Lithium battery SOH universal prediction method under false data injection background

Abstract

The invention provides a lithium battery SOH universal prediction method under a false data injection background, belongs to the technical field of lithium battery management, and solves the problems that BMS measurement data is easily attacked by a network and SOH prediction universality is poor under an actual working condition, and the method specifically comprises the steps of training a GPR model and an FDIA detection model by using lithium battery charging data in a normal operation state, establishing a data driving model based on the trained GPR model; detecting and correcting the charging data of the lithium battery by using the FDIA detection model to obtain sub-normal charging data of the lithium battery; extracting health features of the sub-normal charging data of the lithium battery through a dynamic health feature extraction method; and inputting the health features into the trained data driving model to obtain a predicted value of the SOH at the initial running stage of the lithium battery, fitting the predicted value output by the data driving model by adopting an empirical model to obtain an overall degradation trend of the lithium battery, and fusing a UPF filter to realize online updating of the SOH of the lithium battery and parameters of the empirical model.

Description

Technical Field

[0001] The invention relates to a universal prediction method for lithium battery SOH under the background of false data injection, belonging to the technical field of lithium battery management. Background Art

[0002] As a green and environmentally friendly energy product, lithium batteries have been widely used in electric vehicles and power system energy storage power stations. The battery management system (BMS) can evaluate the status of lithium batteries to ensure the long-term safe and stable operation of lithium batteries. Among them, the accurate estimation of the health state (SOH) of lithium batteries is the key to BMS operation and maintenance, and it is also an important link to ensure the power performance of electric vehicles or the safety performance of energy storage power stations.

[0003] At present, the research on SOH estimation of lithium batteries mainly includes model-based prediction methods, data-driven prediction methods and fusion technology prediction methods. Model-based methods mainly fit the external characteristics of lithium batteries through electrochemical, equivalent circuit or empirical models, and then use filtering methods to identify parameters to achieve SOH estimation. However, these models are usually limited to specific battery systems, and their prediction accuracy depends on the accuracy of the model. Data-driven methods use the external characteristics of lithium batteries as health characteristics (HF), and use machine learning algorithms such as vector machines, Gaussian process regression, neural networks, etc. to construct a nonlinear mapping relationship between HF and SOH to achieve SOH estimation. Data-driven algorithms perform well in the accuracy and real-time performance of lithium battery SOH estimation by learning from a large amount of data and optimizing their own performance. However, it also faces many challenges, such as poor interpretability, easy overfitting and high computational complexity. Therefore, data-driven models are usually used in combination with empirical models or filters to address these problems and improve the stability and reliability of the model. Fusion technology is a method of enhancing performance or improving efficiency by integrating two or more different technical tools. It can be roughly divided into two types. One is the fusion of model-driven, filtering methods and data-driven methods. As mentioned above, it not only reduces the burden on lithium battery monitoring equipment, but also ensures the accuracy of lithium battery SOH prediction; the other is the fusion of multiple data-driven methods. Although this method has high prediction accuracy, it also brings problems such as high cost, complex calculation, and high dependence on training data.

[0004] Although there are many methods for predicting the SOH of lithium batteries, existing research results are often limited to specific environmental conditions in the laboratory, such as constant temperature, standardized application conditions, etc., ignoring the parameter inconsistency of lithium batteries in actual applications and the variability of operating temperature and charging state. The versatility of SOH prediction for lithium batteries under multiple working conditions is poor.

[0005] In addition, as an electric information-physical system, BMS is facing the threat of cyber attacks. The most common one is the false data injection attack (FDIA), which can inject false data into the battery management unit and vector measurement unit of the BMS, thereby interfering with the estimation accuracy of the lithium battery SOH, causing the BMS to misjudge and make wrong decisions, which in turn causes safety hazards. It can be seen that the detection and defense of cyber attacks against BMS has become a major issue facing the development of lithium battery SOH prediction technology.

[0006] Therefore, in view of the problems that BMS measurement data is vulnerable to network attacks and the poor universality of SOH prediction under actual working conditions, the present invention constructs a Gaussian process regression-empirical model-unscented particle filter (GPR-EDM-UPF) framework with false data detection and correction functions, and proposes a generalized prediction method for the health status of lithium batteries based on dynamic peak area characteristics and GPR-EDM-UPF framework. It can effectively improve the universal prediction accuracy of lithium battery SOH under multivariate working conditions such as different rates and wide temperature ranges, and enhance the reliability of battery management systems in the fields of electric vehicles (EVs) and energy storage power stations (ES) under the background of FDIA. Summary of the invention

[0007] The present invention aims to solve the problem that BMS measurement data is vulnerable to network attacks and the versatility of lithium battery SOH prediction under actual working conditions is poor.

[0008] In order to solve the above problems, the present invention proposes a universal prediction method for the health status of lithium batteries based on dynamic peak area characteristics and a GPR-EDM-UPF framework with false data detection and correction functions, comprising the following steps:

[0009] Step 1: Obtain a lithium battery charging data set under normal operating conditions;

[0010] Step 2: Use the lithium battery charging data under normal operating conditions to train the GPR model and the FDIA detection model, obtain the error threshold and the correlation threshold based on the trained FDIA detection model, and establish a data-driven model based on the trained GPR model;

[0011] Step 3: Use the trained FDIA detection model to detect and correct the real-time charging data of the lithium battery to obtain the sub-normal charging data of the lithium battery;

[0012] Step 4: Extract the health features of the subnormal charging data of the lithium battery by using the dynamic health feature extraction method;

[0013] Step 5: Input the health features of the sub-normal charging data of the lithium battery into the trained data-driven model to obtain the predicted value of the SOH of the lithium battery in the initial operation stage. Use the empirical model to fit the predicted value output by the data-driven model to obtain the overall degradation trend of the lithium battery, and integrate the UPF filter to realize the online update of the lithium battery SOH and empirical model parameters.

[0014] Preferably, the step of acquiring the GPR model in step 2 includes:

[0015] Step 2.1: Get the mapping vector of the test set, use the square exponential as the covariance function, and get the GPR model containing noise;

[0016] Step 2.2: Based on the noisy GPR model, obtain the output observations interfered by white noise, calculate the prior distribution of the output observations interfered by white noise and the posterior distribution of Y* in the test set, and use the maximum likelihood estimation to solve the hyperparameter set θ = [δ f ,l,δ n ], and obtain the GPR model;

[0017] The Gaussian process expression of the GPR model is:

[0018] f(X)~GP[m(X),k(X,X * )](1);

[0019] In formula (1), X represents the input sample of the training set, X * represents the input sample of the test set, f(X) is the mapping vector of X; m(X) represents the mean function, and for symmetry, m(X) is usually set to 0, k(X,X * ) represents the covariance function;

[0020] Covariance function k(X,X * ) is:

[0021]

[0022] In formula (2), is the signal variance; l is the characteristic length;

[0023] The expression of the GPR model including noise is:

[0024]

[0025] In formula (3), Y is the output observation value disturbed by white noise; ε is the noise in the actual situation; n is the number of input samples;

[0026] The expression for the prior distribution of output observations subject to white noise is:

[0027]

[0028] In formula (4), I e is the e-dimensional identity matrix;

[0029] The expression of the joint prior distribution of Y in the training set and Y* in the test set is:

[0030]

[0031] The expression of the posterior distribution of Y* is:

[0032] Y * |X,Y,X * ~N(μ,Σ)(6);

[0033] The expressions of the test set mean matrix μ and covariance matrix Σ are:

[0034]

[0035] Preferably, the step of obtaining the subnormal charging data of the lithium battery in step 3 includes:

[0036] Step 3.1: Use the LSTM detection model to perform error analysis on the actual voltage value and the LSTM predicted value. If the error is less than the error threshold, the defense system will not act. If the error is greater than or equal to the error threshold, proceed to step 3.2.

[0037] Step 3.2: Take the actual voltage value as variable X and the LSTM predicted value as variable Y, and use the MIC checker to analyze the correlation between the actual voltage value and the LSTM predicted value. If the correlation is greater than the correlation threshold, the defense system will not take action.

[0038] Step 3.3: If the correlation is less than or equal to the correlation threshold, the defense system is activated, and the false voltage actual value is replaced by the LSTM predicted value to obtain the sub-normal charging data of the lithium battery;

[0039] The calculation formula for the correlation between the actual voltage value and the LSTM predicted value is:

[0040]

[0041] Preferably, the LSTM detection model is an LSTM neural network structure, including three gating mechanisms: a forget gate, an input gate, and an output gate;

[0042] The expressions of forget gate, input gate and output gate are:

[0043] f t =σ(W f [h t-1 ,x t]+b f )(10);

[0044]

[0045] o t =σ(W o [h t-1 ,x t ]+b o )(13);

[0046] h t =o t tanh(c t )(14);

[0047] In formula (10) - formula (14), x t is the input vector; h t-1 h is the hidden layer output vector at the previous moment; t is the hidden layer output vector at the current moment; f t is the output vector of the forget gate; i t Output vector for the input gate; o t Output vector for the output gate; C t-1 is the state vector of the memory unit at the previous moment; C t is the memory unit state vector at the current moment; C t is the candidate update value vector of the memory unit at the current moment; W f is the weight matrix of the forget gate; W i is the weight matrix of the input gate; W c is the weight matrix of the candidate update value layer of the memory unit; W o is the weight matrix of the output gate; b f is the bias vector of the forget gate; b i is the bias vector of the input gate; b c is the bias vector of the candidate update value layer of the memory unit; b o is the bias vector of the output gate; σ() is the sigmoid activation function; tanh is the hyperbolic tangent activation function.

[0048] Preferably, the step of extracting the health features of the sub-normal charging data of the lithium battery by the dynamic health feature extraction method in step 4 includes:

[0049] Step 4.1: Calculate the current charge IC based on the subnormal charging data of the lithium battery i ,IC i-1 =dQ / dU: and record the current voltage U i ;

[0050] Step 4.2: If IC i ≥ICi-1 , then repeat step 4.1 until IC i <IC i-1 ;

[0051] Step 4.3: Get the peak voltage U peak , U peak =U i-1 , select the peak width λ at half peak height in the low voltage direction;

[0052] Step 4.4: According to the peak voltage U peak and peak width λ define the voltage interval [U peak -λ,U peak +λ], rounding the two ends of the voltage interval to get the peak interval [A,B];

[0053] Step 4.5: Calculate the peak interval area S, and use the peak interval area S as a health feature of the sub-normal charging data of the lithium battery.

[0054] Preferably, the step of implementing online update of lithium battery SOH and empirical model parameters in step 5 includes:

[0055] Step 5.1: Use the empirical model to fit the predicted value output by the data-driven model to obtain the global degradation trend of the lithium battery;

[0056] Step 5.2: The empirical model is used as the measurement equation of the UPF filter, the model parameters of the empirical model are the state equation of the UPF filter, and the predicted value output by the data-driven model is the observation value of the UPF filter. When new observation data is obtained at time k, the UPF filter is used to convert the empirical model parameters x at the previous moment into k-1 Update to x k , and then x k Iterate back to the empirical model for known quantities to achieve online update of lithium battery SOH prediction and empirical model, and obtain the optimal estimate of lithium battery SOH at the current moment;

[0057] The expression of the global degradation trend of lithium batteries is:

[0058] SOH n =a×e b×n +c×e d×n (15);

[0059] In formula (15), SOH n is the SOH at the nth cycle, a, b, c, d are the SOH values ​​predicted by the fitted GPR i The obtained parameters are i=1,2,3,..,SP, where SP is the prediction starting point.

[0060] Preferably, in step 5.2, a UPF filter is used to convert the empirical model parameter x at the previous moment intok-1 Update to x k The steps include:

[0061] Step 5.2.1: Establish the state equation of the UPF filter through the empirical model parameters a, b, c, d, and use the state equation of the UPF filter to establish the UPF measurement equation to obtain the state space equation;

[0062] Step 5.2.2: When k = 0, according to the initial distribution N = 500, assigning initial weights to each particle in the empirical model parameters

[0063] Step 5.2.3: Perform importance sampling on each particle;

[0064] Step 5.2.4: Update the importance weight of the particle and normalize it;

[0065] Step 5.2.5: Resample the normalized particles, perform state estimation based on the resampled particles, and complete the online update of the empirical model parameters;

[0066] The state space equation is expressed as:

[0067]

[0068] In formula (16), x k =[a k ,b k ,c k ,d k ] is the exponential model parameter, y k+1 is the observed value, i.e., the GPR predicted value, the subscripts k and k+1 represent the time; w k and v k represents independent Gaussian white noise;

[0069] The expression of the initial distribution of each particle in the empirical model parameters is:

[0070]

[0071] In formula (17), is the covariance matrix, is the Sigma point matrix;

[0072] The expression for updating the importance weight of a particle and normalizing it is:

[0073]

[0074] The expression for resampling the normalized particles is:

[0075]

[0076] In formula (19), N t is the effective particle number threshold;

[0077] The expression of state estimation is:

[0078]

[0079] Preferably, the step of performing importance sampling on each particle in step 5.2.3 comprises:

[0080] Step 5.2.3.1: For k = 1:T, i = 1:N, use the UKF filtering method to generate the importance density function of each particle;

[0081] Step 5.2.3.2: Calculate the Sigma points and weights of the UPF filter;

[0082] Step 5.2.3.3: Use Sigma point calculation to get the predicted value and covariance;

[0083] Step 5.2.3.4: Obtain the observed value based on the predicted value and covariance calculation and weight the observed value;

[0084] Step 5.2.3.5: Calculate the covariance and cross-covariance based on the weighted observations;

[0085] Step 5.2.3.6: Calculate the Kalman gain K based on the covariance and cross-covariance, and update the particle state and covariance;

[0086] Step 5.2.7: Sample the important density function to generate particles;

[0087] The calculation formula for Sigma points and weights is:

[0088]

[0089] In formula (21), is the covariance of the weight, L is the state dimension, β = 2, is the proportional correction coefficient, λ=α 2 (L+κ)-L, α determines the discreteness of the Sigma point, and κ is the secondary sampling factor;

[0090] The formula for calculating the predicted value and covariance is:

[0091]

[0092] The calculation formula for observations and observation weights is:

[0093]

[0094] The formulas for calculating covariance and cross-covariance are:

[0095]

[0096] The expressions of Kalman gain K and updated particle state and covariance are:

[0097]

[0098] The expression for sampling important density functions is:

[0099]

[0100] The beneficial effects of the present invention are:

[0101] (1) The present invention establishes a long short-term neural network-maximum information coefficient (LSTM-MIC) detector, which identifies potential false data in the battery management system measurement data by analyzing the error and correlation between the LSTM estimated value and the actual voltage data. The detection method can monitor the operation status of the battery management system in real time, detect the existence of false data in time, and correct the contaminated measurement values, effectively reducing the interference of FDIA on the accuracy of lithium battery health status estimation, and ensuring the stability and safety of the battery management system.

[0102] (2) The present invention proposes a method for defining a peak interval with the peak point of the capacity increment curve as the center and the half-peak width at both ends, and calculates the peak area of ​​the peak interval online through a recursive update method during the charging process. This method significantly reduces the demand for data storage and processing capabilities of embedded systems, allowing efficient health feature extraction to be achieved in resource-limited embedded environments; at the same time, it is suitable for lithium battery status assessment under multivariate working conditions, laying a solid foundation for the universal prediction of lithium battery SOH.

[0103] (3) The present invention constructs a fusion prediction model for lithium battery SOH that combines a data-driven model with an empirical model using a filter. The method uses a Gaussian process regression model to predict the SOH of the lithium battery in the initial stage of operation, and uses the Gaussian prediction value to establish a double exponential empirical degradation model (EDM) to obtain the global degradation trend of the lithium battery capacity. Finally, an unscented particle filter (UPF) is introduced to iteratively update the empirical model parameters and the lithium battery SOH prediction value. The verification results based on lithium battery data of different models, temperatures and charging rates show that the maximum absolute error of the prediction method proposed by the present invention is less than 3%. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 A flowchart of a universal prediction method for lithium battery SOH under the background of false data injection provided by the present invention;

[0105] Figure 2 A flow chart of false data detection and defense based on LSTM-MIC provided by the present invention;

[0106] Figure 3 This is a flow chart for extracting dynamic health features provided by the present invention. DETAILED DESCRIPTION

[0107] Combination Figure 1-3 This embodiment is described as follows. Figure 1 As shown, the steps of a universal prediction method for lithium battery SOH under the background of false data injection described in this embodiment include:

[0108] S1: Offline period: Use the cycle charging data of multiple types of lithium batteries throughout their life cycle to train the data-driven model and FDIA detection model, establish a data-driven model based on the trained GPR model, and detect and defend against false data based on LSTM-MIC;

[0109] S101: Long Short-Term Memory Neural Networks;

[0110] LSTM neural network is an advanced form of recurrent neural network (RNN). It controls the storage information of time series data by adding storage units in the hidden layer. In order to overcome the problem of gradient disappearance in RNN, this implementation introduces a gating mechanism in LSTM neural network. The gating mechanism includes: forget gate, input gate and output gate, and inserts state units to regulate network operation. The following formula can describe the network structure of LSTM:

[0111] f t =σ(W f [h t-1 ,x t ]+b f )(1);

[0112]

[0113] o t =σ(W o [h t-1 ,x t ]+b o )(4);

[0114] h t =o t tanh(c t )(5);

[0115] In formula (1) to formula (5), x t is the input vector; h t-1h is the hidden layer output vector at the previous moment; t is the hidden layer output vector at the current moment; f t is the output vector of the forget gate; i t Output vector for the input gate; o t Output vector for the output gate; C t-1 is the state vector of the memory unit at the previous moment; C t is the memory unit state vector at the current moment; C t is the candidate update value vector of the memory unit at the current moment; W f is the weight matrix of the forget gate; W i is the weight matrix of the input gate; W c is the weight matrix of the candidate update value layer of the memory unit; W o is the weight matrix of the output gate; b f is the bias vector of the forget gate; b i is the bias vector of the input gate; b c is the bias vector of the candidate update value layer of the memory unit; b o is the bias vector of the output gate; σ() is the sigmoid activation function; tanh is the hyperbolic tangent activation function;

[0116] S102: Obtaining a GPR model;

[0117] S10201: Obtain the mapping vector of the test set, use the square exponential as the covariance function, and obtain the GPR model containing noise;

[0118] S10202: Based on the GPR model containing noise, the output observations interfered by white noise are obtained, the prior distribution of the output observations interfered by white noise and the posterior distribution of Y* in the test set are calculated, and the maximum likelihood estimation is used to solve the hyperparameter set θ = [δ f ,l,δ n ], and obtain the GPR model;

[0119] Gaussian process regression is a special random process with Bayesian inference principle and uncertainty prediction capability. It has been widely used in lithium battery health status assessment. Given a sample D = {(xi,yi)}, the Gaussian process can be described as the Gaussian process expression of the GPR model:

[0120] f(X)~GP[m(X),k(X,X * )](6);

[0121] In formula (1), X represents the input sample of the training set, X *represents the input sample of the test set, f(X) is the mapping vector of X; m(X) represents the mean function, and for symmetry, m(X) is usually set to 0, k(X,X * ) represents the covariance function;

[0122] Commonly used covariance functions include square exponential function, Marton kernel function, and rational quadratic kernel function. This paper uses square exponential function as covariance function. The covariance function k(X,X * ) is:

[0123]

[0124] In formula (2), is the signal variance; l is the characteristic length;

[0125] The expression of the GPR model including noise is:

[0126]

[0127] In formula (3), Y is the output observation value disturbed by white noise; ε is the noise in the actual situation; n is the number of input samples;

[0128] The expression for the prior distribution of output observations subject to white noise is:

[0129]

[0130] In formula (4), I e is the e-dimensional identity matrix;

[0131] The expression of the joint prior distribution of Y in the training set and Y* in the test set is:

[0132]

[0133] The expression of the posterior distribution of Y* is:

[0134] Y * |X,Y,X * ~N(μ,Σ) (11);

[0135] The expressions of the test set mean matrix μ and covariance matrix Σ are:

[0136]

[0137] S103: Maximum information coefficient

[0138] The maximum information coefficient (MIC) is a method based on mutual information that compares the strength of data association horizontally and vertically. It can capture a wide range of associations. When the load changes, the voltage of the microgrid will also fluctuate. How to distinguish whether the fluctuation is caused by a false data injection attack is a task that the MIC detection module needs to complete. This implementation uses MIC to analyze the correlation between the actual voltage and the LSTM predicted voltage to determine whether there is a data error. The calculation formulas for variables X and Y are:

[0139]

[0140] The present invention establishes a long short-term neural network-maximum information coefficient (LSTM-MIC) detector to analyze the error and correlation between the LSTM estimated value and the actual voltage data to identify whether there is false data in the network attack. The detection method can timely identify false data and correct the contaminated measurement value, effectively reduce the impact of FDIA on the accuracy of lithium battery health status estimation, and maintain the stability of the battery management system.

[0141] S2: Prediction period: Use the FDIA detection model to detect and correct the lithium battery charging data to obtain the sub-normal charging data of the lithium battery; then use the dynamic health feature extraction method to extract the IC peak area from the initial moment to the k moment of the sub-normal IC curve, and input it into the trained data-driven model to obtain the predicted value of the SOH in the initial operation of the lithium battery;

[0142] S201: pre-processing the lithium battery charging data to obtain sub-normal charging data of the lithium battery;

[0143] In order to detect and prevent false data in the lithium battery management system, this embodiment proposes a detection mechanism based on LSTM combined with MIC. The process is as follows: Figure 2 As shown in the figure, a lithium battery charging data set obtained in normal operation under offline mode is selected to train the LSTM model and the MIC model to obtain appropriate error thresholds and correlation thresholds for subsequent FDIA inspection. The specific detection and defense steps are as follows:

[0144] S20101: Primary detection: Use the LSTM detection model to perform error analysis on the actual voltage value and the LSTM predicted value. If the absolute error is within a reasonable error range, the defense system will not operate; otherwise, proceed to S10302;

[0145] S20102: Secondary detection: Use the MIC checker to analyze the correlation between the actual voltage value and the LSTM predicted value. If the correlation is greater than the correlation threshold, the defense system will not act; otherwise, proceed to S10303;

[0146] S20103: Defense system: Replace the false actual voltage value with the LSTM predicted value to eliminate the influence of false data and obtain the sub-normal charging data of the lithium battery.

[0147] S202: extracting health characteristics of subnormal charging data of lithium batteries;

[0148] The extraction of health features is the key to establishing a data-driven model. Considering the convenience, rationality and practicality of comprehensive extraction, this project uses the incremental capacity analysis (ICA) method to extract features. This implementation method takes into account the influence of IC peak ranges such as actual ambient temperature, charging rate, and charging mode during lithium battery charging, and uses the recursive update method to determine the peak range in real time during the charging process, and extracts the dynamic peak area as a health feature. The specific process is as follows: Figure 3 As shown, the specific steps are as follows:

[0149] S20201: Calculate the current charge IC based on the subnormal charging data of the lithium battery i ,IC i-1 =dQ / dU: and record the current voltage Ui;

[0150] S20202: If IC i ≥IC i-1 , then repeat step 4.1 until IC i <IC i-1 ;

[0151] S20203: Get peak voltage U peak , U peak =U i-1 , select the peak width λ at half peak height in the low voltage direction;

[0152] S20204: According to the peak voltage U peak and peak width λ define the voltage interval [U peak -λ,U peak +λ], rounding the two ends of the voltage interval to get the peak interval [A,B];

[0153] S20205: Calculate the peak interval area S, and use the peak interval area S as a health feature of the sub-normal charging data of the lithium battery.

[0154] Compared with the traditional method of storing the complete charging curve and then performing offline calculation of the peak area within a fixed voltage range, the method proposed in this paper only needs to store the IC curve within a certain range near the peak point, which greatly saves the storage space of the embedded system and improves the efficiency of data retrieval. It also fully considers the actual situation during the charging process of lithium batteries and realizes the universal extraction of health characteristics.

[0155] S3: Update period: The empirical model is used to fit the data-driven prediction value to provide the overall degradation trend of the lithium battery. At the same time, the fusion filter realizes the online update of the lithium battery SOH and empirical model parameters;

[0156] S301: Using the empirical model to fit the predicted value output by the data-driven model to obtain the global degradation trend of the lithium battery;

[0157] There are many commonly used empirical degradation models, including polynomial models, exponential models, Fourier models, integrated models, etc. Among them, the double exponential empirical model has been widely used in many fields due to its simple structure and excellent fitting performance. Its formula is shown as follows:

[0158] SOH n =a×e b×n +c×e d×n (15);

[0159] In formula (15), SOH n is the SOH at the nth cycle, a, b, c, d are the SOH values ​​predicted by the fitted GPR i The obtained parameters are i=1,2,3,..,SP, where SP is the prediction starting point.

[0160] S302: The empirical model is used as the measurement equation of the UPF filter, the model parameters of the empirical model are the state equation of the UPF filter, the predicted value output by the data-driven model is the observation value of the UPF filter, and when new observation data is obtained at time k, the UPF filter is used to convert the empirical model parameters x at the previous moment into k-1 Update to x k , and then x k Iterate back to the empirical model for known quantities to achieve online update of lithium battery SOH prediction and empirical model, and obtain the optimal estimate of lithium battery SOH at the current moment;

[0161] S30201: In order to solve the problem of unchanged empirical model parameters and improve the prediction accuracy of the empirical model for lithium battery SOH, this implementation adopts UPF filter to update and correct the double exponential empirical model parameters and lithium battery SOH. First, the UPF state equation is established through the lithium battery capacity degradation empirical model parameters a, b, c, d, and then the UPF measurement equation is established using the GPR model prediction value to obtain the state space equation:

[0162] The state space equation is expressed as:

[0163]

[0164] In formula (16), x k =[a k ,b k ,ck ,d k ] is the exponential empirical model parameter, y k+1 is the observed value, i.e., the GPR predicted value, the subscripts k and k+1 represent the time; w k and v k represents independent Gaussian white noise;

[0165] S30202: The UPF algorithm is based on the Unscented Kalman Filter (UKF). Each iteration process introduces new observation information to correct the particle prediction results at the next moment, which greatly improves the prediction accuracy of nonlinear systems. The specific steps are as follows:

[0166] (1) Initialization

[0167] When k = 0, according to the initial distribution N = 500, each particle has an initial weight

[0168]

[0169] In formula (17), is the covariance matrix, is the Sigma point matrix;

[0170] (2) Importance Sampling

[0171] For k=1:T, i=1:N, the UKF filtering method is used to generate the importance density function of each particle.

[0172] a) Calculate Sigma points and weights

[0173]

[0174] In formula (18), is the covariance of the weight, L is the state dimension, β = 2, is the proportional correction coefficient, λ=α 2 (L+κ)-L, α determines the discreteness of the Sigma point, and κ is the secondary sampling factor;

[0175] b) Use Sigma points to get predicted values ​​and covariance:

[0176]

[0177] c) Obtain the observed values ​​and weight them

[0178]

[0179] d) Calculate autocovariance and cross-covariance

[0180]

[0181] e) Calculate the Kalman gain K and update the particle state and covariance

[0182]

[0183] Sampling important density functions to generate particles

[0184]

[0185] (3) Update weights and initialization

[0186] Calculate particle importance weights and normalize:

[0187]

[0188] (4) Resampling

[0189]

[0190] Where N t is the effective particle number threshold, which is generally set to N / 3.

[0191] (5) State estimation:

[0192]

[0193] In summary, the present invention constructs a lithium battery SOH fusion prediction model that uses a filter to combine a data-driven model with an empirical model. This method uses a Gaussian process regression model to predict the SOH of the lithium battery in the initial working stage, and uses the Gaussian prediction value to establish a double exponential empirical degradation model (EDM) to obtain the global degradation trend of the lithium battery capacity. Finally, an unscented particle filter (UPF) is introduced to iteratively update the empirical model parameters and the lithium battery SOH prediction value.

[0194] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.

Claims

1. A universal prediction method for lithium battery SOH under the background of false data injection, characterized in that: The steps of the universal prediction method for lithium battery SOH under the background of false data injection include: Step 1: Obtain a lithium battery charging data set under normal operating conditions; Step 2: Use the lithium battery charging data under normal operating conditions to train the GPR model and the FDIA detection model, obtain the error threshold and the correlation threshold based on the trained FDIA detection model, and establish a data-driven model based on the trained GPR model; Step 3: Use the trained FDIA detection model to detect and correct the lithium battery charging data to obtain the sub-normal charging data of the lithium battery; Step 4: Extract the health features of the subnormal charging data of the lithium battery by using the dynamic health feature extraction method; Step 5: Input the health features of the sub-normal charging data of the lithium battery into the trained data-driven model to obtain the predicted value of the SOH of the lithium battery in the initial operation stage. Use the empirical model to fit the predicted value output by the data-driven model to obtain the overall degradation trend of the lithium battery, and integrate the UPF filter to realize the online update of the lithium battery SOH and empirical model parameters.

2. According to the method for universal prediction of lithium battery SOH under the background of false data injection in claim 1, it is characterized in that: The steps for obtaining the GPR model in step 2 include: Step 2.1: Get the mapping vector of the test set, use the square exponential as the covariance function, and get the GPR model containing noise; Step 2.2: Based on the noisy GPR model, obtain the output observations interfered by white noise, calculate the prior distribution of the output observations interfered by white noise and the posterior distribution of Y* in the test set, and use the maximum likelihood estimation to solve the hyperparameter set θ = [δ f ,l,δ n ], and obtain the GPR model; The Gaussian process expression of the GPR model is: f(X)~GP[m(X),k(X,X * )](1); In formula (1), X represents the input sample of the training set, X * represents the input sample of the test set, f(X) is the mapping vector of X; m(X) represents the mean function, and for symmetry reasons m(X) is usually set to 0, k(X,X * ) represents the covariance function; Covariance function k(X,X * ) is: In formula (2), is the signal variance; l is the characteristic length; The expression of the GPR model including noise is: In formula (3), Y is the output observation value disturbed by white noise; ε is the noise in the actual situation; n is the number of input samples; The expression for the prior distribution of output observations subject to white noise is: In formula (4), I e is the e-dimensional identity matrix; The expression of the joint prior distribution of Y in the training set and Y* in the test set is: The expression of the posterior distribution of Y* is: Y * |X,Y,X * ~N(μ,Σ)(6); The expressions of the test set mean matrix μ and covariance matrix Σ are:

3. The universal prediction method for lithium battery SOH under the background of false data injection according to claim 1 is characterized in that: The steps of detecting and eliminating false data in step 3 to obtain sub-normal charging data of the lithium battery include: Step 3.1: Use the LSTM detection model to perform error analysis on the actual voltage value and the LSTM predicted value. If the error is less than the error threshold, the defense system will not act. If the error is greater than or equal to the error threshold, proceed to step 3.

2. Step 3.2: Take the actual voltage value as variable X and the LSTM predicted value as variable Y, and use the MIC checker to analyze the correlation between the actual voltage value and the LSTM predicted value. If the correlation is greater than the correlation threshold, the defense system will not take action. Step 3.3: If the correlation is less than or equal to the correlation threshold, the defense system is activated, and the false voltage actual value is replaced by the LSTM predicted value to obtain the sub-normal charging data of the lithium battery; The calculation formula for the correlation between the actual voltage value and the LSTM predicted value is:

4. The universal prediction method for lithium battery SOH under the background of false data injection according to claim 3 is characterized in that: The LSTM detection model is an LSTM neural network structure, including three gating mechanisms: forget gate, input gate and output gate; The expressions of forget gate, input gate and output gate are: f t =σ(W f [h t-1 ,x t ]+b f )(10); the t =σ(W o [h t-1 ,x t ]+b o )(13); h t =o t fishy t )(14); In formula (10) - formula (14), x t is the input vector; h t-1 h is the hidden layer output vector at the previous moment; t is the hidden layer output vector at the current moment; f t is the output vector of the forget gate; i t Output vector for the input gate; o t Output vector for the output gate; C t-1 is the state vector of the memory unit at the previous moment; C t is the memory unit state vector at the current moment; C t is the candidate update value vector of the memory unit at the current moment; W f is the weight matrix of the forget gate; W i is the weight matrix of the input gate; W c is the weight matrix of the candidate update value layer of the memory unit; W o is the weight matrix of the output gate; b f is the bias vector of the forget gate; b i is the bias vector of the input gate; b c is the bias vector of the candidate update value layer of the memory unit; b o is the bias vector of the output gate; σ() is the sigmoid activation function; tanh is the hyperbolic tangent activation function.

5. The universal prediction method for lithium battery SOH under the background of false data injection according to claim 1 is characterized in that: The step of extracting the health features of the sub-normal charging data of the lithium battery by the dynamic health feature extraction method in step 4 includes: Step 4.1: Calculate the current charge IC based on the subnormal charging data of the lithium battery i ,IC i-1 =dQ / dU: and record the current voltage U i ; Step 4.2: If IC i ≥IC i-1 , then repeat step 4.1 until IC i <IC i-1 ; Step 4.3: Get the peak voltage U peak , U peak =U i-1 , select the peak width λ at half peak height in the low voltage direction; Step 4.4: According to the peak voltage U peak and peak width λ define the voltage interval [U peak -λ,U peak +λ], rounding the two ends of the voltage interval to get the peak interval [A,B]; Step 4.5: Calculate the peak interval area S, and use the peak interval area S as a health feature of the sub-normal charging data of the lithium battery.

6. The universal prediction method for lithium battery SOH under the background of false data injection according to claim 1 is characterized in that: The steps for implementing online update of lithium battery SOH and empirical model parameters in step 5 include: Step 5.1: Use the empirical model to fit the predicted value output by the data-driven model to obtain the global degradation trend of the lithium battery; Step 5.2: The empirical model is used as the measurement equation of the UPF filter, the model parameters of the empirical model are the state equation of the UPF filter, and the predicted value output by the data-driven model is the observation value of the UPF filter. When new observation data is obtained at time k, the UPF filter is used to convert the empirical model parameters x at the previous moment into k-1 Update to x k , and then x k Iterate back to the empirical model for known quantities to achieve online update of lithium battery SOH prediction and empirical model, and obtain the optimal estimate of lithium battery SOH at the current moment; The expression of the global degradation trend of lithium batteries is: SOH n =a×e b×n +c×e d×n (15); In formula (15), SOH n is the SOH at the nth cycle, a, b, c, d are the SOH values ​​predicted by the fitted GPR i The obtained parameters are i=1,2,3,..,SP, where SP is the prediction starting point.

7. The universal prediction method for lithium battery SOH under the background of false data injection according to claim 6 is characterized in that: In step 5.2, the UPF filter is used to convert the empirical model parameters x k-1 Update to x k The steps include: Step 5.2.1: Establish the state equation of the UPF filter through the empirical model parameters a, b, c, d, and use the state equation of the UPF filter to establish the UPF measurement equation to obtain the state space equation; Step 5.2.2: When k = 0, according to the initial distribution N = 500, assigning initial weights to each particle in the empirical model parameters Step 5.2.3: Perform importance sampling on each particle; Step 5.2.4: Update the importance weight of the particle and normalize it; Step 5.2.5: Resample the normalized particles, perform state estimation based on the resampled particles, and complete the online update of the empirical model parameters; The state space equation is expressed as: In formula (16), x k =[a k ,b k ,c k ,d k ] is the exponential model parameter, y k+1 is the observed value, i.e., the GPR predicted value, the subscripts k and k+1 represent the time; w k and v k represents independent Gaussian white noise; The expression of the initial distribution of each particle in the empirical model parameters is: In formula (17), is the covariance matrix, is the Sigma point matrix; The expression for updating the importance weight of a particle and normalizing it is: The expression for resampling the normalized particles is: In formula (19), N t is the effective particle number threshold; The expression of state estimation is:

8. The universal prediction method for lithium battery SOH under the background of false data injection according to claim 7 is characterized in that: The steps of importance sampling for each particle in step 5.2.3 include: Step 5.2.3.1: For k = 1:T, i = 1:N, use the UKF filtering method to generate the importance density function of each particle; Step 5.2.3.2: Calculate the Sigma points and weights of the UPF filter; Step 5.2.3.3: Use Sigma point calculation to get the predicted value and covariance; Step 5.2.3.4: Obtain the observed value based on the predicted value and covariance calculation and weight the observed value; Step 5.2.3.5: Calculate the covariance and cross-covariance based on the weighted observations; Step 5.2.3.6: Calculate the Kalman gain K based on the covariance and cross-covariance, and update the particle state and covariance; Step 5.2.3.7: Sample the important density function to generate particles; The calculation formula for Sigma points and weights is: In formula (21), is the covariance of the weight, L is the state dimension, β = 2, is the proportional correction coefficient, λ=α 2 (L+κ)-L, α determines the discreteness of the Sigma point, and κ is the secondary sampling factor; The formula for calculating the predicted value and covariance is: The calculation formula for observations and observation weights is: The formulas for calculating covariance and cross-covariance are: The expressions of Kalman gain K and updated particle state and covariance are: The expression for sampling important density functions is:

Images