CN113805064B - Lithium ion battery pack health state prediction method based on deep learning - Google Patents

Abstract

The invention discloses a lithium ion battery pack health state prediction method based on deep learning, which comprises the following steps of: s1, extracting the discharge characteristic parameters of the lithium ion battery pack, and carrying out correlation analysis on the discharge characteristic parameters; s2, obtaining an original feature data set according to the result of the correlation analysis, and performing feature dimension reduction; s3, constructing a node degradation data prediction model; s4, inputting the dimensionality reduction feature set into a node degradation data prediction model to obtain node degradation prediction data of the lithium ion battery pack; and S5, converting the node degradation prediction data into node state distribution prediction data, and acquiring the health state prediction result of the lithium ion battery pack based on the node state distribution prediction data. The method can realize accurate prediction of the health state of the lithium ion battery pack according to the processes of feature extraction, correlation analysis, feature dimension reduction, node state distribution prediction and system health state prediction.

Description

A method for predicting the state of health of lithium-ion battery packs based on deep learning

technical field

The invention relates to the technical field of lithium ion battery state prediction, in particular to a method for predicting the state of health of lithium ion battery packs based on deep learning.

Background technique

Lithium-ion batteries are widely used in the energy power systems of electric vehicles, hybrid ships, drones and other equipment due to their high specific energy, long cycle life, and wide operating temperature range. Due to the low voltage of the single lithium-ion battery, the energy supply needs to be carried out in the form of a battery pack during the power supply process. Therefore, in actual use, hundreds of batteries are connected through various composite connections and various packaged components. Group methods compose complex energy supply systems. From single cells to combined and connected battery packs to complex battery packs, a complete energy power system is formed, which is a typical multi-level system.

On the other hand, in the actual use of lithium-ion battery packs, due to the strong interdependence between batteries, coupled with the initial differences of batteries and the influence of environmental stress, the consistency of the battery pack is difficult to maintain. Under the influence of interdependence, this inconsistency will increasingly aggravate the non-equilibrium effect of the battery pack, which makes the degradation process and characteristic parameters between batteries complex and changeable, and the relationship between them changes dynamically, which is difficult to predict accurately.

SUMMARY OF THE INVENTION

In view of the above problems, the present invention proposes a lithium-ion battery pack health state prediction method based on deep learning, so as to solve the technical problems existing in the prior art, which can be based on experiment-feature extraction-correlation analysis-feature dimensionality reduction-node degradation data prediction —Node state distribution prediction—The process of system health state prediction, the corresponding analysis work has been carried out completely, and the accurate prediction of the health state of the lithium-ion battery pack has been realized.

In order to achieve the above purpose, the present invention provides the following solutions: the present invention provides a method for predicting the state of health of a lithium-ion battery pack based on deep learning, including the following steps:

S1, extracting the discharge characteristic parameters of the lithium-ion battery pack, and performing a correlation analysis on the discharge characteristic parameters;

S2, obtaining an original feature data set according to the result of the correlation analysis, and performing feature dimension reduction on the original feature data set to obtain a dimension reduction feature set;

S3. Acquire the time sequence of the inter-cycle discharge condition of the lithium-ion battery pack according to the correlation between the inter-cycle discharge capacities, and build a node degradation data prediction model based on the time sequence of the inter-cycle discharge condition of the lithium-ion battery pack;

S4. Input the dimensionality reduction feature set into the node degradation data prediction model, and obtain node degradation prediction data of the lithium-ion battery pack;

S5. Convert the node degradation prediction data into node state distribution prediction data, and obtain a lithium-ion battery pack health state prediction result based on the node state distribution prediction data.

Preferably, the discharge characteristic parameters in the S1 include discharge current, discharge voltage, discharge temperature, and discharge capacity data.

Preferably, the process of the correlation analysis in the S1 is as follows: calculating the mean, median and variance of the discharge characteristic parameters, completing statistical analysis, and performing relevant visualization operations according to the results of the statistical analysis to complete the correlation Sexual Analysis.

Preferably, the acquisition process of the original feature data set in S2 is: quantifying the result of the correlation analysis, acquiring the correlation coefficient matrix of the discharge characteristic parameters, and removing all the correlation coefficients less than 0.1 according to the correlation coefficient matrix. Describe the discharge characteristic parameters to obtain the original characteristic data set.

Preferably, the process of feature dimensionality reduction in the S2 includes the following steps:

S2.1, performing grid processing on the original feature data set to obtain a grid feature matrix;

S2.2, build a convolutional neural network model based on the grid feature matrix, and perform model optimization training on the convolutional neural network model;

S2.3. Obtain dimension reduction feature parameters based on the convolutional neural network model after optimization and training of the model, and complete feature dimension reduction.

Preferably, the construction of the node degradation data prediction model in S3 includes the following steps:

S3.1. Acquire a cyclic degradation data set of the lithium-ion battery pack, and obtain a model data set based on the cyclic degradation data set;

S3.2, construct the node degradation data prediction model;

S3.3. Perform parameter optimization training on the node degradation data prediction model based on the model data set, and complete the construction of the node degradation data prediction model.

Preferably, the process of obtaining the predicted result of the state of health of the lithium-ion battery pack in S5 includes the following steps:

S5.1. Perform discretization processing on the node degradation prediction data to complete the multi-state division of the degradation process of the lithium-ion battery pack;

S5.2, build a Bayesian neural network hybrid model;

S5.3. Obtain node state distribution prediction data based on the Bayesian neural network hybrid model, and obtain a lithium-ion battery pack health state prediction result according to the node state distribution prediction data.

The present invention discloses the following technical effects:

The invention uses the multi-level system prediction modeling method to establish a Bayesian network model of the battery system through the multi-feature parameters of battery charge and discharge, and according to feature extraction-correlation analysis-feature dimension reduction-node state distribution prediction-system health In the process of state prediction, the corresponding analysis work is completely carried out, and the accurate prediction of the health state of the lithium-ion battery pack is realized.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the present invention. In the embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative labor.

Fig. 1 is the method flow chart of the present invention;

2 is a schematic diagram of a battery pack test platform according to an embodiment of the present invention;

3 is a schematic structural diagram of a parallel series battery pack according to an embodiment of the present invention;

4 is a schematic diagram of a cycle section of an embodiment of the present invention; wherein, (a) is the 5th cycle of discharge; (b) is the 295th cycle of discharge;

5 is a schematic diagram of feature dimension reduction according to an embodiment of the present invention;

FIG. 6 is an integrated node state probability distribution prediction curve according to an embodiment of the present invention;

7 is a Bayesian network model of a battery pack system according to an embodiment of the present invention;

FIG. 8 is a state prediction result of a lithium-ion battery pack system according to an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

In order to make the above objects, features and advantages of the present invention more clearly understood, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

Referring to FIG. 1 , this embodiment provides a method for predicting the state of health of a lithium-ion battery pack based on deep learning, including the following steps:

S1, extracting discharge characteristic parameters of the lithium-ion battery pack, and performing a correlation analysis on the discharge characteristic parameters.

The discharge characteristic parameters include discharge current, discharge voltage, discharge temperature, and discharge capacity data. The process of correlation analysis is as follows: calculate the mean, median and variance of discharge characteristic parameters, complete statistical analysis, and perform relevant visualization operations according to the results of statistical analysis to complete correlation analysis.

S2. Obtain an original feature data set according to the result of the correlation analysis, and perform feature dimension reduction on the original feature data set through a convolutional neural network to obtain a dimension reduction feature set.

Among them, the acquisition process of the original feature data set is as follows: quantify the results of the correlation analysis, obtain the correlation coefficient matrix of the discharge characteristic parameters, and remove the discharge characteristic parameters whose correlation coefficient is less than 0.1 according to the correlation coefficient matrix to obtain the original feature data set.

The process of feature dimensionality reduction includes the following steps:

S2.1. Perform grid processing on the original feature data set to obtain a grid feature matrix;

S2.2. Construct a convolutional neural network model based on the grid feature matrix, and perform model optimization training on the convolutional neural network model;

S2.3. Obtain dimension reduction feature parameters based on the convolutional neural network model after model optimization and training, and complete feature dimension reduction.

S3. Obtain the time series of the discharge conditions between cycles of the lithium-ion battery pack according to the correlation between the discharge capacities between cycles, and build a node degradation data prediction model based on the time series of the discharge conditions of the lithium-ion battery packs between cycles.

The construction of the node degradation data prediction model includes the following steps:

S3.1. Obtain the cyclic degradation data set of the lithium-ion battery pack, and obtain the model data set based on the cyclic degradation data set;

S3.2. Build a node degradation data prediction model;

S3.3. Perform parameter optimization training on the node degradation data prediction model based on the model data set, and complete the construction of the node degradation data prediction model.

S4. Input the dimension reduction feature set into the node degradation data prediction model, and obtain node degradation prediction data of the lithium-ion battery pack.

S5. Convert the node degradation prediction data into node state distribution prediction data, and obtain a lithium-ion battery pack health state prediction result based on the node state distribution prediction data.

S5.1. Discretize the node degradation prediction data to complete the multi-state division of the degradation process of the lithium-ion battery pack;

S5.2, build a Bayesian neural network hybrid model based on deep learning and program implementation framework;

S5.3. Obtain the node state distribution prediction data based on the Bayesian neural network hybrid model, and obtain the lithium-ion battery pack health state prediction result according to the node state distribution prediction data.

In the cycle degradation test of the lithium-ion battery pack in this embodiment, the tested battery adopts a commercial ICR18650 lithium-ion battery, and four cells are connected in parallel and then in series to form a battery pack. The cycle degradation test platform and test system of the battery pack are shown in the figure. 2 shown. A schematic diagram of the structure of a parallel-connected battery pack is shown in Figure 3, where Cell_1, Cell_2, Cell_3, and Cell_4 are the numbers of the single cells.

In order to evaluate and predict the health state of the battery pack, the degradation test mainly evaluates the performance of each discharge characteristic parameter of the battery pack. The corresponding measurement parameters include the discharge current, discharge voltage, discharge temperature and other characteristic parameters of the battery pack and internal single cells. The overall cycle discharge capacity of the battery pack is measured as an indicator of the state of health of the battery pack. For the transformation content of the degradation data to the state distribution, the target objects are all parallel module integration nodes composed of Cell_1 and Cell_2 (hereinafter referred to as parallel modules).

The parallel module cycle degradation test data set contains 320 cycles of discharge data. The characteristic parameters include the charge and discharge current, charge and discharge voltage, charge and discharge temperature of the parallel module and the single battery inside the parallel module in each cycle, as well as the overall discharge capacity of the parallel module. Data, a schematic diagram of the circulation profile of the parallel module is shown in Figure 4.

In the experiment, the interdependent nature of the parallel modules in the discharge stage is particularly obvious, and with the progress of cycle degradation, the discharge characteristics of the battery pack are significantly different, and the non-equilibrium effect is obviously strengthened. The discharge characteristics can reflect the current discharge performance of the battery pack, which can be very good. As an input parameter of the state of health of the battery pack, and the discharge capacity is the main indicator to reflect the state of health of the battery pack. On this basis, the discharge phase of the cycle profile of the parallel module is intercepted, and the correlation between the characteristics of the discharge phase and the change of the overall health state of the system is analyzed.

The characteristic parameters also change significantly with the progress of the degradation state: with the progress of the cyclic degradation process, the fluctuation of the current parameters increases significantly, the inconsistency between cells is also significantly strengthened, and the influence of the inconsistency also spreads from the end of discharge stage to The entire discharge process; for the voltage parameters, due to the characteristics of the parallel structure, the changes in the early and late stages of degradation are not obvious. It can be initially seen that with the increase of the battery degradation fluctuation, the correlation with the state of the parallel module needs to be further analyzed; the temperature parameters are also Significant changes occur, firstly, the temperature mean value is significantly increased, secondly, the inconsistency is significantly strengthened, and the difference between the two batteries increases. On the other hand, it can also be seen from the cross-sectional views of each discharge feature that the change of the battery state of health also has certain positional characteristics in the change of the characteristic parameters. For example, from the current and temperature data, the degradation process of the battery is mainly Diffusion forward from the end of discharge to capture such feature information will improve the effect of feature dimensionality reduction, which is also one of the main functions of convolution kernels in convolutional neural networks.

In addition, the statistical characteristics of the discharge parameters can also effectively reflect the health status of the battery pack. For example, the average temperature can reflect the change trend of the internal resistance of the battery, and the variance of the current and voltage can reflect the stability of the battery discharge. For the above measured multiple parallel modules The discharge characteristic parameters were statistically analyzed, the mean, median, and variance of each parameter were calculated, and related visualization operations were performed.

There is a strong correlation between each statistical parameter and the degradation process of the parallel module. In order to quantify the correlation, the feature parameter correlation coefficient matrix is calculated. According to the correlation matrix, the features with a correlation coefficient less than 0.1 are eliminated, and the remaining features The discharge current, discharge voltage, and discharge temperature are input as the original feature set, as shown in Table 1 and Table 2.

Table 1

Discharge characteristics Feature data type Cell_1 discharge current Tensor Cell_1 discharge voltage Tensor Cell_1 discharge temperature Tensor Cell_2 discharge current Tensor Cell_2 discharge voltage Tensor Cell_2 discharge temperature Tensor Cell_1 median discharge current scalar Cell_1 discharge current average scalar Cell_1 discharge current variance scalar Cell_2 median discharge current scalar

Table 2

Discharge characteristics Feature data type Cell_2 discharge current average scalar Cell_2 discharge current variance scalar Cell_1 median discharge voltage scalar Cell_1 discharge voltage average scalar Cell_2 median discharge voltage scalar Cell_2 discharge voltage average scalar Cell_1 discharge temperature variance scalar Cell_1 average discharge temperature scalar Cell_2 discharge temperature variance scalar Cell_2 average discharge temperature scalar

Further analysis of the correlation coefficient matrix shows that the characteristic parameters basically have a high correlation with the state of health of the parallel module, indicating that the extracted battery charge and discharge characteristics can be used to predict the state of health of the battery pack, but at the same time there is also a temperature Median, voltage variance value and other parameters with small correlation (<0.1), and it is also found that there is a strong correlation between multi-feature parameters. If the model is directly input, the weight will be shifted and the prediction accuracy of the model will be reduced. , which brings about the overlapping problem of multi-feature data, so further feature dimensionality reduction work is needed.

In order to obtain higher quality input features to improve the model prediction accuracy, for the original feature dataset obtained above, based on the feature dimension reduction method, the convolutional neural network is used to reduce the feature dimension of the discharge feature of the lithium-ion battery pack, as shown in Figure 5 shown.

In order to analyze and evaluate the results of model feature dimensionality reduction, the health state of the parallel module is divided into multiple states, and health, degradation, and failure are used as prediction labels, and Adam algorithm is used as an optimization algorithm to optimize model parameters.

For the evaluation of the multi-feature dimensionality reduction effect, this embodiment evaluates and analyzes the feature dimensionality reduction result from two aspects of dimensionality reduction feature visualization and model classification performance. The first three dimensionality reduction principal components of the dimensionality reduction result are visualized. The original multi-feature data is passed through a convolutional neural network, and the features are nonlinearly combined to obtain dimensionality reduction feature parameters. Analogous to principal component analysis, for the visualization results, the three axes in the image represent the first three principal component dimensions of the dimensionality reduction feature results, each image represents a cycle process of the battery pack, and the change trend of the curve represents the parallel module in the current state. Changes in dimensionality reduction features during discharge cycles.

Taking the model prediction accuracy as the evaluation standard, the model prediction result acc is as follows:

In the formula, m represents the number of correctly predicted samples; n represents the total sample size. The model prediction results are shown in Table 3.

table 3

Total number of samples (n) Number of correct predicted samples (m) Accuracy 320 309 96.6%

The prediction accuracy of the model is 96.6%, which has good prediction performance. It also shows that the dimensionality reduction features obtained by the convolutional neural network can have a good representation of the health status of the system. The degradation process of each cycle is represented by the three principal components extracted by dimensionality reduction. With the progress of the cycle degradation process, the degree of disorder of the discharge process of the parallel modules increases significantly. From the perspective of pattern recognition, the features obtained through the dimensionality reduction of the convolutional neural network can effectively identify the changes in the health status of the battery pack.

On the basis of the above, the degradation characteristic parameters of the parallel module have a strong correlation with the change of the state of health of the battery pack, which can well characterize the degradation state of the battery pack. Therefore, this embodiment takes the health state of the parallel module as the object, and The dimensionality reduction features are used as data input for further battery pack degradation prediction. The cycle discharge capacity is used as the main indicator to characterize the state of health of the battery pack. Therefore, this section mainly focuses on the modeling and prediction of the cycle discharge capacity of the battery pack.

Considering that the health state of the battery pack has continuity in the time series dimension, and the discharge capacity between cycles is correlated, that is, the timing of the discharge conditions of the battery pack between cycles, a hybrid model of convolutional neural network and LSTM recurrent neural network is established. The cyclic discharge capacity of the module is predicted.

The CNN+LSTM hybrid model is used to predict the discharge capacity of the parallel modules, and the model has a good prediction effect on the discharge capacity of the battery pack. Taking root mean square error (RMSE), absolute error (MAE) and R2 value as model evaluation indicators, the corresponding model prediction set evaluation results are shown in Table ⁴ .

Table 4

Evaluation indicators RMSE(mAh) MAE(mAh) R²-value forecast result 9.90 8.68 0.87

From the results in Table 4, it can be seen that the prediction error of the model for the capacity can reach about 10mAh, and the R ² value is > 0.85, which has good prediction performance.

Among them, the calculation formulas of RMSE, MAE, and R ² values are as follows:

表示第i次预测的预测值；i表示测量次数，i＝1,2,……，m。In the formula, y _i represents the ith true value;

Indicates the predicted value of the i-th prediction; i indicates the number of measurements, i=1, 2, ..., m.

On this basis, in order to evaluate the validity of the prediction of the deep learning model more clearly, this embodiment continues to further evaluate the prediction accuracy of the proposed model. In order to analyze the stability of the model prediction, in this embodiment, a box plot is used to perform statistics on the dispersion of prediction accuracy. First, the original prediction result is converted into a relative error calculation. The calculation formula of the relative error is as follows:

表示真实值；y表示预测值。In the formula,

represents the true value; y represents the predicted value.

Through the box line analysis results, it can be concluded that the prediction results of the deep learning prediction model for the degradation of the health state of the parallel modules have small fluctuations, few outliers, and good performance.

The above has realized the prediction of the degradation process of the battery state of health, but in order to realize the inference of the lithium-ion battery system state, it is also necessary to convert the degradation data of the battery pack to the state distribution data, and the Bayesian neural network can be used to convert the degradation data. Distribute data for node states.

First, the degradation process of the battery pack is divided into states. Based on fuzzy theory, the continuous data obtained by the basic components at different times are discretized by fuzzy numbers, that is, the multi-state division of the degradation process data. The finer the state division of the lithium-ion battery capacity degradation data, the more accurate the state can be grasped. However, with the increase of the number of states, the computational complexity increases exponentially. For the lithium-ion battery system, it is suitable to use trapezoidal fuzzy numbers to classify its continuous data. Divide into 3 states (healthy, degraded, failed/normal, degraded, failure) or 4 states (healthy, slightly degraded, severely degraded, failed/normal, slightly degraded, serious degraded, failure) for research.

According to the multi-state division method, the normalized capacity data is discretized. According to the network node state built in the battery degradation process, the computational complexity and accuracy are integrated, and the node state of the established battery system network is divided into 3 Status (health, degradation, failure/normal, degraded, failure), get the multi-state division results of battery system and single node.

Using the characteristics of deep learning and program implementation framework, on the basis of the established CNN+LSTM hybrid model, the Bayesian neural network layer is defined, the prior distribution of hidden neurons is set to Gaussian distribution, and the fully connected network of the original model is replaced. The output dimension of the network model is set according to the division of health status, and a CNN+LSTM+BNN (Bayesian Neural Network) hybrid model is established.

On the basis of state division, a Bayesian neural network model is established, and the state probability distribution prediction curve of the integrated node is shown in Figure 6-8. According to Figure 8, it can be seen that the integration node is about 100 cycles, and the node begins to transition from a healthy state to a degraded state. At about 240 cycles, the probability of a failure state increases sharply, and the battery parallel module quickly enters a degraded state.

The present invention discloses the following technical effects:

The invention uses the multi-level system prediction modeling method to establish a Bayesian network model of the battery system through the multi-feature parameters of battery charge and discharge, and according to feature extraction-correlation analysis-feature dimension reduction-node state distribution prediction-system health In the process of state prediction, the corresponding analysis work is completely carried out, and the accurate prediction of the health state of the lithium-ion battery pack is realized.

Finally, it should be noted that the above embodiments are only specific implementations of the present invention, and are used to illustrate the technical solutions of the present invention, but not to limit them. The protection scope of the present invention is not limited thereto, although with reference to the foregoing embodiments The present invention has been described in detail, and those of ordinary skill in the art should understand that: any person skilled in the art can still modify or modify the technical solutions described in the foregoing embodiments within the technical scope disclosed by the present invention. Changes are easily thought of, or equivalent replacements are made to some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. All should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims (2)

1. The lithium ion battery pack health state prediction method based on deep learning is characterized by comprising the following steps of:

s1, extracting the discharge characteristic parameters of the lithium ion battery pack, and carrying out correlation analysis on the discharge characteristic parameters;

s2, obtaining an original feature data set according to the result of the correlation analysis, and performing feature dimension reduction on the original feature data set to obtain a dimension reduction feature set;

the acquiring process of the original feature data set in S2 is as follows: quantifying the result of the correlation analysis, obtaining a correlation coefficient matrix of the discharge characteristic parameters, removing the discharge characteristic parameters with the correlation coefficient less than 0.1 according to the correlation coefficient matrix, and obtaining an original characteristic data set;

s3, obtaining the time sequence of the discharge condition of the lithium ion battery pack among the cycles according to the incidence relation of the discharge capacity among the cycles, and constructing a node degradation data prediction model based on the time sequence of the discharge condition of the lithium ion battery pack among the cycles;

s4, inputting the dimensionality reduction feature set into the node degradation data prediction model to obtain node degradation prediction data of the lithium ion battery pack;

s5, converting the node degradation prediction data into node state distribution prediction data, and acquiring a health state prediction result of the lithium ion battery pack based on the node state distribution prediction data;

the correlation analysis in S1 includes: calculating the mean value, median and variance of the discharge characteristic parameters to complete statistical analysis, and performing related visual operation according to the result of the statistical analysis to complete correlation analysis;

the feature dimension reduction process in S2 includes the following steps:

s2.1, carrying out gridding processing on the original characteristic data set to obtain a grid characteristic matrix;

s2.2, constructing a convolutional neural network model based on the grid characteristic matrix, and performing model optimization training on the convolutional neural network model;

s2.3, obtaining dimension reduction characteristic parameters based on the convolutional neural network model after model optimization training, and completing characteristic dimension reduction;

the construction of the node degradation data prediction model in the step S3 includes the following steps:

s3.1, acquiring a cycle degradation data set of the lithium ion battery pack, and acquiring a model data set based on the cycle degradation data set;

s3.2, constructing a node degradation data prediction model;

s3.3, performing parameter optimization training on the node degradation data prediction model based on the model data set to complete construction of the node degradation data prediction model;

the process for obtaining the lithium ion battery pack health state prediction result in the step S5 includes the following steps: s5.1, discretizing the node degradation prediction data to finish multi-state division of the degradation process of the lithium ion battery pack;

s5.2, constructing a Bayesian neural network mixed model;

and S5.3, acquiring node state distribution prediction data based on the Bayesian neural network mixed model, and acquiring a health state prediction result of the lithium ion battery pack according to the node state distribution prediction data.

2. The deep learning-based lithium ion battery pack state of health prediction method of claim 1, wherein the discharge characteristic parameters in S1 include discharge current, discharge voltage, discharge temperature, discharge capacity data.

Images