection. Through an adaptive frequency band focusing mechanism, the excitation energy is redistributed to the most sensitive frequency band based on the aging feature sensitivity ranking of each frequency band in the previous round of detection, improving the signal-to-noise ratio. By using a dynamic baseline within the cluster, the statistical distribution of the impedance spectrum of normal cells in the same cluster is used as a baseline reference to eliminate the influence of individual manufacturing deviations. Through an aging trajectory prediction engine, the future aging level change time point is extrapolated based on the morphological feature time series of historical spectral sequences, outputting a forward-looking prediction of "expected to enter L3 in N months." Through an online evolution mechanism for detection configuration, the information contribution of each frequency point is automatically evaluated from historical detection results, dynamically adjusting the frequency point set and excitation amplitude allocation. Attached Figure Description

[0060] Figure 1 This is a diagram of the overall system architecture of the present invention.

[0061] Figure 2 This is a flowchart of the software control process of the present invention.

[0062] Figure 3 A schematic diagram illustrating the morphological characteristics of the Nyquist curve.

[0063] Figure 4 This is a schematic diagram for predicting the aging trajectory of the system.

[0064] Figure 5 This is a comparison chart of the group dynamic baseline and the individual baseline.

[0065] Figure 6 The image shows the test results of the energy storage power station in the example.

[0066] illustrate: Figure 6 The data includes a comprehensive score map of cells within the cluster (a), a predicted distribution map of aging trajectories (b), and a heat map of frequency band sensitivity evolution (c). Detailed Implementation

[0067] The present invention will now be further described with reference to the accompanying drawings.

[0068] The battery aging detection system based on impedance spectroscopy provided by this invention consists of the following ten core modules:

[0069] 1. Adaptive frequency band focusing excitation module: This module is one of the core innovative modules of this invention. Unlike traditional fixed frequency excitation, this module dynamically adjusts the excitation frequency set and the excitation energy distribution of each frequency point according to the aging characteristic sensitivity ranking of each frequency band in the previous round of detection results.

[0070] The frequency band sensitivity is defined as the correlation coefficient between the impedance change and the total aging score change in that frequency band: 60% of the total excitation energy is allocated to the top 30% of the frequency bands in terms of sensitivity; the remaining 40% is allocated to the other frequency bands. The default uniform allocation strategy is used for the first test (when there is no historical data). This module generates multi-frequency superimposed excitation signals, and the initial phase of each frequency point is optimized according to the Schroeder polyphase sequence to reduce the peak factor.

[0071] 2. Synchronous Sampling and Frequency Domain Transformation Module: This module synchronously acquires the time-domain waveforms of the excitation current and response voltage using a high-speed ADC, calculates the complex impedance values ​​at each frequency point through FFT frequency domain transformation, and constructs a digital impedance spectrum. The sampling window duration ensures that the lowest frequency point completes at least a set number of cycles. Synchronization is guaranteed by a unified trigger signal to ensure time alignment of current and voltage sampling.

[0072] 3. Spectral Morphology Feature Encoding Module: This module is the second core innovative module of this invention. This module skips the traditional ECM parameter fitting step and directly extracts non-parametric feature descriptors from the geometric morphology of the Nyquist curve.

[0073] (1) Curvature sequence κ(s): The curvature is calculated along the arc length parameter s of the Nyquist curve to generate a curvature profile sequence. Aging-induced mid-frequency semicircular deformation, asymmetry, and low-frequency wake bending will all produce characteristic changes in the curvature sequence.

[0074] (2) Arc length distribution vector L_band: Divide the frequency range into K equal frequency bands and calculate the proportion of the arc length of the Nyquist curve in each frequency band to the total arc length. Local impedance changes caused by aging will cause the arc length proportion of the corresponding frequency band to shift.

[0075] (3) Semicircle asymmetry γ: The degree to which the mid-frequency semicircle deviates from the ideal semicircle is quantified and defined as the ratio of the area of ​​the upper half to the area of ​​the lower half of the semicircle. Normal cells have a γ close to 1.0, while aged cells have a γ deviation of 1.0.

[0076] This module also retains the traditional ECM parameter features (R0, R_ct, σ_w) as a reference, which together with the morphological features constitute an extended feature vector.

[0077] 4. Cluster Dynamic Baseline Module: This module is the third core innovation of this invention. It utilizes the most recent test results of all cells within the same cluster to calculate the statistical distribution of the cluster impedance spectrum as a dynamic baseline. The baseline construction method involves calculating the truncated mean (the mean after removing the highest and lowest 10%) and truncated standard deviation for each dimension of all cells within the cluster. The aging offset is no longer compared with the "initial test baseline" but with the "current median level of the cluster." When cell data is insufficient (e.g., the first batch of tests for newly commissioned cells), it temporarily reverts to the traditional individual baseline strategy, automatically switching to the cluster baseline once sufficient cluster data has accumulated.

[0078] 5. Temperature Compensation Module: This module performs temperature correction on impedance characteristics based on the Arrhenius temperature compensation model, correcting the characteristic values ​​at the measurement temperature to their equivalent values ​​at the reference temperature. The compensation parameters are determined by calibration experiments.

[0079] 6. Normalized Weighted Scoring Module: This module normalizes the feature vectors after temperature compensation and baseline offset (eliminating dimensional differences), and then calculates the comprehensive aging score through weighted summation. The weighting coefficients consist of two parts: fixed base weights for ECM parameter features and adaptive weights for morphological features. The weights for morphological features are dynamically adjusted by the detection configuration evolution module based on historical sensitivity analysis.

[0080] 7. Aging Grading Module: This module classifies batteries into four levels—L1 (normal), L2 (mild aging), L3 (moderate aging), and L4 (severe aging)—based on the comprehensive aging score and thresholds, and calculates the grading confidence level.

[0081] 8. Aging Trajectory Prediction Module: This module is the fourth core innovative module of this invention. This module maintains the historical detection feature sequence ("spectral timeline") for each cell and uses the trend of the historical feature sequence to extrapolate and predict future aging status. The prediction method is as follows: fit a linear trend (least squares method) to the historical time series of the comprehensive aging score, and calculate the time required for the score to increase from the current value to the threshold of each grade. The output includes: the current aging level, the expected time to enter the next level (in months), and the prediction confidence level. The prediction confidence level is based on the size of the fitting residual: the smaller the residual, the more certain the prediction. When the historical data points are less than the preset minimum value (default 5 tests), it is marked "Insufficient data, no prediction at this time".

[0082] 9. Online Evolution Module for Detection Configuration: This module learns from historical detection results and automatically optimizes the detection configuration. After each round of full-cluster detection, this module calculates the information contribution of each frequency point to the change in aging score (defined as the absolute value of the Pearson correlation coefficient between the change in feature at that frequency point and the change in score). Frequency points with an information contribution below a threshold are marked as "redundant frequency points"; frequency points with high contribution are marked as "critical frequency points". In the next round of detection, critical frequency points are allocated more excitation energy (executed by the adaptive frequency band focusing module), while redundant frequency points can be downgraded or merged into a wider frequency band to shorten the detection time.

[0083] 10. Result Output and Linkage Module: This module summarizes the current aging level, trajectory prediction results, group baseline comparison report, and detection configuration evolution log to generate a comprehensive detection report. It executes linked actions based on the aging level: L2 generates observation alarms, L3 triggers maintenance work orders, and L4 generates emergency alarms. The trajectory prediction results provide forward-looking information such as "expected to enter L3 in N months," assisting in maintenance planning.

[0084] like Figure 1 As shown, Figure 1 This section showcases the connections between ten core modules. It is divided into four sections from top to bottom:

[0085] 1. Excitation and Acquisition Area (light yellow): The left side is the "Adaptive Frequency Band Focusing Excitation Module" (labeled "Sensitivity Ranking → Energy Redistribution"), and the right side is the "Synchronous Sampling and Frequency Domain Transformation Module". The two are connected by a bidirectional arrow indicating "Excitation Current / Response Voltage". A feedback dashed line returns from the bottom "Detection Configuration Evolution Module" to the "Adaptive Focusing" module, labeled "Frequency Band Sensitivity Update ρ_m".

[0086] 2. Feature Encoding Area (light orange): The "Spectral Morphology Feature Encoding Module" is centered, with three output branches labeled: "Curvature Sequence κ_vec", "Arc Length Distribution L_band", and "Asymmetry γ". Next to it are "Traditional ECM Parameter Extraction" (R0 / R_ct / σ_w), indicating that traditional features are retained as a reference. Baseline and Scoring Area (light green): The "Intra-cluster Dynamic Baseline Module" (labeled "Cutoff Mean / Standard Deviation") receives the input from the dashed line on the left, "Other Cell Data within the Cluster". The output is sent to the "Normalized Weighted Scoring Module" and merged with the feature encoding output.

[0087] 3. Prediction and Output Area (light blue): "Aging Grading Judgment Module" → "Aging Trajectory Prediction Engine" (labeled "Historical Sequence → Linear Extrapolation → Estimated Time") → "Result Output and Linkage Module". The bottom independent box "Detection Configuration Evolution Module" receives the full cluster detection results.

[0088] The software control flow of this invention is as follows: Figure 2 As shown, Figure 2 The complete 14-step process is demonstrated. Key differentiating steps are marked with special colors: Step 2 (Adaptive Focusing) in gold, Step 5 (Morphological Encoding) in orange, Step 7 (Population Baseline) in green, Step 10 (Trajectory Prediction) in blue, and Step 12 (Configuration Evolution) in [missing color].

[0089] Purple. Step 12 has a feedback arrow that returns to Step 2.

[0090] The specific process is as follows:

[0091] Step 1: Adaptive Frequency Band Configuration Loading: The system loads the detection configuration parameters. For the initial detection, the default configuration is used: M frequency points are covered at logarithmic intervals from 10mHz to 10kHz, with uniform excitation amplitude distribution across each frequency point. For subsequent detections, the optimized configuration output by the detection configuration evolution module is loaded: the frequency point set, the excitation amplitude distribution ratio for each frequency point, and the frequency band sensitivity sorting.

[0092] Step 2: Generate Adaptive Frequency Band Focusing Excitation Signal: Generate a multi-frequency superimposed excitation signal according to the configuration in Step 1: i_exc(t)=Σ(m=1toM)A_m·sin(2πf_m·t+φ_m); where i_exc(t) is the excitation current signal (unit A); M is the total number of frequency points (positive integer, default 20); A_m is the excitation amplitude of the m-th frequency point (unit A), satisfying A_m≤A_max; f_m is the frequency of the m-th frequency point (unit Hz); φ_m is the initial phase of the Schroeder polyphase sequence optimization (unit rad).

[0093] The adaptive amplitude allocation rule is: A_m = A_total·(λ_m / Σ(m=1 to M)λ_m); where A_total is the total excitation amplitude budget (unit: A, default 100mA); λ_m is the excitation energy allocation coefficient for the m-th frequency point, dimensionless. For the first detection, λ_m = 1 / M (uniform allocation); for subsequent detections, λ_m = max(λ_min, ρ_m^α_focus); where ρ_m is the historical sensitivity value for that frequency point (dimensionless, provided by step 12); α_focus is the focusing index (dimensionless, default 2.0); λ_min is the minimum allocation coefficient (dimensionless, default 0.02), preventing the excitation energy at any frequency point from dropping to zero.

[0094] Step 3: Inject Excitation and Synchronous Sampling: The excitation signal generated in Step 2 is superimposed onto the battery terminal through the excitation injection interface, and the battery terminal voltage response v(t) and actual excitation current i(t) are simultaneously acquired. Sampling preprocessing: outlier removal and Hanning window function weighting.

[0095] Step 4: FFT Frequency Domain Transformation and Impedance Spectrum Construction: Perform FFT on the time-domain sequence to calculate the complex impedance values ​​at each frequency: Z(f_m) = V(f_m) / I(f_m) = |Z(f_m)|·exp(j·θ(f_m)); where Z(f_m) is the complex impedance value at frequency f_m (unit: Ω); V(f_m) is the voltage frequency domain component (unit: V); I(f_m) is the current frequency domain component (unit: A); |Z(f_m)| is the impedance magnitude (unit: Ω); and θ(f_m) is the impedance phase angle (unit: rad). Construct the Nyquist spectrum and perform quality verification.

[0096] Step 5: Spectral Morphology Feature Encoding: This step skips ECM fitting and directly extracts three types of encoded features from the geometric morphology of the Nyquist curve:

[0097] (1) Curvature sequence extraction: The Nyquist curve is parameterized as a function Z(s)=(Re(s),-Im(s)) of arc length s, and the curvature at each point along the curve is calculated as: κ(s)=|d²Z / ds²| / |dZ / ds|³; where κ(s) is the curvature value at the arc length parameter s, in units of Ω⁻¹; d²Z / ds² is the second derivative of the curve with respect to the arc length; dZ / ds is the first derivative of the curve with respect to the arc length. In actual calculations, it is approximated by discrete difference. The curvature sequence is resampled at equal intervals to N_κ points (32 by default), forming a curvature descriptor vector κ_vec=[κ_1,κ_2,...,κ_N_κ].

[0098] (2) Arc length distribution vector: Divide the frequency range into K_band frequency bands (5 by default: UHF, HF, IF, IF, IF) and calculate the proportion of the arc length of the Nyquist curve in each frequency band to the total arc length: L_band_k = L_k / L_total; where L_band_k is the proportion of the arc length of the k-th frequency band (dimensionless, value from 0 to 1); L_k is the arc length of the Nyquist curve in that frequency band (unit: Ω); L_total is the total arc length (unit: Ω). The arc length is calculated by accumulating the Euclidean distances between discrete points.

[0099] (3) Semicircular Asymmetry: Quantify the asymmetric deformation of the mid-frequency semicircle: γ = S_upper / S_lower; where γ is the asymmetry (dimensionless); S_upper is the area enclosed by the upper half of the mid-frequency semicircle (unit Ω²); S_lower is the area enclosed by the lower half (unit Ω²). The axis of symmetry is the diameter of the semicircle (the line connecting R0 to R0+R_ct). For normal cells, γ≈1.0, and for aged cells, γ deviates from 1.0.

[0100] At the same time, traditional ECM parameter features (R0, R_ct, σ_w) are extracted as reference dimensions.

[0101] Step Six: Temperature Compensation Correction: Perform Arrhenius temperature compensation on impedance features: x_j^comp = x_j·exp(E_a_j / R_g·(1 / T_ref-1 / T_meas)); where x_j^comp is the compensated feature value; x_j is the measured value; E_a_j is the activation energy parameter (unit: J / mol); R_g is the ideal gas constant (8.314 J·mol⁻¹·K⁻¹); T_ref is the reference temperature (unit: K, default 298.15 K); T_meas is the measurement temperature (unit: K). Morphological features (curvature sequence and arc length distribution) are compensated using difference normalization.

[0102] Step 7: Establishing the dynamic baseline within the cluster: Collect the detection feature data of all cells in the most recent round within the same cluster, and calculate the population truncation statistic as the dynamic baseline:

[0103] x_j^base_dyn=TrimMean(x_j^(1),x_j^(2),...,x_j^(N_cell),p_trim);

[0104] σ_j^base_dyn=TrimStd(x_j^(1),x_j^(2),...,x_j^(N_cell),p_trim);

[0105] Where x_jbase_dyn is the population dynamic baseline value of feature j; σ_jbase_dyn is the population baseline standard deviation; TrimMean is the truncated mean function; TrimStd is the truncated standard deviation function; N_cell is the total number of cells in the cluster; p_trim is the truncation ratio (dimensionless, default 0.1, i.e., removing the highest and lowest 10%).

[0106] The aging offset is calculated based on the population baseline: Δx_j=(x_j^comp-x_j^base_dyn) / σ_j^base_dyn; where Δx_j is the standardized aging offset (dimensionless), which means "the standard deviation multiple by which the cell deviates from the population median on the j-th feature".

[0107] Step 8: Extended Feature Normalization and Weighted Scoring: Combine the ECM parameter offsets and morphological feature offsets into an extended feature vector, and perform weighted scoring.

[0108] S = Σ(j=1 to J_ecm)w_j·|Δx_j| + Σ(k=1 to J_morph)w_morph_k·d_morph_k; where S is the comprehensive aging score (dimensionless); J_ecm is the number of ECM parameter features (default 3: R0, R_ct, σ_w); w_j is the ECM feature weight (dimensionless); J_morph is the number of morphological features (default 3: curvature sequence distance, arc length distribution offset, asymmetry deviation); w_morph_k is the morphological feature weight (dimensionless, dynamically adjusted by the detection configuration evolution module); d_morph_k is the offset distance of the k-th morphological feature (dimensionless).

[0109] The curvature sequence distance is defined as:

[0110] d_κ=(1 / N_κ)·Σ(i=1toN_κ)(κ_i_current-κ_i_baseline)² / σ_κ_i²; where d_κ is the curvature sequence offset distance (dimensionless); κ_i_current is the i-th curvature value detected currently; κ_i_baseline is the mean curvature of the i-th population baseline; and σ_κ_i is the standard deviation of the i-th curvature of the population baseline.

[0111] Step 9: Aging Classification Determination: Perform a four-level classification determination (L1 / L2 / L3 / L4) based on the comprehensive aging score, and calculate the classification confidence level ρ.

[0112] Step 10: Aging Trajectory Prediction: The aging trajectory prediction engine performs trend extrapolation based on the historical rating sequence {S_1,S_2,...,S_N_hist}. A linear model is fitted to the rating time series: S(t) = S_0 + k_aging·t; where S(t) is the predicted rating at time t; S_0 is the intercept (dimensionless); k_aging is the aging rate (unit: ratings / month); and t is time (unit: month). The predicted time to enter the next level (threshold θ_next) is: t_predict = (θ_next - S_current) / k_aging; where t_predict is the predicted remaining time (unit: month); θ_next is the grading threshold for the next level; and S_current is the current rating.

[0113] Prediction confidence is based on the fitted residuals: Conf_pred=max(0,1-σ_residual / (k_aging·T_horizon)); where Conf_pred is the prediction confidence (dimensionless, 0 to 1); σ_residual is the standard deviation of the linear fit residuals; and T_horizon is the prediction time range (unit: months, default 12 months). When Conf_pred<0.3 or there are fewer than 5 historical data points, it is marked as "Insufficient prediction confidence".

[0114] Step 11: Cluster-level lateral comparison and anomaly localization: Perform a lateral comparison of the current rating and trajectory prediction of all cells in the same cluster. This not only identifies cells with abnormal current ratings, but also cells with abnormally high aging rates (k_aging) (i.e., "potentially risky cells" with normal current ratings but accelerated aging).

[0115] Step 12: Detection Configuration Online Evolution: After each round of full cluster detection, calculate the information contribution of each frequency point: IC_m=|Corr(ΔZ_m,ΔS)|; where IC_m is the information contribution of the m-th frequency point (dimensionless, 0 to 1); ΔZ_m is the impedance change of each cell at frequency point m (based on the offset of the group baseline); ΔS is the aging score of each cell; and Corr is the Pearson correlation coefficient.

[0116] Based on the updated frequency band sensitivity sorting ρ_m=IC_m / max(IC_1,...,IC_M) based on IC_m, it is passed back to the adaptive frequency band focusing excitation module in step two for the redistribution of excitation energy in the next round of detection.

[0117] Step 13: Result Output and Linkage: Output includes the current aging level, level confidence score, aging trajectory prediction (estimated time to enter the next level and predicted confidence score), population baseline comparison report, cluster-level consistency report, and detection configuration evolution log. Linkage alarms are executed according to the level.

[0118] Step Fourteen: Data Recording and Baseline Update: Record the complete detection log. Update the population dynamic baseline (incorporating the latest detection data). Update the historical sequences of the aging trajectory prediction engine.

[0119] like Figure 3 The diagram shows the morphological characteristics of the Nyquist curve. The left side shows the Nyquist curve of a normal cell (solid black line), and the right side shows the curve of the same cell after aging (dashed red line). The following morphological differences are noted: asymmetric deformation of the mid-frequency semicircle (γ changes from 1.0 to 1.15), shift in local high curvature regions (κ peak displacement), and changes in the arc length proportion of each frequency band (increased proportion in the high-frequency band). The small diagram in the lower right corner shows a comparison of curvature columns: the blue line represents the normal curvature profile, and the red line represents the aging curvature profile.

[0120] like Figure 4 The diagram shows a prediction of the aging trajectory. The X-axis represents the number of tests per month, and the Y-axis represents the overall aging score S. Historical data points are represented by blue dots, and linear fitting is represented by a solid blue line. The prediction interval is represented by a light blue shading. Horizontal red dashed lines mark the three thresholds θ1, θ2, and θ3. The intersections of the fitted line with each threshold are labeled "Expected to enter L2 / L3 / L4 in XX months." The upper right corner text box displays "k_aging=0.018 / month, Conf_pred=0.82".

[0121] like Figure 5 The image shows a comparison between the group dynamic baseline and the individual baseline. The left image represents the "individual baseline scheme": using the cell's initial detection value as the baseline (a single horizontal line), the offset may be distorted due to baseline deviation. The right image represents the "group dynamic baseline scheme": using the truncated statistical distribution within the cluster as the baseline (a banded distribution area), the offset is calculated relative to the group median level, eliminating individual bias. The difference in offset between the two schemes for the same aged cell is marked with red arrows.

[0122] Example 1: Cluster-level spectral morphology encoding detection in energy storage power stations:

[0123] 1. Scenario: A 100MWh lithium iron phosphate energy storage power station, with 16 modules × 16 cells / module = 256 cells. The BMS embeds the online detection module of this invention, which performs detection one by one in a polling manner.

[0124] 2. Input configuration: M=20 frequency points (logarithmic intervals from 10mHz to 10kHz), the first round of detection adopts uniform excitation amplitude distribution (A_m=5mA per frequency point), sampling rate 50kHz, detection window 60s.

[0125] 3. Morphological Feature Encoding Process: Taking cell number 37 as an example. After obtaining the 20-point Nyquist curve using FFT:

[0126] (1) Curvature sequence: Resampled to a 32-point curvature descriptor κ_vec. The mid-frequency curvature (points 12-18) of the aged cell increased from 0.85Ω⁻¹ at the baseline to 1.12Ω⁻¹, indicating that the mid-frequency semicircular deformation narrowed.

[0127] (2) Arc length distribution: The arc length ratio of the five frequency bands is [0.08, 0.15, 0.42, 0.22, 0.13] (the baseline of the normal cell is [0.10, 0.18, 0.38, 0.20, 0.14]), and the proportion of the mid-frequency band increases by 0.04.

[0128] (3) Semicircle asymmetry: γ=1.15 (baseline 1.02), the area of ​​the upper half of the mid-frequency semicircle is larger than that of the lower half.

[0129] (4) Simultaneously extract ECM parameters: R0=12.3mΩ, R_ct=8.7mΩ, σ_w=3.2Ω·s⁻¹ / ².

[0130] 4. Population dynamic baseline: The truncated mean R0 of the 256 cell is 9.8 mΩ (truncated by 10%), and the standard deviation of the truncated value is 1.2 mΩ. The standardized offset of R0 at point 37 is (12.3 - 9.8) / 1.2 = 2.08σ. The distance κ_vec is d_κ = 0.35. The γ deviation is |1.15 - 1.02| / 0.05 = 2.6σ.

[0131] 5. Temperature compensation: The detection temperature is 35°C, and the Arrhenius compensation is adjusted to 25°C: R0^comp=10.8mΩ, R_ct^comp=7.5mΩ.

[0132] 6. Scoring and Grading: S = 0.15 × 2.08 + 0.20 × 1.8 + 0.10 × 1.5 + 0.25 × d_κ (1.75 after normalization of 0.35) + 0.15 × L_band offset (1.2) + 0.15 × γ bias (1.3 after normalization of 2.6) = 0.312 + 0.36 + 0.15 + 0.4375 + 0.18 + 0.195 = 1.63 (0.41 after normalization to the 0~1 interval). Based on the thresholds θ1 = 0.25, θ2 = 0.50, and θ3 = 0.75, it is classified as L2 (mild aging), with a confidence level ρ = 0.93.

[0133] 7. Adaptive Frequency Band Focusing Effect: After the first round of detection, the detection configuration evolution module calculates the information contribution of each frequency point. The IC_m for the mid-frequency band (f_8 to f_14) is 0.72 to 0.88 (high contribution), and the IC_m for the high-frequency band (f_17 to f_20) is 0.12 to 0.25 (low contribution). In the second round of detection, 60% of the excitation energy is allocated to the mid-frequency band, 15% to the high-frequency band, and 25% to the low-frequency band. The signal-to-noise ratio of the mid-frequency band is improved by approximately 35%.

[0134] like Figure 6 As shown, (a) is a bar chart of the comprehensive scores of each cell in the cluster, colored according to the level; (b) is a scatter plot of the trajectory prediction results (X=current score, Y=number of months expected to enter L3), marked with "risk zone" (high score + short remaining time); (c) is a heat map of frequency band sensitivity (X=frequency point number, Y=detection round), with the color depth indicating the information contribution, showing how the excitation energy gradually focuses from uniform distribution to key frequency bands.

Claims

1. A battery aging detection method and system based on impedance spectroscopy, characterized in that, The system includes: Adaptive frequency band focusing excitation module: Based on the aging characteristic sensitivity ranking of each frequency band in the previous round of detection results, dynamically adjust the excitation frequency point set and the excitation energy distribution of each frequency point; Synchronous sampling and frequency domain transformation module: The time-domain waveforms of excitation current and response voltage are synchronously acquired through high-speed ADC, and the impedance complex values ​​at each frequency point are calculated by FFT frequency domain transformation to construct a digital impedance spectrum; Spectral morphology feature encoding module: This module directly extracts non-parametric features from the geometric morphology of the Nyquist curve; and retains the traditional ECM parametric features as a reference, which together with the morphological features constitute an extended feature vector; Intra-cluster dynamic baseline module: Calculates the statistical distribution of the group impedance spectrum as a dynamic baseline using the most recent test results of all cells in the same cluster; Temperature compensation module: Based on the Arrhenius temperature compensation model, it performs temperature correction on impedance characteristics, correcting the characteristic values ​​at the measurement temperature to the equivalent values ​​at the reference temperature; Normalized weighted scoring module: Normalizes the feature vectors after temperature compensation and baseline offset, and then calculates the comprehensive aging score by weighted summation; Aging grading module: Based on the comprehensive aging score, the battery is divided into four levels: L1, L2, L3, and L4 according to the threshold, and the grading confidence level is calculated. Aging trajectory prediction module: This module maintains the historical detection feature sequence of each cell and uses the trend of the historical feature sequence to extrapolate and predict the future aging state; Online detection configuration evolution module: This module is responsible for learning from historical detection results and automatically optimizing the detection configuration; Result output and linkage module: This module summarizes the current aging level, trajectory prediction results, population baseline comparison report and detection configuration evolution log to generate a comprehensive detection report; The specific process of the battery aging test method is as follows: S1, Adaptive frequency band configuration loading: The system loads the detection configuration parameters. During the first detection, the default configuration is used: M frequency points are covered at equal logarithmic intervals from 10mHz to 10kHz, and the excitation amplitude of each frequency point is evenly distributed. During subsequent detections, the optimized configuration output by the detection configuration evolution module is loaded: frequency point set, excitation amplitude distribution ratio of each frequency point, and frequency band sensitivity sorting. S2, Adaptive frequency band focusing excitation signal generation: Generate a multi-frequency superimposed excitation signal according to the configuration in step S1: i_exc(t)=Σ(m=1toM)A_m·sin(2πf_m·t+φ_m); where i_exc(t) is the excitation current signal; M is the total number of frequency points; A_m is the excitation amplitude of the m-th frequency point, satisfying A_m≤A_max; f_m is the frequency of the m-th frequency point; φ_m is the initial phase of the Schroeder multiphase sequence optimization; S3, Inject Excitation and Simultaneously Sample: The excitation signal generated in step S2 is superimposed on the battery terminal through the excitation injection interface, and the battery terminal voltage response v(t) and actual excitation current i(t) are simultaneously collected; S4, FFT frequency domain transformation and impedance spectrum construction: Perform FFT on the time domain sequence to calculate the complex impedance values ​​at each frequency point: Z(f_m)=V(f_m) / I(f_m)=|Z(f_m)|·exp(j·θ(f_m)); where Z(f_m) is the complex impedance value at frequency f_m; V(f_m) is the voltage frequency domain component; I(f_m) is the current frequency domain component; |Z(f_m)| is the impedance magnitude; θ(f_m) is the impedance phase angle. Construct the Nyquist spectrum and perform quality verification. S5, Spectral morphology feature encoding: Extract curvature sequence, arc length distribution vector and semicircle asymmetry from the geometric morphology of Nyquist curve, and extract traditional ECM parameter features as reference dimensions. S6, Temperature Compensation Correction: Perform Arrhenius temperature compensation on impedance-type features: x_j^comp=x_j·exp(E_a_j / R_g·(1 / T_ref-1 / T_meas)); where x_j^comp is the compensated feature value; x_j is the measured value; E_a_j is the activation energy parameter; R_g is the ideal gas constant; T_ref is the reference temperature; and T_meas is the measurement temperature. S7, Construction of dynamic baseline within the cluster: Collect detection feature data of all cells in the most recent round within the same cluster, and calculate the population truncation statistic as the dynamic baseline: x_j^base_dyn=TrimMean(x_j^(1),x_j^(2),...,x_j^(N_cell),p_trim); σ_j^base_dyn=TrimStd(x_j^(1),x_j^(2),...,x_j^(N_cell),p_trim); Where x_jbase_dyn is the population dynamic baseline value of feature j; σ_jbase_dyn is the population baseline standard deviation; TrimMean is the truncated mean function; TrimStd is the truncated standard deviation function; N_cell is the total number of cells in the cluster; and p_trim is the truncation ratio. S8, Extended Feature Normalization and Weighted Scoring: Combine ECM parameter offsets and morphological feature offsets into an extended feature vector, and perform weighted scoring. S = Σ(j=1 to J_ecm)w_j·|Δx_j| + Σ(k=1 to J_morph)w_morph_k·d_morph_k; where S is the comprehensive aging score; J_ecm is the number of ECM parameter features; w_j is the ECM feature weight; J_morph is the number of morphological features; w_morph_k is the morphological feature weight; and d_morph_k is the offset distance of the k-th morphological feature. S9, Aging Classification Determination: A four-level classification determination is performed based on the comprehensive aging score, and the classification confidence level ρ is calculated simultaneously. S10, Aging Trajectory Prediction: The aging trajectory prediction engine extrapolates the trend based on the historical rating sequence {S_1,S_2,...,S_N_hist}, fitting a linear model to the rating time series: S(t)=S_0+k_aging·t; where S(t) is the predicted rating at time t; S_0 is the intercept; k_aging is the aging rate; t is time; the predicted time to enter the next level is: t_predict=(θ_next-S_current) / k_aging; where t_predict is the predicted remaining time; θ_next is the grading threshold for the next level; S_current is the current rating; S11, Cluster-level lateral comparison and anomaly localization: Perform a lateral comparison on the current score and trajectory prediction of all cells in the same cluster; S12, Online Evolution of Detection Configuration: After each round of full cluster detection, the information contribution of each frequency point is calculated: IC_m=|Corr(ΔZ_m,ΔS)|; where IC_m is the information contribution of the m-th frequency point; ΔZ_m is the impedance change of each cell at frequency point m; ΔS is the aging score of each cell; and Corr is the Pearson correlation coefficient. S13, Results Output and Linkage: The output includes the current aging level, grade confidence level, aging trajectory prediction, population baseline comparison report, cluster-level consistency report, and detection configuration evolution log; and linkage alarms are executed according to the level. S14, Data Recording and Baseline Update: Record the complete detection log, update the population dynamic baseline, and update the historical sequence of the aging trajectory prediction engine.

2. The battery aging detection method and system based on impedance spectroscopy according to claim 1, characterized in that: In the spectral morphology feature encoding module, non-parametric features include: (1) Curvature sequence κ(s): The curvature is calculated along the arc length parameter s of the Nyquist curve to generate a curvature profile sequence. The mid-frequency semicircular deformation, asymmetry and low-frequency wake bending caused by aging will produce characteristic changes in the curvature sequence. (2) Arc length distribution vector L_band: Divide the frequency range into K frequency bands and calculate the proportion of the arc length of the Nyquist curve in each frequency band to the total arc length. The local impedance change caused by aging will cause the arc length proportion of the corresponding frequency band to shift. (3) Semicircle asymmetry γ: The degree of deviation of the mid-frequency semicircle from the ideal semicircle is quantified and defined as the ratio of the area of ​​the upper half of the semicircle to the area of ​​the lower half. For normal cells, γ is close to 1.0, while for aged cells, γ deviates from 1.

0.

3. The battery aging detection method and system based on impedance spectroscopy according to claim 1, characterized in that: In the intra-cluster dynamic baseline module, the baseline construction method is as follows: (1) Calculate the truncated mean and truncated standard deviation for each dimension of all cells in the cluster; (2) When cell data is insufficient, temporarily revert to the traditional individual baseline strategy, and automatically switch to the group baseline after sufficient group data has been accumulated.

4. The battery aging detection method and system based on impedance spectroscopy according to claim 1, characterized in that: In the normalized weighted scoring module, the weight coefficients consist of two parts: fixed basic weights for ECM parameter features and adaptive weights for morphological features. The weights for morphological features are dynamically adjusted by the detection configuration evolution module based on historical sensitivity analysis.

5. The battery aging detection method and system based on impedance spectroscopy according to claim 1, characterized in that: In the aging trajectory prediction module, the prediction method is as follows: (1) Fit a linear trend to the historical time series of the comprehensive aging score and calculate the time required for the score to grow from the current value to the threshold of each level; the output includes the current aging level, the expected time to enter the next level and the prediction confidence level; (2) The prediction confidence is based on the size of the fitting residuals, that is, the smaller the residuals, the more certain the prediction; (3) When the number of historical data points is less than the preset minimum, mark "Insufficient data, do not predict for now".

6. The battery aging detection method and system based on impedance spectroscopy according to claim 1, characterized in that: In the online evolution module of the detection configuration, after each round of full cluster detection, the online evolution module of the detection configuration calculates the information contribution of each frequency point to the change of aging score. Frequency points with information contribution below the threshold are marked as "redundant frequency points"; frequency points with high contribution are marked as "critical frequency points". In the next round of detection, critical frequency points are allocated more excitation energy, and redundant frequency points can be downgraded or merged into a wider frequency band to shorten the detection time.

7. The battery aging detection method and system based on impedance spectroscopy according to claim 1, characterized in that: In the results output and linkage module, linkage actions are executed according to the aging level: L2 generates observation alarm, L3 triggers maintenance work order, L4 generates emergency alarm, and trajectory prediction results provide forward-looking information such as "expected to enter L3 in N months" to assist in maintenance planning; among them, L1 is normal, L2 is mild aging, L3 is moderate aging, and L4 is severe aging.

8. The battery aging detection method and system based on impedance spectroscopy according to claim 1, characterized in that: In step S2, the adaptive amplitude allocation rule is: A_m = A_total·(λ_m / Σ(m=1 to M)λ_m); where A_total is the total excitation amplitude budget; λ_m is the excitation energy allocation coefficient for the m-th frequency point; for the first detection, λ_m = 1 / M; for subsequent detections, λ_m = max(λ_min, ρ_m^α_focus); where ρ_m is the historical sensitivity value for that frequency point; α_focus is the focusing index; and λ_min is the minimum allocation coefficient to prevent the excitation energy at any frequency point from dropping to zero.

9. The battery aging detection method and system based on impedance spectroscopy according to claim 1, characterized in that: In step S5, (1) Curvature sequence extraction: The Nyquist curve is parameterized as a function Z(s)=(Re(s),-Im(s)) of arc length s, and the curvature at each point along the curve is calculated as: κ(s)=|d²Z / ds²| / |dZ / ds|³; where κ(s) is the curvature value at the arc length parameter s, in units of Ω⁻¹; d²Z / ds² is the second derivative of the curve with respect to the arc length; dZ / ds is the first derivative of the curve with respect to the arc length; (2) Arc length distribution vector: Divide the frequency range into K_band frequency bands and calculate the proportion of the arc length of the Nyquist curve in each frequency band to the total arc length: L_band_k=L_k / L_total; where L_band_k is the proportion of the arc length of the k-th frequency band; L_k is the arc length of the Nyquist curve in that frequency band; and L_total is the total arc length. (3) Semicircle asymmetry: Asymmetric deformation of the quantized mid-frequency semicircle: γ=S_upper / S_lower; where γ is the asymmetry; S_upper is the area enclosed by the upper half of the mid-frequency semicircle; S_lower is the area enclosed by the lower half.

Images