Battery health state prediction method based on two-stage neural Wiener process

Abstract

The invention discloses a battery health state prediction method based on a two-stage neural Wiener process, and the method comprises the following steps: 1, carrying out change point recognition: 11) learning the time sequence characteristics in a battery degradation process through a Bi-GRU model, and predicting the health state in the battery degradation process through the state splicing of forward and reverse propagation hidden layers; 12) introducing a learnable self-adaptive threshold value, and judging whether point change occurs or not; 2, two-stage Wiener process parameter estimation: 21) establishing a two-stage Wiener degradation model comprising a drift coefficient and a diffusion coefficient by taking a change point moment tau as a demarcation point; 22) estimating a drift coefficient and a diffusion coefficient by adopting an EM algorithm; 23) updating the drift coefficient; 3, deducing a life distribution function of the two-stage Wiener process; 4, predicting the state of health of the battery: 41) deducing a state transition probability density function of the neural Wiener process; and 42) taking the mathematical expectation of the state transition probability density function as the battery health state of the current cycle period.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery management and monitoring, and specifically relates to a method for predicting the state of health of a battery based on a two-stage neural Wiener process. Background Art

[0002] Lithium-ion batteries have been widely used in important fields of the national economy. However, with the gradual degradation of battery performance, related safety accidents occur from time to time, which has attracted great attention to battery management and monitoring. Therefore, accurately estimating the state of health (SOH) of a battery and its future degradation trajectory is of great significance for ensuring the safe operation of the battery. This not only helps to improve the operating stability of the battery but also effectively reduces potential safety risks. In this context, due to its excellent mathematical properties, the Wiener process has become an effective tool for quantifying the non-monotonic degradation process of battery performance. This process can capture and describe the complexity of battery performance changes over time in the form of probability.

[0003] Most of the existing studies use a single-stage degradation model with a single degradation rate to predict the SOH of lithium-ion batteries. This model can reasonably explain the actual degradation process of the battery in a specific simplified scenario, including methods based on the Wiener process (WP), methods based on the Gamma process, and methods based on the inverse Gaussian process. Among these methods, the WP model has become an important tool in battery stochastic degradation modeling due to its high efficiency and applicability.

[0004] However, the actual degradation process of lithium-ion batteries often exhibits significant non-linear characteristics and is affected by age and usage status. This means that the degradation trajectories of different batteries under similar conditions may have individual differences, and their change laws may not be consistent. The battery capacity initially decays relatively slowly, but the decay rate accelerates after the change point, indicating a change in the degradation mechanism of the battery before and after the change point. At this time, the single-stage degradation model cannot describe the complex two-stage degradation process due to its assumed single degradation rate. Therefore, it is necessary to study a degradation model that reflects the degradation characteristics of different stages to achieve the prediction of the SOH of batteries with different degradation rates. In the actual degradation process of the two-stage battery, the position of the change point varies for different battery individuals, and the distribution of the change point has great flexibility in time. At the same time, affected by individual differences, batteries in the same batch have different operating conditions and health states, and the degradation processes of different batteries are different from each other, which brings challenges to prediction and management. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method for predicting the state of health of a battery based on a two-stage neural Wiener process, which fully considers the individual differences in the change point positions, is not restricted by the initial distribution form, and can more effectively predict the battery SOH.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for predicting the state of health of a battery based on a two-stage neural Wiener process, comprising the following steps:

[0008] Step 1: Change point identification

[0009] 11) Learn the temporal characteristics during the battery degradation process through a Bi-GRU model, and use the concatenation of the hidden layer states of forward and backward propagation to predict the state of health during the battery degradation process;

[0010] 12) Introduce a learnable adaptive threshold. If the residual between the predicted value and the observed value of the battery state of health exceeds the set adaptive threshold, a change point occurs;

[0011] Step 2: Parameter estimation of the two-stage Wiener process

[0012] 21) Taking the change point moment τ as the demarcation point, establish a two-stage Wiener degradation model including a drift coefficient and a diffusion coefficient;

[0013] 22) Use the EM algorithm to estimate the drift coefficient and the diffusion coefficient in the two-stage Wiener degradation model;

[0014] 23) Use an adaptive gated dual attention unit to dynamically update the drift coefficient;

[0015] Step 3: Deduce the lifetime distribution function of the two-stage Wiener process

[0016] Convert the non-linear degradation process into a standard Brownian motion, use the properties of the standard Brownian motion to deduce the distribution of the remaining lifetime, and deduce the lifetime distribution function according to the independent increment property of the Wiener process;

[0017] Step 4: Battery state of health prediction

[0018] 41) Deduce the state transition probability density function of the neural Wiener process according to Itô's principle;

[0019] 42) Take the mathematical expectation of the state transition probability density function as the state of health of the battery in the current cycle.

[0020] Furthermore, in the step 11), the hidden layer state output by the forward GRU is:

[0021]

[0022] The hidden layer state output by the reverse GRU is:

[0023]

[0024] Concatenating the hidden layer states of forward and backward propagation gives:

[0025]

[0026] Where: Q t is the hidden layer state obtained by concatenation; is the hidden layer state output by the forward GRU; is the hidden layer state output by the reverse GRU; X t―1 is the input data at the previous moment; w n is the embedding vector.

[0027] Furthermore, in step 12), a comprehensive loss function is constructed to optimize the adaptive threshold, and the comprehensive loss function is:

[0028] L total = L change + λL threshold

[0029] Where: L total is the comprehensive loss function; L change is the change point recognition loss function; L threshold is the threshold optimization loss function; λ is the balance coefficient; and:

[0030]

[0031] Where: X t is the true label; is the probability that the Bi-GRU model predicts a change point; T is the current moment;

[0032]

[0033] Where: θ is the adaptive threshold; is the indicator function.

[0034] Furthermore, in step 21), the two-stage Wiener degradation model is expressed as:

[0035]

[0036] Where: X t represents the degradation amount of the lithium-ion battery at time t; X k is the degradation amount of the lithium-ion battery at t k moment; τ is the change point moment; X τis the degradation amount of the lithium-ion battery at the change point; μ1(t1; b1) and μ2(t2−τ; b2) are the drift coefficient functions in the two degradation stages respectively; t1 is the time of the first stage; b1 is the parameter of the first-order drift function; t2 is the time of the second stage; b2 is the drift coefficient function of the second stage; σ1 and σ2 are the diffusion coefficients before and after the change point; B(t) is the standard Brownian motion.

[0037] Further, in the step 22), the method for estimating the drift coefficient and the diffusion coefficient in the two-stage Wiener degradation model by using the EM algorithm is as follows: in the E step, the posterior distribution of the hidden variable is calculated through the Kalman filter, and the optimal degradation state estimation is provided; in the M step, based on the degradation state estimation, the drift coefficient and the diffusion coefficient are estimated by maximizing the log-likelihood function;

[0038] Through the iteration of the EM algorithm, the estimated value of the drift coefficient and the estimated value of the diffusion coefficient

[0039]

[0040] Where: is the degradation state estimation calculated in the E step; T is the current time; R is the variance of the observation noise.

[0041] Further, in the step 23), the adaptive gated dual-attention unit includes a reset gate, an update gate, an attention gate 1, and an attention gate 2;

[0042] The reset gate r t is used to combine the state h at the previous moment t―1 with the current drift coefficient input μ t , expressed as:

[0043] r t = σ(W r [h t―1 , μ t ] + b r )

[0044] Where: σ is the sigmoid activation function; W r is the weight matrix of the reset gate; b r is the bias matrix;

[0045] The update gate z t is used to control the fusion ratio of the state information at the previous moment and the state information at the current moment, expressed as:

[0046] z t = σ(W z [h t―1 , μ t ] + bz )

[0047] Where: W z is the weight matrix of the update gate, and b z is the bias matrix;

[0048] The attention gate 1 is used to control the attention degree of the input at the current moment to the hidden state information at the previous moment, and is expressed as:

[0049] a1 = Tanh(a t ⊙W a1 [h t―1 , μ t )

[0050] Where: a1 represents the attention gate 1; W a1 is the weight matrix of the attention gate 1; a t is the weight converted by the softmax function score; ⊙ represents the dot product operation;

[0051] The attention gate 2 is used to focus on the output states of the reset gate and the update gate, and is expressed as:

[0052] a2 = σ(W s [r t , z t, x c ) ⊙ Tanh(W t [r t , z t, x c )

[0053] Where: a2 represents the attention gate 2; W s and W t are the weight matrices learned by the attention gate 2; x c is the change point position;

[0054] The update formula for the hidden state of the adaptive gated dual attention unit is:

[0055] h h = Tanh(u h μ t + w h (r t ⊙ h t―1 ) + b h

[0056] h t = (1 - z t ) ⊙ h t―1 + z t ⊙ h h+ z t ⊙ a1 + a2

[0057] Where: u h and w h is the weight matrix; b h is the bias matrix; h h Obtained through the Tanh function, it represents the candidate hidden state at time t; h t The current state.

[0058] Furthermore, the output of the adaptive gated dual attention unit is expressed as:

[0059] f(t) 1 = Dropout(h t )

[0060] f(t) 2 = Dropout(Relu(w1f(t) 1 +b1)

[0061] f(t) 3 =w2f(t) 2 +b2

[0062] f(t)=w3(Dropout((f(t) 3 +h t ))+b3

[0063] Where: f(t) is the drift coefficient function updated by the adaptive gated dual attention unit; f(t) 1 is the original hidden state of Dropout regularization; f(t) 2 is the deep feature after ReLU+Dropout; f(t) 3 are the drift coefficient related features after full connection; w1, w2 and w3 are the weight matrices of the output layer; b1, b2 and b3 are bias vectors; Dropout(·) is the overfitting regularization technique of the neural network model, and Relu is the activation function.

[0064] Further, in step 3, during the first stage of the nonlinear degradation process:

[0065]

[0066] During the second stage of the nonlinear degradation process:

[0067]

[0068] After τ time after the first stage, the degradation amount changes from X k To X τ Transition probability m τ (X τ ) is expressed as:

[0069] m τ (X τ ) = Pr{X(τ) = X τ |X(k) = X k , T > τ} * Pr{T > τ}

[0070] Where: Pr{·} represents the posterior probability;

[0071] Based on the FHT definition, the two-stage lifetime distribution function is obtained:

[0072]

[0073] Where: w is the failure threshold; m τ (s τ )ds τ is the transition probability.

[0074] Furthermore, in the said step 41), the neural Wiener process is expressed as:

[0075]

[0076] Where: f(t) represents the drift coefficient function; σ M represents the diffusion coefficient function; W τ represents the standard Brownian motion;

[0077] The state transition probability density function is:

[0078]

[0079] Where: X t represents the historical observation data.

[0080] Furthermore, in the said step 42), the battery health state of the current cycle is:

[0081] E[X t |X t―1 = X t―1 e ―f(t)t

[0082] Where: E[X t |X t―1 is the mathematical expectation of the state transition probability density function.

[0083] The beneficial effects of the present invention are as follows:

[0084] The present invention monitors degradation data using the charge and discharge history of lithium-ion batteries, and proposes a method for predicting the state of health of batteries based on a two-stage neural Wiener process, aiming to achieve a more accurate prediction of the degradation trajectory of lithium-ion batteries. The method of the present invention estimates the drift coefficient and diffusion coefficient in the battery degradation process through the Expectation-Maximization algorithm (EM algorithm), effectively overcoming the limitations of traditional maximum likelihood estimation methods in battery degradation models, and being able to better handle the randomness and unpredictability in the battery degradation process. The method of the present invention fully considers the influence of randomness and change points in the degradation process on the degradation trajectory, and studies an update model of the drift coefficient based on the Adaptive Gated Dual Attention Unit (AGDAU) neural network. The AGDAU neural network model successfully addresses the short-term fluctuations and change point problems in the battery degradation process by adaptively adjusting the attention weights, enabling the model to accurately capture the complex non-linear relationship between the degradation state and the degradation rate.

[0085] Therefore, the method of the present invention has unique advantages. It not only solves the problem that the single-stage degradation model cannot describe the complex two-stage degradation process due to its assumed single degradation rate, but also accurately reflects the degradation model with different degradation characteristics in different stages, realizes the accurate prediction of the state of health (SOH) of batteries with different degradation rates, and has high popularization value and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] In order to make the objectives, technical solutions and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0087] Figure 1 is the flowchart of the method for predicting the state of health of batteries based on the two-stage neural Wiener process of the present invention;

[0088] Figure 2 is the schematic diagram of the Adaptive Gated Dual Attention Unit. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0089] The following further describes the present invention in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments given are not intended to limit the present invention.

[0090] Based on the idea of fusing and driving a random degradation model and a data-driven model, this embodiment proposes a new method for predicting the state of health (SOH) of batteries based on a two-stage neural Wiener process. This method fully considers the individual differences of lithium-ion batteries in the same batch at each stage. To solve the problem that the change point moment is not restricted by the initial distribution form, considering the randomness and short-term change point fluctuations in the degradation process, a change point recognition model is constructed by combining a Bi-GRU network and an adaptive threshold adjustment mechanism, which can adaptively find the change point and optimize the accuracy of change point detection, so as to provide more accurate prediction and decision-making support for battery health management. By designing an adaptive gated dual-attention mechanism, the context information at history, future, and change points is obtained simultaneously. Attention 1 controls the attention degree between the input at the current moment and the hidden state information at the previous moment, and Attention 2 responds to the influence generated by the change point. The drift coefficient function is updated in real time according to the few-failure sample data, enhancing the learning ability and accuracy of the neural Wiener process (WP) model. This model estimates the degradation coefficient based on the EM algorithm, fully considering the uncertainty of the two-stage degradation amount, and uses a neural network to describe the degradation process for each stage at the change point. A more accurate mapping relationship between the drift coefficient and the degradation state is established, overcoming the limitations of existing neural networks, WP models, and single-stage neural WP models in estimating SOH. Compared with other traditional methods, this embodiment fully considers the individual differences at the change point position, and the distribution form is not restricted by the initial conditions, and can more effectively predict the battery SOH. The SOH prediction obtained in this embodiment is closer to the true value.

[0091] Specifically, as Figure 1 shown, the method for predicting the state of health of batteries based on a two-stage neural Wiener process in this embodiment includes the following steps.

[0092] Step 1: Change point recognition

[0093] 11) Learn the temporal features in the battery degradation process through the Bi-GRU model, and use the concatenation of the hidden layer states of forward and backward propagation to predict the state of health in the battery degradation process.

[0094] Traditional two-stage Wiener process change point estimation methods usually assume that the change point follows a specific probability distribution. However, considering the working conditions and individual differences of batteries, the occurrence time of the change point has strong randomness. Therefore, this embodiment learns the temporal features in the battery degradation process through the Bi-GRU model, and uses the concatenation of the hidden layer states of forward and backward propagation to better model the change point in the battery degradation process. The Bi-GRU network is trained using time series data, and the model learns the time-dependent relationship of battery degradation.

[0095] The hidden layer state output by the forward GRU is:

[0096]

[0097] The hidden layer state output by the reverse GRU is as follows:

[0098]

[0099] Concatenating the hidden layer states of forward and backward propagation gives:

[0100]

[0101] Where: Q t is the concatenated hidden layer state; is the hidden layer state output by the forward GRU; is the hidden layer state output by the reverse GRU; X t―1 is the input data at the previous time step; w n is the embedding vector.

[0102] After training is completed, the GRU can predict the battery health state at each time step.

[0103] 12) Introduce a learnable adaptive threshold. If the residual between the predicted value and the observed value of the battery health state exceeds the set adaptive threshold, a change point occurs.

[0104] Specifically, to improve the flexibility of change point detection, we introduce a learnable adaptive threshold θ and adaptively adjust this threshold when optimizing the loss function. By comprehensively considering the identification of change points and the optimization of thresholds, the model can make dynamic adjustments according to different degradation modes.

[0105] In this embodiment, a comprehensive loss function is constructed to optimize the adaptive threshold. The comprehensive loss function is:

[0106] L total = L change + λL threshold

[0107] Where: L total is the comprehensive loss function; L change is the change point identification loss function; L threshold is the threshold optimization loss function; λ is the balance coefficient.

[0108] In this embodiment, the performance of the network in change point detection is measured by the cross-entropy loss. The change point identification loss can be defined as the binary cross-entropy loss, and the formula is:

[0109]

[0110] Where: X t is the true label; is the probability predicted by the Bi-GRU model as the change point; T is the current time.

[0111] By adjusting the threshold θ, the network can better adapt to different degradation modes. The goal of threshold optimization loss is to minimize the error of change point detection and achieve adaptive optimization by introducing this threshold, which is expressed as:

[0112]

[0113] where: θ is the adaptive threshold; is the indicator function, when exceeds the current threshold θ, it will affect the loss function.

[0114] In this embodiment, by combining the Bi-GRU network and the adaptive threshold adjustment mechanism to construct a change point recognition model, the change points in the degradation trajectory of lithium-ion batteries can be accurately detected, thereby providing more accurate prediction and decision support for battery health management.

[0115] Step 2: Two-stage Wiener process parameter estimation

[0116] 21) Taking the change point time τ as the demarcation point, establish a two-stage Wiener degradation model including the drift coefficient and the diffusion coefficient.

[0117] The change of the degradation rate of lithium-ion batteries shows two different stages. In each stage, the degradation rate of the battery is different. For the convenience of research, we assume that these two degradation stages occur in non-overlapping time periods, and each stage follows the definition and characteristics of the Wiener continuous-time stochastic process. Specifically, both of these two stages have the property of stationary independent increments. Based on this assumption, the constructed two-stage Wiener degradation model is expressed as:

[0118]

[0119] where: X t represents the degradation amount of the lithium-ion battery at time t; X k is the degradation amount of the lithium-ion battery at t k time; τ is the change point time; X τ is the degradation amount of the lithium-ion battery at the change point; μ1(t1; b1) and μ2(t2−τ; b2) are the drift coefficient functions in the two degradation stages respectively; t1 is the time of the first stage; b1 is the parameter of the first-order drift function; t2 is the time of the second stage; b2 is the drift coefficient function of the second stage; σ1 and σ2 are the diffusion coefficients before and after the change point respectively; B(t) is the standard Brownian motion.

[0120] 22) Use the EM algorithm to estimate the drift coefficient and the diffusion coefficient in the two-stage Wiener degradation model.

[0121] After extracting the degradation data of lithium batteries, the Expectation-Maximization (EM) algorithm is used to estimate the unknown parameters. In the case of missing data, the EM algorithm is an effective method for parameter estimation. For the Wiener process degradation model, there are some hidden variables, such as the true degradation state Q of the battery at each moment t , which cannot be directly observed and can only be indirectly inferred through the performance indicators of the battery. The core idea of the EM algorithm is to maximize the log-likelihood function of the complete data by iteratively calculating the expectation of the "latent data". Assuming that the degradation process of lithium-ion batteries consists of two stages, each stage can be regarded as an independent Wiener process. The basic steps for estimating the coefficients of the two-stage Wiener degradation model are as follows

[0122] 1. E-step (Expectation step): Calculate the posterior distribution of the hidden variables based on the current parameter estimates

[0123] 2. M-step (Maximization step): Maximize the log-likelihood function based on the expectations of the hidden variables calculated in the E-step to obtain new parameter estimates

[0124] Suppose we have a set of observed data X = {X1, X2, …, X T} of lithium-ion batteries. The relationship between these data and the hidden variables (i.e., the states of the degradation process) can be modeled by the following formula

[0125] State model

[0126]

[0127] where: μ is the drift coefficient; σ is the diffusion coefficient; Δt is the time difference between the previous and the current moment; ∈ t ~N(0, 1), which is a random noise following the standard normal distribution

[0128] Observation model: The relationship between the observed data X t and the hidden state Q t can be expressed by the following formula

[0129] X t = Q t + η t

[0130] where: η t is the observation noise, which is usually assumed to follow the normal distribution η t ~N(0, R), and R is the variance of the observation noise

[0131] Specifically, the method for estimating the drift coefficient and diffusion coefficient in the two-stage Wiener degradation model using the EM algorithm is as follows: In the E step, the posterior distribution of the hidden variable is calculated through Kalman filtering, and the optimal degradation state estimation is provided; in the M step, based on the degradation state estimation, the drift coefficient and diffusion coefficient are estimated by maximizing the log-likelihood function.

[0132] Through the iteration of the EM algorithm, I obtained the estimated value of the final drift coefficient and the estimated value of the diffusion coefficient

[0133]

[0134] Where: is the degradation state estimation calculated in the E step; T is the current time; R is the variance of the observation noise.

[0135] 23) The drift coefficient is dynamically updated using an adaptive gated dual attention unit.

[0136] To enhance the context capture ability so that the model can more effectively process the long-range dependencies in the sequence data, by adding two attention gates, in the presence of change points, the model can more flexibly focus on the influence of historical information on the current state.

[0137] As Figure 2 shown, in this embodiment, the adaptive gated dual attention unit (AGDAU) includes a reset gate, an update gate, attention gate 1, and attention gate 2.

[0138] The reset gate r t is used to combine the state h at the previous time t―1 with the current drift coefficient input μ t , expressed as:

[0139] r t = σ(W r [h t―1 , μ t + b r )

[0140] Where: σ is the sigmoid activation function; W r is the weight matrix of the reset gate; b r is the bias matrix.

[0141] The update gate z t is used to control the fusion ratio of the state information at the previous time and the state information at the current time, expressed as:

[0142] z t = σ(W z [h t―1 , μt +b z )

[0143] Where: W z is the weight matrix of the update gate, b z is the bias matrix.

[0144] Attention gate 1 is used to control the degree of attention to the input at the current moment and the hidden state information at the previous moment, expressed as:

[0145] a1 = Tanh(a t ⊙W a1 [h t―1 , μ t )

[0146] Where: a1 represents attention gate 1; W a1 is the weight matrix of attention gate 1; a t is the weight converted by the softmax function score; ⊙ represents the dot product operation.

[0147] Attention gate 2 is used to focus on the output states of the reset gate and the update gate, expressed as:

[0148] a2 = σ(W s [r t , z t, x c ) ⊙ Tanh(W t [r t , z t, x c )

[0149] Where: a2 represents attention gate 2; W s and W t are the weight matrices learned by attention gate 2; x c is the change point position.

[0150] The update formula for the hidden state of the adaptive gated dual attention unit is:

[0151] h h = Tanh(u h μ t + w h (r t ⊙h t―1 ) + b h

[0152] h t = (1 - z t ) ⊙ h t―1 + z t ⊙ h h+ z t ⊙ a1 + a2

[0153] where: u h and w h are weight matrices; b h is a bias matrix; h h is obtained through the Tanh function and represents the candidate hidden state at time t; h t is the state at the current time.

[0154] Through the outputs of the reset gate, update gate, and attention gates 1 and 2 in AGDAU, a weighted output for calculating the drift coefficient function is finally obtained. The output of the adaptive gating dual attention unit is expressed as:

[0155] f(t) 1 = Dropout(h t )

[0156] f(t) 2 = Dropout(Relu(w1f(t) 1 + b1)

[0157] f(t) 3 = w2f(t) 2 + b2

[0158] f(t) = w3(Dropout((f(t) 3 + h t )) + b3

[0159] where: f(t) is the drift coefficient function updated by the adaptive gating dual attention unit; f(t) 1 is the original hidden state with Dropout regularization; f(t) 2 is the deep feature after ReLU + Dropout; f(t) 3 is the drift coefficient-related feature after full connection; w1, w2, and w3 are all weight matrices of the output layer; b1, b2, and b3 are all bias vectors; Dropout(·) is a neural network model overfitting regularization technique, and Relu is an activation function.

[0160] Step 3: Derive the lifetime distribution function of the two-stage Wiener process

[0161] Convert the non-linear degradation process into a standard Brownian motion, use the properties of the standard Brownian motion to derive the distribution of the remaining lifetime, and derive the lifetime distribution function according to the independent increment property of the Wiener process.

[0162] To derive the remaining life distribution, it is assumed that through an appropriate time - space transformation, the non - linear degradation process can be transformed into a standard Brownian motion B(t). Based on this, the properties of the standard Brownian motion can be used to derive the distribution of the remaining life, and the life distribution function is derived according to the independent increment property of the Wiener process.

[0163] In the first stage of the non - linear degradation process:

[0164]

[0165] In the second stage of the non - linear degradation process:

[0166]

[0167] After τ time from the first stage, the degradation amount changes from X k to X τ The transition probability m τ (X τ ) is expressed as:

[0168] m τ (X τ ) = Pr{X(τ) = X τ |X(k) = X k , T > τ} * Pr{T > τ}

[0169] where: Pr{·} represents the posterior probability.

[0170] Based on the definition of FHT (Fast Hadamard Transform), the two - stage life distribution function is obtained:

[0171]

[0172] where: w is the failure threshold; m τ (s τ )ds τ is the transition probability.

[0173] When the degradation amount of the lithium - ion battery exceeds the failure threshold w before the change point (i.e., the moment when the degradation mechanism changes significantly), it means that the battery has failed before reaching the stage where its degradation mechanism changes. Therefore, it is not necessary to consider the degradation behavior after the change point, and at this time the life T < τ; when the lithium - ion battery fails after the change point, that is, the battery life T > τ, it means that the degradation process of the battery includes two stages: the degradation stage before the change point and the degradation stage after the change point. Once the unknown parameters of the two - stage degradation model are determined, the life distribution function (PDF) can be obtained.

[0174] Step Four: Prediction of the State of Health of the Battery

[0175] 41) Derive the state transition probability density function of the neural Wiener process according to Ito's principle.

[0176] The neural Wiener process is expressed as:

[0177]

[0178] where: f(t) represents the drift coefficient function; σ M is the diffusion coefficient function; W τ is the standard Brownian motion.

[0179] When a historical observation data X t―1 is given, the state transition probability density function (TPDF) about X t can be obtained as:

[0180]

[0181] where: X t represents the historical observation data.

[0182] 42) Take the mathematical expectation of the state transition probability density function as the battery health state of the current cycle.

[0183] Specifically, the battery health state of the current cycle is:

[0184] E[X t |X t―1 = X t―1 e ―f(t)t

[0185] where: E[X t |X t―1 is the mathematical expectation of the state transition probability density function.

[0186] In this embodiment, by monitoring the degradation data based on the charge and discharge history of lithium-ion batteries, a method for predicting the state of health of batteries based on a two-stage neural Wiener process is proposed, aiming to achieve a more accurate prediction of the degradation trajectory of lithium-ion batteries. The method in this embodiment estimates the drift coefficient and diffusion coefficient in the battery degradation process through the Expectation-Maximization algorithm (EM algorithm), effectively overcoming the limitations of traditional maximum likelihood estimation methods in battery degradation models and being able to better handle the randomness and unpredictability in the battery degradation process. The method in this embodiment fully considers the influence of randomness and change points in the degradation process on the degradation trajectory, and studies an update model for the drift coefficient based on an Adaptive Gated Dual Attention Unit (AGDAU) neural network. The AGDAU neural network model successfully addresses the short-term fluctuations and change point problems in the battery degradation process by adaptively adjusting the attention weights, enabling the model to accurately capture the complex non-linear relationship between the degradation state and the degradation rate.

[0187] Therefore, the method in this embodiment has unique advantages. It not only solves the problem that the single-stage degradation model cannot describe the complex two-stage degradation process due to its assumed single degradation rate, but also accurately reflects the degradation model with different degradation characteristics in different stages, achieving an accurate prediction of the State of Health (SOH) of batteries with different degradation rates, and having high popularization value and application value.

[0188] The above-described embodiments are only preferred embodiments given to fully illustrate the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention is subject to the claims.

Claims

1. A battery health status prediction method based on a two-stage neural Wiener process, characterized in that: The steps include: Step 1: Change point identification 11) The Bi-GRU model is used to learn the time series characteristics of the battery degradation process, and the health status of the battery degradation process is predicted by using the hidden layer state splicing of forward and backward propagation; 12) A learnable adaptive threshold is introduced. If the residual between the predicted value and the observed value of the battery health state exceeds the set adaptive threshold, a change point occurs; Step 2: Two-stage Wiener process parameter estimation 21) Taking the change point time τ as the dividing point, a two-stage Wiener degradation model including drift coefficient and diffusion coefficient is established; 22) The EM algorithm is used to estimate the drift coefficient and diffusion coefficient in the two-stage Wiener degradation model; 23) Adopt adaptive gated dual attention units to dynamically update the drift coefficient; Step 3: Derive the lifetime distribution function of the two-stage Wiener process The nonlinear degradation process is transformed into a standard Brownian motion, and the distribution of the remaining life is derived using the properties of the standard Brownian motion. The life distribution function is derived based on the independent increment properties of the Wiener process. Step 4: Battery health status prediction 41) According to Ito's principle, the state transition probability density function of the neural Wiener process is derived; 42) The mathematical expectation of the state transition probability density function is taken as the battery health state of the current cycle.

2. The battery health status prediction method based on the two-stage neural Wiener process according to claim 1 is characterized in that: In step 11), the hidden layer state of the forward GRU output is: The hidden layer state of the reverse GRU output is: Concatenating the hidden layer states of the forward and backward propagation yields: Where: Q t is the hidden layer state obtained by splicing; is the hidden layer state of the forward GRU output; is the hidden layer state of the reverse GRU output; X t―1 is the input data of the previous moment; w n is the embedding vector.

3. The battery health status prediction method based on the two-stage neural Wiener process according to claim 1 is characterized in that: In the step 12), a comprehensive loss function is constructed to optimize the adaptive threshold, and the comprehensive loss function is: THE total =L change +λL threshold Where: L total is the comprehensive loss function; L change is the loss function for change point identification; L threshold is the threshold optimization loss function; λ is the balance coefficient; and: Where: X t is the true label; is the probability of the Bi-GRU model predicting a change point; T is the current time; Where: θ is the adaptive threshold; is the indicator function.

4. The battery health status prediction method based on the two-stage neural Wiener process according to claim 1 is characterized in that: In the step 21), the two-stage Wiener degradation model is expressed as: Where: X t represents the degradation amount of lithium-ion battery at time t; X k For lithium-ion batteries at t k The degradation amount at the moment; τ is the change point moment; X τ is the degradation amount of the lithium-ion battery at the change point; μ1(t1; b1) and μ2(t2-τ); b2) are the drift coefficient functions in the two degradation stages respectively; t1 is the time of the first stage; b1 is the first-order drift function parameter; t2 is the time of the second stage; b2 is the drift coefficient function of the second stage; σ1 and σ2 are the diffusion coefficients before and after the change point respectively; B(t) is the standard Brownian motion.

5. The battery health status prediction method based on the two-stage neural Wiener process according to claim 4 is characterized in that: In the step 22), the method of estimating the drift coefficient and diffusion coefficient in the two-stage Wiener degradation model using the EM algorithm is as follows: the E step calculates the posterior distribution of the latent variable through Kalman filtering and provides the optimal degradation state estimation; the M step estimates the drift coefficient and diffusion coefficient by maximizing the log-likelihood function based on the degradation state estimation; Through the iteration of the EM algorithm, the estimated value of the drift coefficient is obtained and the estimated diffusion coefficient in: is the degradation state estimate calculated in step E; T is the current moment; R is the variance of the observation noise.

6. The battery health status prediction method based on the two-stage neural Wiener process according to claim 5 is characterized in that: In the step 23), the adaptive gated dual attention unit includes a reset gate, an update gate, an attention gate 1 and an attention gate 2; Reset Gate t Used to combine the state h of the previous moment t―1 With the current drift coefficient input μ t , expressed as: r t =σ(W r [h t―1 ,m t ]+b r ) Where: σ is the sigmoid activation function; W r is the weight matrix of the reset gate; b r is the bias matrix; Update gate z t It is used to control the fusion ratio of the state information at the previous moment and the state information at the current moment, expressed as: z t =σ(W z [h t―1 ,m t ]+b z ) Where: W z is the weight matrix of the update gate, b z is the bias matrix; Attention gate 1 is used to control the attention level of the current moment's input and the previous moment's hidden state information, expressed as: a1=Fish(a) t ⊙W a1 [h t―1 ,μ t ]) Where: a1 represents attention gate 1; W a1 is the attention gate 1 weight matrix; a t is the weight converted by the softmax function score; ⊙ represents the dot product operation; Attention gate 2 is used to focus on the output states of the reset gate and update gate, expressed as: a2=σ(W s [r t ,z t, x c ])⊙Tanh(W t [r t ,z t, x c ]) Where: a2 represents attention gate 2; W s and W t is the weight matrix learned by attention gate 2; x c is the change point position; The update formula of the hidden state of the adaptive gated dual attention unit is: h h =Tanh(u h μ t +w h (r t ⊙h t―1 )+b h h t =(1―z t )⊙h t―1 +z t ⊙h h+ z t ⊙a1+a2 Where: u h and w h is the weight matrix; b h is the bias matrix; h h Obtained through the Tanh function, it represents the candidate hidden state at time t; h t The current state.

7. The battery health status prediction method based on the two-stage neural Wiener process according to claim 6 is characterized in that: The output of the adaptive gated dual attention unit is expressed as: f(t) 1 =Dropout(h t ) f(t) 2 =Dropout(Relu(w1f(t) 1 +b1) f(t) 3 =w2f(t) 2 +b2 f(t)=w3(Dropout((f(t) 3 +h t ))+b3 Where: f(t) is the drift coefficient function f(t) updated by the adaptive gated dual attention unit 1 is the original hidden state of Dropout regularization; f(t) 2 is the deep feature after ReLU+Dropout; f(t) 3 are the drift coefficient related features after full connection; w1, w2 and w3 are the weight matrices of the output layer; b1, b2 and b3 are bias vectors; Dropout(·) is the overfitting regularization technique of the neural network model, and Relu is the activation function.

8. The battery health status prediction method based on the two-stage neural Wiener process according to claim 4 is characterized in that: In step 3, during the first stage of the nonlinear degradation process: During the second stage of the nonlinear degradation process: After τ time after the first stage, the degradation amount changes from X k To X τ Transition probability m τ (X τ ) is expressed as: m τ (X τ )=Pr{X(τ)=X τ |X(k)=X k ,T>τ}*Pr{T>τ} Where: Pr{·} represents the posterior probability; Based on the FHT definition, the two-stage life distribution function is obtained: Where: w is the failure threshold; m τ (s τ )ds τ is the transition probability.

9. The battery health status prediction method based on the two-stage neural Wiener process according to claim 1 is characterized in that: In step 41), the neural Wiener process is expressed as: Where: f(t) represents the drift coefficient function; σ M W represents the diffusion coefficient function; τ shows standard Brownian motion; The state transition probability density function is: Where: X t Represents historical observation data.

10. The battery health status prediction method based on the two-stage neural Wiener process according to claim 9 is characterized in that: In step 42), the battery health status of the current cycle is: E[X t |X t―1 ]=X t―1 e ―f(t)t Where: E[X t |X t―1 ] is the mathematical expectation of the state transition probability density function.

Images