SPC quality analysis method for rapid analysis of characteristic factors

Abstract

The invention discloses an SPC quality analysis method for rapid analysis of characteristic factors, and belongs to the technical field of quality analysis, and the analysis method comprises the following specific steps: I, collecting quantitative data in groups, and generating a quality fluctuation chart and a bar-x control chart; iI, finding a specific point through an abnormality judgment rule; iII, mining associated characteristic factors of the quality parameters corresponding to the specific points; iV, carrying out classification processing on the identified characteristic factors; v, detecting the same correlation factors of different specific points; vI, quickly positioning and judging quality defects according to each group of data; according to the method, the accuracy of abnormal point recognition can be improved, the efficiency of abnormal point detection can be improved, data changes can be tracked in real time, quick response can be made, potential quality problems can be prevented, invalid data points used for drawing the control chart are greatly reduced, unnecessary calculation is avoided, control chart generation is more intelligent and efficient, and the method is suitable for large-scale popularization and application. Manual adjustment of algorithm parameters is not needed, and the flexibility and adaptability of quality analysis are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of quality analysis, and in particular to an SPC quality analysis method for rapid analysis of characteristic factors. Background Art

[0002] In modern industrial production processes, quality control is an important part of ensuring product stability and consistency. Traditional statistical process control methods collect and analyze production data to detect anomalies in the production process and take corrective measures. However, as the production environment becomes increasingly complex, product quality is affected by many factors. It is difficult to quickly and accurately identify characteristic factors related to quality problems simply relying on traditional SPC methods. Therefore, how to efficiently and quickly detect anomalies in large-scale data and accurately locate the characteristic factors that cause anomalies has become a major challenge in quality analysis; therefore, it is particularly important to invent an SPC quality analysis method for rapid analysis of characteristic factors.

[0003] The existing SPC quality analysis method for rapid analysis of characteristic factors has low accuracy in identifying outliers, slow outlier detection speed, and is unable to track data changes in real time, which easily leads to potential quality problems. In addition, the existing SPC quality analysis method for rapid analysis of characteristic factors has a large number of invalid data points used to draw control charts, which easily generates a large number of unnecessary redundant calculations and requires manual adjustment of algorithm parameters, reducing the flexibility and adaptability of quality analysis. Therefore, we propose an SPC quality analysis method for rapid analysis of characteristic factors. Summary of the invention

[0004] The purpose of the present invention is to solve the defects in the prior art and to propose an SPC quality analysis method for rapid analysis of characteristic factors.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] An SPC quality analysis method for rapid analysis of characteristic factors, the specific steps of the analysis method are as follows:

[0007] Ⅰ. Collect quantitative data in groups and generate quality fluctuation charts and bar-x control charts;

[0008] Ⅱ. Find singularity through the rule of difference;

[0009] III. Mining the associated characteristic factors of the quality parameters corresponding to each singular point;

[0010] IV. Classify the identified characteristic factors;

[0011] V. Detecting the same correlation factors at different specific points;

[0012] Ⅵ. Quickly locate and judge quality defects based on each set of data.

[0013] As a further solution of the present invention, the specific steps of generating the quality fluctuation diagram and the bar-x control diagram in step I are as follows:

[0014] S1.1: Collect each set of quantitative data from the production process and optimize the selected quantitative data. Then divide the selected data into n groups, each with a data size of m. Use X to represent each measurement value in each set of quantitative data. ij denoted by , where i is the group number, j is the serial number of the measurement within the group, and i = 1, 2, ..., n, j = 1, 2, ..., m;

[0015] S1.2: Calculate the average value of each group of data, and obtain the overall average value of all groups based on the calculated average value of each group of data. Then, based on the difference between the maximum and minimum values ​​of each group of data, obtain the range of each group of data, and calculate the overall average range of all groups;

[0016] S1.3: Calculate the upper and lower control limits based on the overall mean and the overall average range. Plot the mean of each group of data as a data point on the bar-x control chart, draw the global mean as the center line, and mark the upper and lower control limits on the chart respectively. At the same time, draw a quality fluctuation chart of each measurement value over time.

[0017] As a further solution of the present invention, the specific steps of optimizing the selected quantitative data in S1.1 are as follows:

[0018] S2.1: Randomly generate the number of selected individuals and initialize them on different data nodes collected, initialize the pheromone concentration and heuristic information of each data node, and then calculate the selection probability corresponding to the data node through the pheromone concentration and heuristic information of each data order;

[0019] S2.2: Based on the selection probability of each data node, the next data node of each selected individual is selected by random selection, and then a complete data selection path is gradually constructed. After each round of iteration, the pheromone concentration of the corresponding data node is updated according to the representativeness and overall effect of each group of data selection paths constructed;

[0020] S2.3: Repeat the path exploration and pheromone update, and retain the optimal path after each iteration until the preset iteration time is reached, then stop the iteration and select the data point with the highest pheromone concentration to construct a data subset for generating the control chart.

[0021] As a further solution of the present invention, the specific steps of finding the singularity in step II are as follows:

[0022] S3.1: Use the Sigma rule to identify the data that exceeds the upper and lower control limits in the bar-x control chart, and mark the data as abnormal. Then build a set of analysis trees and define each node in the tree as a time point or data point. The branches of the tree represent different abnormal patterns.

[0023] S3.2: Take the data point without abnormality as the initial state, and take the initial state as the root node of the analysis tree. Then, through the UCB selection strategy, select the child node with the highest upper confidence bound as the next selected node, and select downward layer by layer until the selected node is an incompletely expanded node;

[0024] S3.3: Select the best child node to expand, generate a new abnormal point combination or pattern, and according to the SPC rules, make the generated new nodes correspond to an abnormal situation respectively, simulate on the newly expanded nodes, generate a complete path by random sampling until the terminal node is reached, and stop the simulation, score each path to measure its matching degree with the SPC abnormality judgment rule;

[0025] S3.4: After each simulation, the results of the path are backtracked to the upper node, and each node is evaluated based on the number of visits and benefits. In the entire tree structure, the average benefit value of each node is calculated, that is, the frequency of being marked as abnormal in different simulation processes;

[0026] S3.5: According to the average benefit value and the number of visits of all nodes, the paths with higher benefit value and higher number of visits than the preset threshold are selected, and the paths with lower benefit value or lower number of visits than the preset threshold are eliminated. The data points represented by each node are extracted from the selected high benefit paths;

[0027] S3.6: Based on the position of these data points in the control chart, combined with the SPC exception judgment rules, confirm whether they conform to the abnormal pattern. If they do, mark these data points as the most likely abnormal points, and perform final verification on the marked abnormal points. If the characteristics of each data point are consistent with the search results in the analysis book, mark it as the final special point.

[0028] As a further solution of the present invention, the specific steps of mining the associated characteristic factors corresponding to the quality parameters of each singular point in step III are as follows:

[0029] S4.1: Use virtual modeling technology to simulate the real production process, capture the dynamic relationship between each process variable and characteristic factor, input each set of data corresponding to the singularity into the virtual model, train the virtual model with historical data, and calibrate the accuracy of the virtual model through simulation;

[0030] S4.2: Lock the identified singular points, and simulate each group of singular points in the virtual model, simulate the changes in process variables, and extract the corresponding characteristic factors when these singular points appear. Observe the change pattern of the characteristic factors through simulation to determine whether they have a causal relationship with the occurrence of the singular points;

[0031] S4.3: pre-process the extracted characteristic factor data, and make each data entry represent the characteristic factor state and the occurrence of singular points within a time period, and set the support and confidence thresholds;

[0032] S4.4: Starting from a single feature factor, the Apriori algorithm is used to gradually build complex combinations, evaluate the support of each combination, remove combinations with support below the support threshold, retain only frequently occurring feature factor combinations, calculate the confidence of the retained feature factor combinations, and screen out association rules with confidence above the confidence threshold;

[0033] S4.5: Based on the mining results, mark the high-support and high-confidence feature factor combinations related to the singularities and use them as potential causes of the singularities. Use the virtual model to simulate different combinations of these feature factors again to verify the marked associated feature factors. If the singularities are verified to be caused by the corresponding factors, optimize each associated feature factor.

[0034] As a further embodiment of the present invention, the classification process described in step IV

[0035] S5.1: Combine multiple feature factors identified by virtual modeling and association rule mining into a two-dimensional matrix, normalize each group of feature factors, and input the processed data into the pre-trained CNN model;

[0036] S5.2: The input layer of the CNN model receives the normalized feature factor matrix and passes it to the convolutional layer. The convolutional layer then scans the input data through multiple filters to extract features corresponding to specific patterns, and then performs nonlinear activation on the extracted feature data through the ReLU activation function.

[0037] S5.3: Use the pooling layer to perform the maximum pooling operation on each group of feature factors after processing, and repeatedly perform convolution, activation and pooling operations until the preset number of processing times is reached. Then, the feature map after convolution and pooling processing will be flattened into a one-dimensional vector, and classified through the fully connected layer, and the classification result will be input to the output layer;

[0038] S5.4: The output layer uses the Softmax activation function to classify the feature factors, outputs the probability of each category, and classifies each feature factor according to the final category probability. Based on the classification results, the impact of different types of feature factor combinations on the singularity is analyzed.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. The SPC quality analysis method for rapid analysis of characteristic factors constructs a set of analysis trees and defines each node in the tree as a time point or data point. The branches of the tree represent different abnormal patterns. The data point without abnormality is used as the initial state, and the initial state is used as the root node of the analysis tree. Then, through the UCB selection strategy, the child node with the highest upper confidence limit value is selected as the next selected node, until the selected node is an incompletely expanded node, the optimal child node is selected for expansion, and a new abnormal point combination or pattern is generated. According to the SPC rules, the generated new nodes correspond to an abnormal situation respectively, and simulation is performed on the newly expanded nodes. A complete path is generated by random sampling until the terminal node is reached and the simulation is stopped. Each path is scored to measure its matching degree with the SPC abnormality judgment rule. After each simulation, the result of the path is backtracked and updated to the upper node, and according to the access of each node, the result of the path is updated to the upper node. The number of visits and benefits are evaluated. In the entire tree structure, the average benefit value of each node is calculated, that is, the frequency of being marked as abnormal in different simulation processes. According to the average benefit value and the number of visits of all nodes, the path with higher benefit value and higher number of visits than the preset threshold is selected. At the same time, the path with lower benefit or lower number of visits than the preset threshold is eliminated. In the selected high-yield path, the data points represented by each node are extracted. According to the position of these data points in the control chart and the SPC exception judgment rule, it is confirmed whether they conform to the abnormal pattern. If they do, these data points are marked as the most likely abnormal points. The marked abnormal points are finally verified. If the characteristics of each data point are consistent with the search results of the analysis book, they are marked as the final special points, which can improve the accuracy of abnormal point identification, accelerate the efficiency of abnormal point detection, track data changes in real time, and respond quickly to prevent potential quality problems.

[0041] 2. The SPC quality analysis method for rapid analysis of characteristic factors randomly generates the number of selected individuals and initializes them on different data nodes collected, initializes the pheromone concentration and heuristic information of each data node, and then calculates the selection probability corresponding to the data node through the pheromone concentration and heuristic information of each data order. Based on the selection probability of each data node, the next data node of each selected individual is selected by random selection, and then a complete data selection path is gradually constructed. After each round of iteration, the pheromone concentration of the corresponding data node is updated according to the representativeness and overall effect of the constructed data selection path, and the path exploration and pheromone update are repeated. The optimal path is retained after each iteration until the preset iteration time is reached, and the iteration is stopped. The data point with the highest pheromone concentration is selected to construct a data subset for generating control charts, which greatly reduces the invalid data points used for drawing control charts and avoids unnecessary calculations, making control chart generation more intelligent and efficient. There is no need to manually adjust algorithm parameters, which improves the flexibility and adaptability of quality analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0043] Figure 1 The present invention provides a flowchart of an SPC quality analysis method for rapid analysis of characteristic factors. DETAILED DESCRIPTION

[0044] Example 1

[0045] Reference Figure 1 , an SPC quality analysis method for rapid analysis of characteristic factors, the specific steps of the analysis method are as follows:

[0046] Collect quantitative data in groups and generate quality fluctuation graphs and bar-x control charts.

[0047] Specifically, each group of quantitative data from the production process is collected, and the selected quantitative data is optimized, and then the selected data is divided into n groups, each group of data is m in size, and each measurement value in each group of quantitative data is represented by Xij, where i is the group number, j is the serial number of the measurement in the group, and i=1, 2, ..., n, j=1, 2, ..., m, the average value of each group of data is calculated, and the overall average value of all groups is obtained based on the calculated average value of each group of data, and then the range of each group of data is obtained based on the difference between the maximum and minimum values ​​of each group of data, and the overall average range of all groups is calculated, and the upper control limit and the lower control limit are calculated based on the overall average value and the overall average range, and the average value of each group of data is plotted as a data point on a bar-x control chart, the global average value is drawn as the center line, and the upper and lower control limits are marked on the chart respectively, and a quality fluctuation chart of each measurement value over time is plotted.

[0048] It should be further explained that the number of selected individuals is randomly generated and initialized on different data nodes collected, the pheromone concentration and heuristic information of each data node are initialized, and then the selection probability corresponding to the data node is calculated through the pheromone concentration and heuristic information of each data order. Based on the selection probability of each data node, the next data node of each selected individual is selected by random selection, and then the complete data selection path is gradually constructed. After each round of iteration, the pheromone concentration of the corresponding data node is updated according to the representativeness and overall effect of the constructed data selection path, and the path exploration and pheromone update are repeated. The optimal path is retained after each iteration until the preset iteration time is reached, the iteration is stopped, and the data point with the highest pheromone concentration is selected to construct a data subset for generating the control chart.

[0049] Find singularities through the rule of difference.

[0050] Specifically, the Sigma rule is used to identify data that exceeds the upper and lower control limits in the bar-x control chart, and the data is marked as abnormal. Then, a set of analysis trees is constructed, and each node in the tree is defined as a time point or data point. The branches of the tree represent different abnormal patterns. The data point without abnormalities is used as the initial state, and the initial state is used as the root node of the analysis tree. Then, the UCB selection strategy is used to select the child node with the highest upper confidence limit value as the next selected node, and the selection is carried out layer by layer until the selected node is an incompletely expanded node. The optimal child node is selected for expansion to generate a new abnormal point combination or pattern. According to the SPC rules, the generated new nodes correspond to an abnormal situation respectively, and simulation is performed on the newly expanded nodes. A complete path is generated by random sampling until the terminal node is reached and the simulation is stopped. Each path is scored to measure its comparison with SPC. The matching degree of the judgment rule: after each simulation, the result of the path is back-tracked to the upper node, and the evaluation is performed based on the number of visits and the revenue of each node. In the entire tree structure, the average revenue value of each node is calculated, that is, the frequency of being marked as abnormal in different simulation processes. According to the average revenue value and the number of visits of all nodes, the path with higher revenue value and higher number of visits than the preset threshold is selected. At the same time, the path with lower revenue or lower number of visits than the preset threshold is eliminated. In the selected high-return path, the data points represented by each node are extracted. According to the position of these data points in the control chart, combined with the SPC judgment rule, it is confirmed whether they meet the abnormal pattern. If they meet, these data points are marked as the most likely abnormal points, and the marked abnormal points are finally verified. If the characteristics of each data point are consistent with the search results of the analysis book, they are marked as the final special points.

[0051] Example 2

[0052] Reference Figure 1 , an SPC quality analysis method for rapid analysis of characteristic factors, the specific steps of the analysis method are as follows:

[0053] Mining the associated characteristic factors of the quality parameters corresponding to each singular point.

[0054] Specifically, the real production process is simulated by virtual modeling technology, the dynamic relationship between each process variable and characteristic factor is captured, and each group of data corresponding to the singularity is input into the virtual model, and the virtual model is trained by historical data. The accuracy of the virtual model is calibrated by simulation, and the identified singularity is locked. In the virtual model, each group of singularity is simulated and deduced, the change of process variables is simulated, and the corresponding characteristic factors when these singularities appear are extracted. The change pattern of the characteristic factors is observed by simulation to determine whether they have a causal relationship with the occurrence of the singularity. The extracted characteristic factor data is preprocessed, and each data entry represents the characteristic factor state and the singularity within a time period. When a situation occurs, set the support and confidence thresholds, start from a single feature factor through the Apriori algorithm, gradually build complex combinations, evaluate the support of each combination, eliminate combinations with support lower than the support threshold, retain only frequently occurring feature factor combinations, calculate the confidence for the retained feature factor combinations, screen out association rules above the confidence threshold, and mark the feature factor combinations with high support and high confidence related to the singularity according to the mining results, and use them as potential causes of the singularity. Use the virtual model to simulate different combinations of these feature factors again to verify the marked associated feature factors. If the singularity is verified to be caused by the corresponding factor, optimize each associated feature factor.

[0055] Classify the identified characteristic factors.

[0056] Specifically, multiple feature factors identified by virtual modeling and association rule mining are combined in a two-dimensional matrix, and each group of feature factors is normalized. The processed data is input into the pre-trained CNN model. The input layer of the CNN model receives the normalized feature factor matrix and passes it into the convolution layer. The convolution layer then scans the input data through multiple filters to extract features corresponding to specific patterns. The extracted feature data is then nonlinearly activated through the ReLU activation function. The pooling layer performs maximum pooling operations on each group of processed feature factors. Convolution, activation, and pooling operations are repeated until the preset number of processing times is reached. The feature map after convolution and pooling is flattened into a one-dimensional vector and classified through a fully connected layer. The classification results are input into the output layer. The output layer uses the Softmax activation function to classify the feature factors, outputs the probability of each category, and classifies each feature factor according to the final category probability. According to the classification results, the influence of different types of feature factor combinations on the singularity is analyzed.

[0057] Detecting the same correlation factor at different specificities.

[0058] Quickly locate and determine quality defects based on each set of data.

Claims

1. A SPC quality analysis method for rapid analysis of characteristic factors, characterized in that: The specific steps of this analysis method are as follows: Ⅰ. Collect quantitative data in groups and generate quality fluctuation charts and bar-x control charts; Ⅱ. Find singularity through the rule of difference; III. Mining the associated characteristic factors of the quality parameters corresponding to each singular point; IV. Classify the identified characteristic factors; V. Detecting the same correlation factors at different specific points; Ⅵ. Quickly locate and judge quality defects based on each set of data.

2. The SPC quality analysis method for rapid analysis of characteristic factors according to claim 1, characterized in that: The specific steps for generating the quality fluctuation chart and bar-x control chart described in step Ⅰ are as follows: S1.1: Collect each set of quantitative data from the production process and optimize the selected quantitative data. Then divide the selected data into n groups, each with a data size of m. Use X to represent each measurement value in each set of quantitative data. ij denoted by , where i is the group number, j is the serial number of the measurement within the group, and i = 1, 2, ..., n, j = 1, 2, ..., m; S1.2: Calculate the average value of each group of data, and obtain the overall average value of all groups based on the calculated average value of each group of data. Then, based on the difference between the maximum and minimum values ​​of each group of data, obtain the range of each group of data, and calculate the overall average range of all groups; S1.3: Calculate the upper and lower control limits based on the overall mean and the overall average range. Plot the mean of each group of data as a data point on the bar-x control chart, draw the global mean as the center line, and mark the upper and lower control limits on the chart respectively. At the same time, draw a quality fluctuation chart of each measurement value over time.

3. The SPC quality analysis method for rapid analysis of characteristic factors according to claim 2, characterized in that: The specific steps for optimizing the selected quantitative data described in S1.1 are as follows: S2.1: Randomly generate the number of selected individuals and initialize them on different data nodes collected, initialize the pheromone concentration and heuristic information of each data node, and then calculate the selection probability corresponding to the data node through the pheromone concentration and heuristic information of each data order; S2.2: Based on the selection probability of each data node, the next data node of each selected individual is selected by random selection, and then a complete data selection path is gradually constructed. After each round of iteration, the pheromone concentration of the corresponding data node is updated according to the representativeness and overall effect of each group of data selection paths constructed; S2.3: Repeat the path exploration and pheromone update, and retain the optimal path after each iteration until the preset iteration time is reached, then stop the iteration and select the data point with the highest pheromone concentration to construct a data subset for generating the control chart.

4. The SPC quality analysis method for rapid analysis of characteristic factors according to claim 2, characterized in that: The specific steps for finding the singularity in step II are as follows: S3.1: Use the Sigma rule to identify the data that exceeds the upper and lower control limits in the bar-x control chart, and mark the data as abnormal. Then build a set of analysis trees and define each node in the tree as a time point or data point. The branches of the tree represent different abnormal patterns. S3.2: Take the data point without abnormality as the initial state, and take the initial state as the root node of the analysis tree. Then, through the UCB selection strategy, select the child node with the highest upper confidence bound as the next selected node, and select downward layer by layer until the selected node is an incompletely expanded node; S3.3: Select the best child node to expand, generate a new abnormal point combination or pattern, and according to the SPC rules, make the generated new nodes correspond to an abnormal situation respectively, simulate on the newly expanded nodes, generate a complete path by random sampling until the terminal node is reached, and stop the simulation, score each path to measure its matching degree with the SPC abnormality judgment rule; S3.4: After each simulation, the results of the path are backtracked to the upper node, and each node is evaluated based on the number of visits and benefits. In the entire tree structure, the average benefit value of each node is calculated, that is, the frequency of being marked as abnormal in different simulation processes; S3.5: According to the average benefit value and the number of visits of all nodes, the paths with higher benefit value and higher number of visits than the preset threshold are selected, and the paths with lower benefit value or lower number of visits than the preset threshold are eliminated. The data points represented by each node are extracted from the selected high benefit paths; S3.6: Based on the position of these data points in the control chart, combined with the SPC exception judgment rules, confirm whether they conform to the abnormal pattern. If they do, mark these data points as the most likely abnormal points, and perform final verification on the marked abnormal points. If the characteristics of each data point are consistent with the search results in the analysis book, mark it as the final special point.

5. The SPC quality analysis method for rapid analysis of characteristic factors according to claim 4, characterized in that: The specific steps of mining the associated characteristic factors of the quality parameters corresponding to each singular point in step III are as follows: S4.1: Use virtual modeling technology to simulate the real production process, capture the dynamic relationship between each process variable and characteristic factor, input each set of data corresponding to the singularity into the virtual model, train the virtual model with historical data, and calibrate the accuracy of the virtual model through simulation; S4.2: Lock the identified singular points, and simulate each group of singular points in the virtual model, simulate the changes in process variables, and extract the corresponding characteristic factors when these singular points appear. Observe the change pattern of the characteristic factors through simulation to determine whether they have a causal relationship with the occurrence of the singular points; S4.3: pre-process the extracted characteristic factor data, and make each data entry represent the characteristic factor state and the occurrence of singular points within a time period, and set the support and confidence thresholds; S4.4: Starting from a single feature factor, the Apriori algorithm is used to gradually build complex combinations, evaluate the support of each combination, remove combinations with support below the support threshold, retain only frequently occurring feature factor combinations, calculate the confidence of the retained feature factor combinations, and screen out association rules with confidence above the confidence threshold; S4.5: Based on the mining results, mark the high-support and high-confidence feature factor combinations related to the singularities and use them as potential causes of the singularities. Use the virtual model to simulate different combinations of these feature factors again to verify the marked associated feature factors. If the singularities are verified to be caused by the corresponding factors, optimize each associated feature factor.

6. The SPC quality analysis method for rapid analysis of characteristic factors according to claim 5, characterized in that: Classification process described in step IV S5.1: Combine multiple feature factors identified by virtual modeling and association rule mining into a two-dimensional matrix, normalize each group of feature factors, and input the processed data into the pre-trained CNN model; S5.2: The input layer of the CNN model receives the normalized feature factor matrix and passes it to the convolutional layer. The convolutional layer then scans the input data through multiple filters to extract features corresponding to specific patterns, and then performs nonlinear activation on the extracted feature data through the ReLU activation function. S5.3: Use the pooling layer to perform the maximum pooling operation on each group of feature factors after processing, and repeatedly perform convolution, activation and pooling operations until the preset number of processing times is reached. Then, the feature map after convolution and pooling processing will be flattened into a one-dimensional vector, and classified through the fully connected layer, and the classification result will be input to the output layer; S5.4: The output layer uses the Softmax activation function to classify the feature factors, outputs the probability of each category, and classifies each feature factor according to the final category probability. Based on the classification results, the impact of different types of feature factor combinations on the singularity is analyzed.

Images