Lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU

Abstract

The invention provides a lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU, and belongs to the technical field of lithium ion batteries. The technical problem that the traditional SOH prediction error is too large is solved. Comprising the following steps of 1) performing feature extraction on a lithium ion battery charging and discharging data set of NASA by adopting a plurality of complementary feature extraction strategies MSFF, and constructing a comprehensive and reliable feature set; 2) using a random forest algorithm to further screen out high-quality features with importance scores ranking the top ten in the feature set; step 3), constructing a BiLSTM-GRU model of the bidirectional long short-term memory and gating circulation unit hybrid network; step 4) constructing a spider wasp optimization SWO algorithm; and 5) optimizing hyper-parameters of the model by using an SWO algorithm to realize accurate estimation of the SOH. According to the method, high-quality features are screened out and input into the optimized model, and high-precision SOH estimation is achieved.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery technology, and in particular to a lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU. Background Technology

[0002] Lithium-ion batteries are widely used in the new energy vehicle industry due to their high energy density and renewable resource conversion efficiency. State of Health (SOH) is one of the key indicators in battery management systems, providing important insights into the battery's remaining lifespan and overall performance. Accurate estimation of SOH has always been one of the research hotspots in battery management.

[0003] Currently, there are three main methods for SOH estimation: direct measurement, model-based prediction, and data-driven prediction. Direct measurement determines SOH by experimentally measuring the battery's internal resistance and capacity. For example, the battery is fully discharged to its state of charge (SOC), and the integral of the current over time during discharge is recorded to obtain the battery's actual capacity. Direct measurement is relatively easy to implement and achieves acceptable accuracy, but it is highly dependent on the environment and measurement equipment, and can only be performed offline, making it difficult to implement online. Model-based prediction methods mainly include electrochemical models and equivalent circuit models. Electrochemical models have powerful electrochemical theories; they can reflect not only changes in potential and voltage but also describe the internal reaction processes of the battery. However, due to their complexity, they are generally not used for SOH estimation. Equivalent circuit models mainly simulate the battery's internal working state using nonlinear circuit elements and are combined with filtering algorithms. Compared to electrochemical models, their complexity is greatly reduced. However, equivalent circuit models rely heavily on the accuracy of the identified parameters and the reliability of the filtering algorithm, resulting in poor robustness.

[0004] The advantage of data-driven methods lies in their independence from complex mechanistic models; they rely solely on nonlinear relationships learned from historical battery data to estimate the state of harmonics (SOH). Compared to direct measurement and model-based prediction methods, data-driven methods do not require consideration of internal battery chemical reactions and possess higher fitting accuracy and generalization ability. They have been widely applied to battery SOH estimation, but accuracy issues still exist.

[0005] Solving the aforementioned technical problems is the challenge facing this invention. Summary of the Invention

[0006] The purpose of this invention is to provide a lithium battery SOH estimation method based on multi-strategy feature fusion and an improved BiLSTM-GRU. First, features are extracted from the original battery data using multiple complementary feature extraction strategies to construct a comprehensive and reliable feature set. Second, a random forest algorithm is used to analyze the importance of the fused features, further filtering high-quality features strongly correlated with SOH. Then, a SWO algorithm is constructed to optimize hyperparameters. Next, a BiLSTM-GRU model is built to enhance the processing capability of feature information and improve prediction accuracy. Finally, the SWO algorithm is used to optimize the hyperparameters of the BiLSTM-GRU model, achieving accurate SOH estimation.

[0007] This invention is achieved through the following measures: a lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU, comprising the following steps:

[0008] Step 1) Employ multiple complementary feature extraction strategies (MSFF) to extract features from the lithium-ion battery charge-discharge dataset, avoiding the limitations of a single method and constructing a comprehensive and reliable feature set;

[0009] Step 2) Use the Random Forest (RF) algorithm to further filter out the top ten high-quality features in the feature set based on their importance scores;

[0010] Step 3) Construct a bidirectional long short-term memory and gated recurrent unit hybrid network (BiLSTM-GRU) model;

[0011] Step 4) Construct the Spider-Wasp Optimization (SWO) algorithm;

[0012] Step 5) Use the SWO algorithm to optimize the hyperparameters of the model to achieve accurate estimation of SOH.

[0013] As a further optimization of the lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU of the present invention, step 1) specifically includes the following steps:

[0014] Definition of health status:

[0015] Battery health is a key indicator of how much a battery's current performance has degraded relative to its initial state. It reflects the degree of aging and remaining lifespan potential of the battery, expressed as a percentage:

[0016] (1)

[0017] In the formula: Indicates the current capacity. Indicates the rated capacity;

[0018] Multi-strategy feature fusion:

[0019] a. Singular Value Decomposition (SVD)

[0020] Singular value decomposition (SVD) compresses high-dimensional data matrices; the mathematical formula is as follows:

[0021] (1)

[0022] In the formula: It is A real matrix, It is An orthogonal matrix, whose column vectors Known as The left singular vector satisfies , It is A rectangular diagonal matrix, whose diagonal elements These are non-negative real numbers, and these diagonal elements are called... singular values, It is An orthogonal matrix, whose column vectors Called The right singular vector satisfies ;

[0023] b. Wavelet Decomposition (WD)

[0024] This invention employs multi-scale wavelet decomposition technology to extract features of voltage, current, and temperature signals during battery charging and discharging. The Daubechies4 wavelet basis function is selected for three-level decomposition, and its scaling function and wavelet function are defined as follows:

[0025] (3)

[0026] (4)

[0027] Where: scaling function These are the low-frequency basis functions of wavelet decomposition. These are the coefficients of the low-pass filter. This means scaling the function to half its original size. This represents translating the scaled function. Units, wavelet function These are the high-frequency basis functions of wavelet decomposition. These are the coefficients of the high-pass filter;

[0028] The decomposition process is implemented recursively through a filter bank, and the approximation coefficients of the j-th layer... and detail coefficient The formula is obtained by calculating the approximation coefficients of the previous layer (j-1), as follows:

[0029] Approximation coefficient:

[0030] (5)

[0031] Detail factor:

[0032] (6)

[0033] In the formula: n is the index of the coefficient of the previous level, and m is the index of the coefficient of the current level. For the first level decomposition, That is, the original signal ;

[0034] The signal reconstruction formula is:

[0035] (7)

[0036] In the formula: It is a fixed value of 3. This represents the original signal that will be decomposed.

[0037] The voltage, current, and temperature signals of a battery are non-stationary and multi-scale. Wavelet decomposition separates them into low-frequency approximations and high-frequency details. Energy features and entropy features are selected to match the characteristics of battery degradation. The definitions of energy features and entropy features are as follows:

[0038] (8)

[0039] (9)

[0040] (10)

[0041] In the formula: and These are approximate energy and detail energy, respectively. and These are the approximation coefficient length and the detail coefficient length, respectively, p m It is an energy probability distribution;

[0042] c. Mathematical and statistical characteristics of the data (Math)

[0043] The fundamental mathematical features provide an intuitive and clear statistical description, directly reflecting the key statistical characteristics of the signal. They offer direct interpretation of the battery's state of equilibrium (SOH) estimation. The formula is as follows:

[0044] (11)

[0045] (12)

[0046] (13)

[0047] In the formula: The maximum value among all data. The minimum value among all data. yes Length, Indicates the first The value of each sampling point.

[0048] As a further optimization of the lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU of the present invention, step 2) specifically includes the following steps:

[0049] a. Calculation of Out-of-Bag (OOB) Error

[0050] (14)

[0051] In the formula: For the first The OOB sample set of the trees For the first Tree samples The original predicted value, It is the first The true target value of each sample;

[0052] b. Disturbance error calculation

[0053] (15)

[0054] In the formula: Representing the Each feature category Represents the t-th tree in terms of features The predicted value of sample i after its value is shuffled;

[0055] c. Importance score calculation:

[0056] (16)

[0057] In the formula: It is the total number of trees;

[0058] As a further optimization of the lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU of the present invention, step 3) specifically includes the following steps:

[0059] a. Bidirectional Long Short-Term Memory (BiLSTM) Network

[0060] The mathematical formula for the LSTM model is:

[0061] (17)

[0062] (18)

[0063] (19)

[0064] (20)

[0065] (twenty one)

[0066] (twenty two)

[0067] In the formula: For the sigmoid function, Let represent the state of the input gate at time t. It is the hyperbolic tangent function. Input to the current structural cells, , , , All are bias terms. For input gate output, Candidate cell state, The current cell state, For output gate output, , , , Both are weight matrices. For the output of the upper cell structure;

[0068] Combining forward and backward computations, the network structure of BiLSTM is obtained, as shown in the formula:

[0069] (twenty three)

[0070] In the formula: and These are the outputs of the BiLSTM forward and backward propagation, respectively.

[0071] b. Gated Cyclic Unit (GRU)

[0072] GRU significantly reduces the model's complexity and the number of training parameters required by reducing the three gating mechanisms of LSTM to two, thereby effectively improving the model's running efficiency. The mathematical formula is as follows:

[0073] (twenty four)

[0074] (25)

[0075] (26)

[0076] (27)

[0077] In the formula: To reset the value of the gate at time t, To update the value of the gate at time t, Let be the input value at time t. The output value at time t-1 The output value at time t. and To reset the gate weight matrix, and To update the gate weight matrix, It is the sigmoid activation function. To perform a dot product operation at the corresponding positions in the matrix;

[0078] A hybrid model combining BiLSTM and GRU was adopted. By combining bidirectional processing, the long-range memory advantage of LSTM, and the simplicity and efficiency of GRU, it performs exceptionally well in capturing complex contextual information and long-term dependencies in sequence data, significantly improving the accuracy of battery SOH estimation.

[0079] As a further optimization of the lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU of the present invention, step 4) specifically includes the following steps:

[0080] The Spider-Wasp Algorithm is a swarm intelligence optimization method inspired by nature. Its core mechanism draws on the hunting, nest building, and mating strategies exhibited by female spider wasps to ensure the continuation of the population, in order to search for and approximate the optimal solution in the solution space.

[0081] The steps of the spider-wasp optimization algorithm include:

[0082] a. Initial population

[0083] During algorithm initialization, N solution vectors are randomly generated in the D-dimensional domain space to form the initial population. The formula is as follows:

[0084] (28)

[0085] In the formula: represents the position of the z-th individual in the t-th iteration, r is a random value between 0 and 1, and UB and LB represent the lower and upper bounds of the parameter, respectively;

[0086] b. Hunting behavior

[0087] During a hunt, a female spider wasp searches for suitable prey within its spatial domain; this phase is called the search phase. After finding suitable prey, it pursues it; this phase is called the pursuit phase.

[0088] Search phase:

[0089] During this phase, the spider wasp introduces randomness into its search behavior to enhance its global search capabilities; its position update formula is as follows:

[0090] (29)

[0091] (30)

[0092] In the formula: This represents the position of the z-th individual in iteration t+1. and This represents two individuals randomly selected from the population in the t-th iteration. To adjust the step size coefficient, It is a random number between 0 and 1, where rn is a random number generated using a normal distribution;

[0093] When female spider wasps lose track of their prey, they use smaller step sizes for local searches, and the position update formula is:

[0094] (31)

[0095] (32)

[0096] (33)

[0097] In the formula: This represents the female spider wasp randomly selected in the t-th iteration. The coefficient of the step size. It is a random number between 0 and 1, where Q is the control. The coefficient, l, is a random number between -2 and 1;

[0098] In summary, the randomization of the next generation of female spider wasps' positions is implemented as follows:

[0099] (34)

[0100] In the formula: and A random number between 0 and 1;

[0101] Pursuit Phase:

[0102] This phase simulates the behavior of a female spider wasp after finding prey, luring it into her nest, and dragging it there. The position update formula for the spider wasp during the luring process is as follows:

[0103] (35)

[0104] (36)

[0105] In the formula: The female spider wasp is randomly selected in the t-th iteration, and R is a control factor for the change in the female spider wasp's speed. and A random number between 0 and 1, where t is the current iteration number. This represents the maximum number of iterations.

[0106] During the trapping process, the spider will attempt to escape. At this time, the distance between the female wasp and the spider gradually increases. This behavior can be simulated using the following formula:

[0107] (37)

[0108] In the formula: It is a vector generated by the normal distribution [k, -k], where k is... ;

[0109] In summary, the female spider wasp's position was updated as follows during the chase phase:

[0110] (38)

[0111] c. Nest-building behavior

[0112] Female spider wasps exhibit two distinct nest-building behaviors. The first involves dragging prey to the most suitable location and then building the nest there. The location update simulation for this method is as follows:

[0113] (39)

[0114] In the formula: This is the currently obtained local optimal position;

[0115] The second nesting method involves randomly selecting a spider's location within the population to build the nest, and the location update method is as follows:

[0116] (40)

[0117] (41)

[0118] In the formula: The numbers are generated by the Lévy flight distribution. , , Let t be three solutions randomly selected in t iterations. , , K is a random number between 0 and 1, and K is used to represent a binary vector that determines whether to apply a step size to avoid nesting at the same position.

[0119] d. Mating behavior

[0120] This phase simulates the behavior of female and male spider wasps uniformly exchanging operators to incubate eggs and produce offspring. The formula for spider wasp egg production is as follows:

[0121] (42)

[0122] In the formula: Indicates in and The probability of applying a uniform crossover operator between them is called the crossover rate (CR). and Let represent the female and male spider wasps in iteration t, respectively. The formula for generating male spider wasps is:

[0123] (43)

[0124] (44)

[0125] (45)

[0126] In the formula: and These are two numbers randomly generated according to a normal distribution. , and These are three different solutions selected from the population. This indicates the information that determines other solutions in the population. and Features in the new solution;

[0127] The transition between hunting and mating behaviors is based on a predefined factor called the tradeoff rate. ),when When the group is at rest, they engage in hunting and nest building; otherwise, they engage in mating.

[0128] e. Dynamic population strategy

[0129] During the iteration process, the number of female bees in the population changes dynamically. A dynamic population strategy can be used to adaptively adjust the population size at different iteration stages of the algorithm to prevent it from getting trapped in local optima. In the early stages of optimization, the population size is large, and the algorithm has a strong global search capability, capable of discovering potential solutions in the problem space. In the later stages of optimization, the population size gradually decreases, and the algorithm focuses more on the precise search for the currently found optimal solution, making it easier to converge to the global optimum. The length of the new population is updated as follows:

[0130] (46)

[0131] In the formula: N represents the minimum population size used at different stages of the optimization process to avoid getting trapped in local optima; k = 1 - t / t max The dynamic adjustment rate of the population size is controlled by rapidly decreasing it in the early stages of iteration to speed up the global search, while in the later stages of iteration, the rate of change of the population size is slowed down, enabling the algorithm to converge to the optimal solution more stably.

[0132] As a further optimization of the lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU of the present invention, step 5) specifically includes the following steps:

[0133] Step 5-1) Optimization of hyperparameters

[0134] The SWO algorithm was used to optimize the hyperparameters of the BiLSTM-GRU hybrid model, including the number of neurons in BiLSTM and GRU, the initial learning rate, and the learning rate reduction factor.

[0135] Step 5-2) Estimation of SOH

[0136] After fusing features from the raw battery data using multiple strategies, a random forest algorithm is used to select high-quality features related to battery state of health (SOH). These features are then input into the optimized model to obtain accurate and stable SOH predictions.

[0137] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0138] 1. This invention proposes a method that integrates multiple feature extraction strategies. Singular value decomposition excels at revealing the structural relationships and main change patterns within data, and can extract hidden latent variables. Wavelet decomposition excels at analyzing signals simultaneously in the time and frequency domains, providing multi-scale analysis. Its basic mathematical features are easy to calculate and understand, and have direct interpretability. Integrating the features extracted by these three methods can cover more comprehensive and richer information in the data, avoiding important features that might be missed by a single method, and providing a more reliable basis for feature selection.

[0139] 2. This invention uses the random forest algorithm to select features and performs importance analysis on the fused feature set. This allows for the selection of high-quality features that are strongly correlated with SOH, which can be used as input for estimating battery SOH, thus avoiding model performance degradation due to feature redundancy.

[0140] 3. This invention constructs a BiLSTM-GRU model, which integrates the BiLSTM model and the GRU model. By combining bidirectional processing, the long-range memory advantage of LSTM and the simplicity and efficiency of GRU, it can perform very well in capturing complex contextual information and long-term dependencies in sequence data, and significantly improve the accuracy of battery SOH estimation.

[0141] 4. This invention proposes to use the SWO algorithm to optimize the hyperparameters of the BiLSTM-GRU model. Compared with traditional swarm intelligence optimization algorithms, the SWO algorithm has the characteristics of fast convergence speed and high accuracy of the identified parameters, making the SOH estimation of the battery more accurate. Attached Figure Description

[0142] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0143] Figure 1 This is the overall flowchart of the present invention.

[0144] Figure 2 This is a flowchart of the random forest algorithm provided by the present invention.

[0145] Figure 3 This is a structural diagram of the LSTM model provided by the present invention.

[0146] Figure 4 This is a structural diagram of the BiLSTM model provided by the present invention.

[0147] Figure 5 This is a structural diagram of the GRU model provided by the present invention.

[0148] Figure 6 This is a comparison curve of SOH estimation between the feature fusion method provided by this invention and a single method.

[0149] Figure 7 This is the SOH estimation error curve of the feature fusion method provided by this invention and a single method.

[0150] Figure 8 This is a comparison curve of SOH estimation between the model optimized by the SWO algorithm and the unoptimized model provided by this invention.

[0151] Figure 9This is a graph showing the SOH estimation error curves of the model optimized by the SWO algorithm and the model without optimization, provided by this invention.

[0152] Figure 10 This is a comparison curve of SOH estimation between the SWO algorithm provided by this invention and traditional optimization algorithms.

[0153] Figure 11 This is a graph showing the SOH estimation error of the SWO algorithm provided by this invention compared to traditional optimization algorithms. Detailed Implementation

[0154] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0155] Example 1

[0156] This embodiment studies an 18650 lithium-ion battery with a rated capacity of 2Ah, conducted at 24°C. The charging process is divided into CC charging mode and CV charging mode. First, the battery is charged at 1.5A in CC charging mode until the battery voltage reaches 4.2V. Then, the battery is charged in CV charging mode until the battery current drops to 20mA. After a 2-hour rest period, the battery is discharged at a constant current of 2A until the battery voltage drops to 2.7V, 2.5V, and 2.2V.

[0157] See Figures 1 to 11 This invention provides a lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU, comprising the following steps:

[0158] Step 1) Employ multiple complementary feature extraction strategies (MSFF) to extract features from the lithium-ion battery charge-discharge dataset, avoiding the limitations of a single method and constructing a comprehensive and reliable feature set;

[0159] Step 2) Use the Random Forest (RF) algorithm to further filter out the top ten high-quality features in the feature set based on their importance scores;

[0160] Step 3) Construct a bidirectional long short-term memory and gated recurrent unit hybrid network (BiLSTM-GRU) model;

[0161] Step 4) Construct the Spider-Wasp Optimization (SWO) algorithm;

[0162] Step 5) Use the SWO algorithm to optimize the hyperparameters of the model to achieve accurate estimation of SOH.

[0163] Step 1) specifically includes the following steps:

[0164] Definition of health status:

[0165] Battery health is a key indicator of how much a battery's current performance has degraded relative to its initial state. It reflects the degree of aging and remaining lifespan potential of the battery, expressed as a percentage:

[0166] (1)

[0167] In the formula: Indicates the current capacity. Indicates the rated capacity;

[0168] Multi-strategy feature fusion:

[0169] a. Singular Value Decomposition (SVD)

[0170] Singular value decomposition (SVD) compresses high-dimensional data matrices; the mathematical formula is as follows:

[0171] (2)

[0172] In the formula: It is A real matrix, It is An orthogonal matrix, whose column vectors Known as The left singular vector satisfies , It is A rectangular diagonal matrix, whose diagonal elements These are non-negative real numbers, and these diagonal elements are called... singular values, It is An orthogonal matrix, whose column vectors Called The right singular vector satisfies ;

[0173] b. Wavelet Decomposition (WD)

[0174] Multi-scale wavelet decomposition is employed to extract features of voltage, current, and temperature signals during battery charging and discharging. The Daubechies4 wavelet basis function is selected for three-level decomposition, and its scaling and wavelet functions are defined as follows:

[0175] (3)

[0176] (4)

[0177] Where: scaling function These are the low-frequency basis functions of wavelet decomposition. These are the coefficients of the low-pass filter. This means scaling the function to half its original size. This represents translating the scaled function. Units, wavelet function These are the high-frequency basis functions of wavelet decomposition. These are the coefficients of the high-pass filter;

[0178] The decomposition process is implemented recursively through a filter bank, and the approximation coefficients of the j-th layer... and detail coefficient The formula is obtained by calculating the approximation coefficients of the previous layer (j-1), as follows:

[0179] Approximation coefficient:

[0180] (5)

[0181] Detail factor:

[0182] (6)

[0183] In the formula: n is the index of the coefficient of the previous level, and m is the index of the coefficient of the current level. For the first level decomposition, That is, the original signal ;

[0184] The signal reconstruction formula is:

[0185] (7)

[0186] In the formula: It is a fixed value of 3. This represents the original signal that will be decomposed.

[0187] The voltage, current, and temperature signals of a battery are non-stationary and multi-scale. Wavelet decomposition breaks them down into low-frequency approximations and high-frequency details. This invention selects energy features and entropy features to match the characteristics of battery degradation. The definitions of energy features and entropy features are as follows:

[0188] (8)

[0189] (9)

[0190] (10)

[0191] In the formula: and These are approximate energy and detail energy, respectively. and These are the approximation coefficient length and the detail coefficient length, respectively, p m It is an energy probability distribution;

[0192] c. Mathematical and statistical characteristics of the data (Math)

[0193] The fundamental mathematical features provide an intuitive and clear statistical description, directly reflecting the key statistical characteristics of the signal. They offer direct interpretation of the battery's state of equilibrium (SOH) estimation. The formula is as follows:

[0194] (11)

[0195] (12)

[0196] (13)

[0197] In the formula: The maximum value among all data. The minimum value among all data. yes Length, Indicates the first The value of each sampling point.

[0198] Step 2) specifically includes the following steps:

[0199] a. Calculation of Out-of-Bag (OOB) Error

[0200] (14)

[0201] In the formula: For the first The OOB sample set of the trees For the first Tree samples The original predicted value, It is the first The true target value of each sample;

[0202] b. Disturbance error calculation

[0203] (15)

[0204] In the formula: Representing the Each feature category Represents the t-th tree in terms of features The predicted value of sample i after its value is shuffled;

[0205] c. Importance score calculation:

[0206] (16)

[0207] In the formula: It represents the total number of trees.

[0208] Step 3) specifically includes the following steps:

[0209] a. Bidirectional Long Short-Term Memory (BiLSTM) Network

[0210] The mathematical formula for the LSTM model is:

[0211] (17)

[0212] (18)

[0213] (19)

[0214] (20)

[0215] (twenty one)

[0216] (twenty two)

[0217] In the formula: For the sigmoid function, Let represent the state of the input gate at time t. It is the hyperbolic tangent function. Input to the current structural cells, , , , All are bias terms. For input gate output, Candidate cell state, The current cell state, For output gate output, , , , Both are weight matrices. For the output of the upper cell structure;

[0218] Combining forward and backward computations, the network structure of BiLSTM is obtained, as shown in the formula:

[0219] (twenty three)

[0220] In the formula: and These are the outputs of the BiLSTM forward and backward propagation, respectively.

[0221] b. Gated Cyclic Unit (GRU)

[0222] GRU significantly reduces the model's complexity and the number of training parameters required by reducing the three gating mechanisms of LSTM to two, thereby effectively improving the model's running efficiency. The mathematical formula is as follows:

[0223] (twenty four)

[0224] (25)

[0225] (26)

[0226] (27)

[0227] In the formula: To reset the value of the gate at time t, To update the value of the gate at time t, Let be the input value at time t. The output value at time t-1 The output value at time t. and To reset the gate weight matrix, and To update the gate weight matrix, It is the sigmoid activation function. To perform a dot product operation at the corresponding positions in the matrix;

[0228] A hybrid model combining BiLSTM and GRU was adopted. By combining bidirectional processing, the long-range memory advantage of LSTM, and the simplicity and efficiency of GRU, it performs exceptionally well in capturing complex contextual information and long-term dependencies in sequence data, significantly improving the accuracy of battery SOH estimation.

[0229] Step 4) specifically includes the following steps:

[0230] The Spider-Wasp Algorithm is a swarm intelligence optimization method inspired by nature. Its core mechanism draws on the hunting, nest building, and mating strategies exhibited by female spider wasps to ensure the continuation of the population, in order to search for and approximate the optimal solution in the solution space.

[0231] The steps of the spider-wasp optimization algorithm include:

[0232] a. Initial population

[0233] During algorithm initialization, N solution vectors are randomly generated in the D-dimensional domain space to form the initial population. The formula is as follows:

[0234] (28)

[0235] In the formula: represents the position of the z-th individual in the t-th iteration, r is a random value between 0 and 1, and UB and LB represent the lower and upper bounds of the parameter, respectively;

[0236] b. Hunting behavior

[0237] During a hunt, a female spider wasp searches for suitable prey within its spatial domain; this phase is called the search phase. After finding suitable prey, it pursues it; this phase is called the pursuit phase.

[0238] Search phase:

[0239] During this phase, the spider wasp introduces randomness into its search behavior to enhance its global search capabilities; its position update formula is as follows:

[0240] (29)

[0241] (30)

[0242] In the formula: This represents the position of the z-th individual in iteration t+1. and This represents two individuals randomly selected from the population in the t-th iteration. To adjust the step size coefficient, It is a random number between 0 and 1, where rn is a random number generated using a normal distribution;

[0243] When female spider wasps lose track of their prey, they use smaller step sizes for local searches, and the position update formula is:

[0244] (31)

[0245] (32)

[0246] (33)

[0247] In the formula: This represents the female spider wasp randomly selected in the t-th iteration. The coefficient of the step size. It is a random number between 0 and 1, where Q is the control. The coefficient, l, is a random number between -2 and 1;

[0248] In summary, the randomization of the next generation of female spider wasps' positions is implemented as follows:

[0249] (34)

[0250] In the formula: and A random number between 0 and 1;

[0251] Pursuit Phase:

[0252] This phase simulates the behavior of a female spider wasp after finding prey, luring it into her nest, and dragging it there. The position update formula for the spider wasp during the luring process is as follows:

[0253] (35)

[0254] (36)

[0255] In the formula: The female spider wasp is randomly selected in the t-th iteration, and R is a control factor for the change in the female spider wasp's speed. and A random number between 0 and 1, where t is the current iteration number. This represents the maximum number of iterations.

[0256] During the trapping process, the spider will attempt to escape. At this time, the distance between the female wasp and the spider gradually increases. This behavior can be simulated using the following formula:

[0257] (37)

[0258] In the formula: It is a vector generated by the normal distribution [k, -k], where k is... ;

[0259] In summary, the female spider wasp's position was updated as follows during the chase phase:

[0260] (38)

[0261] c. Nest-building behavior

[0262] Female spider wasps exhibit two distinct nest-building behaviors. The first involves dragging prey to the most suitable location and then building the nest there. The location update simulation for this method is as follows:

[0263] (39)

[0264] In the formula: This is the currently obtained local optimal position;

[0265] The second nesting method involves randomly selecting a spider's location within the population to build the nest, and the location update method is as follows:

[0266] (40)

[0267] (41)

[0268] In the formula: The numbers are generated by the Lévy flight distribution. , , Let t be three solutions randomly selected in t iterations. , , K is a random number between 0 and 1, and K is used to represent a binary vector that determines whether to apply a step size to avoid nesting at the same position.

[0269] d. Mating behavior

[0270] This phase simulates the behavior of female and male spider wasps uniformly exchanging operators to incubate eggs and produce offspring. The formula for spider wasp egg production is as follows:

[0271] (42)

[0272] In the formula: Indicates in and The probability of applying a uniform crossover operator between them is called the crossover rate (CR). and Let represent the female and male spider wasps in iteration t, respectively. The formula for generating male spider wasps is:

[0273] (43)

[0274] (44)

[0275] (45)

[0276] In the formula: and These are two numbers randomly generated according to a normal distribution. , and These are three different solutions selected from the population. This indicates the information that determines other solutions in the population. and Features in the new solution;

[0277] The transition between hunting and mating behaviors is based on a predefined factor called the tradeoff rate. ),when When the group is at rest, they engage in hunting and nest building; otherwise, they engage in mating.

[0278] e. Dynamic population strategy

[0279] During the iteration process, the number of female bees in the population changes dynamically. A dynamic population strategy can be used to adaptively adjust the population size at different iteration stages of the algorithm to prevent it from getting trapped in local optima. In the early stages of optimization, the population size is large, and the algorithm has a strong global search capability, capable of discovering potential solutions in the problem space. In the later stages of optimization, the population size gradually decreases, and the algorithm focuses more on the precise search for the currently found optimal solution, making it easier to converge to the global optimum. The length of the new population is updated as follows:

[0280] (46)

[0281] In the formula: N represents the minimum population size used at different stages of the optimization process to avoid getting trapped in local optima; k = 1 - t / t max The dynamic adjustment rate of the population size is controlled by rapidly decreasing it in the early stages of iteration to speed up the global search, while in the later stages of iteration, the rate of change of the population size is slowed down, enabling the algorithm to converge to the optimal solution more stably.

[0282] Step 5) specifically includes the following steps:

[0283] Step 5-1) Optimization of hyperparameters

[0284] The SWO algorithm was used to optimize the hyperparameters of the BiLSTM-GRU hybrid model, including the number of neurons in BiLSTM and GRU, the initial learning rate, and the learning rate reduction factor.

[0285] Step 5-2) Estimation of SOH

[0286] After fusing features from the raw battery data using multiple strategies, a random forest algorithm is used to select high-quality features related to battery state of health (SOH). These features are then input into the BiLSTM-GRU model to obtain an estimated SOH curve. The SOH estimates from individual methods are then compared. Figure 6 As shown, the error is as follows Figure 7 As shown, the feature fusion method of this invention has high accuracy and a stable error variation curve. After optimizing the hyperparameters of the BiSLTM-GRU model using the SWO algorithm, it is trained again, the SOH is estimated, and compared with the unoptimized model, as shown... Figure 8 As shown, the error variation curve is as follows: Figure 9As shown, the SWO algorithm-optimized model exhibits improved SOH estimation accuracy compared to the unoptimized model, demonstrating the necessity of using the SWO algorithm for optimization. Furthermore, the PSO and GA algorithms are employed for parameter optimization and compared with the SWO algorithm constructed in this invention. Figure 10 As shown, the error variation curve is as follows: Figure 11 As shown, the SWO algorithm has significant advantages over traditional swarm intelligence algorithms.

[0287] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU, characterized in that, Includes the following steps: Step 1) Use multiple complementary feature extraction strategies (MSFF) to extract features from the lithium-ion battery charge and discharge dataset to construct a comprehensive and reliable feature set; Step 2) Use the Random Forest (RF) algorithm to select the top ten features in the feature set based on their scores; Step 3) Construct a BiLSTM-GRU model, a hybrid network of bidirectional long short-term memory and gated recurrent units; Step 4) Construct the Spider-Wasp Optimized SWO Algorithm; Step 5) Use the SWO algorithm to optimize the hyperparameters of the model to achieve accurate estimation of SOH.

2. The lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU according to claim 1, wherein, Step 1) includes the following steps: 2.1 Definition of Health Status Battery health is a key indicator that measures the degree of performance degradation of a battery relative to its initial state. It reflects the degree of aging and remaining lifespan potential of the battery, expressed as a percentage: (1)； wherein: represents the current capacity, represents the rated capacity; 2.2 Multi-strategy feature fusion.

3. The lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU according to claim 2, characterized in that, Section 2.2 includes the following steps: 2.2a. Singular Value Decomposition (SVD) Singular value decomposition (SVD) compresses high-dimensional data matrices; the mathematical formula is as follows: (2)； In the formula: It is A real matrix, It is An orthogonal matrix, whose column vectors Known as The left singular vector satisfies , It is A rectangular diagonal matrix, whose diagonal elements These are non-negative real numbers, and these diagonal elements are called... singular values, It is An orthogonal matrix, whose column vectors Called The right singular vector satisfies ; 2.2b. Wavelet Decomposition (WD) Multi-scale wavelet decomposition technology is used to extract the features of voltage, current, and temperature signals during battery charging and discharging. The Daubechies4 wavelet basis function is selected for three-level decomposition, and its scaling function and wavelet function are defined as follows: (3)； (4)； Where: scaling function These are the low-frequency basis functions of wavelet decomposition. These are the coefficients of the low-pass filter. This means scaling the function to half its original size. This represents translating the scaled function. Units, wavelet function These are the high-frequency basis functions of wavelet decomposition. These are the coefficients of the high-pass filter; The decomposition process is implemented recursively by a filter bank, the approximation coefficients of the j-th level and the detail coefficients are computed from the approximation coefficients of the previous level (j-1) by the following formula: Approximation coefficient: (5)； Detail factor: (6)； where n is the index of the previous layer coefficient and m is the index of the current layer coefficient, for the first layer decomposition, is the original signal ; The signal reconstruction formula is: (7)； In the formula: is a fixed value of 3, denotes the original signal to be decomposed; The voltage, current, and temperature signals of a battery are non-stationary and multi-scale. Wavelet decomposition separates them into low-frequency approximations and high-frequency details. Energy features and entropy features are selected to match the characteristics of battery degradation. The definitions of energy features and entropy features are as follows: (8)； (9)； (10)； where: and are the approximation and detail energies, respectively, and are the approximation and detail coefficient lengths, respectively, p m is the energy probability distribution; 2.2c. Mathematical and Statistical Characteristics of Data (Math) The fundamental mathematical characteristics reflect the key statistical properties of the signal and have a direct interpretative effect on the SOH estimation of the battery. The formula is as follows: (11)； (12)； (13)； where: is the maximum value among all data, is the minimum value among all data, is the length of denotes the value of the th sample point.​ 4. The lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU according to claim 1, characterized in that, Step 2) includes the following steps: 2a. Calculation of Out-of-Bag (OOB) Error for Reference Bag Data (14)； In the formula: For the first The OOB sample set of the trees For the first Tree samples The original predicted value, It is the first The true target value of each sample; 2b. Calculation of disturbance error (15)； In the formula: Representing the Each feature category Represents the t-th tree in terms of features The predicted value of sample i after its value is shuffled; 2c. Importance score calculation: (16)； In the formulae: is the total number of trees.

5. The lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU according to claim 1, characterized in that, Step 3) includes the following steps: 3a. Bidirectional Long Short-Term Memory Network (BiLSTM) The mathematical formula for the LSTM model is: (17)； (18)； (19)； (20)； (21)； (22)； In the formula: For the sigmoid function, Let represent the state of the input gate at time t. It is the hyperbolic tangent function. Input to the current structural cells, , , , All are bias terms. For input gate output, Candidate cell state, The current cell state, For output gate output, , , , Both are weight matrices. For the output of the upper cell structure; Combining forward and backward computations, the network structure of BiLSTM is obtained, as shown in the formula: (23)； wherein: and are the BiLSTM forward and backward propagation outputs, respectively; 3b. Gated Loop Unit (GRU) The gated recurrent unit (GRU) reduces the three gating mechanisms of the original LSTM to two, as shown in the formula: (24)； (25)； (26)； (27)； In the formula: To reset the value of the gate at time t, To update the value of the gate at time t, Let be the input value at time t. The output value at time t-1 The output value at time t. and To reset the gate weight matrix, and To update the gate weight matrix, It is the sigmoid activation function. This is to perform a dot product operation at the corresponding positions in the matrix.

6. The lithium battery SOH estimation method based on multi-policy feature fusion and improved BiLSTM-GRU according to claim 1, characterized in that, Step 4) includes the following steps: The Spider Wasp Algorithm draws on the hunting, nest-building, and mating strategies exhibited by female spider wasps to ensure the continuation of their population, and is used to search for and approximate the optimal solution in the solution space. The steps of the spider-wasp optimization algorithm include: 4a. Initial population During algorithm initialization, N solution vectors are randomly generated in the D-dimensional domain space to form the initial population, expressed by the formula: (28)； In the formula: represents the position of the z-th individual in the t-th iteration, r is a random value between 0 and 1, and UB and LB represent the lower and upper bounds of the parameter, respectively; 4b. Hunting behavior During the hunting process, female spider wasps search for suitable prey in their spatial domain. This stage is called the search stage. After finding suitable prey, they will chase after it. This stage is called the chase stage. 4c. Nest-building behavior Female spider wasps exhibit two distinct nest-building behaviors. The first involves dragging prey to a suitable location and then building a nest there. The location update simulation for this method is as follows: (29)； In the formula: This is the currently obtained local optimal position; The second nesting method involves randomly selecting a spider's location within the population to build the nest, and the location update method is as follows: (30)； (31)； In the formula: The numbers are generated by the Lévy flight distribution. , , Let t be three solutions randomly selected in t iterations. , , K is a random number between 0 and 1, and K is used to represent a binary vector that determines whether to apply a step size to avoid nesting at the same position. 4d. Mating behavior This stage simulates the behavior of female and male spider wasps uniformly exchanging operators to incubate eggs and produce offspring. The formula for spider wasp egg production is as follows: (32)； In the formula: Indicates in and When a uniform crossover operator is applied between two points, the probability is called the crossover rate CR. and Let represent the female and male spider wasps in iteration t, respectively. The formula for generating the male spider wasp is: (33)； (34)； (35)； In the formula: and These are two numbers randomly generated according to a normal distribution. , and These are three different solutions selected from the population. This indicates the information that determines other solutions in the population. and Features in the new solution; The transition between hunting and mating behaviors is based on a predefined factor called the tradeoff. ,when When the group is at rest, they engage in hunting and nest building; otherwise, they engage in mating. 4e. Dynamic Population Strategy During the iteration process, the number of female bees in the population changes dynamically. The dynamic population strategy is used to adaptively adjust the population size at different iteration stages of the algorithm to prevent the algorithm from getting trapped in local optima. The length of the new population is updated as follows: (36)； where N represents the minimum population size used at different stages of the optimization process to avoid getting trapped in local optima; k = 1 - t / t max controls the speed of dynamic adjustment of the population size.

7. The lithium battery SOH estimation method based on multi-strategy feature fusion and improved BiLSTM-GRU according to claim 1, characterized in that, Step 5) includes the following steps: Step 5-1) Optimization of hyperparameters The hyperparameters of the BiLSTM-GRU hybrid model were optimized using the SWO algorithm, including the number of neurons in BiLSTM and GRU, the initial learning rate, and the learning rate reduction factor. Step 5-2) Estimation of SOH After fusing features from the raw battery data using multiple strategies, a random forest algorithm is used to select high-quality features related to battery SOH. These features are then input into the optimized model to obtain accurate and stable SOH predictions. 8.The lithium battery SOH estimation method based on multi-policy feature fusion and improved BiLSTM-GRU according to claim 1, wherein, The hunting behavior described in 4b. includes the following stages: Search phase: During this phase, the spider wasp introduces randomness into its search behavior to enhance its global search capabilities; its position update formula is as follows: (37)； (38)； In the formula: This represents the position of the z-th individual in iteration t+1. and This represents two individuals randomly selected from the population in the t-th iteration. To adjust the step size coefficient, It is a random number between 0 and 1, where rn is a random number generated using a normal distribution; When the female spider wasp loses track of its prey, it uses a step size for local search, and the position update formula is: (39)； (40)； (41)； In the formula: This represents the female spider wasp randomly selected in the t-th iteration. The coefficient of the step size. It is a random number between 0 and 1, where Q is the control. The coefficient, l, is a random number between -2 and 1; The randomization of the next generation of female spider wasps' positions is implemented as follows: (42)； In the formula: and A random number between 0 and 1; Pursuit Phase: The chase phase simulates the behavior of a female spider wasp after finding prey, luring it into her nest, and then dragging it there. The position update formula for the spider wasp during this luring process is as follows: (43)； (44)； In the formula: The female spider wasp is randomly selected in the t-th iteration, and R is a control factor for the change in the female spider wasp's speed. and A random number between 0 and 1, where t is the current iteration number. This represents the maximum number of iterations. During trapping, the spider will attempt to escape, increasing the distance between the female wasp and the spider. This behavior can be simulated using the following formula: (45)； wherein: is a vector of normal distribution [k,-k], k is ; During the chase phase, the female spider wasp's position is updated as follows: (46)。

Images