CN113094989B - Unmanned aerial vehicle battery life prediction method based on random configuration network - Google Patents
Unmanned aerial vehicle battery life prediction method based on random configuration network Download PDFInfo
- Publication number
- CN113094989B CN113094989B CN202110371138.2A CN202110371138A CN113094989B CN 113094989 B CN113094989 B CN 113094989B CN 202110371138 A CN202110371138 A CN 202110371138A CN 113094989 B CN113094989 B CN 113094989B
- Authority
- CN
- China
- Prior art keywords
- representing
- layer node
- network
- hidden
- value
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/36—Arrangements for testing, measuring or monitoring the electrical condition of accumulators or electric batteries, e.g. capacity or state of charge [SoC]
- G01R31/367—Software therefor, e.g. for battery testing using modelling or look-up tables
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/04—Ageing analysis or optimisation against ageing
Abstract
The invention discloses a method for predicting the service life of an unmanned aerial vehicle battery based on a random configuration network, which comprises the following steps: the capacity of a lithium ion battery for the unmanned aerial vehicle is used as a direct health factor, the discharge voltage of the battery is used as an indirect health factor, model training and parameter debugging are carried out by adopting a random configuration network, and a service life prediction model of the lithium ion battery is obtained for prediction. The problems of low precision, high training data requirement and the like in the existing data-driven prediction method are effectively avoided, the characteristics of strong SCN autonomy, high convergence rate, low network cost and the like are fully exerted, compared with other neural networks, the RMSE value of the SCN is minimum, the method has lower training loss and better network fitting effect, and is an effective RUL prediction algorithm for the lithium battery.
Description
Technical Field
The invention belongs to the technical field of unmanned aerial vehicle battery health management, and relates to a method for predicting the service life of an unmanned aerial vehicle battery based on a random configuration network.
Background
The lithium ion battery generates current by depending on the flow of internal lithium ions, and has the advantages of high power density, wide working temperature range, long service life and the like. Although improvement and optimization are performed for many years, performance degradation of the lithium ion battery still occurs after long-term use of the lithium ion battery, so that the service Life and the use safety of the lithium ion battery are affected, and even serious property, personnel and other losses are caused, so that it is very necessary to predict the Remaining service Life (Remaining Useful Life, RUL) of the lithium ion battery, so as to realize health management, avoid various operations which damage the battery Life, and reasonably manage and use the lithium ion battery safely and effectively.
The lithium ion Battery internal system is a dynamic system which changes along with time, and the development status of the RUL of the Battery is systematically summarized by Huxiaosong, xu Le, lin xianke, et. Al Battery Life sciences [ J ]. Cross ref (2): 310-346; huxiasong, xule, lin Xiong, etc.; and Battery life prediction [ J ]. Cross ref (2): 310-346) of the university vehicle team on the basis of analyzing external influence factors of the Battery life attenuation and internal physical and chemical factors, and based on three types of algorithms of models, data driving and fusion. The RUL prediction method of the battery based on the model is to establish a battery life degradation mathematical model, such as a physical model or an electrochemical equation of the internal structure of the battery, by using professional related knowledge and experience, and to deeply discuss the mathematical mechanism and the degradation rule in the internal structure of the battery, so that the RUL prediction method has the advantages of stable performance, high prediction precision and the like, but has the defects that the RUL prediction method needs to know the accurate electrochemical and battery physical action equations and is not suitable for off-line detection. The battery RUL prediction method based on data is characterized in that potential relation between implicit data and data is mined by utilizing an algorithm through collecting relevant failure data of battery life degradation, an approximation model of the battery life degradation is established under the support of technologies such as statistical analysis and the like, and the model is extrapolated, so that the prediction of the battery RUL is realized. The fusion-based battery RUL prediction method fully exerts the advantages of various methods, develops the advantages of various methods, better extracts and uses information and data characteristics in data by fusing and integrating models and algorithms, algorithms and algorithms to obtain more excellent network model robustness and prediction accuracy, but the fusion-based method has the problems of high consumption, large data demand and the like and limits the application range and generalization capability of the fusion-based battery RUL prediction method.
The biggest advantage of the data-driven prediction algorithm is that the prediction algorithm can be completed only by a certain amount of degradation data without knowing an accurate model of the battery, so that the data-driven prediction method of the battery RUL is more popular and widely applied, and is slowly the mainstream method for predicting the service life of the battery. The data-driven battery RUL prediction method mainly comprises an artificial neural network, a support vector machine, support vector regression, particle filtering and the like, wherein the artificial neural network aims at simulating the operation of a human brain nervous system, and is greatly different in the battery life prediction problem due to the advantages of strong nonlinear processing capability, self-adaptability, self-learning capability and the like, but the problems of slow training, excessive resource consumption, weak generalization and the like caused by the artificial neural network are slowly shown in the use process, so that the problems become a difficulty which hinders the development of the artificial neural network in the lithium ion battery life prediction direction, and the problems are gradually solved when incremental learning appears.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: the unmanned aerial vehicle battery life prediction method based on the random configuration network is provided to solve the technical problems in the prior art.
The technical scheme adopted by the invention is as follows: a unmanned aerial vehicle battery life prediction method based on a random configuration network comprises the following steps: the capacity of a lithium ion battery for the unmanned aerial vehicle is used as a direct health factor, the discharge voltage of the battery is used as an indirect health factor, a random configuration network is adopted to carry out model training and parameter debugging, and a life prediction model of the lithium ion battery is obtained for prediction.
The random configuration network algorithm comprises the following specific steps:
(1) For a given objective function f:
assume that a single layer feedforward network with L-1 hidden nodes has been established, as shown in (1).
Wherein f is L-1 (x) Represents a hypothetical single-layer feedforward network with L-1 hidden nodes, beta j Representing hidden layer output weights, g j It is shown that the activation function is,
input weights representing hidden layers, b j An input bias representing a hidden layer;
error is expressed as (2)
e L-1 ＝f-f L-1 ＝[e L-1,1 ,...,e L-1,m ] (2)
Wherein e L-1 Representing the current error, f represents the network value currently having L nodes, f L-1 Representing the network value of the last stage with L-1 nodes, e L-1,1 Indicating the error of L-1 hidden layer node to the first output layer node, \8230;, e L-1,m The error of the L-1 hidden layer node to the mth output layer node;
if error e L-1 If the preset error limit is not reached, the algorithm aims to add beta by an incremental learning method L ,g L (wherein g is L Has a weight of w L Offset is b L ) Update f L ＝f L-1 +β L g L Until the error eL-1 reaches a given error limit;
(2) Suppose for
Are all 0 < | g | < b g Given 0 < r < 1 and a sequence of non-negative real numbers mu L In which μ L Satisfy the requirement of
And mu L Less than or equal to (1-r). Hidden node L =1, 2.. Is represented as:
wherein delta L Represents the value of node L, δ L,q A value representing that the Lth hidden layer node corresponds to the q-th output layer node, r is a self-set numerical value, mu L Is a sequence of non-negative real numbers, e L-1,q Representing the error of the qth output layer node corresponding to the L-1 hidden layer node, | | | | | represents the norm calculation;
if the generated random basis function gL satisfies the formula (4)
Wherein e L-1,q Representing the error of the q-th output layer node corresponding to the L-1 th hidden layer node, g L Representing random basis functions, b g Representing the bias, δ, of the random basis function L,q A value representing that the lth hidden layer node corresponds to the qth output layer node;
and the output weight is expressed by formula (5)
Wherein beta is L,q Representing the output weight, e L-1,q Represents the error of the L-1 hidden layer node corresponding to the q output layer node, g L Representing a random basis function, | | | | | initiative norm calculation;
if the calculation is obtained, the conclusion can be drawn
Wherein:
thereby deriving an approximator with strong generalization.
Unlike other methods of maximizing the objective function, the supervision mechanism in equation (4) calculates the optimal weight w L And bias b L Obtaining a suitable new node, thereby further achieving the mathematical condition satisfying the formula (7):
<e L-1 ,g L > 2 /||g L || 2 (7)
wherein e L-1,q Representing the error of the q-th output layer node corresponding to the L-1 th hidden layer node, g L RepresentRandom basis function, | mush | represents norm calculation;
obvious type (8)
β L ＝[β L,1 ,...,β L,m ] T (8)
Wherein beta is L Representing the output weight, beta L,1 Output weights of \ 8230;. Beta.of the L-th hidden layer node corresponding to the 1 st output layer node L,m An output weight representing that the Lth hidden layer node corresponds to the mth output layer node;
analysis by the formula (9)
Wherein beta is L,q Representing the output weight of the qth output layer node corresponding to the lth hidden layer node, e L-1,q Representing the error of the q-th output layer node corresponding to the L-1 th hidden layer node, g L Representing a random basis function, i. representing norm calculation;
the formula (9) is analyzed and evaluated, and is kept unchanged in the network for adding further nodes. However, this approach may result in slow convergence of the construction process. Therefore, the calculation scheme of the output weight is reconsidered, i.e. g is calculated according to (4) j (j =1, 2.. Multidot., L), β is calculated by minimizing the global residual 1 ,β 2 ,...,β L The value of (c).
The establishment of the random configuration network algorithm is divided into two steps: 1) Configuration of hidden node parameters: randomly assigning input weights and offsets to satisfy equation (4); then generating a new hidden node and adding the new hidden node into the current learner model; 2) Calculating an output weight: the system of linear equations is solved by using a least squares method or a regularized version thereof.
The random configuration network algorithm adopts a sigmoid function as a hidden node activation function, wherein the parameter Y = { lambda = { [ lambda ]) 1 :Δλ:λ max For adaptively determining the range of random parameters w and b, selecting a desired error margin e or by a preset maximum implicitNumber of hidden nodes L max To terminate the algorithm, the maximum number of hidden nodes L max Determined using single variable experiments.
The randomly configured network takes the time-varying value of r, in which case r needs to be set very close to 1 (where the value of r is 0 < r < 1, r is an uncertain value, and needs to be adjusted in real time according to the use condition of the algorithm, and the meaning of very close to 1 is similar to the values of 0.99999, 0.99999999 and 0.9999999999999), so as to make the formula (10) approach to zero, thereby making xi as much as possible L ≥0：
When r is set to be close to 1, the ratio of e to the equation (11) becomes larger L-1 2 Much smaller:
δ L ＝(1-r-μ L )e L-1 2 。 (11)
the unmanned aerial vehicle battery life prediction method based on the random configuration network further comprises a reliability analysis method for selecting health factors, wherein the reliability analysis method for selecting the health factors adopts grey correlation degree analysis, and the specific method comprises the following steps: the method comprises the following steps of setting n influencing elements in a gray system, setting an important element X, knowing that other n-1 elements have influence on X in different degrees, calculating the influence degree of the other n-1 elements on X by using multi-element statistical analysis, calculating a GRA value between two data elements by using the following formula, wherein the higher the tightness degree of the two elements is, the higher the GRA value is:
xi in formula (12) i (k) For the calculated gray correlation coefficient of the kth influence element sequence, y (k) is a reference element sequence, namely a battery capacity sequence X; x is the number of i (k) Comparing element sequences, namely n-1 discharge voltage sequences; rho is a self-defined resolution coefficient (generally 0.6); finally, the average of the discharge voltage sequence is determined according to the formula (13)And (3) obtaining a grey correlation value of discharge voltage to battery capacity:
wherein gamma is i Representing a grey correlation value, ξ i (k) Representing the correlation coefficient of element k of the total n elements.
The random configuration network training data uses the root mean square error value as the performance evaluation label of the random configuration network, the calculation of the root mean square error value is as the formula (14), wherein
And y i Predicted capacity and real capacity, respectively:
training data requires removing the first M cycles and the last N cycles of the data, and training the randomly configured network by using the middle L cycles of data.
The invention has the beneficial effects that: compared with the prior art, the invention can effectively avoid the problems of low precision, large training data requirement and the like in the existing data-driven prediction method, fully exerts the characteristics of strong SCN autonomy, high convergence rate, low network cost and the like, has the minimum RMSE value of SCN compared with other neural networks, has lower training loss and better network fitting effect, and is an effective RUL prediction algorithm for the lithium battery.
Drawings
FIG. 1 is a graph of battery voltage versus capacity fade images;
FIG. 2 is a graph of a 100 cycle data fit to a training set;
FIG. 3 is a graph of a 168 cycle data fit to a training set;
FIG. 4 is a comparison graph of a network test fit image of a training set of 100 and 168 cycle data;
fig. 5 is a comparative experimental plot of single variables of γ;
FIG. 6 is a fitting graph of SCN-based battery life prediction;
FIG. 7 is a graph of training RMSE attenuation
FIG. 8 is a training fit graph
FIG. 9 is a graph of a control experiment of SCN and various artificial neural networks.
Detailed Description
The invention is further described with reference to the accompanying drawings and specific embodiments.
The Stochastic Configuration Networks (SCN) is a neural network adopting an incremental learning mode, hidden nodes are added in an incremental mode, inequalities are used for constraining parameters and weights of the Stochastic Configuration nodes, and compared with neural networks such as CNN and ANN, the neural network has the advantages of being strong in network generalization, simple in structure, strong in real-time performance, fast in training and the like. Therefore, aiming at the problems of the neural network in the battery life prediction, the invention takes the battery capacity as a direct health factor and the battery discharge voltage as an indirect health factor, and carries out model training and parameter debugging on SCN by using an NASA lithium ion battery data set to carry out a life prediction experiment of the lithium ion battery; finally, various neural networks such as FNN, CNN, LSTM and the like are compared with the method for verification, the usability and the superiority of the method are proved, scientific basis is provided for the accurate application and popularization of a random configuration algorithm in the battery life prediction, and scientific guarantee is provided for the safety application of the lithium ion battery in various fields such as aerospace, electronic equipment, automobiles and the like.
Example 1: a unmanned aerial vehicle battery life prediction method based on a random configuration network comprises the following steps: the capacity of the lithium ion battery for the unmanned aerial vehicle is used as a direct health factor, the discharge voltage of the battery is used as an indirect health factor, and a random configuration network is adopted to carry out model training and parameter debugging, so that the service life prediction of the lithium ion battery is realized.
The random configuration network algorithm comprises the following specific steps:
(1) For a given objective function f:
assume that a single-layer feedforward network with L-1 hidden nodes has been established, as shown in (1).
Error is shown as (2)
e L-1 ＝f-f L-1 ＝[e L-1,1 ,...,e L-1,m ] (2)
If the error e L-1 If the preset error limit is not reached, the algorithm aims to add beta by an incremental learning method L ,g L (wherein g is L Has a weight of w L Offset is b L ) Update f L ＝f L-1 +β L g L Up to error e L-1 Reaching a given error limit;
(2) Suppose for
Are all 0 < | g | < b g Given 0 < r < 1 and a sequence of non-negative real numbers [ mu ] L In which μ L Satisfy the requirement of
And mu L Less than or equal to (1-r). Hidden node L =1,2.. Is represented as:
if the generated random basis function gL satisfies the formula (4)
And the output weight is expressed by formula (5)
If the calculation is obtained, the conclusion can be drawn
Wherein:
thereby deriving an approximator with strong generalization.
Unlike other methods of maximizing the objective function, the supervision mechanism in equation (4) works by computing the optimal weight w L And bias b L Obtaining a suitable new node, thereby further achieving the mathematical condition satisfying the formula (7):
<e L-1 ,g L > 2 /||g L || 2 (7)
obvious type (8)
β L ＝[β L,1 ,...,β L,m ] T (8)
By passing
The formula (9) is analyzed and evaluated, and is kept unchanged in the network for adding further nodes. However, this approach may result in slow convergence of the construction process. Therefore, the calculation scheme of the output weight is reconsidered, i.e. g is calculated according to (4) j (j =1,2, \8230;, L), β is calculated by minimizing the global residual 1 ,β 2 ,…,β L The value of (c).
The establishment of the random configuration network algorithm comprises two steps: 1) Configuration of hidden node parameters: randomly assigning input weights and deviations to satisfy equation (4); then generating a new hidden node and adding the new hidden node into the current learner model; 2) Calculating an output weight: the system of linear equations is solved by using a least squares method or a regularized version thereof.
The random configuration network algorithm adopts a sigmoid function as a hidden node activation function, wherein the parameter Y = { lambda = 1 :Δλ:λ max For the self-adaptive determination of the range of random parameters w and b, selecting a desired error margin epsilon or a preset maximum number L of hidden nodes max To terminate the algorithm, the maximum number of hidden nodes L max Determined using single variable experiments.
The randomly configured network takes a time-varying value of r, in which case r needs to be set very close to 1 to drive equation (10) towards zero, so that ξ is made as much as possible L ≥0：
When r is set to be close to 1, the ratio of e to the equation (11) becomes larger L-1 2 Much smaller:
δ L ＝(1-r-μ L )e L-1 2 。 (11)
the unmanned aerial vehicle battery life prediction method based on the random configuration network further comprises a reliability analysis method for selecting health factors, wherein the reliability analysis method for selecting the health factors adopts grey correlation degree analysis, and the specific method comprises the following steps: the gray system is provided with n influencing elements and an important element X, the influence of other n-1 elements on X is known to have different degrees, the influence degree of the other n-1 elements on X is calculated by using multi-element statistical analysis, the GRA value between two data elements is calculated by using the following formula, and the GRA value is calculated to be higher if the tightness of the two elements is higher:
xi in formula (12) i (k) For the calculated gray correlation coefficient of the kth influence element sequence, y (k) is a reference element sequence, namely a battery capacity sequence X; x is a radical of a fluorine atom i (k) In order to compare the sequences of elements,namely n-1 discharge voltage sequences; rho is a self-defined resolution coefficient (generally 0.6); and finally, calculating the average value of the discharge voltage sequence according to a formula (13) to obtain a grey correlation value of the discharge voltage to the battery capacity:
the random configuration network training data uses the root mean square error value as the performance evaluation label of the random configuration network, the calculation of the root mean square error value is as the formula (14), wherein
And y i Predicted capacity and real capacity, respectively:
training data requires removing the first M cycles and the last N cycles of data, and training the randomly configured network with the middle L cycles of data.
The method takes the battery capacity as a direct health factor and the battery discharge voltage as an indirect health factor, distributes random parameters through inequality constraint, adaptively selects the range of the random parameters, and establishes a prediction model of the residual life of the lithium battery; the prediction model was trained using the NASA battery dataset.
To illustrate the effects of the present invention, the following simulations were performed:
selecting an experimental data set: data of three batteries B0005, B0006 and B0007 in a NASA battery data set PCoE databases are used as input data of the SCN to carry out a cross validation experiment, a discharge voltage data sequence and a battery capacity data sequence of one battery are used for training the SCN, and discharge voltage data of the other two batteries are respectively used for predicting the self capacity, so that the residual effective life of the lithium ion battery is predicted. The PCoE data set carries out charging and discharging treatment on the battery at different temperatures, records data such as charging and discharging voltage, current, temperature, impedance and the like of the battery, is a good data set for a data driving algorithm, and the specific information of NASA battery data is listed in Table 1.
TABLE 1 NASA Battery data
The selection of the network parameters comprises the selection of health factors, training data and training parameters:
(1) Health factors: the battery discharge voltage is selected as an indirect health factor, the battery capacity is used as a health factor, the battery discharge voltage is used as network input data, a data-driven battery capacity prediction model is built, and then the battery capacity is predicted and the health state of the lithium ion battery is evaluated. And Grey correlation Analysis (GRA) was used to prove the reliability of the health factor selection. The grey relevance analysis is used for calculating the influence degree of other elements in the system on a certain element. Assuming that there are n influencing elements in a gray system and an important element X, knowing that the other n-1 influencing elements all have different degrees of influence on X, the strength of the influence of the other n-1 influencing elements on X can be calculated by using multi-element statistical analysis. The GRA value between two data elements can be calculated using equation (12), with the higher the closeness of the two elements, the higher the GRA value that is calculated. Xi in equation (12) i (k) For the calculated gray correlation coefficient of the kth influencing element sequence, y (k) is the reference element sequence, i.e. the above-mentioned element sequence X of interest, which in the present invention is the battery capacity sequence; x is the number of i (k) For comparison of the element sequences, i.e. the N-1 other element sequences mentioned above, in the present case the discharge voltage sequences; ρ is the custom resolution coefficient (typically 0.6). And finally, calculating the average value of the discharge voltage sequence according to a formula (13) to obtain a grey correlation value of the discharge voltage to the battery capacity.
Fig. 1 is a voltage vol and capacity cap fade contrast image of battery B0007 after pretreatment. The relevance of the two system elements can be more intuitively observed according to the discharge voltage and the attenuation image of the battery capacity.
(2) Training data: root-Mean-Square Error (RMSE) values were chosen as performance evaluation labels for the SCNs used, each RMSE value being the Mean of 20 experiments in the following. The RMSE is calculated as shown in formula (14), wherein
And y i Predicted capacity and real capacity, respectively; considering the unstable voltage condition of the lithium ion battery caused by chemical reaction at the beginning and the end of charging and discharging, comparing 168 times of cycle data of the battery, we remove the first 30 cycles and the end 38 cycles of the data, and train the SCN with the middle 100 times of cycle data. The following are the resulting network training RMSE, training set fitted images and test set fitted image comparisons, where the network experiment used a combination of B0005 cells as the training set and B0007 cells as the test set.
TABLE 2 network RMSE values
| Number of | RMSE | |
| 100 times (twice) | 0.0192 | |
| 168 times (a) | 0.0495 |
Comparing the 100-time cycle data with the 168-time cycle data to respectively make a test fitting image of a network training data set, and the RMSE mean value in table 2, it can be seen that removing the first 30-time cycle and the last 38-time cycle of the voltage data can effectively reduce the influence of unstable voltage data on network training, thereby enhancing the network training effect, and therefore, the 100-time cycle data of the battery discharge voltage, which removes the first 30-time cycle and the last 38-time cycle, is finally adopted as the training input data of the SCN.
(3) Training parameters: after the training set data volume is determined, next, battery B0005 is used as a training set battery B0007 as a test set, and a single variable experiment is used for finding out the optimal L max I.e. the maximum number of nodes in the hidden layer. As can be seen from Table 3, the network training effect is enhanced with the increase of the maximum node number of the hidden layer, where the maximum node number is L max The network training RMSE is 0.0192 at 100, and the network fitting effect is the best at this moment. The article also sets the node number to 200, 500 and 1000 respectively to carry out comparison experiments, and the experimental result is that the network training is all at L max When the number is less than or equal to 100, the training is automatically cut off, and the maximum node number L of the hidden layer is shown max The value of (c) is chosen to be 100 for best network training.
TABLE 3 network L max Number and corresponding RMSE
| L max Number of | RMSE |
| 5 | 0.0270 |
| 10 | 0.0261 |
| 50 | 0.0243 |
| 100 | 0.0192 |
In SCN, γ is an important parameter. Y is a series of positive scalars used to randomly configure the network weights ω and offsets b. From 0.5, upsizing to 10000 was carried out, and a single variable experiment was carried out by using a combination of battery B0005 as a training set and battery B0007 as a test set, and it was found that upsizing of upsizing also resulted in a strong and humanized overfitting condition of good performance of the training set and the test set, as shown in fig. 5. Considering the results of multiple experiments, the γ was finally set to a maximum value of 200 and a minimum value of 0.5.
Simulation result and comparative analysis:
after several comparative experiments, the final parameter settings for SCN are shown in table 4.
TABLE 4 SCN parameter settings
| Maximum number of hidden |
100 |
| Margin of error | 0.00 |
| Number of |
100 |
| Υ | Maximum 200, minimum 0.5 |
| r | Increasing by 0.9-0.99999 |
| batch size | 1 |
Fig. 6 is a battery life prediction fit image of SCN, where the black curve is the true decay value of battery life and the red curve is the predicted value of SCN. It is obvious that the SCN fitting curve fluctuates slightly only in the early stage of the network, but because the nodes are fewer at the beginning of the network construction, the curve fitting condition gradually becomes good as the network nodes increase.
Fig. 7 is a battery training RMSE attenuation image with the battery data of B0005, B0006, B0007 as the training data sets, respectively.
Fig. 8 is a training fit image of three pieces of battery data, where the solid line is the true capacity fade image of the battery and the dashed line is the model predicted fade image.
A plurality of commonly used battery RUL prediction algorithms and a battery RUL prediction algorithm based on SCN are selected to carry out comparison experiments so as to verify the usability and effectiveness of the randomly configured network for predicting the residual life of the battery. A comparison experiment is carried out on two Feedforward Neural Networks (FNN) with different network structures, two Convolutional Neural Networks (CNN) with different network structures and a long-time and short-time memory network LSTM; data set B0005 data was used as training set B0007 data as a combination of test sets. The first full-connection layer of FNN1 adopts output dimension 10 and normal distribution initialization weight, the activation layer adopts a Leaky ReLU activation function, the second full-connection layer adopts output dimension 1 and the activation function adopts a linear activation function, the network optimizer adopts an Adam optimizer, and a Dropout layer is not arranged; the first full-connection layer of FNN2 adopts output dimension 40 and normal distribution initialization weight, the activation layer adopts a Leaky ReLU activation function, the second full-connection layer adopts output dimension 1 and the activation function adopts a linear activation function, the network optimizer adopts an Adam optimizer, and a Dropout layer is arranged between the two full-connection layers; CNN1 adopts a small convolution kernel and a Leaky ReLU activation function, CNN2 adopts a large convolution kernel and a Leaky ReLU activation function, and Dropout layers are respectively added between a first layer of convolution and a second layer of convolution and between a second layer of convolution and a full connection layer. The LSTM network also employs the leak ReLU activation function and adds a Dropout layer.
From table 5 comparing RMSE means of the SCN algorithm and other algorithms, it can be seen that the training rms error of SCN is minimal. The test set fits most well on SCN. Compared with other artificial neural networks, the training RMSE value of SCN is the smallest, with a value of 0.0192, 43% of FNN1, 41% of FNN2, 42% of CNN1, and 34% of CNN2, respectively.
TABLE 5 network RMSE value comparison table
| Training algorithm | RMSE mean value |
| FNN1 | 0.0445 |
| FNN2 | 0.0462 |
| CNN1 | 0.0460 |
| CNN2 | 0.0553 |
| LSTM | 0.0254 |
| SCN | 0.0186 |
Figure 9 is an image of a control experiment fit of SCN to various neural networks, where the black bold line is the true value and the red bold line is the SCN fit value. Obviously, the overall fitting condition of the SCN is better than that of other networks, and the fitting effect of the network stability is slightly worse than that of the LSTM network due to less hidden nodes in the early prediction stage and is similar to that of the CNN 1; after 50 cycles, the fitting effect of SCN is significantly better than the rest of the network, almost overlapping with the real life decay image. Therefore, the lithium ion battery service life prediction method based on the random configuration network has practicability, and the accuracy is superior to that of various existing neural network methods.
And (4) simulation conclusion: the invention provides a lithium ion battery RUL prediction method based on a random configuration network. The battery RUL prediction experiment is carried out by utilizing an incremental random neural network SCN, firstly, training set data of the SCN is determined, the data set adopts a NASA PCoE battery data set, firstly, 100 cycles of stable data are intercepted to serve as training data of the SCN, secondly, comparison experiments are carried out to successively determine network parameter values, and finally, two FNN networks, two CNN networks, an LSTM network and the SCN are used for carrying out comparison experiments.
The above description is only an embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily think of the changes or substitutions within the technical scope of the present invention, and therefore the scope of the present invention shall be subject to the protection scope of the claims.
Claims (4)
1. A method for predicting the service life of an unmanned aerial vehicle battery based on a random configuration network is characterized by comprising the following steps: the method comprises the following steps: taking the capacity of a lithium ion battery for the unmanned aerial vehicle as a direct health factor and the discharge voltage of the battery as an indirect health factor, and performing model training and parameter debugging by adopting a random configuration network to obtain a service life prediction model of the lithium ion battery for prediction; the random configuration network algorithm comprises the following specific steps:
(1) For a given objective function f:
suppose a single-layer feedforward network with L-1 hidden nodes is established, as shown in equation (1):
wherein f is L-1 (x) Represents a hypothetical single-layer feedforward network with L-1 hidden nodes, beta j Representing hidden layer output weights, g j It is shown that the activation function is,
input weights representing hidden layers, b j An input bias representing a hidden layer;
error is expressed as formula (2)
e L-1 ＝f-f L-1 ＝[e L-1,1 ,...,e L-1,m ] (2)
Wherein e L-1 Representing the current error, f represents the network value currently having L nodes, f L-1 Representing the network value of the last stage with L-1 nodes, e L-1,1 Indicating the error of the L-1 hidden layer node to the first output layer node, \ 8230;, e L-1,m The error of the L-1 hidden layer node to the mth output layer node;
if the error e L-1 If the preset error limit is not reached, adding beta by using an incremental learning method L ,g L Wherein g is L Has a weight of w L Offset is b L Update f L ＝f L-1 +β L g L Up to an error e L-1 Reaching a given error limit;
(2) Suppose for
Are all 0 < | g | < b g Given 0 < r < 1 and a sequence of non-negative real numbers [ mu ] L In which μ L Satisfy the requirement of
And mu L ≦ (1-r), hidden node L =1, 2.
Wherein delta L Represents the value of node L, δ L,q A value representing the q output layer node corresponding to the L hidden layer node, r is a self-set numerical value, mu L Is a sequence of non-negative real numbers, e L-1,q Representing the error of the qth output layer node corresponding to the L-1 hidden layer node, | | | | | represents the norm calculation;
if the random basis function g is generated L Satisfaction formula (4)
Wherein e L-1,q Representing the error of the q-th output layer node corresponding to the L-1 th hidden layer node, g L Representing random basis functions, b g Representing the bias, δ, of the random basis function L,q A value representing that the lth hidden layer node corresponds to the qth output layer node;
and the output weight is represented by formula (5)
Wherein beta is L,q Represents the output weight, e L-1,q Representing the error of the q-th output layer node corresponding to the L-1 th hidden layer node, g L Representing a random basis function, i. representing norm calculation;
if the calculation is obtained, the conclusion is reached
Wherein:
the supervision mechanism in equation (4) calculates the optimal weight w L And bias b L Obtaining a proper newly added node to meet the mathematical condition of the formula (7):
<e L-1 ,g L > 2 /||g L || 2 (7)
wherein e L-1,q Representing the error of the q-th output layer node corresponding to the L-1 th hidden layer node, g L Representing a random basis function, i. representing norm calculation;
obvious type (8)
β L ＝[β L,1 ,...,β L,m ] T (8)
Wherein beta is L Representing the output weight, beta L,1 Output weights, \ 8230, and beta, indicating that the Lth hidden layer node corresponds to the 1 st output layer node L,m Representing the output weight of the Lth hidden layer node corresponding to the mth output layer node;
analysis by formula (9):
wherein beta is L,q Representing the output weight of the qth output layer node corresponding to the lth hidden layer node, e L-1,q Represents the error of the L-1 hidden layer node corresponding to the q output layer node, g L Representing a random basis function, | | | | | initiative norm calculation;
the formula (9) is analyzed and evaluated, and g is obtained by calculating the calculation scheme of the output weight according to the formula (4) j (j =1, 2.. Multidot., L), β is calculated by minimizing the global residual 1 ,β 2 ,…,β L A value of (d);
the establishment of the random configuration network algorithm is divided into two steps: 1) Configuration of hidden node parameters: randomly assigning input weights and deviations to satisfy equation (4); then generating a new hidden node and adding the new hidden node into the current learner model; 2) Calculating an output weight: solving a system of linear equations by using a least squares method or a regularized version thereof;
the random configuration network algorithm adopts a sigmoid function as a hidden node activation function, wherein the parameter Y = { lambda = { [ lambda ]) 1 :Δλ:λ max For the self-adaptive determination of the range of random parameters w and b, selecting a desired error margin epsilon or a preset maximum number L of hidden nodes max To terminate the algorithm;
the randomly configured network takes a time-varying value of r, which in this case needs to be set very close to 1 to get equation (10) to approach zero, making ξ L ≥0：
When r is set to be close to 1, the ratio of e to the equation (11) becomes larger L-1 2 Small:
δ L ＝(1-r-μ L )e L-1 2 (11)。
2. the unmanned aerial vehicle battery life prediction method based on the random configuration network as claimed in claim 1, wherein: the method also comprises a reliability analysis method for selecting the health factors, wherein the reliability analysis method for selecting the health factors adopts grey correlation degree analysis, and the specific method comprises the following steps: setting n influencing elements in a gray system, setting an important element X, knowing that other n-1 elements have influence on X in different degrees, calculating the influence degree of the other n-1 elements on X by using multi-element statistical analysis, and calculating the GRA value between two data elements by using the following formula:
xi in the formula (12) i (k) For the calculated k-th sequence of influencing elements in greyThe correlation coefficient, y (k) is a reference element sequence, namely a battery capacity sequence X; x is the number of i (k) Comparing element sequences, namely n-1 discharge voltage sequences; rho is a self-defined resolution coefficient; and finally, calculating the average value of the discharge voltage sequence according to a formula (13) to obtain a grey correlation value of the discharge voltage to the battery capacity:
wherein gamma is i Representing a grey correlation value, ξ i (k) Representing the correlation coefficient of element k of the total n elements.
3. The method for predicting the life of the unmanned aerial vehicle battery based on the randomly configured network as claimed in claim 1, wherein: the random configuration network training data uses the root mean square error value as the performance evaluation label of the random configuration network, the calculation of the root mean square error value is as the formula (14), wherein
And y i Predicted capacity and real capacity, respectively:
wherein
y i Respectively representing a predicted value and a true value;
training data requires removing the first M cycles and the last N cycles of the data, and training the randomly configured network by using the middle L cycles of data.
4. The unmanned aerial vehicle battery life prediction method based on the random configuration network as claimed in claim 1, wherein: number of large hidden nodes L max Determined using single variable experiments.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110371138.2A CN113094989B (en) | 2021-04-07 | 2021-04-07 | Unmanned aerial vehicle battery life prediction method based on random configuration network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110371138.2A CN113094989B (en) | 2021-04-07 | 2021-04-07 | Unmanned aerial vehicle battery life prediction method based on random configuration network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113094989A CN113094989A (en) | 2021-07-09 |
| CN113094989B true CN113094989B (en) | 2022-12-09 |
Family
ID=76674324
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110371138.2A Active CN113094989B (en) | 2021-04-07 | 2021-04-07 | Unmanned aerial vehicle battery life prediction method based on random configuration network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113094989B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113777496B (en) * | 2021-09-06 | 2023-10-24 | 北京化工大学 | Lithium ion battery residual life prediction method based on time convolution neural network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP5815134B2 (en) * | 2013-08-28 | 2015-11-17 | 株式会社日立製作所 | Li-ion battery life prediction method and system |
| CN109061504A (en) * | 2018-08-28 | 2018-12-21 | 中北大学 | Same type difference lithium ion battery remaining life prediction technique and system |
| CN110568374A (en) * | 2019-09-02 | 2019-12-13 | 东北电力大学 | Lithium ion battery residual service life prediction method based on consideration of internal and external characteristics |
| CN111443294A (en) * | 2020-04-10 | 2020-07-24 | 华东理工大学 | Method and device for indirectly predicting remaining life of lithium ion battery |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108549036A (en) * | 2018-05-03 | 2018-09-18 | 太原理工大学 | Ferric phosphate lithium cell life-span prediction method based on MIV and SVM models |
-
2021
- 2021-04-07 CN CN202110371138.2A patent/CN113094989B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP5815134B2 (en) * | 2013-08-28 | 2015-11-17 | 株式会社日立製作所 | Li-ion battery life prediction method and system |
| CN109061504A (en) * | 2018-08-28 | 2018-12-21 | 中北大学 | Same type difference lithium ion battery remaining life prediction technique and system |
| CN110568374A (en) * | 2019-09-02 | 2019-12-13 | 东北电力大学 | Lithium ion battery residual service life prediction method based on consideration of internal and external characteristics |
| CN111443294A (en) * | 2020-04-10 | 2020-07-24 | 华东理工大学 | Method and device for indirectly predicting remaining life of lithium ion battery |
Non-Patent Citations (2)
| Title |
|---|
| Reliable Life Prediction and Evaluation Analysis of Lithium-ion Battery Based on Long-Short Term Memory Model;Xuejiao Zhao et al.;《2019 IEEE 19th International Conference on Software Quality, Reliability and Security Companion (QRS-C)》;20191007;507-509 * |
| 无人机故障诊断技术研究进展概述;苗建国 等;《仪器仪表学报》;20200930(第9期);56-69 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113094989A (en) | 2021-07-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109991542B (en) | Lithium ion battery residual life prediction method based on WDE optimization LSTM network | |
| CN112241608B (en) | Lithium battery life prediction method based on LSTM network and transfer learning | |
| CN109659933B (en) | Electric energy quality prediction method for power distribution network with distributed power supply based on deep learning model | |
| CN111539515B (en) | Complex equipment maintenance decision method based on fault prediction | |
| US20190265768A1 (en) | Method, system and storage medium for predicting power load probability density based on deep learning | |
| CN112036084B (en) | Similar product life migration screening method and system | |
| CN111736084B (en) | Valve-regulated lead-acid storage battery health state prediction method based on improved LSTM neural network | |
| Wang et al. | Rapid ultracapacitor life prediction with a convolutional neural network | |
| CN113361197B (en) | Method and system for predicting remaining service life of lithium battery | |
| CN113406521B (en) | Lithium battery health state online estimation method based on feature analysis | |
| CN112734002B (en) | Service life prediction method based on data layer and model layer joint transfer learning | |
| CN113156325A (en) | Method for estimating state of health of battery | |
| CN111680786B (en) | Time sequence prediction method based on improved weight gating unit | |
| CN112381320A (en) | Community personnel flow prediction method, system, storage medium and equipment | |
| CN113094981B (en) | Lithium ion battery reliability evaluation method based on grey neural network model and self-service method | |
| CN114035098A (en) | Lithium battery health state prediction method integrating future working condition information and historical state information | |
| Venugopal et al. | Analysis of optimal machine learning approach for battery life estimation of Li-ion cell | |
| CN113094989B (en) | Unmanned aerial vehicle battery life prediction method based on random configuration network | |
| CN112115636A (en) | Method and system for predicting insulation aging life of power cable in advance | |
| CN113361692B (en) | Lithium battery remaining life combined prediction method | |
| WO2024023211A1 (en) | A method for characterizing the evolution of state of health of a device with duration of operation | |
| CN113884936B (en) | ISSA coupling DELM-based lithium ion battery health state prediction method | |
| CN111061708A (en) | Electric energy prediction and restoration method based on LSTM neural network | |
| CN115221731A (en) | Transformer life evaluation method based on data fusion and Wiener model | |
| CN115453871A (en) | Non-linear system modeling method based on IDE extended multidimensional Taylor network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |