Lithium ion battery monitoring method based on improved particle filtering and electronic equipment

Abstract

The invention provides a lithium ion battery monitoring method based on improved particle filtering and electronic equipment, and belongs to the field of lithium battery performance evaluation. According to the method, a capacity degradation model of the lithium ion battery is constructed, and a posterior probability density function is estimated by utilizing Bayesian formula reasoning and combining Monte Carlo sampling. A real probability density function is approximated by introducing mixed Gaussian distribution, particles are corrected and subjected to weight normalization, and Gaussian distribution parameters are updated by using an EM algorithm. And estimating a state quantity probability density function by using an enhanced joint Gaussian mixture EUMG-PF, and adaptively adjusting a covariance by using capacity second-order difference information and a capacity difference. Based on a large number of battery degradation data sets, a battery degradation cut-off capacity range is determined, Gaussian component sampling with the maximum weight is selected, and a constraint equation is constructed to solve an optimal state quantity mean value. According to the method, the accuracy of predicting the residual life of the lithium ion battery can be remarkably improved while the capacity estimation accuracy in a short time is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery performance evaluation, and more particularly to a lithium ion battery monitoring method and electronic equipment based on improved particle filtering. Background Art

[0002] Lithium-ion batteries (lithium batteries) have the advantages of long cycle life, low self-discharge rate, and high energy density. They have been widely used in new energy vehicles, consumer electronics, and grid energy storage. Due to repeated charging and discharging of the battery, the activity of the positive and negative electrode materials inside the battery will gradually decrease, and the electrolyte capacity will continue to decrease, resulting in a shortening of the battery capacity and remaining useful life (Remaining Useful Life, RUL). The remaining useful life refers to the number of cycles from the current to the capacity decay to 80%. Accurate and rapid estimation of battery capacity and remaining useful life is conducive to timely replacement or maintenance of the battery, ensuring the safe and reliable operation of the system.

[0003] At present, there are many methods for estimating the capacity of lithium batteries. Among them, the method based on the empirical degradation model starts from the data perspective and seeks the battery degradation law. The model is easy to obtain and has a wide range of applications. Particle filter (PF) has the optimal solution for nonlinear non-Gaussian degradation problems and has excellent performance in battery capacity estimation. However, there are the following problems:

[0004] (1) There is a serious particle degradation problem during the repeated iteration process, which affects the accuracy of the estimation;

[0005] (2) The state variables of the battery degradation model should be jointly distributed with multiple Gaussians, rather than independently distributed with a single Gaussian.

[0006] (3) There is correlation and obvious fluctuation in adjacent discharge conditions of actual lithium batteries;

[0007] (4) Imbalance between long-term and short-term capacity predictions. That is, more accurate predictions of state quantities can achieve more effective short-term capacity estimation, but long-term capacity estimation results may have serious deviations, resulting in extremely poor RUL prediction accuracy.

[0008] Therefore, how to provide a lithium-ion battery monitoring method and electronic equipment based on an improved particle filter that can solve the above problems is a problem that those skilled in the art urgently need to solve. Summary of the invention

[0009] In view of this, the present invention provides a lithium-ion battery monitoring method and electronic equipment based on an improved particle filter to solve the technical problems existing in the above-mentioned prior art, and further improve the accuracy of remaining life prediction while ensuring accurate short-term capacity estimation.

[0010] In order to achieve the above object, the present invention provides the following technical solutions:

[0011] A lithium-ion battery monitoring method based on an improved particle filter comprises:

[0012] Constructing a capacity degradation model of a lithium-ion battery, wherein the capacity degradation model of a lithium-ion battery includes a state equation and an observation equation;

[0013] Use the Bayesian formula to infer the state equation and observation equation to obtain the posterior state probability density function;

[0014] Monte Carlo sampling is introduced to estimate the posterior probability density function using particles and particle weights;

[0015] The joint probability density function of the mixed Gaussian is introduced to approximate the true probability density function of the state quantity, particle correction is performed, low-weight particles are removed, weights are normalized, and the mean, covariance and weight of the Gaussian distribution are updated and calculated using the EM algorithm;

[0016] The probability density function of the state quantity is estimated and sampled based on the enhanced joint mixed Gaussian EUMG-PF model;

[0017] Adaptively adjust the covariance using the second-order difference information of capacity and capacity difference;

[0018] Obtain a large number of battery degradation data sets, determine the battery degradation cutoff capacity range based on a large number of battery degradation data, select the Gaussian component with the largest weight for sampling, and construct a constraint equation to obtain the optimal state quantity mean;

[0019] Output the optimal state quantity mean to realize particle filter lithium-ion battery monitoring.

[0020] Preferably, the expression of the state equation is:

[0021]

[0022] where x k =[a k b k c k d k ] T represents the system state at time k; represents the process noise of the multidimensional mixed Gaussian distribution at time k-1, Φ represents the normal distribution, M represents the number of mixed Gaussian distributions in the state equation of the system, m represents the mth component in the mixed Gaussian distribution, Φ(μ,∑) represents the Gaussian distribution with mean μ and covariance ∑, represents the mean of the mth Gaussian distribution at time k-1; represents the covariance matrix of the mth Gaussian distribution at time k-1, πm Represents the weight coefficient of the mth Gaussian distribution component;

[0023] The expression of the observation equation is:

[0024]

[0025] Where Q k Represents x k The observed value at time k, a k and c k is a parameter related to the internal resistance of lithium batteries, b k and d k It is a parameter related to the degradation rate of lithium batteries.

[0026] Preferably, the state equation and the observation equation are inferred using the Bayesian formula to obtain the posterior state probability density function, including:

[0027] p(x 0:k |Q 1:k-1 )=∫p(x k |x 0:k-1 )p(x 0:k-1 |Q 1:k-1 )dx 0:k-1 ;

[0028] p(x 0:k |Q 1:k )∝p(Q k |x k )p(x k |x k-1 )p(x 0:k-1 |Q 1:k-1 );

[0029] Where p(x 0:k |Q 1:k-1 ) is the prior probability density function at time k, p(x k ,x k-1 ) is the posterior probability model of the evolution of the system state equation, p(Q k |x k ) is the likelihood function, and ∝ indicates a positive correlation.

[0030] Preferably, Monte Carlo sampling is introduced to estimate the posterior probability density function using particles and their weights, including:

[0031] Introducing Monte Carlo sampling, using a set of random particle samples The weight associated with Estimate the posterior probability density function p(x k |Q 1:k-1), introduce importance sampling to calculate the weight of the particle, and then normalize the weight. The formula is:

[0032]

[0033] Where p vk Indicates v k The probability density function form of ;

[0034] Use the weighted particles to estimate the state x k :

[0035]

[0036] In the formula, Q k represents the observed value at time k, is the observation model function, which represents the system state After observing the prediction results of model h.

[0037] Preferably, the joint probability density function of the mixed Gaussian is expressed as:

[0038]

[0039] Where G represents the number of Gaussian probability functions, ω g is the weight of the g-th Gaussian component, μ g is the mean vector of the g-th Gaussian component D*D, Σ g is the D*D dimensional covariance matrix of the g-th Gaussian component, |Σ g |isΣ g The determinant of , x represents the observation value, and T is the transpose of the matrix.

[0040] Preferably, the EM algorithm is used to update and calculate the mean, covariance and weight of the Gaussian distribution, including:

[0041] right The parameters are initialized respectively, and the best mixed Gaussian function is obtained through E-step and M-step iterations, where E-step:

[0042]

[0043] In the formula, x i is the i-th state data sample point, ω g is the weight of the g-th Gaussian component, μ g is the mean vector of the g-th Gaussian component D*D, Σ g is the D*D dimensional covariance matrix of the g-th Gaussian component, |Σ g |isΣ g The determinant of x represents the observation value, T is the transpose of the matrix, r igRepresents a given sample x i After that, it comes from the probability generated by the g-th Gaussian distribution;

[0044] M-step:

[0045]

[0046] Where N g Represents the sum of the probabilities that each sample belongs to the g-th Gaussian component;

[0047] Compute the log-likelihood function:

[0048]

[0049] The E-step and M-step are iterated until the convergence ε is less than the given threshold ε th , update the mean, covariance and weight of the Gaussian distribution.

[0050] Preferably, estimating the probability density function of the state quantity based on the enhanced joint mixed Gaussian EUMG-PF model and sampling includes:

[0051] Let the capacity parameter be defined as:

[0052] ΔQ(k)=Q(k)-Q(k-1);

[0053] Δ2Q(k)=ΔQ(k)-ΔQ(k-1);

[0054] Where k is the number of cycles, Q(k) is the battery capacity at time k, ΔQ(k) is the first-order difference of the battery capacity at time k, and Δ2Q(k) is the second-order difference of the battery capacity at time k.

[0055] Preferably, the covariance is adaptively adjusted using the second-order difference information of the capacity and the capacity difference, and the adjusted covariance is:

[0056]

[0057] Where a and b are correction coefficients, is the optimized covariance at time k, Σ(k) is the original covariance at time k, Q i (k) is the estimated SOH value of the i-th particle at time k; Indicates the difference between the estimated capacity and the actual capacity.

[0058] Preferably, a large number of battery degradation data sets are obtained, a battery degradation cutoff capacity range is determined based on the large number of battery degradation data, a Gaussian component with the largest weight is selected for sampling, and a constraint equation is constructed to obtain the optimal state quantity mean, including:

[0059]

[0060]

[0061] in The term represents the generation of N sample points Gaussian distribution, N is the total number of sample points; The term represents the estimated state quantity μ g The calculated mean square error between the historical capacity forecast and the actual capacity value, λ1 is the correction weight.

[0062] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of a lithium-ion battery monitoring method based on an improved particle filter are implemented.

[0063] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a lithium-ion battery monitoring method and electronic device based on an improved particle filter. First, a battery capacity and remaining life prediction model is established based on a double exponential empirical degradation model and a particle filter algorithm; then, a joint probability density function of state parameters estimated by a mixed Gaussian combined with a second-order difference algorithm is constructed to solve the problem of capacity correlation and obvious fluctuations under multiple working conditions; finally, transfer learning is combined with a constraint equation to optimize the state quantity from the joint probability density function, and the accuracy of the remaining life prediction is further improved under the premise of ensuring the accuracy of short-term capacity estimation. The specific beneficial effects are:

[0064] (1) The present invention uses a mixed joint Gaussian method to estimate the joint probability density function of state variables, replacing the resampling method, maintaining the diversity of particles and approximating the real particle distribution;

[0065] (2) The present invention calculates the distribution characteristics of the joint probability density function of multiple cycles, and provides a second-order difference algorithm method to optimize the joint probability density function, solving the problem that there is correlation and obvious fluctuation between two adjacent discharge conditions;

[0066] (3) The present invention uses transfer learning and constraint equation methods to determine the distribution range of RUL through historical similar data sets and existing current data sets, and uses RUL distribution to construct constraint equations to optimize the joint probability density function to ensure the balance between short-term capacity and long-term RUL estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0068] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0069] FIG2( a ) is a distribution diagram of 200 particles at the time k=10 provided in Example 1 of the present invention;

[0070] FIG2( b ) is a schematic diagram of predicted capacity corresponding to 200 particles at the time k=10 provided in Example 1 of the present invention;

[0071] FIG2( c ) is a schematic diagram of the weights of 200 particles at the time k=10 provided in Example 1 of the present invention;

[0072] FIG2(d) is a probability density distribution diagram of k at time k=10 provided in Example 1 of the present invention;

[0073] FIG2(e) is a distribution diagram of 200 particles at the time k=11 provided in Example 1 of the present invention;

[0074] FIG3( a ) is a schematic diagram of a situation where adjacent capacities have a large difference, provided in Embodiment 1 of the present invention;

[0075] FIG3( b ) is a schematic diagram of the estimated SOH value and the actual SOH of a particle provided in Example 1 of the present invention;

[0076] FIG4( a ) is a schematic diagram of RUL estimation comparison (NASA05) when k=90 provided in Example 2 of the present invention;

[0077] FIG4( b ) is a schematic diagram of RUL estimation comparison (NASA05) when k=98 provided in Example 2 of the present invention; ...

[0078] FIG4( c ) is a schematic diagram of RUL estimation comparison (NASA05) when k=104 provided in Example 2 of the present invention;

[0079] FIG5( a ) is a schematic diagram of RUL estimation comparison (NASA07) when k=90 provided in Example 2 of the present invention;

[0080] FIG5( b ) is a schematic diagram of RUL estimation comparison (NASA07) when k=98 provided in Example 2 of the present invention;

[0081] FIG5( c ) is a schematic diagram of RUL estimation comparison (NASA07) when k=90 provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0082] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0083] Embodiment 1:

[0084] See attached Figure 1 As shown, Embodiment 1 of the present invention discloses a lithium-ion battery monitoring method based on an improved particle filter, comprising:

[0085] A capacity degradation model of a lithium-ion battery, wherein the capacity degradation model of a lithium-ion battery includes a state equation and an observation equation;

[0086] Use the Bayesian formula to infer the state equation and observation equation to obtain the posterior state probability density function;

[0087] Monte Carlo sampling is introduced to estimate the posterior probability density function using particles and particle weights;

[0088] The joint probability density function of the mixed Gaussian is introduced to approximate the true probability density function of the state quantity, particle correction is performed, low-weight particles are removed, weights are normalized, and the mean, covariance and weight of the Gaussian distribution are updated and calculated using the EM algorithm;

[0089] Estimation of the probability density function of the state quantity and sampling based on the enhanced joint mixed Gaussian EUMG-PF;

[0090] Adaptively adjust the covariance using the second-order difference information of capacity and capacity difference;

[0091] Obtain a large number of battery degradation data sets, determine the battery degradation cutoff capacity range based on a large number of battery degradation data, select the Gaussian component with the largest weight for sampling, and construct a constraint equation to obtain the optimal state quantity mean;

[0092] Output the optimal state quantity mean to realize particle filter lithium-ion battery monitoring.

[0093] In a specific embodiment, a model based on a combination of mixed Gaussian temporal attention and transfer learning includes:

[0094] Lithium battery degradation model based on particle filtering:

[0095] Among them, the capacity degradation model of lithium-ion batteries is a typical dynamic system, which includes state equations and observation equations. The state equation is as follows:

[0096]

[0097] where x k =[a k b k c k d k ] T represents the system state at time k; represents the process noise of the multidimensional mixed Gaussian distribution at time k-1, Φ represents the normal distribution, M represents the number of mixed Gaussian distributions in the state equation of the system, m represents the mth component in the mixed Gaussian distribution, Φ(μ,∑) represents the Gaussian distribution with mean μ and covariance ∑, represents the mean of the mth Gaussian distribution at time k-1; represents the covariance matrix of the mth Gaussian distribution at time k-1, π m Represents the weight coefficient of the mth Gaussian distribution component;

[0098] The observation equation is as follows:

[0099]

[0100] Where Q k Represents x k The observed value at time k, a k and c k is a parameter related to the internal resistance of lithium batteries, b k and d k It is a parameter related to the degradation rate of lithium batteries.

[0101] Specifically, by performing state Bayesian reasoning on formula (1) and formula (2) using the Bayesian formula, we can obtain the posterior state probability density function, which can be divided into two steps: prediction (3) and update (4):

[0102] p(x 0:k |Q 1:k-1 )=∫p(x k |x 0:k-1 )p(x 0:k-1 |Q 1:k-1 )dx 0:k-1 (3)

[0103] p(x 0:k |Q 1:k )∝p(Q k |x k )p(x k |x k-1 )p(x 0:k-1 |Q 1:k-1 ) (4)

[0104] Where p(x 0:k |Q 1:k-1) is the prior probability density function at time k, p(x k ,x k-1 ) is the posterior probability model of the evolution of the system state equation, p(Q k |x k ) is the likelihood function, ∝ indicates a positive correlation, the prediction step is to estimate the posterior probability density function at the current moment based on the last posterior probability density function, and the update step uses the latest measurement value Q k Update the posterior probability density function.

[0105] Specifically, for general nonlinear non-Gaussian models, it is difficult to obtain an analytical solution for the posterior probability. In order to solve this problem, Monte Carlo sampling is introduced. Using a set of random samples (particles) The weight associated with Estimate the posterior probability density function p(x k |Q 1:k-1 ), introduce importance sampling to calculate the weight of particles, and then normalize the weight, that is, formula (5):

[0106]

[0107] In the formula Indicates v k The probability density function form, Q k represents the observed value at time k, is the observation model function, which represents the system state After observing the prediction results of model h

[0108] Finally, use the particles with calculated weights to estimate the state x k :

[0109]

[0110] Specifically, the detailed algorithm flow of the traditional particle filter is as follows:

[0111] Traditional particle filtering has the following problems in predicting lithium battery capacity and remaining life: (1) Traditional methods assume that each component of the multivariate state variable is independently Gaussian distributed, but in fact the probability density distribution of the state quantity of the battery degradation model is jointly non-Gaussian, and the traditional resampling method cannot accurately estimate the subsequent state quantity; (2) In actual lithium batteries, adjacent discharge conditions are correlated and significantly different. When the adjacent state variables vary greatly, even if the probability density distribution of the state quantity of the joint non-Gaussian is accurately estimated, there may be significant errors in sampling and calculating the subsequent capacity of the distribution; (3) The traditional particle filter estimates the capacity at time k+1 from time k with high accuracy, but the long-term RUL prediction curve will have extremely large errors or even complete failure, that is, the imbalance between long-term RUL estimation and short-term capacity estimation.

[0112] In a specific embodiment, the state quantity probability density function estimation based on the mixed joint Gaussian includes:

[0113] Traditional particle filtering assumes that each component of the multivariate state variable is an independent Gaussian distribution and that resampling has problems such as particle degradation. The probability density function of the state quantity of the actual lithium battery capacity degradation model is composed of multiple Gaussians and each Gaussian is multivariate joint. In order to improve the estimation accuracy of the subsequent state quantity, this section introduces a mixed joint Gaussian function to approximate the true probability density function of the state quantity. This method first corrects the particles, that is, removes particles with weights below the threshold and normalizes the particles to avoid the interference of low-weight particle estimates on the overall estimation; then the expectation maximization (EM) algorithm is used to update and calculate the mean, covariance and weight of each multivariate joint Gaussian distribution, so that the state quantity distribution is closer to the true distribution and the diversity of particles can be maintained.

[0114] (1)Particle correction

[0115] After the prediction update step of the particle filter, the weights are less than a given threshold w th Particles are eliminated and the weights are normalized by the number of remaining particles.

[0116] (2) Probability density function estimation

[0117] Assuming that the mixed Gaussian is a linear combination of multiple joint Gaussian functions, the joint probability density function of the mixed Gaussian is expressed as:

[0118]

[0119] Where G represents the number of Gaussian probability functions, ω g is the weight of the g-th Gaussian component, μ g is the mean vector of the g-th Gaussian component D*D, Σ g is the D*D dimensional covariance matrix of the g-th Gaussian component, |Σ g |isΣ g The determinant of , x represents the observation value, and T is the transpose of the matrix.

[0120] The EM algorithm is used to optimize the particle distribution. The parameters are assigned initial values, and then the best mixed Gaussian function is obtained through E-step and M-step iterations.

[0121] Step E:

[0122]

[0123] Where xi is the i-th state data sample point, r ig Given a sample x i Finally, it comes from the probability generated by the g-th Gaussian distribution.

[0124] M-step:

[0125]

[0126] Where N g represents the sum of the probabilities that each sample belongs to the g-th Gaussian component,

[0127] Compute the log-likelihood function:

[0128]

[0129] Specifically, the E-step and M-step are iterated repeatedly until the convergence ε is less than the given threshold ε th .

[0130] In order to describe the principle of the proposed joint probability density algorithm based on mixed Gaussian, in this embodiment 1, the state parameter of the state equation of formula (1) is set to x k =[a k b k ], the observation equation of formula (2) is selected as When the joint probability density function of a and b at time k-1 is known, the mixed Gaussian method is used to estimate the joint probability density function of state quantities a and b. The experimental results are shown in Figure 2(a)-(e). From the probability density distribution estimation process based on the mixed Gaussian method in Figure 2(a)-(e), it can be seen that: 1) x at time k=10 k The probability density function is composed of two Gaussians; 2) for x with k = 10 k The probability density function is used to sample N particles, and N SOH estimates are estimated by observing the equation. The distribution of the SOH estimate is non-Gaussian. 3) The distribution of the weighted state quantities a and b is calculated according to the true SOH value, and the particle diversity is obviously degraded; 4) The particles are corrected using the method in this section and the joint probability density function of a and b is estimated using the EM method. The probability density a and b at k = 10 maintains the diversity of particles well, which can effectively avoid the problem of particle degradation. 5) There are differences in the particle distribution at k = 10 and k = 11, but they are also correlated.

[0131] However, if the center position of the real SOH is very different from that of the estimated SOH value, it means that there is a huge difference in the probability density distribution between time k-1 and time k, that is, there is a huge estimation error problem. The main reason for this phenomenon is that there is a significant change in the real SOH value between time k-1 and time k.

[0132] In a specific embodiment, the probability density function estimation optimization based on multiple working conditions includes:

[0133] Embodiment 1 of the present invention accurately estimates the probability density function of the state quantity and samples it based on the UMG-PF method, effectively solving the particle degradation problem. Since the SOH of lithium batteries has a gradual decline at first and then a rapid decline, and there is capacity regeneration, as the number of cycles increases, there is a significant difference in the difference between adjacent SOH values ​​at different times. Assume that the capacity parameter is defined as:

[0134] ΔQ(k)=Q(k)-Q(k-1) (12)

[0135] Δ2Q(k)=ΔQ(k)-ΔQ(k-1) (13)

[0136] Where k is the number of cycles, Q(k), ΔQ(k), and Δ2Q(k) are the battery capacity, first-order difference of capacity, and second-order difference of capacity at time k. Partial diagram of the area where capacity decreases rapidly in Figure 3(a)-(b). It can be seen from Figure (a): 1) There is a significant difference in capacity between time k and k+1, that is, ΔQ(k) is not zero. 2) There is also a significant trend change in the adjacent capacity change rate, that is, Δ2Q(k) is not zero. It can be seen from Figure (b): when k=180~190, the capacity changes relatively slowly; when k=218~228, the capacity decreases relatively quickly. Therefore, although the probability density of time k estimated based on UMG-PF is extremely accurate, there is still a large deviation in sampling and estimating the capacity at k+1 based on this probability density. In order to reduce this deviation, this section uses the second-order difference information of capacity and the difference between the estimated capacity and the actual capacity to adaptively scale the covariance. At the same time, considering the high complexity of the mixed Gaussian and the limited number of Gaussians, in order to prevent the variance from being too large, a correction coefficient b is set to improve the estimation accuracy. The optimized covariance is:

[0137]

[0138] Where a and b are correction coefficients, Σ(k), Q i (k) is the optimized covariance, original covariance and estimated SOH value of the i-th particle at time k. Indicates the difference between the estimated capacity and the actual capacity.

[0139] In a specific embodiment, the RUL estimation based on the probability density function of the long-short time balance includes:

[0140] In the traditional particle filter estimation method based on the empirical degradation model, there is usually an imbalance between long-term prediction and short-term prediction, that is, the short-term capacity estimation is more accurate, but the long-term prediction is unstable and has serious deviations and extremely poor accuracy. This is because when future observation data is unavailable, the state quantity cannot be updated dynamically, and the current state quantity can only be used to estimate all subsequent capacity values ​​to determine the RUL. The short-term capacity estimation processed by EUMG-PF (Enhanced Unit Mixture Gaussian) is extremely accurate, but there is also the problem of serious deviation in long-term capacity estimation, especially RUL estimation. To this end, this section determines the battery degradation cutoff capacity Q by statistically analyzing a large number of battery degradation data sets of the same type. EOL The value range of EOL is [k min ,k max ]. In this paper, we select the weighted maximum Gaussian component of the mixed Gaussian probability density function calculated by EUMG-PF to obtain N sample points w k =1 / N, considering the long-term capacity, especially RUL, the state quantity mean μ is concerned g , this section restricts the covariance to Σ g =I, construct constraint equation to find the optimal state quantity mean μ g

[0141]

[0142] in The term represents the generation of N sample points The most likely Gaussian distribution, the larger the term, the more likely the current state μ g Generate N sample points The higher the probability; The term represents the estimated state quantity μ g The mean square error between the calculated historical capacity prediction and the actual capacity value is small, indicating that the obtained state quantity μ g Comply with the degradation law of the battery; constraints 1 and 2 ensure that the state quantity μ g It is estimated that the EOL is within the range given for big data statistics.

[0143] On the other hand, embodiment 1 of the present invention also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of a lithium-ion battery monitoring method based on an improved particle filter are implemented.

[0144] Embodiment 2:

[0145] In order to better estimate the capacity and RUL of lithium-ion batteries, the data set of this embodiment 2 comes from NASA and Tianneng Investment Group, and the type is 18650 lithium-ion battery. The relevant parameters are shown in Table 1. Both types of data sets are charged in a constant current mode of 1.5A during the charging phase until the battery voltage reaches 4.2V, and then switched to constant voltage discharge until the current drops to 20mA, and charging stops. During the discharge phase, the current is discharged at a constant current of 1C until the operating voltage drops to the minimum cut-off voltage, and the discharge stops.

[0146] Table 1 Battery Type

[0147]

[0148] In a specific embodiment, the results and analysis include the following:

[0149] Based on the mixed Gaussian particle filtering method of long-short time balance, in order to verify the effectiveness of the ERGMM-PF-RUL algorithm in this embodiment 2, the ERGMM-PF-RUL algorithm, the ERGMM-PF algorithm and the PF algorithm are used to compare the RUL prediction. This embodiment 2 uses two data sets, NASA05 and NASA07, for testing, and sets the number of algorithm particles N p is 1000, the maximum number of iterations M is 1000, and the threshold ε th is 0.001, and EOM is set to k = 90, 89, 104. Battery degradation cutoff capacity Q of NASA05 dataset EOL is 70%, and the EOL value range is [160,162]; the battery degradation cutoff capacity Q of the NASA07 dataset EOL is 75%, and the range of EOL is [148,157]. RMSE, MAE and RUL are selected to evaluate the performance of the algorithm. The experimental results of the dataset NASA05 are shown in Figure 4(a)-Figure 4(c) As shown in Table 2, the experimental results of the NASA07 dataset are as follows Figure 5(a)-Figure 5(c) As shown in Table 3. From the chart, we can see that the traditional PF algorithm occasionally has serious deviations and extremely poor accuracy, such as Figure 4(a) and 5(c) ; The EUMG-PF algorithm has already greatly improved this situation, but the estimation accuracy is still low; the ERGMM-PF-RUL algorithm statistically analyzes a large number of battery degradation data sets of the same type, constructs constraint equations to find the optimal state quantity, and improves the estimation accuracy of RUL while ensuring the balance between long-term and short-term predictions. The performance of the ERGMM-PF-RUL algorithm is better than that of PF and ERGMM-PF.

[0150] The RMSE and MAE performance of EUMG-PF algorithm are improved compared with PF.

[0151] The experimental results comparison with two existing technologies are added in Tables 2 and 3. The algorithm of Example 2 performs well in the RMSE index, which is better than the PF-RT Model algorithm and better than the FA-ARIMAX algorithm overall. The performance of RUL estimation on the NASA05 dataset is also better than PF-RTModel and FA-ARIMAX, and the performance of the algorithm of Example 2 on the NASA07 dataset is slightly inferior. The mixed Gaussian particle filtering method based on long-short time balance in Example 2 is generally better than the PF-RTModel and FA-ARIMAX algorithms.

[0152] Table 2 Comparison of indicators of 5 particle filtering methods (1)

[0153] algorithm <![CDATA[RMSE(*10 -2 )]]> <![CDATA[MAE(*10 -2 )]]> <![CDATA[E RUL (k=90)]]> <![CDATA[E RUL (k=98)]]> <![CDATA[E RUL (k=104)]]> PF 0.849 0.562 129 28 48 EUMG-PF 0.658 0.501 22 15 8 EUMG-PF-RUL 0.658 0.501 2 2 0 PF-RTModel

[15] 1.89 - 5 2 0 FA-ARIMAX

[19] 1.00 - 5 4 1

[0154] Table 3 Comparison of indicators of 5 particle filtering methods (2)

[0155] algorithm <![CDATA[RMSE(*10 -2 )]]> <![CDATA[MAE(*10 -2 )]]> <![CDATA[E RUL (k=90)]]> <![CDATA[E RUL (k=98)]]> <![CDATA[E RUL (k=104)]]> PF 0.712 0.429 43 28 - EUMG-PF 0.547 0.423 26 24 9 EUMG-PF-RUL 0.547 0.423 12 6 3 PF-RTModel

[15] 2.7 - 3 1 2 FA-ARIMAX

[19] 0.53 -

[0156] Specifically, in this embodiment 2, considering the large capacity fluctuation and poor accuracy of remaining life prediction under multiple battery working conditions, a method for predicting the capacity and remaining life of lithium-ion batteries based on improved particle filtering is proposed, which can achieve accurate and effective monitoring of lithium-ion batteries. It can be seen from the above embodiments that:

[0157] 1) In view of the traditional particle filter's assumption that each component of the multivariate state variable is independently Gaussian distributed and that resampling has problems such as particle degradation, a mixed joint Gaussian function is introduced to approximate the true probability density function of the state variable, and a multivariate joint Gaussian distribution is used to obtain a state variable distribution that is closer to the real state. Experiments show that this method performs better than the UG-PF, IG-PF and PF methods.

[0158] 2) There is a significant difference in the difference between adjacent SOH values ​​of lithium batteries at different moments, and there is still a large deviation when probability density sampling estimates the capacity at the next moment. In order to reduce this deviation, this embodiment 2 uses the second-order difference information of capacity and the difference between the estimated capacity and the actual capacity to propose a method of adaptively scaling the covariance. Experiments show that this method performs better than the UMG-PF method.

[0159] 3) In view of the problem that the traditional particle filter has long-term unstable prediction, serious deviation and extremely poor accuracy, this embodiment 2 determines the value range of the battery degradation cutoff capacity as [k min ,k max ], the constraint equation is used to find the optimal state quantity mean μ g, which improves the accuracy of remaining life prediction. The experimental data is better than the traditional PF method and achieves a balance between short-term capacity and long-term remaining life prediction performance.

[0160] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0161] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the present embodiments may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown in the present embodiment, but will conform to the widest range consistent with the principles and novel features disclosed in the present embodiment.

Claims

1. A lithium-ion battery monitoring method based on improved particle filtering, characterized in that: include: Constructing a capacity degradation model of a lithium-ion battery, wherein the capacity degradation model of a lithium-ion battery includes a state equation and an observation equation; Use the Bayesian formula to infer the state equation and observation equation to obtain the posterior state probability density function; Monte Carlo sampling is introduced to estimate the posterior probability density function using particles and particle weights; The joint probability density function of the mixed Gaussian is introduced to approximate the true probability density function of the state quantity, particle correction is performed, low-weight particles are removed, weights are normalized, and the mean, covariance and weight of the Gaussian distribution are updated and calculated using the EM algorithm; The probability density function of the state quantity is estimated and sampled based on the enhanced joint mixed Gaussian EUMG-PF model; Adaptively adjust the covariance using the second-order difference information of capacity and capacity difference; Obtain a battery degradation data set, determine the battery degradation cutoff capacity range based on the battery degradation data set, select the Gaussian component with the maximum weight for sampling, and construct a constraint equation to obtain the optimal state quantity mean; Output the optimal state quantity mean to realize particle filter lithium-ion battery monitoring.

2. The improved particle filter lithium ion battery monitoring method according to claim 1, characterized in that: The expression of the state equation is: where x k =[a k b k c k d k ] T represents the system state at time k; represents the process noise of the multidimensional mixed Gaussian distribution at time k-1, Φ represents the normal distribution, M represents the number of mixed Gaussian distributions in the state equation of the system, m represents the mth component in the mixed Gaussian distribution, Φ(μ,∑) represents the Gaussian distribution with mean μ and covariance ∑, represents the mean of the mth Gaussian distribution at time k-1; represents the covariance matrix of the mth Gaussian distribution at time k-1, π m Represents the weight coefficient of the mth Gaussian distribution component; The expression of the observation equation is: Where Q k Represents x k The observed value at time k, a k and c k is a parameter related to the internal resistance of lithium batteries, b k and d k It is a parameter related to the degradation rate of lithium batteries.

3. The improved particle filter based lithium ion battery monitoring method according to claim 1, characterized in that: The Bayesian formula is used to infer the state equation and observation equation to obtain the posterior state probability density function, including: Where p(x 0:k |Q 1:k-1 ) is the prior probability density function at time k, p(x k ,x k-1 ) is the posterior probability model of the evolution of the system state equation, p(Q k |x k ) is the likelihood function, and ∝ indicates a positive correlation.

4. The improved particle filter-based lithium-ion battery monitoring method according to claim 1, characterized in that: Monte Carlo sampling is introduced to estimate the posterior probability density function using particles and their weights, including: Introducing Monte Carlo sampling, using a set of random particle samples The weight associated with Estimate the posterior probability density function p(x k |Q 1:k-1 ), introduce importance sampling to calculate the weight of the particle, and then normalize the weight. The formula is: In the formula Indicates v k The probability density function form of ; Use the weighted particles to estimate the state x k : In the formula, Q k represents the observed value at time k, is the observation model function, which represents the system state After observing the prediction results of model h.

5. The improved particle filter-based lithium-ion battery monitoring method according to claim 1, characterized in that: The joint probability density function of the mixed Gaussian is expressed as: Where G represents the number of Gaussian probability functions, ω g is the weight of the g-th Gaussian component, μ g is the mean vector of the g-th Gaussian component D*D, Σ g is the D*D dimensional covariance matrix of the g-th Gaussian component, |Σ g |isΣ g The determinant of , x represents the observation value, and T is the transpose of the matrix.

6. The improved particle filter-based lithium-ion battery monitoring method according to claim 5, characterized in that: Use the EM algorithm to update and calculate the mean, covariance, and weight of the Gaussian distribution, including: right The parameters are initialized respectively, and the best mixed Gaussian function is obtained through E-step and M-step iterations, where E-step: In the formula, x i is the i-th state data sample point, ω g is the weight of the g-th Gaussian component, μ g is the mean vector of the g-th Gaussian component D*D, Σ g is the D*D dimensional covariance matrix of the g-th Gaussian component, |Σ g |isΣ g The determinant of x represents the observation value, T is the transpose of the matrix, r ig Represents a given sample x i After that, it comes from the probability generated by the g-th Gaussian distribution; M-step: Where N g Represents the sum of the probabilities that each sample belongs to the g-th Gaussian component; Compute the log-likelihood function: The E-step and M-step are iterated repeatedly until the convergence ε is less than the given threshold ε th , update the mean, covariance and weight of the Gaussian distribution.

7. The lithium-ion battery monitoring method based on improved particle filtering according to claim 1, characterized in that: The probability density function of the state quantity is estimated and sampled based on the enhanced joint mixed Gaussian EUMG-PF model, including: Let the capacity parameter be defined as: ΔQ(k)=Q(k)-Q(k-1); Δ2Q(k)=ΔQ(k)-ΔQ(k-1); Where k is the number of cycles, Q(k) is the battery capacity at time k, ΔQ(k) is the first-order difference of the battery capacity at time k, and Δ2Q(k) is the second-order difference of the battery capacity at time k.

8. The improved particle filter-based lithium-ion battery monitoring method according to claim 1, characterized in that: The covariance is adaptively adjusted using the second-order difference information of capacity and the capacity difference. The adjusted covariance is: Where a and b are correction coefficients, is the optimized covariance at time k, Σ(k) is the original covariance at time k, Q i (k) is the estimated SOH value of the i-th particle at time k; Indicates the difference between the estimated capacity and the actual capacity.

9. The lithium-ion battery monitoring method based on improved particle filtering according to claim 1, characterized in that: Obtain a large number of battery degradation data sets, determine the battery degradation cutoff capacity range based on a large number of battery degradation data, select the Gaussian component with the largest weight for sampling, and construct a constraint equation to obtain the optimal state quantity mean, including: in The term represents the generation of N sample points Gaussian distribution, N is the total number of sample points; The term represents the estimated state quantity μ g The calculated mean square error between the historical capacity forecast and the actual capacity value, λ1 is the correction weight.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of a lithium-ion battery monitoring method based on an improved particle filter as described in any one of claims 1 to 9 are implemented.

Images