A kind of battery life status recognition methods
Technical field
The invention belongs to the technology such as electrokinetic cell, battery life evaluation, particularly relate to a kind of electrokinetic cell service life state recognition methods.
Background technology
Along with expanding economy, coal, oil equal energy source critical shortage, development and utilization new forms of energy become the only way which must be passed of countries in the world sustainable development.So, at present electrodynamic research is become mainstream research direction.Battery is as the core component of electric power product, and its level of development directly affects the development of electric power industry, and the electrokinetic cell that only technology maturation, low cost, safety are high just can make electric product be widely developed.On the one hand we should research and develop the electrokinetic cell of high performance and long service life;On the other hand electrokinetic cell lifetime estimation method and life model should be set up, the evaluation of science and prediction battery life.Battery life status identification problem is one of the most key problem of battery system, and the length of battery life can be represented by battery equivalent internal resistance or circulating battery access times, according to this 2 point, a kind of simply the most effectively judge that the method for battery life status seems extremely important.
Currently for the research of electrokinetic cell life problems, major part is to set up battery service life model.Some researchs occur decline in various degree to set up battery life equivalent model according to factors such as battery impedance, capacity, energy, power, then according to the factor identifications such as electric current, voltage, temperature, equivalent internal resistance, prediction battery life status;Also having some researchs is to set up battery cycle life model, and this method can predict the cycle-index that battery is following, but can not identify current cycle-index or current service life state.And this two classes method is required for building an effective battery system model, but a battery system model needs to consider substantial amounts of parameter, increasing so that model becomes complicated of parameter, builds battery equivalent model and also becomes complicated.
Analyzing according to above, by building battery equivalent model, to analyze the method for battery life more complicated.Along with the development of computer technology, recognition methods based on data-driven is greatly developed, and battery life status identification technology the most inherently develops to direction based on data-driven.And, the identification technology of current HMM has had the application of maturation at a lot of aspects, so the present invention proposes a kind of method utilizing HMM identification battery life status.
Summary of the invention
For problem present in above-mentioned background technology, the present invention proposes a kind of battery life status recognition methods, solves the problem that physical model statistic property configuration is complicated, solves in line computation and the problem of electrokinetic cell ONLINE RECOGNITION meanwhile.
The technical solution used in the present invention step is as follows:
A kind of battery life status recognition methods, for realizing in line computation and state estimation battery life status, including step:
A. gather electrokinetic cell system terminal voltage under each battery life status and charging and discharging currents, the two parameter set is carried out pretreatment;
B. pretreated data are carried out characteristics extraction, the input feature vector value sequence of HMM after characteristic value normalization, will be obtained;
C. set up characteristic value sequence mixture gaussian modelling, determine HMM parameter, set up the HMM being suitable for battery life status;
D. gather observed data, after feature extraction, import HMM, calculate forward direction probability;
E. compare the probit of each model output, be identified result, then jump to step D, carry out the identification of next group observed data.
Described step A battery life status is divided into 4 states according to equivalent internal resistance, and standard internal resistance represents battery life original state, two times, three times of standard internal resistances represent battery life intermediateness, four times of standard internal resistances represent battery life failure state.
Described step A gathers electrokinetic cell system terminal voltage under each state and charging and discharging currents, after having gathered, data is carried out pretreatment, first terminal voltage and charging and discharging currents is divided by, and with voltage, the result being divided by is formed new data set.
Described step B carries out characteristics extraction to pretreated data, extracts the mean-square value of U, virtual value, average, median and the average of U/I, coefficient of dispersion.U is made maximum normalization, U/I is made [-1,1] normalization.After the characteristics extraction of certain one piece of data completes, more New Data Segment, extracts the eigenvalue of lower one piece of data.
Described step C sets up characteristic value sequence mixture gaussian modelling, the parameter of Markov Chain with mixed Gaussian probability is combined, determine HMM parameter, after training data is carried out characteristics extraction, utilize Forward-backward algorithm and Baum-Welch algorithm to estimate HMM parameter, and then obtain the model parameter corresponding to each state.
Described step D gathers observed data, after feature extraction, imports HMM, calculates forward direction probability, utilizes HMM parameter to calculate forward direction probability P (O | λ) formula as follows:
αi(1)=πibi(o1), 1≤i≤N
P (O | λ)=[α1(T), α2(T) ..., αN(T)]
Wherein πiRepresent the probability of initial time state i, bi(ot) represent o under state itProbability distribution, aijRepresenting the probability that observation sequence shifts to state j from state i, N represents status number.
Described step E compares the probit of each model output, is identified result.The P (O | λ) obtained in step D is compared, the battery life status corresponding to probit maximum in P (O | λ) is recognition result, preserve recognition result, then jump to step D and carry out the identification of next data segment, until completing the identification of all observation sequences, realize by the way at line computation and state estimation.
Accompanying drawing explanation
Fig. 1 is the HMM flow chart that the present invention sets up battery life status.
Fig. 2 is the flow chart of identification battery life status of the present invention.
Detailed description of the invention
Below, in conjunction with accompanying drawing, the detailed description of the invention of the present invention is described further.
As shown in Figures 1 and 2, specific implementation process and the operation principle of the present invention are as follows:
A. gather electrokinetic cell system terminal voltage under each battery life status and charging and discharging currents, and the two parameter set is carried out pretreatment;
B. pretreated data are carried out characteristics extraction, the input feature vector value sequence of HMM after characteristic value normalization, will be obtained;
C. according to battery life feature, set up mixed Gaussian probability Distribution Model and characterize the probability distribution of battery data characteristic value sequence, utilize the data training HMM gathered, it is thus achieved that model parameter;
D. gather observed data, after feature extraction, import each state HMM, calculate forward direction probability;
E. compare the probit of each model output, be identified result, then jump to step D, carry out the identification of next group observed data.
Step A gathers electrokinetic cell system terminal voltage under each state and charging and discharging currents, more simply too much, after having gathered, by terminal voltage divided by electric current than physical model statistic property configuration.After pretreatment, the data of each state are made up of two parts, are that terminal voltage and voltage are divided by electric current respectively.
The step B feature by analytical data, extracts the temporal signatures value of data, is the mean-square value of voltage, virtual value, average, median and voltage respectively divided by the average of electric current, coefficient of dispersion.The eigenvalue of voltage segment uses maximum normalization to process, and voltage uses [-1,1] normalized divided by the normalization of current segment, obtains characteristic value sequence after normalization.
Step C sets up the HMM of battery life status identification, and Fig. 1 introduces the process of modeling.The use process of battery is a cell degradation process, and state to bad, is selected L-R type Markov Chain by well, and eigenvalue probability distribution utilizes three Gaussian mixtures to represent, formula is as follows
Wherein bj(ot) represent o under state jtProbability distribution, Q represents the number of Gauss module, Q=3,Be state j correspondingThe meansigma methods of individual Gauss distribution,Be state j correspondingThe covariance of individual Gauss distribution,Be state j correspondingWeight shared by individual Gauss distribution.One HMM can be expressed asWherein π is probability, and A is state-transition matrix.Utilize Forward-backward algorithm and Baum-Welch algorithm to estimate the parameter of each HMM, when reaching the condition of convergence, stop parameter estimation, preservation model parameter, obtained the HMM parameter of each state by said method.
Battery life status identification correspondence step D and E in Fig. 2.Step D and E gather observed data, and the state of observed data changes to four times of standard equivalent internal resistances from standard equivalent internal resistance.Step D often collects one piece of data, after completing feature extraction, characteristic value sequence is imported in HMM, obtains forward direction probit P (O | λ) that each model calculates, and computing formula is as follows:
αi(1)=πibi(o1), 1≤i≤N
P (O | λ)=[α1(T), α2(T) ..., αN(T)]
Wherein πiRepresent the probability of initial time state i, bi(ot) represent o under state itProbability distribution, aijRepresenting the probability that shift from state i of observation sequence to state j, N represents status number, and N=4, expression P (O | λ) comprise the probability calculation value of 4 kinds of battery life status.
Step E compares forward direction probit, find out the probit of maximum, using battery life status corresponding for this probit as recognition result, preserve recognition result, then jump to step D and carry out the identification of next data segment, until completing the identification of all observation sequences, realize by the way at line computation and state estimation.This recognition methods can also carry out battery life status identification with this recognition methods without battery off-line, operating battery system.