The background technology is as follows:
in the large background of exhaustion of fossil energy and sustainable development, the pure electric automobile has been rapidly developed and widely used in recent years. But the actual capacity of the power cell as a source of energy for providing it will drop below 80% of nominal after 3-5 years of operation, the power cell requiring decommissioning and replacement. The retired power battery directly recovers resources in a material mode, so that the resource waste is caused, and the power battery does not accord with the green economy development principle. In order to improve the economy of power battery utilization, the method is applied to other application scenes with low energy and power requirements, and the gradient utilization of the power battery is of great significance. However, the power cells have significant performance differences after retirement, which creates significant difficulties in cell pack balancing, power distribution, and battery management design for subsequent applications. Therefore, research such as screening of power batteries after retirement, estimation of health state indexes and the like is carried out, and is very important to improving the application safety and rationality of the power batteries.
Currently, assessment of power cell state of health is based mainly on experimental analysis, modeling estimation, and data driving three approaches: in the experimental analysis method, a direct discharge method, an internal resistance measurement method and the like are typical, and the method has the advantages that more accurate and direct power battery health state evaluation indexes are obtained through strict and careful experimental processes and results, but the efficiency is low when screening a large number of power batteries because of strict experimental requirements, complicated processes and complex analysis and calculation; the modeling estimation method is more typical, such as Kalman filtering, fuzzy logic and the like, has the advantages that the health state of the power battery can be estimated qualitatively and quantitatively in real time, standard experimental processes and calculation are carried out without dismantling the power battery, and the power battery is usually used for real-time health of the running state of the power battery, but the problems of accumulated errors, high precision requirements of the method design on a power battery model, large amount of human experience and the like exist; the data driving estimation method, such as a support vector machine, a neural network and the like, has the advantages that based on a large number of deep factors influencing the health state of the power battery, the correlation is formed by reflecting the object essence through data, the estimation of the power battery can be quickly and automatically realized after modeling, the detection efficiency is improved, the labor cost is reduced, but the method depends on the quality and the scale of the data, the requirements on the data comprehensiveness and granularity are high, and the model failure can be caused by the data with one-sided ambiguity. Therefore, a new state estimation method is needed.
The invention comprises the following steps:
in order to improve the effectiveness of power battery health state estimation, enhance the robustness of an estimation method, reduce the influence of manual setting and experience on the accuracy of the method, accelerate the detection speed of the power battery health state, save the labor cost and realize the efficient estimation of the power battery health state, the invention provides the power battery health state estimation method under integrated clustering. The invention adopts the technical scheme that:
a power battery health state evaluation method under integrated clustering comprises the following steps:
step 1, mining historical data characteristics to construct a sample set; the method comprises the following steps:
step 1.1, obtaining power battery health state values under different health states for n power batteries of the same model as the power battery to be evaluated
Simultaneously measuring voltage, current and temperature signals in the charging and discharging experiment process of the power batteries;
step 1.2, extracting m key features describing the state of the power battery based on the existing power battery voltage, current and temperature feature extraction method, and carrying out normalization processing on each feature, wherein the feature vector of the ith sample is as follows
Forming a plurality of sample sets of power battery sample multidimensional feature vectors and health status values:
step 2, forming a plurality of estimated values of the state of health of the power battery to be evaluated; the method comprises the following steps:
step 2.1, according to the feature extraction method of step 1.2, calculating the voltage, current and temperature feature values of the power battery to be evaluated in the charge and discharge processes to form feature vectors describing the power battery to be evaluated
Setting the integration subset size N required by the method, and enabling t=1;
step 2.2, randomly sampling the sample set A obtained in the step 1 to obtain a subset B with a determined capacity t And guarantee subset B t Mutual exclusion of all samples in the system;
step 2.3, randomly defining a clustering quantity parameter, and utilizing a kmeans clustering method to cluster the subset B according to the Euclidean distance t The samples with the multidimensional feature vectors are gathered into a plurality of groups, and the centers of the groups are obtained;
step 2.4, calculating the characteristic vector X of the power battery to be detected
(test) And each group center
Is the Euclidean distance of (2)
In subset B
t Judging that the power battery to be detected belongs to the group center and the nearest group R (t);
step 2.5 computing at subset B t The weight of each sample health state value in the group to which the power battery to be measured belongs is determined;
step 2.5.1 calculating the feature vector X of the power battery to be detected
(test) And subset B
t The Euclidean distance of each sample in the group R (t) of interest, where s is to the group R (t)
R(t) The distance of each sample is
S
R(t) Representing the number of power cell samples in the population R (t);
step 2.5.2 the weights of the samples are normalized to define the weight of the samples in such a way that the closer the distance the greater the weight, i.e. the s-th
R(t) The power battery sample state of health value is
And the weight is +.>
Forming weight vectors
Step 2.6 according to the current subsetB
t Health status value of each sample in the following group R (t)
Weighting of
Obtaining an estimate of the power cell state of health value to be detected for the subset:
step 2.7, judging whether t is larger than or equal to a set integration subset scale N; if yes, entering a step 3, if not, t=t+1, and returning to the step 2.2 to continue to generate a new subset to estimate the state of health value of the power battery to be detected;
step 3, counting the mean value and standard deviation of the estimated value of the state of health of the power battery to be tested of each subset;
counting the subsets obtained in the
step 2 to obtain an estimated value set of the state of health of the power battery to be tested
Calculate its mean +.>
Represents the estimated result of the state of health of the final power battery to be detected, its standard deviation +.>
And (5) characterizing the error of the estimated value and finishing test estimation.
Preferably, the step 2.3 specifically includes the following steps:
step 2.3.1, randomly defining a clustering quantity parameter K (t), randomly selecting K (t) from n power battery samples to form initialized K (t) population centers,
step 2.3.2 computing subset B
t Euclidean distance between the feature vector of the rest power battery samples and the center of the K (t) th group, wherein the Euclidean distance between the (r) th sample and the center of the K (t) th group is
Judging the distance from each sample to the center of the K (t) th group, and assigning the distance to the nearest group;
step 2.3.3, updating the group center according to the sample conditions in each group:
wherein the j-th feature of the k (t) -th population center is +.>
S
k(t) Representing the number of samples in the kth (t) th population;
and 2.3.4, judging whether the clustering process converges to the requirement, if so, entering the step 2.4, and if not, returning to the step 2.3.2.
Compared with the closest prior art, the invention has the following excellent effects:
according to the technical scheme, after the power battery health state estimated values under a plurality of different kmeans clusters are generated by randomly generating the differential sample subsets and the kmeans parameters, the final power battery health state estimated value and the estimation error are obtained in a statistical mean mode. Compared with the existing power battery state of health evaluation mode, the method and the device avoid the influence of the random initialization class center on the clustering result in the kmeans clustering method, and simultaneously reduce the influence of noise and singular samples on the estimation result. Meanwhile, the power battery health state estimation results of the mean value and the standard deviation are finally used, so that the overall cognition of the evaluation level of the power battery health state is facilitated, and the reliability of the method is improved.
In the technical scheme of the invention, the control parameters in the kmeans clustering method are randomly generated, the optimization requirement of the parameters is reduced by utilizing the integration process, the influence of the number of the kmeans clusters on the final result is small after the number of the kmeans clusters reaches a certain scale, and meanwhile, weighted average or average calculation is carried out on all subsets and the final power battery health state estimation result based on the weight defined by the Euclidean distance. Compared with the existing power battery health state evaluation method based on the regression prediction model, the method does not need to assume a regression model function form and additional data set to optimize and evaluate parameters, and the method has the advantages that the required set parameters are few, the subjective influence of human experience is reduced, the method is insensitive to the parameters, and the robustness and the adaptability of the method are improved.
The specific embodiment is as follows:
examples:
a power battery health state evaluation method under integrated clustering comprises the following steps:
step 1, mining historical data characteristics to construct a sample set; the method comprises the following steps:
step 1.1, obtaining power battery health state values under different health states for n power batteries of the same model as the power battery to be evaluated
Simultaneously measuring voltage, current and temperature signals in the charging and discharging experiment process of the power batteries;
step 1.2, extracting descriptive power based on the existing power battery voltage, current and temperature characteristic extraction methodM key features of the battery state and carrying out normalization processing on each feature, wherein the feature vector of the ith sample is as follows
Forming a plurality of sample sets of power battery sample multidimensional feature vectors and health status values:
step 2, forming a plurality of estimated values of the state of health of the power battery to be evaluated; the method comprises the following steps:
step 2.1, according to the feature extraction method of step 1.2, calculating the voltage, current and temperature feature values of the power battery to be evaluated in the charge and discharge processes to form feature vectors describing the power battery to be evaluated
Setting the integration subset size N required by the method, and enabling t=1;
step 2.2, randomly sampling the sample set A obtained in the step 1 to obtain a subset B with a determined capacity t And guarantee subset B t Mutual exclusion of all samples in the system;
step 2.3, aggregating subset B by using a kmenas clustering method t A sample set forming a plurality of groups;
step 2.3.1, randomly defining a clustering quantity parameter K (t), randomly selecting K (t) from n power battery samples to form initialized K (t) population centers,
step 2.3.2 computing subset B
t Euclidean distance between the feature vector of the rest power battery samples and the center of the K (t) th group, wherein the Euclidean distance between the (r) th sample and the center of the K (t) th group is
Judging the distance from each sample to the center of the K (t) th group, and assigning the distance to the nearest group;
step 2.3.3, updating the group center according to the sample conditions in each group:
wherein the j-th feature of the k (t) -th population center is +.>
S
k(t) Representing the number of samples in the kth (t) th population;
step 2.3.4, judging whether the clustering process converges to the requirement, if the degree of change of the group centers is smaller than a set threshold, if so, entering step 2.4, and if not, returning to step 2.3.2;
step 2.4, calculating the characteristic vector X of the power battery to be detected
(test) And each group center
Is the Euclidean distance of (2)
In subset B
t Judging that the power battery to be detected belongs to the group center and the nearest group R (t);
step 2.5 computing at subset B t The weight of each sample health state value in the group to which the power battery to be measured belongs is determined;
step 2.5.1 calculating the feature vector X of the power battery to be detected
(test) And subset B
t The Euclidean distance of each sample in the group R (t) of interest, where s is to the group R (t)
R(t) The distance of each sample is
S
R(t) Representing the number of power cell samples in the population R (t);
step 2.5.2 the weights of the samples are normalized to define the weight of the samples in such a way that the closer the distance the greater the weight, i.e. the s-th
R(t) The power battery sample state of health value is
And the weight is +.>
Forming weight vectors
Step 2.6 according to the current subset B
t Health status value of each sample in the following group R (t)
Weighting of
Obtaining an estimate of the power cell state of health value to be detected for the subset:
step 2.7, judging whether t is larger than or equal to a set integration subset scale N; if yes, entering a step 3, if not, t=t+1, and returning to the step 2.2 to continue to generate a new subset to estimate the state of health value of the power battery to be detected;
step 3, counting the mean value and standard deviation of the estimated value of the state of health of the power battery to be tested of each subset;
counting the subsets obtained in the
step 2 to obtain an estimated value set of the state of health of the power battery to be tested
Calculate its mean +.>
Health care representative of final power battery to be detectedHealth state estimation result, standard deviation->
And (5) characterizing the error of the estimated value and finishing test estimation.