Battery health estimation method and electronic device based on wavelet fuzzy threshold denoising

Abstract

The present invention discloses a battery health estimation method and electronic device based on wavelet fuzzy threshold denoising, which belongs to the field of new energy power battery health management. This method addresses the problems of severe noise interference, poor model generalization and high overfitting risk in the existing technology, and proposes the following technical solutions: first, the lithium-ion battery aging data is preprocessed; secondly, the wavelet fuzzy threshold denoising method is used to achieve adaptive denoising and build a case library; then, a regularized stacked long short-term memory network model is constructed, and the objective function is optimized through parameter constraints and upper and lower bounds of regularization coefficients to suppress overfitting; finally, the denoised case library is input into the regularized stacked long short-term memory network model, and after iterative optimization, the trained optimal prediction model is output to estimate the health status of the lithium-ion battery. Experiments show that this method can significantly improve the accuracy and robustness of battery health status prediction, providing a reliable basis for new energy battery health management.

Description

Technical Field

[0001] The present invention belongs to the field of new energy power battery health management, and specifically relates to a battery health estimation method and electronic equipment based on wavelet fuzzy threshold denoising. Background Art

[0002] Lithium-ion batteries, with their significant advantages of high energy storage density and long service life, have become a core focus in the new energy sector. Predicting remaining useful life (RUL) and estimating state of health (SOH) are key research topics in battery health management and reliability engineering. Accurate battery health estimation methods can optimize equipment maintenance strategies, extend service life, and reduce lifecycle costs, making them a research hotspot for electronic system health management and fault diagnosis.

[0003] As a core component of the new energy industry, traditional prediction models for lithium-ion batteries are based on physical factors such as electrical impedance. However, in practice, uncertainties such as the internal environment (impedance parameters) and external environment (operating environment) prevent accurate modeling of the battery's full lifecycle. This directly results in traditional prediction technologies often only being able to predict lithium-ion batteries under a single operating condition and lacking effective processing of raw data.

[0004] Existing methods for preprocessing battery degradation data have significant shortcomings. Traditional denoising techniques, such as sliding average filtering and Fourier transforms, can suppress high-frequency noise but are unable to effectively handle non-stationary noise (such as random fluctuations during the charge and discharge process) and sudden spikes (such as abnormal signals caused by transient sensor failures or internal battery micro-shorts) in lithium-ion battery degradation data. Sliding average filtering, due to its fixed window, causes signal delay and has limited ability to smooth sudden spikes. Fourier transforms, due to their reliance on global frequency decomposition, struggle to capture local time-domain features, leading to signal distortion. Furthermore, existing wavelet threshold denoising methods (such as hard thresholding and soft thresholding) have the following problems: the hard threshold function is discontinuous at the threshold point, easily introducing pseudo-Gibbs phenomena and causing oscillations in the reconstructed signal. The soft threshold function, while continuous, has a fixed global threshold and cannot dynamically adjust the threshold based on local signal characteristics (such as local energy and variance). This can oversmooth useful signals (such as the capacity decay inflection point) or leave residual low-frequency noise. This problem is particularly prominent in the battery capacity decay curve. For example, when the battery is nearing the end of its life, the capacity drop signal may be misjudged as noise and filtered out, directly affecting the prediction accuracy.

[0005] Existing battery health prediction models have poor generalization capabilities and perform poorly across operating conditions. This is primarily due to the following: physical models (such as equivalent circuit models based on electrochemical impedance spectroscopy) rely on precise internal battery parameters (such as SEI film thickness and electrolyte concentration). However, in actual applications, these parameters change dynamically due to factors such as temperature and cycle number, making it difficult to update the model in real time. Traditional long-short-term memory networks (LSTMs) have weak generalization capabilities across battery types or different charge and discharge modes. Furthermore, existing methods are mostly trained based on single-operating condition data and fail to consider the nonlinear effects of multivariate coupling on battery degradation in actual applications, causing the model to fail under complex operating conditions.

[0006] Deep learning models face the following challenges in battery health prediction: Large parameter size: Deep networks like stacked LSTMs have high parameter counts (e.g., millions of parameters), making them prone to overfitting with limited training data. Traditional regularization methods also have shortcomings: L2 norm regularization insufficiently penalizes sparse parameters, and because the objective function is non-differentiable at zero, it can easily lead to vanishing gradients. Existing methods (such as principal component analysis) rely on linear transformations and cannot effectively capture the nonlinear dynamic characteristics of battery degradation.

[0007] While hybrid models proposed recently (such as CNN-LSTM and physics-data fusion models) attempt to combine the strengths of multiple approaches, they suffer from several challenges. Firstly, their structure is complex. CNN-LSTM requires simultaneous training of convolution kernels and LSTM units, resulting in complex parameter interactions and a 2-3x increase in training time compared to single models. Secondly, hybrid models are computationally intensive (for example, real-time inference latency exceeds 1 second), making them difficult to meet the real-time requirements of vehicle battery management systems (BMSs).

[0008] Existing technologies for processing battery degradation data have significant shortcomings in noise suppression, model generalization, overfitting risk, and practicality. Therefore, research into battery health estimation methods based on wavelet fuzzy threshold denoising and regularized stacked LSTMs is of great significance in systematically improving prediction accuracy and engineering applicability. Summary of the Invention

[0009] The purpose of the present invention is to provide a battery health estimation method and electronic equipment based on wavelet fuzzy threshold denoising, systematically solving the problems of severe noise interference, poor model generalization, and high overfitting risk in lithium-ion battery health prediction, and providing a high-precision and high-reliability technical solution for new energy battery management.

[0010] To achieve the above object, the present invention adopts a technical solution: a battery health estimation method based on wavelet fuzzy threshold denoising, comprising the following steps:

[0011] S1. Preprocessing the original lithium-ion battery aging data;

[0012] S2. Smoothing the pre-processed data using wavelet fuzzy threshold denoising method to eliminate noise and spikes, and constructing a denoised case library;

[0013] S3. Constructing a regularized stacked long short-term memory network model, including stacked LSTM layers and a regularization optimization module, wherein the regularization optimization module optimizes the objective function by introducing parameter constraints and upper and lower bounds of regularization coefficients. The regularization coefficients optimize the norm function of the model parameters while regularizing the loss function to suppress overfitting;

[0014] S4. Input the denoised case library data into the regularized stacked long short-term memory network model for training, evaluate the model performance, adjust the model hyperparameters until the optimal performance is achieved, output the trained optimal prediction model, and use the optimal prediction model to estimate the health status of the lithium-ion battery.

[0015] Furthermore, step S1 specifically includes: eliminating battery data with a cycle number less than 500 and abnormal values caused by the test circuit or sensor; when the battery capacity decays to 80% of the rated capacity, it is determined that the battery life has ended.

[0016] Furthermore, in step S2, a wavelet fuzzy threshold denoising method is used for smoothing, which specifically includes:

[0017] S21, using Laplace Gaussian wavelet to perform multi-scale decomposition on the battery attenuation signal to obtain an approximate part and a noise part;

[0018] S22, dynamically adjusting the threshold based on an adaptive fuzzy soft threshold method, using the absolute value of the wavelet coefficient, the local standard deviation, and the local energy as fuzzy inputs through an adaptive fuzzy rule base, outputting a dynamic threshold level, and performing defuzzification using a weighted average method to obtain an attenuated signal after threshold processing;

[0019] S23, sorting the attenuated signals after threshold processing according to their absolute values;

[0020] S24. Reconstruct the arranged attenuated signal sequence using inverse wavelet transform to obtain a noise-processed signal.

[0021] Furthermore, in step S21, the expression of Laplace Gaussian wavelet is:

[0022]

[0023] Where ψ(t) represents the Laplace-Gaussian wavelet function, t represents the independent variable, and e is a constant.

[0024] Furthermore, in step S22, the specific implementation process of the adaptive fuzzy soft threshold method is as follows:

[0025] First, a triangular fuzzy function is used to divide the variables into three levels, and an adaptive fuzzy rule base is established. The adaptive fuzzy rule base includes the following rules: if the absolute value of the wavelet coefficient is small, the local standard deviation is low, and the local energy is weak, the threshold level is high; if the absolute value of the wavelet coefficient is medium and the local standard deviation is medium, the threshold level is medium; if the absolute value of the wavelet coefficient is large, the local standard deviation is high, or the local energy is strong, the threshold level is low;

[0026] Then, the weighted average method is used for defuzzification, and its expression is:

[0027]

[0028] In the formula, λ is the threshold value output by weighted average method, u i is the activation degree of each rule, c i is the central value of the corresponding output membership function;

[0029] Finally, the fuzzy system output threshold λ(w) is applied to each wavelet coefficient w to obtain the wavelet coefficient after soft threshold processing:

[0030]

[0031] Where W λ is the wavelet coefficient after soft threshold processing; sign(x) is the sign function, which is defined as: when x>0, sign(x)=1, when x=0, sign(x)=0, when x<0, sign(x)=-1; w is the original coefficient; λ(w) is the threshold output by the fuzzy system corresponding to the coefficient w.

[0032] Furthermore, the objective function of regularization optimization in step S3 is:

[0033]

[0034] Where N is the number of samples, f(x n ; n) is the depth prediction model to be trained, θ is its parameter, f(n) is the true output of the nth sample, x n is the input of the nth sample; D is the dimension of parameter θ, d is the dth component of parameter θ, ε is a small positive number; p and q are both regularization coefficients, p+q=1, p>0, q>0.

[0035] Furthermore, the parameters of the regularized stacked long short-term memory network model in step S3 include: input data length is 50, the number of stacking layers is 5, the number of LSTM layer units is 30, the learning rate is 0.001, the optimizer is Adam, and the loss function is MSE.

[0036] Furthermore, in step S4, the hyperparameter adjustment adopts the grid search method and five-fold cross validation, and the iteration termination condition is that the MSE is less than 0.1.

[0037] Furthermore, in step S4, the mean square error is used to evaluate the model performance, and its calculation formula is:

[0038]

[0039] Where, and y i are the predicted and actual values of the i-th battery tuple, and n represents the number of battery tuples.

[0040] The present invention also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the battery health estimation method based on wavelet fuzzy threshold denoising is implemented.

[0041] The beneficial effects of the above scheme are:

[0042] (1) The present invention effectively improves the noise suppression capability and retains key signal features. The wavelet fuzzy threshold denoising method is used to dynamically adjust the threshold according to the signal's local energy, standard deviation and other characteristics, solving the "unclear boundary" problem caused by the traditional fixed threshold, effectively distinguishing noise signals from normal signals, weakening useless parts and enhancing useful parts, obtaining smooth data after eliminating noise and spikes, and improving data quality; the Laplace Gaussian wavelet is used for wavelet decomposition, and its continuous and smooth characteristics can accurately detect abnormal attenuation points in the battery continuous attenuation signal. Experimental data show that ( Figure 5 ), the battery capacity decay curve after denoising is smooth and the mutation points (such as capacity "diving") are completely preserved, which strongly proves the effectiveness and superiority of this method in data processing.

[0043] (2) The present invention can enhance the generalization ability of the model and adapt to multiple working conditions. This method performs regularization optimization on the parameters in the stacked long short-term memory network model. Innovatively, while optimizing the norm function, it also regularizes the loss function and imposes upper and lower bounds on the regularization coefficient, effectively suppressing overfitting and making the model more stable during training, thereby improving the generalization ability of the model. Comparative experiments show that the model prediction MSE error is significantly lower than that of the traditional CNN-LSTM across battery types and different charge and discharge rate scenarios. In addition, the data processing method and powerful model structure of the present invention stabilize the mean absolute error of the model under complex working conditions, can adapt to the diversity and complexity of battery data under different working conditions, and improve its practicality in actual scenarios.

[0044] (3) Through the above-mentioned technological innovations, the present invention effectively improves the prediction accuracy and reliability, and can accurately capture the nonlinear dynamic characteristics of battery degradation data. From the actual experimental results, the prediction model of the present invention has a model evaluation index (error value MSE) that is lower than other models in the comparative experiment. In comparison with the convolutional active residual network (CNN-ATResNet) and the convolutional long short-term memory network (CNN-LSTM), the method proposed by the present invention has a significantly lower MSE on samples with different battery numbers. Experimental verification: in the test of 100 sets of battery data, the average MSE of the method of the present invention is 20.75, which is significantly lower than that of CNN-ATResNet (average MSE 33.75) and CNN-LSTM (average MSE 60.5).

[0045] (4) The present invention can reduce computational complexity and improve engineering practicality. By setting specific model parameters, combined with the use of grid search and five-fold cross-validation, the model hyperparameter adjustment time is significantly shortened, and the number of training iterations is reduced to 300. The optimized regularized stacked LSTM model meets the real-time requirements of the vehicle battery management system. Through high-precision SOH prediction, battery maintenance or replacement strategies can be planned in advance to avoid sudden failures. In the field of batteries, especially in the field of new energy vehicle power batteries, the present invention has very broad market value and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the battery health estimation method of the present invention;

[0047] Figure 2 This is a flow chart of the wavelet fuzzy threshold denoising of the present invention;

[0048] Figure 3 is a graph showing the capacity of the lithium battery in Example 1 attenuated with the number of cycles;

[0049] Figure 4In embodiment 1, a triangular fuzzy function is used to perform fuzzy division on the fuzzy threshold parameters;

[0050] Figure 5 This is a comparison chart of the effects of using wavelet fuzzy soft threshold denoising in Example 1;

[0051] Figure 6 is a training flow chart of the regularized stacked long short-term memory network model in Example 1;

[0052] Figure 7 is the prediction result of the regularized stacked long short-term memory network model of the present invention in Example 1;

[0053] Figure 8 This is the comparison result between the regularized stacked long short-term memory network model of the present invention and other prediction methods. DETAILED DESCRIPTION

[0054] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] It should be noted that unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0056] Example 1:

[0057] like Figure 1 As shown, a battery health estimation method based on wavelet fuzzy threshold denoising includes the following steps:

[0058] S1. Preprocessing the raw lithium-ion battery aging data; the preprocessing includes eliminating battery data with less than 500 cycles, using the triple standard deviation principle to eliminate outliers caused by test circuit or sensor failures, and performing z-score normalization on the data.

[0059] S2. The preprocessed data is smoothed using a wavelet fuzzy threshold denoising method to eliminate noise and spikes, and a denoised case library is constructed. The wavelet fuzzy threshold denoising method uses Laplace Gaussian wavelets to perform multi-scale decomposition of the battery attenuation signal, dynamically adjusts the threshold based on local energy and local standard deviation, and outputs a dynamic threshold level according to an adaptive fuzzy rule library.

[0060] S3. Construct a regularized stacked long short-term memory network model, including stacked LSTM layers and a regularization optimization module, wherein the regularization optimization module optimizes the objective function by introducing parameter constraints and upper and lower bounds of the regularization coefficient. The regularization coefficient optimizes the norm function of the model parameters while regularizing the loss function to suppress overfitting.

[0061] S4. Input the denoised case library data into the regularized stacked long short-term memory network model for training, evaluate the model performance, adjust the model hyperparameters until the optimal performance is achieved, output the trained optimal prediction model, and use the optimal prediction model to estimate the health status of the lithium-ion battery.

[0062] The present invention adopts the wavelet fuzzy threshold denoising method to obtain smooth data after eliminating noise and spikes; at the same time, to address the problem of difficulty in extracting features due to the uncertainty and high nonlinearity of battery degradation data, a regularized stacked long short-term memory network model (deep prediction model) is introduced; the processed data is input into the regularized stacked long short-term memory network model for model training and prediction, which effectively avoids overfitting while improving the prediction accuracy and reliability, and realizes the estimation of the health status of lithium-ion batteries under different working conditions.

[0063] Each step of the present invention is described in detail below:

[0064] Step S1: Preprocessing the original lithium-ion battery aging data

[0065] Data source: A123 Systems' APR18650M1A lithium-ion battery pack, rated capacity 1.1Ah, rated voltage 3.3V, and charge and discharge temperature set at 30°C. High-precision sensors collect battery capacity decay and cycle life data, including charge and discharge current, voltage, temperature, and capacity decay curves.

[0066] Battery data with fewer than 500 cycles (considered incomplete aging) was removed. Outliers caused by test circuit or sensor failures were identified and removed using the triple standard deviation method. Voltage and current data were z-score normalized. The battery life was determined to be complete when capacity decayed to 80% of the rated capacity. After data cleaning, interference between capacity decay data and battery cycle life was eliminated.

[0067] In this embodiment, 100 groups of valid data are retained after cleaning, of which 80 groups are used for training and 20 groups are used for testing.

[0068] Step S2: Use the wavelet fuzzy threshold denoising method to process the data, eliminate noise and spikes, and build a case library.

[0069] Figure 2 This is the flowchart of wavelet threshold denoising. Wavelet fuzzy threshold denoising method is used to smooth the data, eliminate noise and spikes, and build a denoised case library.

[0070] The specific implementation process of smoothing the pre-processed data in this embodiment is as follows:

[0071] S21. Use Laplace Gaussian wavelet to perform multi-scale decomposition on the battery attenuation signal to obtain an approximate part and a noise part.

[0072] The battery capacity attenuation curve that needs to be smoothed is defined as f(t):

[0073]

[0074] Where t represents the time t of the decay sequence, x i Indicates the discharge capacity at time i of the battery decay cycle.

[0075] The present invention uses the Laplace Gaussian wavelet (Mexican Hat) for wavelet transform. This wavelet function has the characteristics of continuity and smoothness. When processing the continuous attenuation signal of the battery, it can detect abnormal attenuation points and is suitable for soft threshold processing. Its mathematical expression is:

[0076]

[0077] Where ψ(t) represents the Laplace-Gaussian wavelet function, t is the independent variable, and e is a constant.

[0078] The original signal f(t) is transformed by wavelet to obtain the wavelet basis function W f (t), expressed as:

[0079]

[0080] Where A I represents the approximation coefficient, D represents the detail coefficient, I represents the number of signal decomposition layers, and c represents the coefficient.

[0081] The approximate part of the battery attenuation signal f(t) after wavelet decomposition is stored in cA I The noise part is stored in cD i In this embodiment, the signal decomposition layer is four. The approximation coefficient (low-frequency signal) reflects the overall attenuation trend; the detail coefficient reflects the high-frequency noise of the signal, reflecting the signal details and local changes. In battery signals, the detail coefficient can capture detailed information such as rapid fluctuations and sudden changes in parameters such as voltage and current.

[0082] S22. Dynamically adjust the threshold based on the adaptive fuzzy soft threshold method, use the amplitude, local standard deviation and local energy of the wavelet coefficient as fuzzy input through the adaptive fuzzy rule base, output the dynamic threshold level, and use the weighted average method to defuzzify, and obtain the attenuated signal after threshold processing.

[0083] The expression of the soft threshold function is expressed as:

[0084]

[0085] Where w is the wavelet coefficient and λ is the threshold.

[0086] In order to solve the problems of "unclear boundaries" and "uncertain judgment" in traditional soft threshold functions, the present invention introduces an adaptive fuzzy soft threshold method, which can dynamically adjust the threshold according to signal characteristics (local energy, local standard deviation, etc.).

[0087] First, the variables are divided into three levels using a triangular fuzzy function, and an adaptive fuzzy rule base is established. The adaptive fuzzy rule base includes the following rules: if the absolute value of the wavelet coefficient is small, the local standard deviation is low, and the local energy is weak, the threshold level is high; if the absolute value of the wavelet coefficient is medium and the local standard deviation is medium, the threshold level is medium; if the absolute value of the wavelet coefficient is large, the local standard deviation is high, or the local energy is strong, the threshold level is low.

[0088] In the specific implementation, triangular fuzzy functions are used to divide the variables into three levels, and the fuzzy variable membership functions are defined as shown in Table 1.

[0089] Table 1 Definition of fuzzy variable membership function

[0090]

[0091] Table 1 , is the absolute value (amplitude) of the wavelet coefficient, V=σ local is the local standard deviation of the window where the wavelet coefficient is located, Q=Q local is the local energy of the wavelet coefficient window.

[0092] The absolute value of the wavelet coefficient is divided into three levels: Small, Medium, and Large; the local standard deviation is divided into three levels: Low, Medium, and High; the local energy is divided into three levels: Weak, Medium, and Strong; and the fuzzy soft threshold is divided into three levels: Low, Medium, and High.

[0093] in, The expression is:

[0094]

[0095] Where n is the number of wavelet coefficients in the window, w i is the i-th wavelet coefficient in the window.

[0096] Figure 4In order to use the triangular fuzzy function to perform fuzzy division on the fuzzy threshold parameters, in this embodiment, the particle swarm optimization method is used to further optimize the parameters of the triangular fuzzy function. After the particle swarm optimization, the membership function is divided as follows:

[0097] A optimization parameters are {Small(0, 0, 0.4), Medium(0.3, 0.5, 0.65), Large(0.5, 0.9, 1.0);

[0098] The optimized parameters for V are {Low(0, 0, 0.3), Medium(0.3, 0.6, 0.7), High(0.5, 0.8, 1.0)};

[0099] The optimized parameters for Q are {Weak(0, 0, 0.5), Medium(0.3, 0.6, 0.7), Strong(0.5, 0.9, 1.0)};

[0100] The λ optimization parameters are {Low(0, 0.1, 0.3), Medium(0.3, 0.5, 0.6), High(0.5, 0.9, 1.0)}.

[0101] According to the rules of triangular fuzzy functions, different numerical ranges are mapped to corresponding fuzzy sets to achieve fuzzy representation of signal characteristics and threshold levels, providing a basis for subsequent fuzzy rule-based processing.

[0102] Afterwards, an adaptive fuzzy rule base is established. In this embodiment, a total of 5 core rules are set, as shown in Table 2.

[0103] Table 2 Adaptive fuzzy rule base

[0104]

[0105] Specifically, the application scenario of R1 is: the wavelet coefficients have low amplitude, low volatility and low energy, which are judged as strong noise, with a membership of 0.9, and strong threshold suppression is performed; the application scenario of R2 is: medium amplitude and medium volatility, which are judged as a mixture of signal and noise, with a membership of 0.5, balancing denoising and information retention to avoid transition smoothing; the application scenario of R3 is: high amplitude, high volatility or high energy, which are judged as the main components of the signal, with a membership of 0.1, and weak threshold retention to avoid loss of useful signals; the application scenario of R4 is: low amplitude and high volatility, which may contain noise, and a medium threshold is required to suppress abnormal spikes; the application scenario of R5 is: medium amplitude and low energy, which may be a transition state, and a medium threshold is required for robust processing.

[0106] In the battery signal processing scenario, different signal characteristics reflect the different states and noise conditions of the battery. Through the adaptive fuzzy rule base, the threshold can be dynamically adjusted according to the actual signal conditions to achieve more accurate signal processing and analysis, solve the "unclear boundary" problem of traditional fixed thresholds, and more effectively remove noise and extract useful signal features.

[0107] Then, the weighted average method is used for defuzzification and the dynamic threshold is calculated, which is expressed as:

[0108]

[0109] In the formula, λ is the threshold value output by weighted average method, u i is the activation degree of each rule, c i is the central value of the corresponding output membership function.

[0110] Finally, the fuzzy system output threshold λ(w) is applied to each wavelet coefficient w, thereby obtaining the wavelet coefficient processed by soft thresholding. The fuzzy system is an adaptive fuzzy soft thresholding method based on an adaptive fuzzy rule base.

[0111]

[0112] Where W λ is the wavelet coefficient after soft threshold processing; sign(x) is the sign function, which is defined as: when x>0, sign(x)=1, when x=0, sign(x)=0, when x<0, sign(x)=-1; w is the original coefficient; λ(w) is the threshold output by the fuzzy system corresponding to the coefficient w.

[0113] S23. Sort the attenuated signals after threshold processing according to their absolute values.

[0114] Let F(k) be the sum of W f The absolute values of the elements in (t) are arranged from small to large and the attenuated signal sequence is expressed as:

[0115]

[0116] Where k is the index of the sorted coefficient and L is the total number of sorted coefficients.

[0117] After thresholding, the coefficients corresponding to noise are suppressed, but residual noise or weak signals may still exist in the retained coefficients. Sorting by absolute value concentrates the primary signal components at the beginning of the sequence, while placing secondary signal components at the end. By selectively retaining the top N largest coefficients, sparsity can be further enhanced, improving the quality and efficiency of signal reconstruction.

[0118] S24, reconstruct the arranged attenuated signal sequence using inverse wavelet transform to obtain a noise-processed signal. The wavelet reconstructed signal is expressed as:

[0119]

[0120] Where, To reconstruct the wavelet basis function, F a (k) The decay signal sequence associated with scale a, where a and b are the scaling factor and the displacement factor along the time axis of the battery decay sequence, respectively.

[0121] After wavelet fuzzy threshold reconstruction, the useless part of the original battery decay sequence is weakened and the useful part is enhanced. By performing wavelet fuzzy threshold processing on the battery decay sequence signal, the noise signal and the normal signal are distinguished, thereby achieving the purpose of deleting or weakening the noise signal.

[0122] Figure 3 The voltage and current decay curve of the battery during one cycle with the number of cycles. The horizontal axis is the charge and discharge time node, the left vertical axis is the actual current change value, and the right vertical axis is the actual voltage change value. Figure 3 It can be seen that the characteristic curve after wavelet fuzzy threshold denoising is smooth and flat, which is convenient for extracting the corresponding attenuation features in the later stage. Figure 5 This is the curve after denoising using fuzzy threshold. It can be clearly seen from the figure that the battery capacity decay curve after noise removal is smooth and the mutation points are intact, which strongly proves the effectiveness and superiority of this method in data processing.

[0123] Step S3: Optimize the stacked long short-term memory network and build a prediction model.

[0124] A regularized and optimized stacked long short-term memory network is used as a battery prediction model. This model avoids the overfitting risk of traditional deep learning models and, through regularization optimization, achieves excellent robustness and accuracy.

[0125] The parameters of the regularized stacked long short-term memory network model are shown in Table 3. To enhance the extraction of hidden features, the short-term memory network model structure in the embodiment of the present invention includes: an input layer, a convolution layer, a pooling layer, an expansion layer, a validation layer, and an output layer. A time-distributed packaging layer technology is used in the validation layer, and the output layer outputs the battery health status prediction value through a fully connected layer.

[0126] Table 3 Parameters of the regularized stacked long short-term memory network model of the present invention

[0127]

[0128] The specific process of constructing a prediction model by regularizing and optimizing the stacked long short-term memory network is as follows:

[0129] S31. To solve the problem of difficult feature extraction in battery aging data, a stacking method is used to improve learning ability. In this network structure, the weight gradient and bias vector gradient at time step t are expressed as:

[0130]

[0131]

[0132] Where E represents the loss function, W represents the weight matrix in the neural network, and b represents the bias vector. is the error term of the jth unit at time step t, J is the number of units, t is the time step, and T is the total duration.

[0133] S32. Regularize and optimize the parameters in the prediction model to solve the problem of large number of parameters in the network training process. The stacked long short-term memory network causes the number of parameters to increase exponentially, which easily leads to overfitting.

[0134] In the traditional long short-term memory network training process, there is a phenomenon that the norm is not differentiable at zero, which will cause the training gradient to disappear and the network training to fail. In the traditional regularization process, only the norm function is constrained, which easily leads to the target optimization function being overly affected by the regularization coefficient. Therefore, the present invention innovatively optimizes the norm function while regularizing the target loss function to prevent overfitting, and the regularization coefficient is used as upper and lower bound constraints.

[0135] First, in order to avoid the norm being undifferentiable at zero during the depth prediction model training process, the parameter ε is introduced, and its restriction conditions are: ε>0, lim(ε)=0. At this time, the following norm function is introduced for parameter constraint:

[0136]

[0137] Then, the objective function is further regularized and optimized, and its objective function is:

[0138]

[0139] Where N is the number of samples, f(x n ; n) is the depth prediction model to be trained, θ is its parameter, f(n) is the true output of the nth sample, x n is the input of the nth sample; D is the dimension of parameter θ, d is the dth component of parameter θ, ε is a small positive number; p and q are both regularization coefficients, p+q=1, p>0, q>0.

[0140] In the process of selecting the regularization coefficient for the above objective function, if the value is too large or too small, it will lead to problems such as failure of the training network. It is necessary to increase the optimization range of the regularization coefficient at the same time. Therefore, the present invention adds the regularization coefficient constraint condition as follows:

[0141]

[0142] Afterwards, according to the method in step S32, a regularized stacked long short-term memory network model is obtained to predict the health status of the lithium-ion battery.

[0143] Step S4: Evaluate the performance of the prediction model, optimize the model hyperparameters and output the optimal prediction model to estimate the health status of the lithium-ion battery. Figure 6 .

[0144] S41. Input the case library data (training set) after wavelet fuzzy threshold denoising into the regularized stacked long short-term memory network model. The model calculates neuron parameters through back propagation to perform battery health status prediction analysis, and uses mean square error (MSE) to evaluate model performance. The calculation formula is:

[0145]

[0146] Where, and y i are the predicted and actual values of the i-th battery tuple, and n represents the number of battery tuples.

[0147] S42. According to the method of step S41, the regularized stacked long short-term memory network model is trained multiple times. The learning rate is dynamically adjusted through the learning rate warmup strategy. The Adam optimizer improves the model parameters. The prediction results are recorded and evaluated. The model hyperparameters are continuously adjusted according to the results until the optimal performance is achieved. The trained optimal prediction model is output.

[0148] S43. Take the trained optimal prediction model as the final model for battery health state estimation to perform lithium-ion battery health state estimation.

[0149] It's important to emphasize that for different types of battery data, a grid search method is required to adjust parameters to obtain the most appropriate network model parameters. Network model parameters are not universally applicable. The grid search method uses an exhaustive search to try every possible parameter combination among all candidate parameter choices. This method permutes and combines the possible values of each parameter to form a "grid." Each combination is then used for model training, and performance is evaluated using cross-validation. After the fitting function has tried all parameter combinations, it returns the classifier and automatically adjusts to the optimal parameter combination.

[0150] In this embodiment, five-fold cross validation is used to ensure the robustness of the model results. The iteration standard set in the embodiment of the present invention is: when MSE < 0.1, the iteration is stopped and the optimal prediction model is output.

[0151] Comparative experiment:

[0152] The battery health status prediction results of the two objects in the case library are as follows: Figure 7 As shown, Figure 7 (a) is the final prediction effect of the training set, Figure 7 (b) shows the prediction results on the test set. The results show that the method of the present invention achieves better lifespan prediction for batteries with relatively short lifespans (<800 cycles), with an overall MSE <30 cycles. For samples with longer lifespans, the prediction results remain at an MSE <70 cycles, with some improvement. The validation results of the present invention on the test set are comparable to those on the training set (MSE error <33 cycles), demonstrating the robustness of the method.

[0153] The method proposed in the present invention is compared with the convolutional active residual network (CNN-ATResNet) and the convolutional long short-term memory network (CNN-LSTM). It can be clearly seen from Table 4 that the mean value of the model evaluation index (MSE error) of the model using the present invention is much smaller than that of other models in the comparative experiment.

[0154] Table 4 Comparative experimental results of the regularized stacked long short-term memory network model of the present invention and other models

[0155]

[0156] Figure 8 The comparison results of the present invention with the CNN-ATResNet and CNN-LSTM prediction methods are shown in Figure 2. The horizontal axis represents the number of battery cycles, and the vertical axis represents the predicted indicator of battery health status. The prediction value of the present invention is represented by the orange line, and the prediction values of CNN-ATResNet and CNN-LSTM are represented by the yellow and purple lines respectively. Figure 8 It can be seen that the predicted value of the present invention is almost the same as the true value. Therefore, the regularized stacked long short-term memory network model of the present invention has better prediction accuracy and is significantly better than the other two methods.

[0157] The batteries used in the implementation examples of this invention are all automotive power batteries. Its charge-discharge strategy simulates different new energy power battery usage scenarios (fast and slow charging, power battery discharge scenarios, etc.). Verified by 100 sets of different charge and discharge scenarios, the invention demonstrates excellent prediction accuracy and robustness, adapting to the diversity and complexity of battery data under different operating conditions. This technology has broad market and application value in the battery field, especially in the field of new energy vehicle power batteries.

[0158] Example 2:

[0159] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the battery health estimation method based on wavelet fuzzy threshold denoising described in Example 1 is implemented.

[0160] Finally, it should be noted that the parts of the present invention that are not described in detail are all prior art. Those skilled in the art will understand that the above description is only a preferred embodiment of the invention and is not intended to limit the invention. Although the invention has been described in detail with reference to the above examples, those skilled in the art can still modify the technical solutions described in the above examples or replace some of the technical features therein with equivalents. Any modifications, equivalent replacements, etc. made within the spirit and principles of the invention should be included in the scope of protection of the invention.

Claims

1. A battery health estimation method based on wavelet fuzzy threshold denoising is characterized by: The following steps are involved: S1. Preprocessing the original lithium-ion battery aging data; S2. Smoothing the pre-processed data using wavelet fuzzy threshold denoising method to eliminate noise and spikes, and constructing a denoised case library; S3. Constructing a regularized stacked long short-term memory network model, including stacked LSTM layers and a regularization optimization module, wherein the regularization optimization module optimizes the objective function by introducing parameter constraints and upper and lower bounds of regularization coefficients. The regularization coefficients optimize the norm function of the model parameters while regularizing the loss function to suppress overfitting; S4. Inputting the denoised case library data into the regularized stacked long short-term memory network model for training, evaluating the model performance, adjusting the model hyperparameters until optimal performance is achieved, outputting the trained optimal prediction model, and using the optimal prediction model to estimate the health status of the lithium-ion battery; In step S2, the wavelet fuzzy threshold denoising method is used for smoothing, which specifically includes: S21, using Laplace Gaussian wavelet to perform multi-scale decomposition on the battery attenuation signal to obtain an approximate part and a noise part; S22. Dynamically adjust the threshold based on an adaptive fuzzy soft threshold method, use the absolute value of the wavelet coefficient, the local standard deviation, and the local energy as fuzzy inputs through an adaptive fuzzy rule base, output a dynamic threshold level, and use a weighted average method to perform defuzzification to obtain an attenuated signal after threshold processing; wherein the specific implementation process of the adaptive fuzzy soft threshold method is as follows: First, a triangular fuzzy function is used to divide the variables into three levels, and an adaptive fuzzy rule base is established. The adaptive fuzzy rule base includes the following rules: if the absolute value of the wavelet coefficient is small, the local standard deviation is low, and the local energy is weak, the threshold level is high; if the absolute value of the wavelet coefficient is medium and the local standard deviation is medium, the threshold level is medium; if the absolute value of the wavelet coefficient is large, the local standard deviation is high, or the local energy is strong, the threshold level is low; Then, the weighted average method is used for defuzzification, and its expression is: In the formula, λ is the threshold value output by weighted average method, u i is the activation degree of each rule, c i is the central value of the corresponding output membership function; Finally, the fuzzy system output threshold λ(w) is applied to each wavelet coefficient w to obtain the wavelet coefficient after soft threshold processing: Where W λ is the wavelet coefficient after soft threshold processing; sign(x) is the sign function, which is defined as: when x>0, sign(x)=1, when x=0, sign(x)=0, when x<0, sign(x)=-1; w is the original coefficient; λ(w) is the threshold output by the fuzzy system corresponding to the coefficient w S23, sorting the attenuated signals after threshold processing according to their absolute values; S24. Reconstruct the arranged attenuated signal sequence using inverse wavelet transform to obtain a noise-processed signal.

2. The battery health estimation method based on wavelet fuzzy threshold denoising according to claim 1 is characterized in that: Step S1 specifically includes: eliminating battery data with a cycle count of less than 500 and abnormal values caused by the test circuit or sensor; when the battery capacity decays to 80% of the rated capacity, it is determined that the battery life has ended.

3. The battery health estimation method based on wavelet fuzzy threshold denoising according to claim 1, characterized in that: In step S21, the expression of Laplace Gaussian wavelet is: Where ψ(t) represents the Laplace-Gaussian wavelet function, t represents the independent variable, and e is a constant.

4. The battery health estimation method based on wavelet fuzzy threshold denoising according to claim 1, characterized in that: The objective function of regularization optimization in step S3 is: Where N is the number of samples, f(x n ; n) is the depth prediction model to be trained, θ is its parameter, f(n) is the true output of the nth sample, x n is the input of the nth sample; D is the dimension of parameter θ, d is the dth component of parameter θ, ε is a small positive number; p and q are both regularization coefficients, p+q=1, p>0, q>0.

5. The battery health estimation method based on wavelet fuzzy threshold denoising according to claim 1, characterized in that: The parameters of the regularized stacked long short-term memory network model described in step S3 include: input data length is 50, the number of stacking layers is 5, the number of LSTM layer units is 30, the learning rate is 0.001, the optimizer is Adam, and the loss function is MSE.

6. The battery health estimation method based on wavelet fuzzy threshold denoising according to claim 1, characterized in that: In step S4, the hyperparameter adjustment adopts the grid search method and five-fold cross validation, and the iteration termination condition is that the MSE is less than 0.

1.

7. The battery health estimation method based on wavelet fuzzy threshold denoising according to claim 1, characterized in that: In step S4, the mean square error is used to evaluate the model performance, and its calculation formula is: Where, and y i are the predicted and actual values of the i-th battery tuple, and n represents the number of battery tuples.

8. An electronic device, characterized in that: The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the battery health estimation method based on wavelet fuzzy threshold denoising according to any one of claims 1 to 7 is implemented.

Images