A method for estimating the state of charge of a lithium battery through multi-sensor information fusion

Abstract

The present invention discloses a method for estimating the state of charge of a lithium battery by fusing multi-sensor information. Aiming at the problem that the traditional lithium battery SOC estimation method only uses a single sensor, if the sensor fails, it will seriously affect the estimation effect. First, the extended Kalman filter algorithm is improved by introducing an adaptive function, and the least squares method with a forgetting factor is used to identify the parameters of the lithium battery, so that the algorithm continuously corrects the parameters during the process of estimating the lithium battery SOC, reduces the estimation error caused by the change of battery parameters, and improves the accuracy. Then, the improved D-S evidence theory is introduced to update the weight of the fused information of the multi-sensor in real time, overcoming the disadvantage that the traditional extended Kalman filter algorithm gives and remains unchanged by experience when allocating weights to the fused information. The combination of the two improvements can greatly improve the robustness of the estimation algorithm.

Description

Technical Field

[0001] The present invention relates to a method for estimating the state of charge (SOC) of a lithium battery, belonging to the field of new energy technologies, and in particular to a method for estimating the state of charge of a lithium battery by multi-sensor information fusion. Background Art

[0002] The current automotive industry is booming. As an important real economy industry, it plays an important role in the national economy and social development. The new energy vehicle industry, as a strategic emerging industry, will undoubtedly become the future development direction of automobiles. The lithium battery is one of the core components of an electric vehicle and is a complex chemical system that requires precise management by a battery management system. The estimation of its SOC can provide a prediction of the remaining driving range and is one of the key links in the battery management system.

[0003] Currently, many methods have been proposed to estimate the SOC of a lithium battery, such as the open-circuit voltage method, the ampere-hour method, and the Kalman series algorithms. The open-circuit voltage method has a relatively high accuracy in estimating the state of charge of a lithium battery, but this method requires the lithium battery to be static for a long enough time and is not suitable for online estimation. The ampere-hour method calculates the remaining charge by integrating the charging and discharging current over a certain period of time. This method is simple to calculate and easy to implement, but due to its open-loop nature, the integration error cannot be corrected in a timely manner, and the algorithm is prone to divergence after long-term operation. The Kalman Filter (KF) series algorithms are efficient autoregressive filtering algorithms. Using these algorithms to estimate the state of charge of a lithium battery has become one of the mainstream directions. When using the Extended Kalman Filter (EKF) algorithm for estimation, the nonlinear part in the battery model needs to be linearized, but this will introduce linearization errors, thereby affecting the estimation accuracy of the state of charge of the lithium battery. The Unscented Kalman Filter (UKF) algorithm uses the unscented transform to approximately obtain the statistical quantities of the mean and variance of the state, avoiding the errors caused by linearization and overcoming the limitations of the extended Kalman filter algorithm, with high estimation accuracy. However, the unscented Kalman filter algorithm will diverge if the positive definiteness of the covariance matrix cannot be guaranteed during use. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for estimating the state of charge of a lithium battery by multi-sensor information fusion, using the FFRLS algorithm to identify the lithium battery parameters online, and making the two links of parameter identification and SOC estimation work on different time scales (first estimating the battery parameters and then estimating the SOC of the battery), so that the algorithm continuously corrects the parameters during the SOC estimation process, reducing the estimation error caused by the change of battery parameters and improving the accuracy of the algorithm.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: an estimation method for the state of charge of a lithium battery with multi-sensor information fusion, comprising the following steps:

[0006] Step S1: Establish a second-order RC model of the lithium battery based on the second-order RC model, determine the number of voltage and current sensors used, collect data of all groups of voltage and current sensors, and perform online identification of the lithium battery parameters using the least squares method with a forgetting factor;

[0007] Step S2: Send the data of all groups of voltage and current sensors into the sub-filter, and use the DS-AFEKF algorithm to estimate the SOC of the sub-filter. Each sub-filter works in parallel to obtain different local optimal values, and then send these values into the main filter. Next, the main filter fuses the information of these local optimal values to obtain the global optimal estimation value and output it;

[0008] Step S3: Assign respective weights to each voltage and current sensor according to their credibility. In each algorithm iteration, divide all voltage and current sensors into target sensors and evidence sensors. According to the relative size relationship between the evidence and the target sensors, define the basic belief assignment of the new evidence, and use the improved D-S evidence theory to update the fusion weights of all voltage and current sensors in the AFEKF algorithm in real time;

[0009] Step S4: Output the basic belief assignment weights of the new evidence to each sub-filter for the next algorithm iteration;

[0010] Step S5: When estimating the SOC of the lithium battery, repeatedly perform the same optimization and iteration operations as in Steps S2 to S4 until the iteration ends, output the estimated SOC value, and plot its change curve.

[0011] A further improvement of the technical solution of the present invention lies in that: the expression of the second-order RC model of the lithium battery in Step S1 is:

[0012]

[0013]

[0014] where, u oc is the open-circuit voltage of the battery, u is the terminal voltage of the battery, i is the current in the battery, R S is the ohmic internal resistance of the battery, R τ and R L respectively represent the concentration polarization internal resistance and the electrochemical polarization internal resistance, C τ and C L respectively represent the concentration polarization capacitance and the electrochemical polarization capacitance of the battery, u τand u L respectively represent the concentration polarization voltage and the electrochemical polarization voltage of the battery, η is the Coulomb efficiency, Q C is the rated capacity of the battery, t 0 is the initial time, and t is the current time.

[0015] A further improvement of the technical solution of the present invention lies in: in the step S1, the second-order RC model of the lithium battery is discretized and linearized, the voltages of two RC circuits and the state of charge of the battery are selected as state variables, and random disturbances are considered to obtain the discrete state equation and the observation equation as follows:

[0016]

[0017] where X(k) = [u τ (k) u L (k) SOC(k)] T , U(k) = i(k), Z(k) = u(k),

[0018]

[0019]

[0020] where the partial derivative term is obtained by performing a first-order Taylor expansion of H[SOC(k)] around SOC(k). D = R S , T S is the sampling period, W(k) is the process noise, V(k) is the observation noise, and k is the time; the noise statistical characteristics of the system are described as:

[0021]

[0022] where E[·] represents the mathematical expectation of the relevant noise, Q(k) is the process noise variance, R(k) is the observation noise variance, and η(k,j) is the Kronecker function, which is defined as:

[0023]

[0024] A further improvement of the technical solution of the present invention lies in: in the step S2, the i-th sub-filter is described as:

[0025]

[0026] The time update and the prediction of the observed quantity of the i-th sub-filter and the main filter are:

[0027]

[0028] where, when i = 1, 2,..., M, it represents the sub-filter, and when i = g, it represents the main filter;

[0029] The measurement updates of the i-th sub-filter and the main filter are as follows:

[0030] K i = P i (k + 1|k)C T (CP i (k + 1|k)C T + R i ) -1

[0031]

[0032] P i (k + 1|k + 1) = P i (k + 1|k)-P i (k + 1|k)K i C T

[0033] The covariance and state assignment formulas of the i-th sub-filter and the main filter are as follows:

[0034]

[0035]

[0036]

[0037] Where Q i is the observation noise variance of the sub-filter, Q g is the observation noise variance of the main filter, P i is the sub-filter system covariance matrix, P g is the system covariance matrix of the main filter, β i is the feedback weight value after each filtering and must satisfy the information conservation principle: And 0 ≤ β i ≤ 1;

[0038] Summarize the local optimal values of each filter to the main filter for information fusion. The fusion method is as follows:

[0039]

[0040] The global optimal state update can be obtained from the following formula:

[0041]

[0042] The global state covariance matrix can be obtained from the following formula:

[0043]

[0044] A further improvement of the technical solution of the present invention lies in that: in the steps S1 and S2, the FFRLS online identification system parameters and SOC estimation are alternately performed to obtain an adaptive combined extended Kalman filter algorithm. The specific formula of the adaptive combined extended Kalman filter algorithm is as follows:

[0045]

[0046] In the formula, is the parameter variable to be identified, P(k) is the estimated covariance matrix of the system, h(k) is the gain matrix of the system, is the system data variable, n 1 n 2 are positive numbers, and λ is the introduced forgetting factor, and the value range is 0.95 - 1.

[0047] A further improvement of the technical solution of the present invention lies in that: the process of assigning respective weights to the credibility of each voltage and current sensor in the step S3 is the basic belief assignment, and the formula is as follows:

[0048]

[0049] In the formula, m(·) is the basic belief assignment function on the recognition frame Θ. Moreover, α such that m(α)>0 is called the focal element, and φ is the empty set.

[0050] A further improvement of the technical solution of the present invention lies in that: the process of improving the D - S evidence theory to update the fusion weights of all voltage and current sensors in the AFEKF algorithm in real - time in the step S3 is as follows:

[0051] Suppose that at a certain moment, the set composed of the measurement values of each sub - filter voltage sensor is Τ:

[0052] Τ = [Z 1 (k),…,Z i (k),…,Z M (k)],

[0053] 1) Randomly select one of the sensors Z i (k) as the target sensor set T o , then the remaining all sensors form the evidence sensor set T m :

[0054]

[0055] 2) Define the size between the measurement values of the target sensor and the evidence sensor as d. Then the size between the measurement values of the target sensor and each evidence sensor is:

[0056] d j =|Zi (k)-Z j (k)|,j=1,2,…,M,j≠i

[0057] 3) Obtain the basic confidence allocation function of each voltage sensor corresponding to the sub-filter:

[0058]

[0059] In the formula, α i , i=1,2,…,M, is the i-th focal element;

[0060] 4) All evidences are integrated to obtain the distribution weights of the sub-filters corresponding to each voltage sensor:

[0061]

[0062] In the formula, is the average trust distribution of all evidences, β i is the weight of probability distribution, γ represents the conflict coefficient between evidences, and its value represents the degree of correlation between evidences.

[0063] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0064] 1. The present invention uses the FFRLS algorithm to identify the lithium battery parameters online, and makes the parameter identification and SOC estimation work on different time scales (first perform battery parameter estimation and then perform battery SOC estimation), so that the algorithm continuously corrects the parameters during the SOC estimation process, reduces the estimation error caused by the battery parameter changes, and improves the accuracy of the algorithm;

[0065] 2. When improving the DS evidence theory, the present invention has created a new basic trust allocation process according to the technical field to be used, specifically: (1) The voltage sensors are randomly allocated as evidence sensors and target sensors. Randomness is guaranteed, and the situation where the estimation accuracy is affected if the target sensor is damaged when a single target sensor is set is avoided; (2) The characteristics of sensors in the battery field are fully combined: because the data between the sensors are consistent, their measurement results can be verified with each other, so they can be assigned respective weights according to the credibility of each sensor. If a sensor has abnormal values ​​one after another, it is considered that this sensor does not have a high credibility. Taking into account that if a voltage sensor fails, its measurement value may fluctuate around the true value, in each algorithm iteration, all sensors are divided into target sensors and evidence sensors, and the basic trust allocation of new evidence can be defined according to the relative size relationship between the evidence and the target sensor;

[0066] 3. The present invention introduces an improved D-S evidence theory to dynamically allocate the weights of each sub-filter, reducing the impact of sensor failures on the estimation results. In some cases where the voltage sensor cannot be replaced in a timely manner, the SOC estimation method proposed by the present invention can still achieve good estimation accuracy and has stronger robustness compared to other SOC estimation algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is the algorithm structure diagram of the present invention;

[0068] Figure 2 is the second-order RC model of the lithium battery of the present invention;

[0069] Figure 3 is the parameter identification result diagram of the present invention;

[0070] Figure 4 is the estimation result diagram of the estimation method of the present invention and other methods;

[0071] Figure 5 is the estimation result diagram of the present method and other methods when it is assumed that a sensor is damaged in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0072] The object of the present invention is to invent a SOC estimation method for multi-sensor information fusion in view of the problem that traditional lithium battery SOC estimation methods only use a single sensor, and if the sensor fails, it will seriously affect the estimation effect. First, the Federal Extended Kalman Filter (FEKF) algorithm is improved by introducing an adaptive function, and the least squares method with a forgetting factor is used to identify the parameters of the lithium battery, so that the algorithm continuously corrects the parameters during the process of estimating the SOC of the lithium battery, reducing the estimation error caused by the change of battery parameters and improving the accuracy. Then, an improved D-S evidence theory is introduced to update the weights of the fusion information of multiple sensors in real time, overcoming the disadvantage that the traditional joint extended Kalman filter algorithm gives and keeps the weights of the fusion information by experience. The combination of the two improvements can greatly improve the robustness of the estimation algorithm.

[0073] To achieve the above object, the technical solution of the present invention is as follows:

[0074] As Figure 1 shown, a method for estimating the state of charge of a lithium battery with multi-sensor information fusion includes the following steps:

[0075] Step S1: Establish a second-order RC model for the lithium battery based on the second-order RC model, determine the number of voltage and current sensors to be used, collect the data of all groups of voltage and current sensors, and perform online identification of the lithium battery parameters using the least squares method with a forgetting factor.

[0076] The second-order RC model of the lithium battery is as Figure 2 shown, and the expression is as follows:

[0077]

[0078]

[0079] where, u oc is the open-circuit voltage of the battery, u is the terminal voltage of the battery, i is the current in the battery, R S is the ohmic internal resistance of the battery, R τ and R L represent the concentration polarization internal resistance and the electrochemical polarization internal resistance respectively, C τ and C L represent the concentration polarization capacitance and the electrochemical polarization capacitance of the battery respectively, u τ and u L represent the concentration polarization voltage and the electrochemical polarization voltage of the battery respectively, η is the Coulomb efficiency, Q C is the rated capacity of the battery, t 0 is the initial moment, and t is the current moment.

[0080] Discretize and linearize the second-order RC model of the lithium battery, select the voltages of two RC circuits and the state of charge of the battery as state variables, and consider random disturbances to obtain the discrete state equation and the observation equation as follows:

[0081]

[0082] In the formula, X(k) = [u τ (k) u L (k) SOC(k)] T , U(k) = i(k), Z(k) = u(k),

[0083]

[0084]

[0085] where the partial derivative terms are obtained by performing a first-order Taylor expansion of H[SOC(k)] around SOC(k). D = R S , T S is the sampling period, W(k) is the process noise, V(k) is the observation noise, and k is the moment; the noise statistical characteristics of the system are described as:

[0086]

[0087] In the formula, E[·] represents the mathematical expectation of the correlation noise, Q(k) is the process noise variance, R(k) is the observation noise variance, and η(k,j) is the Kronecker function, which is defined as:

[0088]

[0089] Step S2: Send all groups of voltage and current sensor data into the sub-filter, and use the DS-AFEKF algorithm to estimate the SOC of the sub-filter. Each sub-filter works in parallel to obtain different local optimal values respectively, and then send these values into the main filter. Next, the main filter fuses the information of these local optimal values to obtain the global optimal estimated value and output it. Each sub-filter assigns weights according to the different information it has, distributes the variance updated by the main filter, and uses it for the next operation. This method of first performing block processing and then global fusion makes the most of the available data to improve the accuracy of the final estimation.

[0090] For the lithium battery system mathematically modeled according to the present invention, its i-th sub-filter can be described as:

[0091]

[0092] The time update and prediction of the observed quantity of the i-th sub-filter and the main filter are:

[0093]

[0094] In the formula, when i = 1, 2,..., M, it represents the sub-filter, and when i = g, it represents the main filter;

[0095] The measurement update of the i-th sub-filter and the main filter is:

[0096] K i =P i (k + 1|k)C T (CP i (k + 1|k)C T +R i ) -1

[0097]

[0098] P i (k + 1|k + 1)=P i (k + 1|k)-P i (k + 1|k)K i C T

[0099] The covariance and state assignment formulas for the i-th sub-filter and the main filter are as follows:

[0100]

[0101]

[0102]

[0103] Where Q i is the observation noise variance of the sub-filter, Q g is the observation noise variance of the main filter, P i is the system covariance matrix of the sub-filter, P g is the system covariance matrix of the main filter, β i is the feedback weight value after each filtering, and must satisfy the information conservation principle: And 0 ≤ β i ≤ 1.

[0104] Summarize the local optimal values of each filter to the main filter for information fusion. The fusion method is as follows:

[0105]

[0106] The global optimal state update can be obtained from the following formula:

[0107]

[0108] The global state covariance matrix can be obtained from the following formula:

[0109]

[0110] As Figure 3 shown, it shows the results of online identification of lithium battery parameters using the least squares method with a forgetting factor of the present invention. Considering that the battery parameters will change with the aging of the battery, the present invention alternately performs FFRLS online identification system parameters and SOC estimation to obtain an adaptive joint extended Kalman filtering algorithm. The specific formula of the adaptive joint extended Kalman filtering algorithm is as follows:

[0111]

[0112] Where is the parameter variable to be identified, P(k) is the estimated covariance matrix of the system, h(k) is the gain matrix of the system, is the system data variable, n 1 , n 2 are positive numbers, and λ is the introduced forgetting factor, and the value range is 0.95 - 1.

[0113] Step S3: Assign respective weights to each voltage and current sensor according to their credibility. Considering that if a voltage sensor fails, its measured value may fluctuate around the true value. In each algorithm iteration, all voltage and current sensors are divided into target sensors and evidence sensors. According to the relative magnitude relationship between the evidence and the target sensors, define the basic belief assignment of the new evidence, and use the improved D-S evidence theory to update the fusion weights of all voltage and current sensors in the AFEKF algorithm in real time;

[0114] The weight β assigned by the FEKF algorithm i is generally given by experience and is fixed, while each β i value has a great influence on the estimation result. The improved D-S evidence theory of the invention updates the weights assigned to the information in real time. The D-S evidence theory is an important theory that can handle multi-source information. Its greatest feature is that it uses "interval estimation" rather than "point estimation" to describe uncertain information. It shows great flexibility in distinguishing between ignorance and uncertainty and accurately reflecting evidence collection. For objective evidence and subjective estimation, neither result should be favored in information collection, but their importance should be noted. The D-S theory can fully retain the uncertainty of both results. For all possible sets of uncertainties, which are called the frame of discernment in the theory and are denoted by Θ. The elements inside are mutually exclusive, and all possible combinations of problems are 2 Θ , and the D-S theory assigns probabilities to all of them, that is, it has the ability to retain and process multi-source information. The assignment process is called the basic belief assignment (BBA), generally referred to as the mass function. In the power set of the frame of discernment, there is:

[0115]

[0116] where m(·) is the basic belief assignment function on the frame of discernment Θ, and moreover, α for which m(α)>0 is called the focal element, and φ is the empty set.

[0117] The process of the improved D-S evidence theory updating the fusion weights of all voltage and current sensors in the AFEKF algorithm in real time is as follows:

[0118] Because the data between each sensor is consistent and their measurement results can be mutually verified, respective weights can be assigned to them according to the credibility of each sensor. If a certain sensor continuously shows abnormal values, it is considered that this sensor does not have a high credibility. Considering that if a voltage sensor fails, its measured value may fluctuate around the true value. In each algorithm iteration, all sensors are divided into target sensors and evidence sensors, and the basic belief assignment of the new evidence can be defined according to the relative magnitude relationship between the evidence and the target sensors. The weight adjustment based on the D-S theory is as follows:

[0119] At a certain moment, the set composed of the measured values of each sub-filter voltage sensor is Τ:

[0120] Τ = [Z 1 (k), …, Z i (k), …, Z M (k)],

[0121] 1) Randomly select one of the sensors Z i (k) as the target sensor set T o , and the remaining all sensors form the evidence sensor set T m :

[0122]

[0123] 2) Define the magnitude between the measured values of the target sensor and the evidence sensors as d, then the magnitude between the measured value of the target sensor and each evidence sensor is:

[0124] d j = |Z i (k) - Z j (k)|, j = 1, 2, …, M, j ≠ i

[0125] 3) Obtain the basic belief assignment function of each voltage sensor corresponding to the sub-filter:

[0126]

[0127] In the formula, α i , i = 1, 2, …, M, is the i-th focal element;

[0128] 4) Fuse all the evidences to obtain the assignment weight values of each voltage sensor corresponding to the sub-filter:

[0129]

[0130] In the formula, is the average belief assignment of all evidences. β i is the weight value of the probability assignment. γ represents the conflict coefficient between evidences, and its value represents the correlation degree between evidences. The larger the conflict value, the greater the dissimilarity between the measurement evidences; on the contrary, it indicates that the correlation between the measurement evidences is greater. Therefore, the data fusion of the measurement evidences can be modified according to the actual situation to obtain the fused weight value β i .

[0131] Step S4: Output the basic belief assignment weight value of the new evidence to each sub-filter and perform the next algorithm iteration;

[0132] Step S5: Repeatedly perform the same optimization and iteration operations as in Steps S2 - S4 until the iteration ends when estimating the SOC of the lithium battery. Output the estimated SOC value and plot its change curve.

[0133] As Figure 4 shown, to demonstrate the advantages of the algorithms, the AUKF algorithm (using only 1 set of sensors, i.e., one voltage and one current sensor) and the FEKF algorithm (using 3 sets of sensors) are selected for comparison. The initial value of the SOC is uniformly set to 0.9. Among them, the estimation error of the DS - AFEKF algorithm is the smallest, within 1.12%. The estimation error of the FEKF algorithm is second only to that of the DS - AFKF, within 1.94%. The estimation error of the AUKF is the largest, within 2.11%. Compared with the AUKF and FEKF, the DS - AFEKF algorithm has better estimation accuracy due to adopting the optimal fusion criterion with improved D - S weighting.

[0134] As Figure 5 shown, assume that 1 out of the 3 voltage sensors fails, that is, a perturbation (a constant value decreased by 0.12V) is added to the voltage data it collects. When 1 voltage sensor fails, the estimation error of the DS - AFEKF algorithm is the smallest, within 5.87%. The estimation error of the FEKF algorithm is the second, within 9.13%. And the estimation error of the AUKF algorithm is the largest, within 16.59%. It can be seen that in actual operation, if a sensor fails, the measured data deviates greatly from the actual value, and having only one sensor will have a great impact on the result (such as the large estimation error of the AUKF algorithm). While the DS - AFKF and FEKF algorithms have 3 sensors, compared with the AUKF algorithm using only one filter, they have better anti - interference ability. At the same time, the DS - AFEKF algorithm, based on the improved D - S theory, can dynamically allocate more reasonable weights, and has higher estimation accuracy compared with the FEKF algorithm with fixed weight allocation.

Claims

1. A method for estimating the state of charge of a lithium battery through multi-sensor information fusion, characterized in that: It includes the following steps: Step S1: Establish a second-order RC model of the lithium battery based on the second-order RC model, determine the number of voltage and current sensors to be used, collect data of all groups of voltage and current sensors, and perform online identification of the lithium battery parameters using the least squares method with a forgetting factor; Step S2: Send the data of all groups of voltage and current sensors into the sub-filter, and use the DS-AFEKF algorithm to estimate the SOC of the sub-filter. Each sub-filter works in parallel, respectively obtaining different local optimal values, and sending these values into the main filter. Then, the main filter fuses this information of these local optimal values to obtain the global optimal estimated value and output it; Step S3: Assign respective weights to each voltage and current sensor according to their credibility. During each algorithm iteration, divide all voltage and current sensors into target sensors and evidence sensors. According to the relative magnitude relationship between the evidence and the target sensors, define the basic belief assignment of the new evidence, and use the improved D-S evidence theory to update the fusion weights of all voltage and current sensors in the AFEKF algorithm in real time; Step S4: Output the basic belief assignment weights of the new evidence to each sub-filter for the next algorithm iteration; Step S5: When estimating the SOC of the lithium battery, repeatedly perform the same optimization and iteration operations as in Steps S2 to S4 until the iteration ends, output the estimated SOC value, and plot its change curve.

2. A method for estimating the state of charge of a lithium battery through multi-sensor information fusion according to claim 1, characterized in that: The expression of the second-order RC model of the lithium battery in Step S1 is: Among them, u oc is the open-circuit voltage of the battery, u is the terminal voltage of the battery, i is the current in the battery, and R S is the ohmic internal resistance of the battery, and R τ and R L represent the concentration polarization internal resistance and the electrochemical polarization internal resistance respectively, C τ and C L represent the concentration polarization capacitance and the electrochemical polarization capacitance of the battery respectively, u τ and u L represent the concentration polarization voltage and the electrochemical polarization voltage of the battery respectively, η is the Coulomb efficiency, Q C is the rated capacity of the battery, t 0 is the initial moment, and t is the current moment.

3. A method for estimating the state of charge of a lithium battery through multi-sensor information fusion according to claim 2, characterized in that: In Step S1, the second-order RC model of the lithium battery is discretized and linearized. Select the voltages of two RC circuits and the state of charge of the battery as state variables, and consider random disturbances to obtain the following discrete state equation and observation equation: where X(k) = [u τ (k)u L (k)SOC(k)] T , U(k) = i(k), Z(k) = u(k), The partial derivative term is obtained by performing a first-order Taylor expansion of H[SOC(k)] around SOC(k) as D = R S , T S is the sampling period, W(k) is the process noise, V(k) is the observation noise, and k is the time instant; the noise statistical characteristics of the system are described as: In the formula, E[·] represents the mathematical expectation of the relevant noise, Q(k) is the process noise variance, R(k) is the observation noise variance, and η(k,j) is the Kronecker function, defined as:

4. A method for estimating the state of charge of a lithium battery through multi-sensor information fusion according to claim 3, characterized in that: The description of the i-th sub-filter in Step S2 is: The time update and prediction of the observed quantity of the i-th sub-filter and the main filter are: In the formula, when i = 1, 2,..., M, it represents the sub-filter, and when i = g, it represents the main filter; The measurement update of the i-th sub-filter and the main filter is: K i = P i (k + 1|k)C T (CP i (k + 1|k)C T + R i ) -1 , P i (k + 1|k + 1) = P i (k + 1|k) - P i (k + 1|k)K i C T , The covariance and state assignment formula of the i-th sub-filter and the main filter are: where Q i is the observation noise variance of the sub-filter, and Q g is the observation noise variance of the main filter, P i is the system covariance matrix of the sub-filter, and P g is the system covariance matrix of the main filter, β i is the feedback weight after each filtering, and must satisfy the information conservation principle: and 0 ≤ β i ≤ 1; Summarize the local optimal values of each filter to the main filter for information fusion. The fusion method is: The global optimal state update can be obtained from the following formula: The global state covariance matrix can be obtained from the following formula:

5. A method for estimating the state of charge of a lithium battery through multi-sensor information fusion according to claim 4, characterized in that: In the steps S1 and S2, the least squares method with forgetting factor is used to alternately estimate the lithium battery parameters and SOC, and an adaptive joint extended Kalman filter algorithm is obtained. The specific formula of the adaptive joint extended Kalman filter algorithm is as follows: In the formula, is the parameter variable to be identified, P(k) is the estimated covariance matrix of the system, h(k) is the gain matrix of the system, is the system data variable, n 1 n 2 are positive numbers, λ is the introduced forgetting factor, and its value range is 0.95 to 1.

6. A method for estimating the state of charge of a lithium battery by multi-sensor information fusion according to claim 5, characterized in that: In the step S3, the process of assigning respective weights to the credibility of each voltage and current sensor is the basic belief assignment, and the formula is as follows: In the formula, m(·) is the basic belief assignment function on the identification frame Θ. Moreover, α for which m(α)>0 is called the focal element, and φ is the empty set.

7. A method for estimating the state of charge of a lithium battery by multi-sensor information fusion according to claim 6, characterized in that: In the step S3, the process of improving the D-S evidence theory to update the fusion weights of all voltage and current sensors in the AFEKF algorithm in real time is as follows: Suppose that at a certain moment, the set composed of the measured values of the voltage sensors of each sub-filter is Τ: T = [Z 1 (k), …, Z i (k), …, Z M (k)], 1) Randomly select one of the sensors Z i (k) is the target sensor set T o , then all the remaining sensors form the evidence sensor set T m : 2) Define the magnitude between the measured values of the target sensor and the evidence sensor as d, then the magnitude between the measured values of the target sensor and each evidence sensor is: d j = |Z i (k) - Z j (k)|, j = 1, 2, …, M, j ≠ i; 3) Obtain the basic belief assignment function of each sub-filter corresponding to the voltage sensor: where α i , i = 1, 2, …, M, is the i-th focal element; 4) Fuse all the evidences to obtain the assignment weights of each sub-filter corresponding to the voltage sensors: Wherein, is the average belief assignment of all evidences, β i is the weight of the probability assignment, and γ represents the conflict coefficient between evidences, and its value represents the correlation degree between evidences.

Images