Cigarette quality control chart mode recognition method based on interpretable graph convolutional neural network

Abstract

The invention relates to the technical field of cigarette production monitoring, in particular to a cigarette quality control chart mode recognition method based on an interpretable graph convolutional neural network. The method comprises the following steps: acquiring a plurality of cigarette control chart training samples in different modes; extracting training features of each cigarette control chart training sample and generating a feature test set; constructing a plurality of classification network models, and respectively inputting the feature test set corresponding to each cigarette control chart training sample into the plurality of classification network models to obtain a plurality of classification prediction result data and ROC curve graphs; screening according to the ROC curve graph to obtain an optimal classification network model, and inputting the optimal classification network model into an SHAP interpreter; calculating the marginal contribution of each feature in the feature test set to the classification prediction result data, and generating a high-value feature subset according to a Shapley value; training and obtaining a target image convolutional neural network by using the high-value feature subset; and obtaining an actually collected cigarette control chart sample, and inputting the sample into the target chart convolutional neural network to complete pattern recognition.

Description

Technical Field

[0001] The invention relates to the technical field of cigarette production monitoring, and in particular to a control chart pattern recognition method. Background Art

[0002] With the rapid development of intelligent manufacturing and information technology, control chart abnormal pattern recognition has become an effective method for intelligent quality monitoring of production processes. In the cigarette industry, it is common to collect and analyze cigarette quality data in real time during the production process, draw control charts to identify abnormal patterns in the production process, and then monitor the operation status of the production process. Combined with the fault library for tracing, it accelerates diagnosis and achieves rapid recovery of the production process.

[0003] The quality of cigarettes is one of the important indicators to measure whether the final cigarette product is qualified or not. It can reflect the influence of various factors such as the process parameters of the rolling equipment and the processing technology on the cigarettes during the processing. Therefore, identifying abnormal patterns in the cigarette quality control chart plays an important role in monitoring the production process.

[0004] In the process of identifying abnormal patterns in control charts, machine learning methods are needed to classify different abnormal patterns in control charts according to features. Existing methods often focus on feature extraction steps to improve recognition accuracy. However, the contribution of the extracted features to the classification results is often difficult to accurately quantify, which restricts the scientificity and effectiveness of cigarette quality assessment. Summary of the invention

[0005] The purpose of the present invention is to provide a cigarette quality control chart pattern recognition method based on an interpretable graph convolutional neural network. The method can interpret the shape characteristics or statistical characteristics of the control chart, thereby specifying different features for pattern classification, and can accurately quantify the contribution of various feature quantities to the classification model results, thereby improving the scientificity and effectiveness of cigarette quality assessment.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A cigarette quality control graph pattern recognition method based on an interpretable graph convolutional neural network, characterized in that the method comprises:

[0008] S01. Obtaining cigarette control chart training samples in multiple different modes;

[0009] S02, extracting training features of each cigarette control chart training sample; the training features include multiple shape features and multiple statistical features;

[0010] S03, generating a feature test set according to the training features; constructing multiple classification network models, inputting the feature test set corresponding to each cigarette control chart training sample into the multiple classification network models respectively, obtaining multiple classification prediction result data, and obtaining corresponding ROC curve graphs;

[0011] S04, screening and obtaining an optimal classification network model according to the ROC curve graph; inputting the optimal classification network model, the feature test set and the classification prediction result data output by the optimal classification network model into the SHAP interpreter;

[0012] S05, the SHAP interpreter calculates the marginal contribution of each feature in the feature test set to the classification prediction result data, generates multiple Shapley values; generates a high-value feature subset according to the Shapley values;

[0013] S06. Inputting the high-value feature subset into the constructed graph convolutional neural network, training the model and obtaining a target graph convolutional neural network;

[0014] S07. Obtain samples of the actual collected cigarette control graph, and input them into the target graph convolutional neural network to perform pattern recognition to obtain cigarette quality fluctuation results.

[0015] Therefore, in this method, it is first necessary to obtain a plurality of cigarette control chart training samples under different modes. Since it is difficult to obtain abnormal mode data in batches during the actual production process, and most of them require destructive experiments to expand the sample size of abnormal data, which leads to practical problems such as high data acquisition cost and poor practicality, the present invention intends to use Monte Carlo simulation algorithm (Monte Carlo, MC) to design different parameters to generate different types of mean trend charts as the original training samples of the model. Secondly, analyze the quality characteristics of cigarette quality, judge that these continuous variables should use measurement value control charts, and select the conventional and effective control chart-mean control chart.

[0016] According to different abnormal factors, the mean control chart can be divided into several different modes, such as normal mode (NOR), cycle mode (CYC), downward trend mode (DT), upward trend mode (UT), downward step mode (DS) and upward step mode (US).

[0017] Then, the training features of each cigarette control chart training sample are extracted; the training features include multiple shape features and multiple statistical features. The present invention adopts 9 conventional shape features with high classification efficiency, namely: the slope of the least squares line, the average number of crossings (nc1), the number of least squares line crossings (nc2), the cycle member (cmember), the average slope of the line segment (as), the slope difference (sd) between the least squares line and the line segment representing the pattern, the area between the pattern and the mean line (apm), the area between the pattern and its least squares line (aps), and the area between the least squares line and the line segment (ass); and 6 statistical features, namely, mean, cumulative sum statistic (cusum), skewness, kurtosis, mean square value, and standard deviation.

[0018] Then, a feature test set is generated according to the training features; each feature test set corresponds to a cigarette control chart training sample, which includes 15 features extracted from the above cigarette control chart training sample.

[0019] Next, multiple classification network models are constructed, such as classic machine learning classification network models such as decision tree, random forest, support vector machine, GBDT, XGBoost and logistic regression for preliminary classification; specifically, the feature test sets corresponding to each cigarette control chart training sample are respectively input into multiple classification network models to obtain multiple classification prediction result data, and corresponding ROC curve graphs are obtained.

[0020] The ROC curve contains multiple curves, each of which represents a different classification network model. The larger the area under the curve, the better the classification effect of the model. Therefore, the optimal classification network model can be screened according to the ROC curve.

[0021] Afterwards, the above-screened optimal classification network model, feature test set and classification prediction result data output by the optimal classification network model are input into the SHAP interpreter.

[0022] The SHAP interpreter calculates the marginal contribution of each feature in the feature test set to the classification prediction result data, generates multiple Shapley values, and generates a high-value feature subset based on the Shapley value. SHAP is based on the Shapley value principle in game theory and calculates the marginal contribution of each feature to the classification prediction result data of the optimal classification network model. The Shapley value quantifies the actual impact of the feature on the model output based on other features, while also taking into account the interaction between features.

[0023] Then, the high-value feature subset is input into the constructed graph convolutional neural network, the model is trained and the target graph convolutional neural network is obtained. Graph convolutional neural network is a neural network in the prior art that processes graph structure data, which can effectively capture the association between nodes (that is, samples). For control graph data, the similarity between samples can be constructed as the edge of the graph, and finally the classification result Y is output through the activation function.

[0024] Finally, the dynamic data stream of cigarette quality is monitored using the filtered features and trained models. Specifically, samples of actual cigarette control charts are obtained and input into the target graph convolutional neural network to complete pattern recognition and find abnormal quality fluctuations in the cigarette production process.

[0025] In summary, in this method, with the help of the SHAP interpreter, the shape characteristics and statistical characteristics of the control chart were interpreted, the Shapley value was obtained, the contribution of different features to the classification model results was accurately quantified, and a high-value feature subset was generated. The model was trained with this high-value feature subset and the target graph convolutional neural network was obtained, which greatly improved the prediction accuracy of the target graph convolutional neural network for abnormal pattern recognition of cigarette quality control charts, and ultimately comprehensively improved the scientificity and effectiveness of cigarette quality assessment.

[0026] As a preferred embodiment of the present invention, in S01, the method for obtaining cigarette control chart training samples under multiple different modes is specifically: based on the control chart mode simulation formula, using the Monte Carlo method to simulate and obtain cigarette control chart training samples under multiple different modes.

[0027] Since abnormal pattern data are difficult to obtain in batches in the actual production process, and most of them require destructive experiments to expand the sample size of abnormal data, which leads to practical problems such as high data acquisition cost and poor practicality, the present invention intends to use Monte Carlo simulation algorithm (Monte Carlo, MC) to design different parameters to generate different types of mean trend charts as the original training samples of the model.

[0028] As a preferred embodiment of the present invention, the modes of the cigarette control chart training samples include: normal mode (NOR), cycle mode (CYC), downward trend mode (DT), upward trend mode (UT), downward step mode (DS) and upward step mode (US).

[0029] Among them, normal mode (NOR): indicates that the machine is properly maintained, the material quality is stable, there is no major fluctuation, and it is in a stable state; cyclic mode (CYC): periodic fluctuations of equipment, such as periodic mechanical wear or vibration of the rollers, gears or transmission devices of the rolling machine group, resulting in periodic fluctuations in cigarette quality; downward trend mode (DT) and upward trend mode (UT): the change in tobacco volume caused by the gradual wear of machine parts, or the increase in environmental humidity, resulting in a gradual decrease or increase in cigarette quality; downward step mode (DS) and upward step mode (US): indicate that the temperature or humidity of environmental factors suddenly changes, resulting in changes in tobacco moisture, sudden adjustments to equipment or changes in raw material batches, resulting in changes in cigarette indicators.

[0030] As a preferred embodiment of the present invention, the shape features include: the slope of the least squares line (slope), the average number of crossings (nc1), the number of least squares line crossings (nc2), the cycle member (cmember), the average slope of the line segment (as), the slope difference (sd) between the least squares line and the line segment representing the pattern, the area between the pattern and the mean line (apm), the area between the pattern and its least squares line (aps) and the area between the least squares line and the line segment (ass).

[0031] Shape features provide information about the appearance and shape of a graphic by describing its geometric shape and structure, and can help qualitatively describe and analyze the specific shape of the graphic. The nine shape features used in the present invention are as follows:

[0032] The slope of the least squares line: represents the trend of the pattern. For normal and periodic patterns, the absolute value of the slope is close to zero, while for step and trend patterns, the absolute value of the slope is greater than zero.

[0033] Average number of crossings (nc1): indicates the number of intersections between the pattern and the mean line. The normal pattern has the most intersections, the step and trend patterns have the least, and the cycle pattern is in between.

[0034] Number of least squares line crossings (nc2): indicates the number of times the pattern crosses the least squares line;

[0035] Cyclic member (cmember): indicates the similarity between the pattern and the periodic pattern. For periodic patterns, this value is positive, and for other patterns it is negative;

[0036] Average slope of line segments (as): represents the average slope of each line segment in the pattern, and the average slope of the trend pattern is the highest;

[0037] Slope difference (sd) between the least squares line and the line segment representing the pattern: the difference between the slope of the least squares line and the average slope of the line segment;

[0038] Area between mode and mean line (apm): The area of ​​this feature for normal mode is lowest, so this feature is different between normal mode and other modes;

[0039] Area between the mode and the least squares line (aps): This feature has higher aps values ​​in periodic and step modes than in normal and trend modes, so aps can distinguish between modes such as periodic and step;

[0040] Area between the least squares line and the line segment (ass): This feature is approximately 0 in the trend mode and may be higher for other modes, so this feature can distinguish the trend mode from other modes.

[0041] As a preferred embodiment of the present invention, the statistical features include: mean, cumulative sum statistic (cusum), skewness, kurtosis, mean square value and standard deviation.

[0042] Statistical features provide information such as data distribution, central tendency and dispersion through mathematical calculations, which can help quantitatively describe and analyze the overall properties of the graph. The six statistical features used in this invention are as follows:

[0043] Mean: represents the location of the data center, used to determine whether the data fluctuates around a certain central value and show the central trend of the data;

[0044] Cumulative sum statistic (cusum): A statistic that detects changes by accumulating deviations between data points and target values. It is used to monitor process changes, especially small, sustained shifts. Cumulative sum positive statistics are used to detect upward mean shifts, and cumulative sum negative statistics are used to detect downward mean shifts.

[0045] Skewness: reflects the symmetry of data distribution. Positive skewness indicates that the data is right-skewed, and negative skewness indicates that the data is left-skewed. It is used to identify the process trend of the inclination, deviation or asymmetry of data distribution.

[0046] Kurtosis is used to describe the sharpness of the data distribution. A high kurtosis indicates that the data distribution has a sharp peak, while a low kurtosis indicates a flat distribution. It is used to detect outliers or extreme values.

[0047] The mean square value is the average of the square values ​​of all data points. It is used to measure the energy or degree of variation of a data set. It can help detect the specific trend of data changes, especially in control charts, reflecting volatility and instability. The larger the value, the greater the surface fluctuation.

[0048] The standard deviation indicates the degree of dispersion between data points and the mean, and is used to measure data fluctuation and dispersion. Compared with the mean square value, the standard deviation reflects more how the data fluctuates around the mean.

[0049] As a preferred embodiment of the present invention, in S03, the multiple classification network models constructed are specifically: a decision tree model, a random forest model, a logistic regression model, a support vector machine model, an XGBoost model and a GBDT model.

[0050] As a preferred embodiment of the present invention, in S05, the specific method of generating a high-value feature subset according to the Shapley value is: calculating the average absolute value of each Shapley value and sorting them according to the cumulative proportion, and selecting features with the first 90% of the cumulative proportion of Shapley values ​​to form a high-value feature subset.

[0051] As a preferred embodiment of the present invention, in S06, a back propagation algorithm is used to train the constructed graph convolutional neural network and obtain the target graph convolutional neural network.

[0052] In summary, the present invention has the following beneficial effects:

[0053] In this method, the SHAP interpreter is used to interpret the shape characteristics and statistical characteristics of the control chart, and the Shapley value is obtained, which accurately quantifies the contribution of the feature quantity to the classification model results and generates a high-value feature subset. The model is trained with this high-value feature subset and the target graph convolutional neural network is obtained, which greatly improves the prediction accuracy of the target graph convolutional neural network in identifying abnormal patterns in the cigarette quality control chart, and ultimately comprehensively improves the scientificity and effectiveness of cigarette quality assessment.

[0054] The contribution differences of 9 shape features and 6 statistical features to different control chart modes are studied, with a wide range of considerations, thereby improving the scientific nature of the features selected by the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A flow chart of the method of the invention;

[0056] Figure 2 This is a control diagram in the normal mode of this embodiment;

[0057] Figure 3 This is a control diagram in the periodic mode of this embodiment;

[0058] Figure 4 This is a control diagram in the upward step mode in this embodiment;

[0059] Figure 5 This is the control diagram in the downward step mode in this embodiment;

[0060] Figure 6 This is the control chart in the upward trend mode in this embodiment;

[0061] Figure 7 This is the control chart in the downward trend mode in this embodiment;

[0062] Figure 8 It is the ROC curve diagram in this embodiment;

[0063] Fig. 9 Schematic diagram of a bee colony in this embodiment;

[0064] Fig.10 : is a topological structure diagram of the graph convolutional neural network in this embodiment;

[0065] Fig.11 is a loss curve diagram in this embodiment;

[0066] Fig.12 It is the confusion matrix diagram in this embodiment;

[0067] Fig.13 This is the cigarette quality mean control chart in this embodiment. DETAILED DESCRIPTION

[0068] The technical solutions of the embodiments of the present invention are explained and described below in conjunction with the drawings of the embodiments of the present invention, but the following embodiments are only preferred embodiments of the present invention, not all. Based on the embodiments in the implementation mode, other embodiments obtained by those skilled in the art without creative work are all within the protection scope of the present invention.

[0069] The terms "first", "second", etc. in the description and claims of this specification and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the term "comprising" and any variation thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices.

[0070] First, if Figure 1 As shown, cigarette control chart training samples under multiple different modes are obtained.

[0071] Specifically, a metering value control chart, i.e., a mean control chart, is selected according to the characteristics of cigarette quality. Since it is difficult to obtain abnormal mode data in batches during actual production, and most of them require destructive experiments to expand the sample size of abnormal data, which leads to practical problems such as high data acquisition cost and poor practicality, the present invention intends to use Monte Carlo simulation algorithm (MC) to design different parameters to generate different types of mean trend charts, such as Figure 2-7 The graphs of the six modes are shown, and the simulation formulas for the different modes are as follows:

[0072] Where x(t) is the quality characteristic value of the product at time t, u is the statistical mean under controlled conditions, n is the total number of moments, and σ is the standard deviation of process capability. γ is the fluctuation caused by random factors, which obeys the Gaussian white noise sequence and cannot be removed or the cost of removal is high. d(t) is the fluctuation caused by abnormal factors at time t, the cause of which can be identified and eliminated. The simulation formula for each mode is shown in Table 1:

[0073] Table 1 Simulation formula table for each mode

[0074]

[0075]

[0076] Based on the above control chart mode simulation formula, the embodiment of the present invention uses the Monte Carlo method to simulate 6 modes of data. And according to the frequency of collecting a cigarette sample every 6 seconds in the actual production line, 5 minutes (50 cigarettes in 300 seconds) is used as the time span for drawing the control chart, that is, 50 quality characteristic observation points are used to draw a control chart sample. The advantage of using 5 minutes as the time span of a control chart is that it can fully capture short-term fluctuations in the production process, improve the sensitivity of detecting abnormalities, and help to discover and correct potential abnormal patterns at an early stage. Therefore, during the Monte Carlo simulation, the observation points of each control chart sample are 50, and 200 are simulated for each mode, totaling 1200 samples, 80% of which are selected as training sets and the rest as test sets.

[0077] According to different abnormal factors, the mean control chart can be divided into several different modes, specifically, including:

[0078] Normal mode (NOR): indicates that the machine is properly maintained, the material quality is stable, there are no major fluctuations, and it is in a stable state;

[0079] Cyclic mode (CYC): Cyclic fluctuations in equipment, such as periodic mechanical wear or vibration of the rollers, gears or transmission devices of the cigarette making machine, which leads to periodic fluctuations in cigarette quality;

[0080] Downward trend mode (DT) Upward trend mode (UT): due to the gradual wear of machine parts causing changes in tobacco volume, or increased ambient humidity leading to a gradual decrease or increase in cigarette mass;

[0081] Downward step mode (DS) and upward step mode (US): Indicates that sudden changes in temperature or humidity of environmental factors lead to changes in tobacco moisture, sudden adjustments in equipment or changes in raw material batches, which in turn lead to changes in cigarette indicators.

[0082] Then, the training features of each cigarette control chart training sample are extracted; the training features include 9 shape features and 6 statistical features.

[0083] Among them, shape features provide information about the appearance and form of a graphic by describing its geometric form and structure, and can help qualitatively describe and analyze the specific shape of the graphic. The nine shape features used in the present invention are as follows:

[0084] The slope of the least squares line: represents the trend of the pattern. For normal and periodic patterns, the absolute value of the slope is close to zero, while for step and trend patterns, the absolute value of the slope is greater than zero.

[0085] Average number of crossings (nc1): indicates the number of intersections between the pattern and the mean line. The normal pattern has the most intersections, the step and trend patterns have the least, and the cycle pattern is in between.

[0086] Number of least squares line crossings (nc2): indicates the number of times the pattern crosses the least squares line;

[0087] Cyclic member (cmember): indicates the similarity between the pattern and the periodic pattern. For periodic patterns, this value is positive, and for other patterns it is negative;

[0088] Average slope of line segments (as): represents the average slope of each line segment in the pattern, and the average slope of the trend pattern is the highest;

[0089] Slope difference (sd) between the least squares line and the line segment representing the pattern: the difference between the slope of the least squares line and the average slope of the line segment;

[0090] Area between mode and mean line (apm): The area of ​​this feature for normal mode is lowest, so this feature is different between normal mode and other modes;

[0091] Area between the mode and the least squares line (aps): This feature has higher aps values ​​in periodic and step modes than in normal and trend modes, so aps can distinguish between modes such as periodic and step;

[0092] Area between the least squares line and the line segment (ass): This feature is approximately 0 in the trend mode and may be higher for other modes, so this feature can distinguish the trend mode from other modes.

[0093] In addition, six statistical features are extracted for each cigarette control chart training sample, specifically:

[0094] Mean: represents the location of the data center, used to determine whether the data fluctuates around a certain central value and show the central trend of the data;

[0095] Cumulative sum statistic (cusum): A statistic that detects changes by accumulating deviations between data points and target values. It is used to monitor process changes, especially small, sustained shifts. Cumulative sum positive statistics are used to detect upward mean shifts, and cumulative sum negative statistics are used to detect downward mean shifts.

[0096] Skewness: reflects the symmetry of data distribution. Positive skewness indicates that the data is right-skewed, and negative skewness indicates that the data is left-skewed. It is used to identify the process trend of the inclination, deviation or asymmetry of data distribution.

[0097] Kurtosis is used to describe the sharpness of the data distribution. A high kurtosis indicates that the data distribution has a sharp peak, while a low kurtosis indicates a flat distribution. It is used to detect outliers or extreme values.

[0098] The mean square value is the average of the square values ​​of all data points. It is used to measure the energy or degree of variation of a data set. It can help detect the specific trend of data changes, especially in control charts, reflecting volatility and instability. The larger the value, the greater the surface fluctuation.

[0099] The standard deviation indicates the degree of dispersion between data points and the mean, and is used to measure data fluctuation and dispersion. Compared with the mean square value, the standard deviation reflects more how the data fluctuates around the mean.

[0100] Then, a feature test set is generated according to the training features; each feature test set corresponds to a cigarette control chart training sample, which includes 15 features extracted from the above cigarette control chart training sample.

[0101] Next, multiple classification network models are constructed.

[0102] In order to avoid the bias caused by the prediction of a single model, six conventional machine learning models were selected for preliminary classification prediction.

[0103] Traditional machine learning models mainly focus on how to improve the prediction accuracy of the model, but ignore the interpretability of the prediction results. Therefore, the SHAP interpretable model can be used to quantitatively explain the contribution of shape features and statistical features to the control chart pattern recognition ability.

[0104] The classification network models in this embodiment are specifically: decision tree, random forest, support vector machine, GBDT, XGBoost and logistic regression machine learning classification network models.

[0105] Specifically, we first train the cigarette control chart samples under multiple different modes; divide the training set and test set into the designated models decision tree, random forest, logistic regression, support vector machine, XGBoost and GBDT for preliminary classification. These six algorithms represent different learning methods and data processing capabilities, and are suitable for comprehensive analysis of feature contribution. After the models are trained, we compare them and obtain a comparison chart of ROC curves of multiple models, as shown in the figure below. Figure 8 shown.

[0106] The ROC curve contains multiple curves, each of which represents a different classification network model. The larger the area under the curve, the better the classification effect of the model. Figure 8 It can be seen that by comparing the ROC curve graph, it can be concluded that the support vector machine is used as the best classification network model to input the subsequent calculation of the SHAP model.

[0107] Afterwards, the above-screened support vector machine classification network model, feature test set and classification prediction result data output by the optimal classification network model are input into the SHAP interpreter.

[0108] The SHAP interpreter then calculates the marginal contribution of each feature in the feature test set to the classification prediction result data and generates multiple Shapley values; these Shapley values ​​show the positive or negative contribution of each feature to the classification prediction result. The SHAP interpreter traverses all possible combinations of features and calculates the contribution value of each feature, thus providing a fair evaluation. The generated Shapley values ​​are displayed through visual graphics, such as Fig. 9 As shown in the bee colony diagram, the diagram vividly and concretely represents the impact of each feature on the model decision. By summarizing these values, the most important features for control chart pattern recognition can be identified, thereby optimizing feature selection and improving the accuracy of pattern recognition. SHAP not only improves the interpretability of model prediction results, but also provides empirical evidence to guide feature screening and model optimization.

[0109] Fig. 9Each point in corresponds to a sample. The horizontal axis in the figure is the Shapley value (the data without units are dimensionless data). The positive and negative values ​​represent the positive or negative correlation between the feature value and the classification prediction result. The larger the Shapley value, the stronger the correlation. The vertical axis is the contributing feature, which is arranged in descending order from top to bottom according to the contribution of the feature. The points of a certain feature are mainly concentrated in the positive direction of the Shapley value (right side), which means that when the feature value is high, the probability of the model predicting a certain category is increased, and vice versa.

[0110] Although the nine shape features and six statistical features contribute to different control chart patterns, Fig. 9 From the SHAP beeswarm diagram, we can see that the contribution of different features varies greatly. Therefore, we select the first few features with a cumulative Shapley value of 90% to construct a high-value feature subset. Because the cumulative share of the first 90% can usually capture most of the important information while reducing the interference of unimportant features. By calculating the average absolute value of each feature Shapley value and sorting by cumulative share, we can determine which features have a major impact in the model, thereby effectively screening out the most important features for control chart pattern recognition. Doing so not only improves the efficiency of feature selection, but also ensures that the selected features can maximize the retention of the model's predictive power and interpretability.

[0111] In this embodiment, the 9 features in the high-value feature subset that can be determined are mean, mean square value, cumulative and positive statistics, kurtosis, skewness, area between the least squares line and the line segment, area between the mode and the least squares line, standard deviation, and slope of the least squares line.

[0112] Then, the high-value feature subset is input into the constructed graph convolutional neural network, the model is trained and the target graph convolutional neural network is obtained.

[0113] Graph convolutional neural network is a neural network that processes graph structure data. It can effectively capture the association between nodes (that is, samples). For control chart data, the similarity between samples can be constructed as the edges of the graph. Therefore, the input of the model is the node vector matrix X composed of the filtered features and the adjacency matrix P between control chart samples at different time sequences: Input = (X, P), where X is an n×d feature matrix, where n is the number of input control chart samples and d is the number of features filtered by the SHAP model. The adjacency matrix represents the similarity measure between historical and current samples in matrix form. The output of the model is different control chart patterns Y. The graph convolutional neural network model is weightedly aggregated through the convolution operation of two graph convolutional layers to obtain a new feature representation. And two dense layers further extract global information, and finally output the classification result Y through the activation function. Fig.10 Represents the topological structure of a graph convolutional neural network.

[0114] The high-value feature subsets are divided into training sets and test sets, and input into the graph convolutional neural network model for training. The training process uses the backpropagation algorithm, which mainly includes two steps, forward excitation propagation and backward weight update. For forward excitation propagation, the input variables are processed by multiple graph convolutional layers and then passed to the dense layer. The pattern type of the sample is output by the dense layer, and the classification result and the actual result are used to calculate the loss function (error). For backward weight update, the weight of each layer is updated by the gradient descent method. When the set number of iterations is reached, the test set is used to evaluate the performance of the graph convolutional neural network, such as Fig.11 The trend of the loss function in the embodiment as the number of iterations increases is shown.

[0115] Can be obtained from Fig.11 It can be seen that in the early stage of the training process, the loss function of the training set decreases rapidly with the increase of the number of iterations. When the number of iterations is greater than 200, the loss function tends to a constant and no longer continues to decrease, indicating that the network has converged and is relatively stable at this time. In order to ensure the convergence of the network. The graph convolutional neural network is used to control the classification of graph patterns after 500 iterations. After the model is adjusted and trained, the next step is to evaluate and verify the model performance. In order to ensure the scientificity and reliability of the evaluation, common classification evaluation indicators are used, including accuracy, recall and confusion matrix. Specifically, the accuracy is: The recall rate is: Among them, TP (TruePositives) represents true positive examples, which are predicted as positive examples and are actually positive examples, FP (FalsePositives) represents false positive examples, which are predicted as positive examples but are actually negative examples, FN (FalseNegatives) represents false negative examples, which are predicted as negative examples but are actually positive examples, and TN (TrueNegatives) represents true negative examples, which are predicted as negative examples and are actually negative examples. Therefore, the accuracy rate represents the ratio of all correctly predicted samples to the total number of samples, and the recall rate represents the proportion of samples that are correctly predicted by the model as a certain category among all samples that are actually of a certain category. In this embodiment, the training set and test set of the graph convolutional neural network for control chart pattern recognition are shown in Table 2:

[0116]

[0117] And the confusion matrix of the test set is as follows Fig.12 As shown. Fig.12 It can be seen that the recognition accuracy of the graph convolutional neural network for the control chart pattern in this embodiment reaches 94.3% on average, which can realize early warning of abnormal patterns caused by different faults in the production process.

[0118] After the training is completed, the target graph convolutional neural network is obtained, which is a network with optimal parameters set.

[0119] Finally, the dynamic data stream of cigarette quality is monitored using the filtered features and trained models. Specifically, samples of actual cigarette control charts are obtained and input into the target graph convolutional neural network to complete pattern recognition and find abnormal fluctuations in the cigarette production process.

[0120] Specifically, the longer the time span of the data is, the greater the lag in identifying abnormal quality characteristics; secondly, in the real-time production process, when the data source is continuously and dynamically changing, the sliding window technology can ensure that short-term changes are captured. Therefore, the window sequence of the measured data is set as the time span of the Monte Carlo simulation. That is, 50 cigarette quality characteristic values ​​are a mean control chart sample, and a sliding window with a window moving step of 1 is used to monitor the production process in real time. That is, for each additional cigarette quality value, the window moves back one position, and extracts the 9 features selected by the SHAP interpretable model, and then inputs the features into the target graph convolutional neural network for pattern classification.

[0121] Fig.13It is the real-time measurement value of the dynamic data stream of cigarette quality. When the sliding window moves to the 56th data, the recognition model alarm reminds that a downward trend pattern (DT) appears in the cigarette manufacturing process, and looks for abnormal fluctuations in the cigarette production process. The abnormal factors of the corresponding control chart pattern can be checked to find that it is because of changes in environmental temperature or humidity, or because the wear of the equipment continues to affect the process, which leads to a downward trend in cigarette quality. Therefore, targeted solutions and other measures can be proposed.

[0122] The above descriptions are only preferred embodiments disclosed in this application and descriptions of the technical principles used. Those skilled in the art should understand that the scope of protection involved in this disclosure is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in this disclosure (but not limited to) to form a technical solution.

[0123] In addition, although several specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Certain features described in the context of a separate embodiment can also be implemented in a single embodiment in combination. On the contrary, the various features described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable sub-combination.

Claims

1. A cigarette quality control graph pattern recognition method based on an interpretable graph convolutional neural network, characterized in that: Methods include: S01. Obtaining cigarette control chart training samples in multiple different modes; S02, extracting training features of each cigarette control chart training sample; The training features include multiple shape features and multiple statistical features; S03, generating a feature test set according to the training features; constructing multiple classification network models, inputting the feature test set corresponding to each cigarette control chart training sample into the multiple classification network models respectively, obtaining multiple classification prediction result data, and obtaining corresponding ROC curve graphs; S04, screening and obtaining an optimal classification network model according to the ROC curve graph; inputting the optimal classification network model, the feature test set and the classification prediction result data output by the optimal classification network model into the SHAP interpreter; S05, the SHAP interpreter calculates the marginal contribution of each feature in the feature test set to the classification prediction result data, generates multiple Shapley values; generates a high-value feature subset according to the Shapley values; S06. Inputting the high-value feature subset into the constructed graph convolutional neural network, training the model and obtaining a target graph convolutional neural network; S07. Obtain samples of the actual collected cigarette control graph, and input them into the target graph convolutional neural network to perform pattern recognition to obtain cigarette quality fluctuation results.

2. The cigarette quality control graph pattern recognition method based on interpretable graph convolutional neural network according to claim 1 is characterized in that: In S01, the method for obtaining cigarette control chart training samples under multiple different modes is specifically: based on the control chart mode simulation formula, using the Monte Carlo method to simulate and obtain cigarette control chart training samples under multiple different modes.

3. The cigarette quality control graph pattern recognition method based on interpretable graph convolutional neural network according to claim 1, characterized in that: The modes of the cigarette control chart training samples include: normal mode (NOR), cycle mode (CYC), downward trend mode (DT), upward trend mode (UT), downward step mode (DS) and upward step mode (US).

4. The cigarette quality control graph pattern recognition method based on interpretable graph convolutional neural network according to claim 1, characterized in that: The shape features include: the slope of the least squares line (slope), the average number of crossings (nc1), the number of least squares line crossings (nc2), the cycle member (cmember), the average slope of the line segment (as), the slope difference (sd) between the least squares line and the line segment representing the pattern, the area between the pattern and the mean line (apm), the area between the pattern and its least squares line (aps), and the area between the least squares line and the line segment (ass).

5. The cigarette quality control graph pattern recognition method based on interpretable graph convolutional neural network according to claim 1, characterized in that: The statistical features include: mean, cumulative sum statistic (cusum), skewness, kurtosis, mean square value and standard deviation.

6. The cigarette quality control graph pattern recognition method based on interpretable graph convolutional neural network according to claim 1, characterized in that: In S03, the multiple classification network models constructed are specifically: decision tree model, random forest model, logistic regression model, support vector machine model, XGBoost model and GBDT model.

7. The cigarette quality control graph pattern recognition method based on interpretable graph convolutional neural network according to claim 6 is characterized in that: In S05, a specific method for generating a high-value feature subset according to the Shapley value is: calculating the average absolute value of each Shapley value and sorting them according to the cumulative proportion, and selecting features with the first 90% of the cumulative proportion of Shapley values ​​to form a high-value feature subset.

8. The cigarette quality control graph pattern recognition method based on interpretable graph convolutional neural network according to claim 7 is characterized in that: In S06, a back propagation algorithm is used to train the constructed graph convolutional neural network and obtain the target graph convolutional neural network.

Images