Method for detecting health state of household energy storage lithium battery

Abstract

The invention relates to the technical field of energy storage lithium batteries, in particular to a method for detecting the health state of a household energy storage lithium battery. The method for detecting the state of health of the household energy storage lithium battery is characterized by comprising the following specific steps: S1, preprocessing charge-discharge cycle data; s2, performing multi-domain feature extraction; s3, an SA-XGBoost model oriented to static data and a TSO-LSTM-AM model oriented to dynamic data are constructed respectively, and SOH estimation is carried out on the SA-XGBoost model and the TSO-LSTM-AM model; and S4, finally, performing compression optimization on the model through knowledge distillation and quantification technologies to meet a lightweight deployment requirement. Compared with the prior art, in combination with historical charging and discharging parameters such as battery voltage, current and temperature, feature vectors are constructed through adaptive data cleaning and multi-domain feature extraction, the SOH estimation capability of static and dynamic data is optimized by adopting machine learning and a deep learning model respectively, and the deployment requirement of practical application is met by utilizing a model compression technology.

Description

Technical Field

[0001] This invention relates to the field of energy storage lithium battery technology, specifically a method for detecting the health status of household energy storage lithium batteries. Background Technology

[0002] With the steady progress towards global carbon neutrality and the widespread adoption of home energy storage systems, lithium batteries, with their advantages of high energy density, long cycle life, and low self-discharge rate, have become a core component of modern energy storage. The State of Health (SOH) of lithium batteries, as a key indicator characterizing the degree of performance degradation, directly determines the operational reliability, lifespan, and electrical safety of energy storage systems. Traditional SOH estimation methods mostly rely on physical-electrochemical modeling or empirical formula derivation, which not only involves complex modeling processes requiring specialized domain knowledge but also suffers from poor adaptability to operating conditions and insufficient real-time update capabilities.

[0003] In recent years, data-driven methods, represented by machine learning, have provided a new solution for SOH estimation. These methods avoid explicit modeling of the complex electrochemical mechanisms within the battery, directly extracting battery aging characteristics and degradation patterns from historical operating data, demonstrating good generalization performance. However, their practical application still faces several key challenges: the raw charge-discharge data contains a large amount of noise and outliers, affecting the accuracy of feature extraction; there is a lack of effective cross-cycle, multi-scale fusion mechanisms between dynamic operating data and static characteristic data, failing to fully exploit the complementary potential of multi-source information; and the constructed models are often complex in structure, have a large number of parameters, and high computational overhead, making it difficult to meet the stringent requirements of real-time performance and lightweight deployment for battery management systems (BMS) or cloud deployments. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a method for detecting the state of health (SOH) of household energy storage lithium batteries. It combines historical charge and discharge parameters such as battery voltage, current, and temperature, constructs feature vectors through adaptive data cleaning and multi-domain feature extraction, optimizes the SOH estimation capabilities of static and dynamic data using machine learning and deep learning models respectively, and utilizes model compression technology to meet the deployment requirements of practical applications.

[0005] To achieve the above objectives, a method for detecting the health status of household energy storage lithium batteries is designed, characterized by the following specific method flow:

[0006] S1. First, the charge and discharge cycle data are preprocessed: adaptive window optimization and dual-condition screening are used to remove outliers for static and dynamic data respectively, and abnormal curves are removed based on dynamic thresholds.

[0007] S2, then multi-domain feature extraction is performed: Pearson correlation coefficient analysis is used for static data, and grey relational analysis is used for dynamic data to screen key health features;

[0008] S3, and then construct the SA-XGBoost model for static data and the TSO-LSTM-AM model for dynamic data respectively to estimate SOH;

[0009] S4. Finally, the model is compressed and optimized using knowledge distillation and quantization techniques to meet the requirements for lightweight deployment.

[0010] The specific process of step S1 is as follows:

[0011] S11, for static data, firstly, adaptive window optimization is performed based on local statistical characteristics. By analyzing the fluctuation characteristics of the data neighborhood in real time, the window size is dynamically adjusted to achieve the global optimum of noise suppression and detail preservation. The adaptive window optimization method is based on the current battery capacity C. i Centered on the variable, a sliding window is used to calculate the local variance: Where n is the total number of data points in the window, m is the baseline window half-width, μ is the mean of the data points in the window, i is the current point index, k is the cycle index, and C k It is the battery capacity value in the k-th cycle;

[0012] S12, The adaptive window optimization method establishes the dynamic adjustment equation for the window size: Among them, W base σ is the reference window size. 2 (·) represents the local window variance, θ is the variance threshold, α and β are adjustment coefficients, and W min For the minimum window size constraint, W max Maximum window size constraint;

[0013] S13, by monitoring the local variance σ 2 This enables intelligent adjustment of the window width;

[0014] S14 introduces two key indicators, the volatility significance ratio (R1) and the median deviation (R2), for outliers, using a dual-condition screening approach. The volatility significance ratio measures the degree to which the current point deviates from the window mean, while the median deviation assesses the degree to which the current point deviates from the window mean. The formula is described as follows: Wherein, μ(W i ) is the window W i The mean, range(W) i ' ) to remove C i The range of the back window data, median(W) i ) is the window W iThe median, μ(W) i ' ) to remove C i The mean of the data in the back window;

[0015] S15. For dynamic data, abnormal curves are removed based on a dynamic threshold method with two feature points, i.e., t1 = 500s and t2 = 1000s. Data cleanliness is ensured through residual detection.

[0016] In step S15, the voltage values ​​at two feature points for each charging cycle are extracted to construct feature sequences V. A and V B Extract V A and V B The first 20 maxima and the last 20 minima are used to construct extreme value sample sets S. top and S bottom : In this context, Top20(·) represents the first 20 maxima in the sequence, and Bottom20(·) represents the last 20 minima in the sequence.

[0017] In step S2, a feature extraction framework is constructed for static data using three dimensions: time domain, frequency domain, and entropy features. Features highly correlated with capacity decay are then selected by calculating the Pearson correlation coefficient.

[0018] The time-domain method described above extracts six feature values, defined and described by formulas as follows:

[0019] (1) Capacity decay rate (Decay_Rate): Defined as the differential value of the capacity change between adjacent cycles, the formula is as follows: Among them, Q n Q represents the charging capacity during the nth cycle. n-1 This represents the charging capacity during the (n-1)th cycle.

[0020] (2) Cumulative decay Cum Decay: Defined as the integral of the capacity decay rate, the formula is Cum Decay = Q0 - Q n Where Q0 is the initial charging capacity;

[0021] (3) Standard deviation of voltage fluctuation σ V Defined as the degree of dispersion of the local voltage signal calculated based on the sliding window method, the formula is: Where W is the size of the sliding window, k is the right boundary of the sliding window, and V i Let be the voltage value in the i-th cycle. The average voltage within the window;

[0022] (4) Voltage slope β: Defined as a parameter calculated based on linear regression analysis using a sliding window, the formula is as follows: Among them, t i A sequence of loop counts within the window;

[0023] (5) Internal resistance increment ΔR: defined as the absolute change in ohmic internal resistance, the formula is ΔR = R n -R0; where R n R is the internal resistance of the nth cycle, and R0 is the initial internal resistance;

[0024] (6) Internal resistance acceleration α R Defined as the second derivative of the internal resistance increment, the formula is:

[0025] The time-domain signal was converted to the frequency domain using Fourier transform to reveal the frequency-domain energy distribution characteristics of the electrochemical reactions inside the battery. Three feature values ​​were extracted using the frequency domain method. A sliding window processing method was applied to the voltage signal V(t) and the internal resistance signal R(t) to calculate the center frequency, frequency variance, and root mean square frequency. i The frequency components obtained by Fourier transform calculation, P(f i If is the power spectral density, then the formula for calculating the above eigenvalues ​​is:

[0026] Multiscale arrangement entropy quantifies the changes in the dynamic characteristics of battery signals by analyzing the complexity of the arrangement patterns of time series at different time scales. A decrease in multiscale arrangement entropy indicates enhanced signal order, implying a more stable electrode material structure. An increase in multiscale arrangement entropy indicates increased signal complexity, suggesting that the SEI film of the battery material ruptures, triggering a nonlinear response. Its calculation formula is as follows: Where S is the maximum scale factor. is the permutation entropy value at scale S, and MPE is the multi-scale permutation entropy.

[0027] In step S2, for dynamic data, voltage and current related features are extracted, the key point positioning accuracy is improved by cubic spline interpolation, and the following high correlation features are selected by gray relational analysis: (1) the time t1 when the voltage reaches 4.0V;

[0028] (2) The moment t2 when the voltage reaches 4.2V is the constant current charging time;

[0029] (3) The inflection point location and time t3;

[0030] (4) The time difference Δt1 for the voltage to rise from 4.0V to 4.2V.

[0031] In step S3, for static data, the XGBoost algorithm is used as the basic estimation model, and simulated annealing is introduced to achieve adaptive optimization of key hyperparameters. The Metropolis criterion is used to accept degraded solutions with a certain probability. The Metropolis criterion and temperature update equation are as follows:

[0032] T k+1 =αT k (0<α<1); where ΔE represents the change in the objective function, T is the current temperature, and α is the cooling coefficient.

[0033] In step S3, for dynamic data, an attention mechanism is introduced based on a long short-term memory network, and hyperparameters are tuned using the tuna swarm optimization algorithm.

[0034] Compared with existing technologies, this invention provides a method for detecting the state of health (SOH) of household energy storage lithium batteries. It combines historical charge and discharge parameters such as battery voltage, current, and temperature, constructs feature vectors through adaptive data cleaning and multi-domain feature extraction, optimizes the SOH estimation capability of static and dynamic data using machine learning and deep learning models respectively, and utilizes model compression technology to meet the deployment requirements of practical applications. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the framework of the present invention.

[0036] Figure 2 This is a schematic diagram of the adaptive window optimization method.

[0037] Figure 3 This is a schematic diagram of a simulated annealing algorithm. Detailed Implementation

[0038] The present invention will now be further described with reference to the accompanying drawings.

[0039] The technical solution adopted in this invention is as follows: Figure 1 As shown, the charge-discharge cycle data is first preprocessed. Adaptive window optimization and dual-condition screening are used to remove outliers for both static and dynamic data, as well as anomaly curve removal based on dynamic thresholds. Next, multi-domain feature extraction is performed. Pearson correlation coefficient analysis is used for static data, and grey relational analysis is used for dynamic data to screen key health features. Based on this, SA-XGBoost models for static data and TSO-LSTM-AM models for dynamic data are constructed for SOH estimation. Finally, knowledge distillation and quantization techniques are used to compress and optimize the models to meet lightweight deployment requirements.

[0040] I. Data Preprocessing and Outlier Removal:

[0041] Charge-discharge cycle data generally includes periodically sampled dynamic data and static data sampled and processed during the resting phase. Dynamic data is larger in volume, revealing microscopic mechanisms and providing highly sensitive features for anomaly detection; static data reflects macroscopic performance degradation and provides input for lifetime models. This paper preprocesses the static data from the CALCE dataset and the dynamic data from the NASA dataset, removing outlier data points and curves to obtain cleaned charge-discharge cycle data for subsequent feature extraction.

[0042] For static data, this invention proposes a method for filtering outliers in battery data based on adaptive window adjustment of local variance and dual outlier conditions. This method first performs adaptive window optimization based on local statistical features. By analyzing the fluctuation characteristics of the data neighborhood in real time, it dynamically adjusts the window size to achieve a globally optimal balance between noise suppression and detail preservation. The adaptive window optimization method uses the current battery capacity C as the starting point. i Centered on the variable, a sliding window is used to calculate the local variance: Where n is the total number of data points in the window, m is the baseline window half-width, μ is the mean of the data points in the window, i is the current point index, k is the cycle index, and C k It is the battery capacity value in the k-th cycle.

[0043] The adaptive window optimization method establishes a dynamic adjustment equation for the window size: Among them, W base σ is the reference window size. 2 (·) represents the local window variance, θ is the variance threshold, α and β are adjustment coefficients, and W min For the minimum window size constraint, W max This is a constraint on the maximum window size.

[0044] This mechanism monitors the local variance σ. 2 This enables intelligent adjustment of the window width. For example... Figure 2 As shown, when σ 2 When σ > θ, the current window is determined to be a high-fluctuation region, and the window is reduced to improve resolution; when σ > θ, the current window is determined to be a high-fluctuation region, and the window is reduced to improve resolution. 2 When the value is less than θ, the current window is determined to be in a low-fluctuation region, and the window is enlarged to enhance the trend-following capability.

[0045] To address outliers, this invention introduces two key indicators for dual-condition screening: the volatility significance ratio (R1) and the median deviation (R2). The volatility significance ratio measures the degree to which the current point deviates from the window mean, while the median deviation assesses the degree to which the current point deviates from the window mean. Formula description: Wherein, μ(W i ) is the window W i The mean, range(W) i ') is to remove Ci The range of the back window data, median(W) i ) is the window W i The median, μ(W) i ') is to remove C i The mean of the data in the back window.

[0046] When R1>k and R2>k (the optimal threshold k = 0.685 was determined experimentally), label C. i This is an outlier. Median replacement was used to correct it.

[0047] The median was chosen over the mean as a replacement value primarily because of its robustness: in the presence of non-Gaussian noise or residual outliers, the median can effectively avoid secondary contamination and ensure the physical rationality of the corrected data.

[0048] For dynamic data, this invention proposes a dynamic thresholding method based on two feature points (t1 = 500s, t2 = 1000s) to remove outlier curves, and ensures data cleanliness through residual detection. Specifically, this method first extracts the voltage values ​​at the two feature points for each charging cycle to construct feature sequences V. A and V B Extract V A and V B The first 20 maxima and the last 20 minima are used to construct extreme value sample sets S. top and S bottom : In this context, Top20(·) represents the first 20 maxima in the sequence, and Bottom20(·) represents the last 20 minima in the sequence.

[0049] Calculate and compare the extreme value sample sets S top K and S bottom K The residuals R between each extreme point and the sample median i The mean of the residuals between the current point and the median at all other points. Anomaly detection is performed using dynamic threshold criteria. Formula description: R i =|v i -median(S)|(v i ∈S); Where median(S) is the median of the extreme value sample set S, v i R represents a voltage extreme value in the sample set S. j It is the residual between the j-th (j≠i) voltage extreme point in the sequence and the sample median. Where σ = 4 is the optimal threshold coefficient determined through a large number of experiments.

[0050] II. Multi-domain Feature Extraction and Correlation Analysis:

[0051] Static data contains limited battery information and relies on multiple iterations for SOH estimation. Therefore, for static data, this invention constructs a feature extraction framework from three dimensions: time domain, frequency domain, and entropy features. By calculating the Pearson correlation coefficient, features highly correlated with capacity decay are selected.

[0052] Time-domain analysis primarily extracts statistical features from signal waveforms. This invention employs a time-domain method to extract six feature values, defined and described by the formulas below:

[0053] 1. Capacity decay rate (Decay_Rate): Defined as the differential value of the capacity change between adjacent cycles, the formula is as follows: Among them, Q n Q is the charging capacity for the nth cycle. n-1 This represents the charging capacity during the (n-1)th cycle.

[0054] 2. Cumulative Decay (Cum Decay): Defined as the integral of the capacity decay rate, with the formula Cum Decay = Q0 - Q n Where Q0 is the initial charging capacity.

[0055] 3. Standard deviation of voltage fluctuation σ V Defined as the degree of dispersion of the local voltage signal calculated based on the sliding window method, the formula is: Where W is the size of the sliding window, k is the right boundary of the sliding window, and V i Let be the voltage value in the i-th cycle. This represents the average voltage within the window.

[0056] 4. Voltage slope β: Defined as a parameter calculated based on linear regression analysis using a sliding window, the formula is as follows: Among them, t i This is a sequence of the number of loop iterations within the window.

[0057] 5. Internal resistance increment ΔR: Defined as the absolute change in ohmic internal resistance, the formula is ΔR = R n -R0; where R n R is the internal resistance of the nth cycle, and R0 is the initial internal resistance.

[0058] 6. Internal resistance acceleration α R Defined as the second derivative of the internal resistance increment, the formula is:

[0059] By converting time-domain signals to the frequency domain using Fourier transform (FFT), the frequency-domain energy distribution characteristics of electrochemical reactions within a battery can be revealed. This invention employs a frequency-domain method to extract three feature values, applies a sliding window processing to the voltage signal V(t) and internal resistance signal R(t), and calculates the center frequency (FC), frequency variance (FV), and root mean square frequency (FRMS).

[0060] f i For the frequency components obtained from FFT calculation, P(f i If is the power spectral density, then the formula for calculating the above eigenvalues ​​is:

[0061] To further capture the nonlinear characteristics of the signal, this invention introduces multi-scale permutation entropy (MPE) analysis. MPE quantifies the changes in the dynamic characteristics of the battery signal by analyzing the complexity of the permutation patterns of the time series at different time scales. A decrease in MPE indicates enhanced signal order, which may mean that the electrode material structure is becoming more stable. An increase in MPE indicates increased signal complexity, which may mean that the SEI film of the battery material is ruptured, triggering a nonlinear response. The calculation formula is as follows: Where S is the maximum scale factor. Let S be the permutation entropy value at scale S.

[0062] Dynamic data contains point-to-point characteristic information at various moments during the battery charge-discharge cycle, and is rich in information. Therefore, for dynamic data, this invention extracts voltage and current-related features, improves the positioning accuracy of key points (4.0V voltage point, 4.2V voltage point, 0.2A current point, 1.0A current point) through cubic spline interpolation, and uses grey relational analysis (GRA) to select the following high-correlation features:

[0063] 1. The moment t1 when the voltage reaches 4.0V;

[0064] 2. The moment t2 when the voltage reaches 4.2V is the constant current charging time;

[0065] 3. The inflection point at time t3;

[0066] 4. The time difference Δt1 for the voltage to rise from 4.0V to 4.2V.

[0067] III. Machine Learning Modeling:

[0068] To achieve high-precision SOH estimation, this invention designs optimal machine learning models for static and dynamic data respectively, forming a two-branch modeling solution.

[0069] To address the characteristics of static data, this invention employs the XGBoost algorithm as the basic estimation model and introduces simulated annealing to achieve adaptive optimization of key hyperparameters, thereby further improving model accuracy and generalization ability.

[0070] XGBoost is a machine learning algorithm based on the gradient boosting framework. It iteratively generates a series of decision trees, each tree aiming to fit the residual of the prediction result of the previous tree, and finally outputs the prediction result in a weighted ensemble. Its objective function combines the loss function and the regularization term, which effectively suppresses overfitting while improving the fitting ability.

[0071] To optimize the hyperparameters of the XGBoost model, this invention introduces a simulated annealing algorithm. Simulated annealing is a global optimization algorithm inspired by the solid-state annealing process, and its core mechanism lies in dynamically adjusting the search strategy by controlling the temperature parameter. Figure 3 As shown, in the early stage of the search (high temperature stage), the Metropolis criterion accepts the inferior solution with a certain probability; as the iteration proceeds, the temperature gradually decreases, the probability of the algorithm accepting the inferior solution decreases, and finally converges to the global optimum.

[0072] Metropolis Criterion and Temperature Update Equation Formula:

[0073] T k+1 =αT k (0<α<1); where ΔE represents the change in the objective function, T is the current temperature, and α is the cooling coefficient.

[0074] Compared to traditional optimization methods such as grid search and random search, simulated annealing can effectively escape local optima due to its probabilistic acceptance mechanism, and its temperature decay mechanism can achieve a balance between global search and local optimization, making it suitable for continuous and discrete mixed parameter spaces.

[0075] In this invention, three key hyperparameters affecting model performance in XGBoost were selected as optimization variables: the number of decision trees, the maximum tree depth, and the learning rate. Initial solutions were chosen with 100 decision trees, a maximum tree depth of 3, and a learning rate of 0.1. The mean squared error (MSE) of the model on the validation set was used as the fitness function, and simulated annealing was employed for parameter tuning. After optimization, the SA-XGBoost model showed significant reductions in both mean absolute error (MAE) and root mean square error (RMSE) on the CALCE dataset, and a significant improvement in SOH estimation accuracy, as detailed in Table 1 (Performance metrics of the SA optimized model).

[0076] Table 1

[0077]

[0078] To address the characteristics of dynamic data, this invention is based on a Long Short-Term Memory (LSTM) network, introduces an attention mechanism (AM), and uses the Tuna Swarm Optimization (TSO) algorithm for hyperparameter tuning.

[0079] LSTM, by introducing gating mechanisms (input gate, forget gate, output gate), can solve the gradient vanishing and gradient exploding problems encountered by standard recurrent neural networks during training, thus better handling long-term dependencies in long sequence data. To further enhance the model's ability to perceive the importance of different time steps in the input sequence, this invention introduces an additive attention mechanism after the LSTM layer. This mechanism uses a small neural network to calculate the correlation score between the hidden state and the final state at each time step, normalizes it into attention weights, and finally performs a weighted sum of the hidden states at all time steps to form a context vector. This allows the model to focus on the key time segments most relevant to SOH decay, effectively improving the model's performance and interpretability.

[0080] To optimize the hyperparameters of LSTM-AM models, this invention introduces a tuna swarm optimization algorithm. Tuna swarm optimization is a metaheuristic optimization algorithm that simulates the cooperative foraging behavior of tuna groups, possessing advantages such as few parameters, fast convergence speed, and strong global search capability. Its optimization process mainly consists of two strategies: spiral foraging and parabolic cooperative foraging, maintaining a good balance between exploration and exploitation.

[0081] This invention applies the TSO algorithm to optimize three key hyperparameters in the LSTM-AM model: learning rate, hidden layer dimension, and attention dimension. After optimization, the TSO-LSTM-AM model shows significant reductions in MAE and RMSE on the NASA dataset, and an improvement in the coefficient of determination. Its performance is significantly better than other mainstream single models, as shown in Table 2 (Comparison of SOH estimation algorithms for dynamic data).

[0082] Table 2

[0083]

[0084] IV. Model Compression and Deployment Optimization:

[0085] To apply the high-precision model to resource-constrained residential energy storage BMS, this invention lightweights the computationally intensive dynamic SOH estimation model (TSO-LSTM-AM) and proposes a model compression scheme that combines knowledge distillation and INT8 quantization.

[0086] Knowledge distillation is a model compression technique that aims to transfer knowledge from a large and complex "teacher model" to a lightweight "student model." This invention uses a pre-trained TSO-LSTM-AM model as the teacher model and constructs a simpler LSTM network (e.g., fewer layers, fewer hidden units) as the student model. Model distillation is set with 180 iterations, a temperature of 120, and a weight coefficient of 0.5. Based on the mean squared error loss, the losses between the student model and the teacher model outputs, and between the student model and the ground truth labels, are calculated. These losses are then weighted and summed to obtain the total loss. The Adam optimizer is used to update the student model parameters, allowing the student model to learn knowledge from the teacher model, thus reducing scale and computational complexity. To further compress the model, this invention introduces INT8 quantization after knowledge distillation, quantizing the fully connected layer parameters in the model into INT8 type, reducing storage and computational resource consumption.

[0087] Table 3

[0088]

[0089] Table 3 shows the changes in deployment metrics and estimation errors before and after model compression (comparison of deployment metrics and estimation errors before and after model compression). After model compression, the number of parameters decreased from 48.63K to 34.92K, a decrease of 28%; the computational complexity decreased from 3.412MFLOPs to 2.851MFLOPs, a decrease of 16%; and the model size decreased from 0.189MB to 0.153MB, a decrease of 19%. Regarding error metrics, MAE increased by 0.051%, an increase of 8%, and RMSE increased by 0.013%, an increase of 1.5%. Experiments show that the compression scheme adopted in this invention significantly reduces key deployment metrics such as the number of parameters, computational complexity, and model size while maintaining low estimation accuracy loss, meeting the real-time requirements of practical applications.

[0090] The advantages of this invention are:

[0091] (1) A static data cleaning method based on adaptive window optimization and dual anomaly screening mechanism is proposed, and combined with a dynamic data cleaning method based on dynamic threshold discrimination of dual feature points, which can effectively retain the key trend of characterization capacity decay in a strong noise environment and avoid false rejection and information loss.

[0092] (2) A two-branch machine learning modeling method is proposed to improve estimation accuracy and generalization ability: Based on static features, a static SOH estimation model is constructed based on XGBoost and optimized by simulated annealing (SA); based on dynamic features, a dynamic SOH estimation model is constructed based on long short-term memory network (LSTM-AM) with attention mechanism and optimized by tuna swarm optimization (TSO).

Claims

1. A method for detecting the health status of household energy storage lithium batteries, characterized in that, The specific method and process are as follows: S1. First, the charge and discharge cycle data are preprocessed: adaptive window optimization and dual-condition screening are used to remove outliers for static and dynamic data respectively, and abnormal curves are removed based on dynamic thresholds. S2, then multi-domain feature extraction is performed: Pearson correlation coefficient analysis is used for static data, and grey relational analysis is used for dynamic data to screen key health features; S3, and then construct the SA-XGBoost model for static data and the TSO-LSTM-AM model for dynamic data respectively to estimate SOH; S4. Finally, the model is compressed and optimized using knowledge distillation and quantization techniques to meet the requirements for lightweight deployment.

2. The method for detecting the health status of household energy storage lithium batteries according to claim 1, characterized in that, The specific process of step S1 is as follows: S11, for static data, firstly, adaptive window optimization is performed based on local statistical characteristics. By analyzing the fluctuation characteristics of the data neighborhood in real time, the window size is dynamically adjusted to achieve the global optimum of noise suppression and detail preservation. The adaptive window optimization method is based on the current battery capacity C. i Centered on the variable, a sliding window is used to calculate the local variance: Where n is the total number of data points in the window, m is the baseline window half-width, μ is the mean of the data points in the window, i is the current point index, k is the cycle index, and C k It is the battery capacity value in the k-th cycle; S12, The adaptive window optimization method establishes the dynamic adjustment equation for the window size: Among them, W base σ is the reference window size. 2 (·) represents the local window variance, θ is the variance threshold, α and β are adjustment coefficients, and W min For the minimum window size constraint, W max Maximum window size constraint; S13, by monitoring the local variance σ 2 This enables intelligent adjustment of the window width; S14 introduces two key indicators, the volatility significance ratio (R1) and the median deviation (R2), for outliers, using a dual-condition screening approach. The volatility significance ratio measures the degree to which the current point deviates from the window mean, while the median deviation assesses the degree to which the current point deviates from the window mean. The formula is described as follows: Wherein, μ(W i ) is the window W i The mean, range(W) i ' ) to remove C i The range of the back window data, median(W) i ) is the window W i The median, μ(W) i ' ) to remove C i The mean of the data in the back window; S15. For dynamic data, abnormal curves are removed based on a dynamic threshold method with two feature points, i.e., t1 = 500s and t2 = 1000s. Data cleanliness is ensured through residual detection.

3. The method for detecting the health status of household energy storage lithium batteries according to claim 2, characterized in that: In step S15, the voltage values ​​at two feature points for each charging cycle are extracted to construct feature sequences V. A and V B Extract V A and V B The first 20 maxima and the last 20 minima are used to construct extreme value sample sets S. top and S bottom : In this context, Top20(·) represents the first 20 maxima in the sequence, and Bottom20(·) represents the last 20 minima in the sequence.

4. The method for detecting the health status of household energy storage lithium batteries according to claim 1, characterized in that, In step S2, a feature extraction framework is constructed for static data using three dimensions: time domain, frequency domain, and entropy features. Features highly correlated with capacity decay are then selected by calculating the Pearson correlation coefficient.

5. The method for detecting the health status of household energy storage lithium batteries according to claim 4, characterized in that, The time-domain method described above extracts six feature values, defined and described by formulas as follows: (1) Capacity decay rate (Decay_Rate): Defined as the differential value of the capacity change between adjacent cycles, the formula is as follows: Among them, Q n Q represents the charging capacity during the nth cycle. n-1 This represents the charging capacity during the (n-1)th cycle. (2) Cumulative decay Cum Decay: Defined as the integral of the capacity decay rate, the formula is Cum Decay = Q0 - Q n Where Q0 is the initial charging capacity; (3) Standard deviation of voltage fluctuation σ V Defined as the degree of dispersion of the local voltage signal calculated based on the sliding window method, the formula is: Where W is the size of the sliding window, k is the right boundary of the sliding window, and V i Let be the voltage value in the i-th cycle. The average voltage within the window; (4) Voltage slope β: Defined as a parameter calculated based on linear regression analysis using a sliding window, the formula is as follows: Among them, t i A sequence of loop counts within the window; (5) Internal resistance increment ΔR: defined as the absolute change in ohmic internal resistance, the formula is ΔR = R n -R0; where R n R is the internal resistance of the nth cycle, and R0 is the initial internal resistance; (6) Internal resistance acceleration α R Defined as the second derivative of the internal resistance increment, the formula is:

6. The method for detecting the health status of household energy storage lithium batteries according to claim 4, characterized in that: The time-domain signal was converted to the frequency domain using Fourier transform to reveal the frequency-domain energy distribution characteristics of the electrochemical reactions inside the battery. Three feature values ​​were extracted using the frequency domain method. A sliding window processing method was applied to the voltage signal V(t) and the internal resistance signal R(t) to calculate the center frequency, frequency variance, and root mean square frequency. i The frequency components obtained by Fourier transform calculation, P(f i If is the power spectral density, then the formula for calculating the above eigenvalues ​​is:

7. The method for detecting the health status of household energy storage lithium batteries according to claim 4, characterized in that: Multiscale arrangement entropy quantifies the changes in the dynamic characteristics of battery signals by analyzing the complexity of the arrangement patterns of time series at different time scales. A decrease in multiscale arrangement entropy indicates enhanced signal order, implying a more stable electrode material structure. An increase in multiscale arrangement entropy indicates increased signal complexity, suggesting that the SEI film of the battery material ruptures, triggering a nonlinear response. Its calculation formula is as follows: Where S is the maximum scale factor. is the permutation entropy value at scale S, and MPE is the multi-scale permutation entropy.

8. The method for detecting the health status of household energy storage lithium batteries according to claim 1, characterized in that: In step S2, for dynamic data, voltage and current related features are extracted, cubic spline interpolation is used to improve the positioning accuracy of key points, and grey relational analysis is used to select the following high-correlation features: (1) The moment t1 when the voltage reaches 4.0V; (2) The moment t2 when the voltage reaches 4.2V is the constant current charging time; (3) The inflection point location and time t3; (4) The time difference Δt1 for the voltage to rise from 4.0V to 4.2V.

9. The method for detecting the health status of household energy storage lithium batteries according to claim 1, characterized in that: In step S3, for static data, the XGBoost algorithm is used as the basic estimation model, and the simulated annealing algorithm is introduced to achieve adaptive optimization of key hyperparameters, and the Metropolis criterion is used to accept the degraded solution with a certain probability. Metropolis Criterion and Temperature Update Equation Formula: T k+1 = αT k (0 < α < 1); where ΔE represents the change in the objective function, T is the current temperature, and α is the cooling coefficient.

10. The method for detecting the health status of household energy storage lithium batteries according to claim 1, characterized in that: In step S3, for dynamic data, an attention mechanism is introduced based on a long short-term memory network, and hyperparameters are tuned using the tuna swarm optimization algorithm.

Images