A method and system for predicting the adaptability of a cigarette brand on a machine
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HONGYUN HONGHE TOBACCO (GRP) CO LTD
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-19
Smart Images
Figure CN122241354A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of cigarette manufacturing technology, specifically to a method and system for predicting the adaptability of cigarette trademarks for machine manufacturing. Background Technology
[0002] With the ever-expanding production scale of the cigarette industry, more and more high-speed printing units are being introduced and applied to production to meet market demand and improve production capacity. The use of high-speed units has significantly increased production speed; however, this also places more stringent demands on all aspects of cigarette production. On the one hand, high-speed production requires greater stability and consistency in raw and auxiliary materials; on the other hand, different cigarette production equipment, due to differences in parameters and manufacturing processes, has varying adaptability to cigarette label paper produced by different manufacturers. This "machine adaptability" directly manifests in whether paper jams, blockages, poor folding, or uneven packaging occur during production, affecting production efficiency and leading to increased material consumption.
[0003] Traditional methods for predicting the adaptability of cigarette labels for machine use employ a strategy based on past production experience and fixed binding. Specifically, based on previous production experience, it is determined that labels produced by a particular supplier are well-suited for a specific machine, with a lower probability of equipment malfunctions and quality issues. Then, the supplier's labels are bound to that machine, meaning it's assumed that all labels supplied by that supplier will be used on that machine. If problems such as paper jams, inaccurate positioning, or poor forming occur after actual use, then labels from other suppliers are tried.
[0004] However, this prediction method relies heavily on the operator's subjective experience and is easily affected by accidental factors such as equipment status at a specific time and fluctuations in raw material batches. It is difficult to accurately reflect the intrinsic relationship between key parameters of trademark paper and equipment performance, resulting in poor prediction accuracy. Summary of the Invention
[0005] The main purpose of this application is to provide a method and system for predicting the adaptability of cigarette labels on a machine, aiming to solve the technical problem of poor prediction accuracy in traditional methods for predicting the adaptability of cigarette labels on a machine.
[0006] To achieve the above objectives, this application provides the following technical solution:
[0007] A method for predicting the adaptability of cigarette labels for machine operation, comprising:
[0008] S1, obtain the multi-dimensional physical parameters of historical batches of trademark paper and the corresponding machine adaptability results;
[0009] S2, construct interactive features based on the multidimensional physical parameters, and train machine learning prediction models for different machines by taking the multidimensional physical parameters and interactive features as inputs and the machine adaptability results as outputs.
[0010] S3. Obtain the multidimensional physical parameters of the trademark paper to be tested, construct interactive features in the same way as in step S2, input the multidimensional physical parameters and interactive features into the trained machine learning prediction model, and output the machine adaptive prediction result.
[0011] Optionally, in step S1, the multidimensional physical parameters include indentation height, indentation width, indentation parallelism, static friction coefficient, dynamic friction coefficient, and indentation stiffness;
[0012] The indentation height, indentation width, and indentation parallelism are obtained using a high-precision laser measuring instrument; the static friction coefficient and dynamic friction coefficient are measured using a friction measuring instrument; and the indentation stiffness is obtained using a stiffness meter.
[0013] Optionally, in step S1, the on-machine adaptability result is obtained in the following way:
[0014] Obtain the actual production indicators of historical batches of trademark paper on the corresponding machine, classify the actual production indicators based on preset indicator thresholds, and obtain the machine adaptability label as the machine adaptability result.
[0015] The machine adaptability labels include good, molding defect, easy to clog, and surface whitening.
[0016] Optionally, in step S2, the interaction features are constructed in the following way:
[0017] The height and width of each indentation on the historical batch of trademark paper are obtained. Based on the indentation height and width, the interaction features are determined. The interaction features include the aspect ratio of each indentation, the height difference and width difference between adjacent indentations, the mean and standard deviation of all indentation heights, and the mean and standard deviation of all indentation widths.
[0018] Optionally, after the interaction features are constructed, the multidimensional physical parameters and interaction features are subjected to data standardization processing to eliminate the dimensional differences between the features in the multidimensional physical parameters and interaction features, so that the features are on the same scale.
[0019] Optionally, after the data standardization process, the standardized features are filtered, and the filtering includes:
[0020] Calculate the correlation between each feature and the on-machine adaptability result, and remove redundant features and features with low correlation;
[0021] Principal component analysis was performed on the remaining features after correlation screening to screen the original features that made significant contributions to the principal components;
[0022] Importance scores of the original features are extracted based on a random forest model. The top few original features whose cumulative importance reaches a preset importance threshold are selected as the key parameter set of the historical batch of trademark paper.
[0023] Optionally, the training of the machine learning prediction model includes:
[0024] A training dataset is constructed using the aforementioned key parameter set and the corresponding on-machine adaptation labels;
[0025] The training dataset is divided into a training set and a test set according to a preset ratio, and stratified sampling is performed according to the on-machine adaptability label to ensure that the sample ratio of each category in the training set and the test set is consistent with that in the training dataset.
[0026] Within the training set, cross-validation is used to tune the hyperparameters of various machine learning algorithms, and the optimal combination of hyperparameters for each algorithm is determined by the average performance score of cross-validation.
[0027] The various machine learning algorithms mentioned include support vector machines, decision trees, random forests, neural networks, and XGBoost.
[0028] Optionally, after the cross-validation is completed, the classification accuracy on the test set is used as the primary indicator to select the optimal machine learning prediction model for each machine from multiple algorithms.
[0029] When classification accuracy is similar, the machine learning prediction model with stronger interpretability should be selected as the final model for the machine.
[0030] Optionally, after step S3, the method further includes:
[0031] Based on the machine adaptability prediction results, the machine with the highest matching degree with the trademark paper to be tested is recommended;
[0032] Obtain the actual production indicators of the trademark paper to be tested in actual production, update the training dataset based on the actual production indicators, and iteratively optimize the machine learning prediction model based on the updated training dataset.
[0033] A cigarette label machine adaptability prediction system, applied to the cigarette label machine adaptability prediction method described above, includes:
[0034] The data import module is used to import the multidimensional physical parameters of historical batches of label paper and the corresponding machine adaptability results, as well as the multidimensional physical parameters of the label paper to be tested;
[0035] The data processing and storage module is used to construct interactive features based on the multidimensional physical parameters of the historical batches of trademark paper and store the processed data; and to construct interactive features based on the multidimensional physical parameters of the trademark paper to be tested using the same method.
[0036] The model prediction module is used to train a machine learning prediction model based on the stored processed data, and input the interactive features of the trademark paper to be tested into the trained model, and output the machine adaptive prediction result.
[0037] The result output module is used to output the prediction results and recommend the optimal matching machine.
[0038] The visualization module is used to visualize the prediction results, recommendation information, and system status.
[0039] The result feedback module is used to obtain the actual production indicators after the test trademark paper is put into the machine, process the actual production indicators and store them in the database to update the processed data, and iteratively optimize the machine learning prediction model.
[0040] The technical solution provided in this application can include the following beneficial effects: This application constructs a quantitative mapping relationship between the physical properties of trademark paper and its performance on different machines by collecting multi-dimensional physical parameters of historical batches of trademark paper and their corresponding machine adaptability results, and uses this relationship to train a machine learning prediction model. In subsequent predictions, the physical parameters of the trademark paper to be tested are input into the trained model, which can then output its adaptability prediction results on different machines. This method transforms traditional experience-based judgment into model prediction based on historical data training, which can more accurately reflect the intrinsic relationship between key parameters of trademark paper and equipment performance, while reducing the uncertainty caused by subjective factors, thereby improving prediction accuracy. Attached Figure Description
[0041] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0042] Figure 1 This is a simplified flowchart of the method for predicting the adaptability of cigarette labels on a machine;
[0043] Figure 2 This is a detailed flowchart of the method for predicting the adaptability of cigarette labels on machines;
[0044] Figure 3 This is a schematic diagram of the architecture of the cigarette label onboard adaptability prediction system. Detailed Implementation
[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the described embodiments are merely some, not all, of the embodiments of this application. Unless otherwise specified, the embodiments and features described in this application can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0046] Example 1:
[0047] See Figure 1 A method for predicting the adaptability of cigarette labels on a machine, comprising:
[0048] S1, obtain the multi-dimensional physical parameters of historical batches of trademark paper and the corresponding machine adaptability results;
[0049] S2, construct interactive features based on the multidimensional physical parameters, and train machine learning prediction models for different machines by taking the multidimensional physical parameters and interactive features as inputs and the machine adaptability results as outputs.
[0050] S3. Obtain the multidimensional physical parameters of the trademark paper to be tested, construct interactive features in the same way as in step S2, input the multidimensional physical parameters and interactive features into the trained machine learning prediction model, and output the machine adaptive prediction result.
[0051] Specifically, in step S1, multiple historical batches of label paper are collected, and the multidimensional physical parameters of each batch of label paper are measured. At the same time, these historical batches of label paper are actually produced on different cigarette rolling machines, and the actual production performance of each batch on each machine is recorded. Based on preset evaluation criteria, these performances are converted into machine adaptability results.
[0052] This step yields a set of pairwise data for each historical batch and each machine: the physical parameters of that batch and its suitability results on that machine. This set of data implies an intrinsic link between physical parameters and machine performance—physical parameters are inherent properties of the label paper, while machine performance is the external manifestation of these properties during the production process; the two establish a correspondence through the actual production process.
[0053] In step S2, the original physical parameters cannot fully reflect all the information related to the machine performance. Therefore, new features (i.e., interactive features) are derived from the original physical parameters through mathematical transformation. These interactive features can more comprehensively characterize the physical properties of the label paper. The original physical parameters and interactive features are merged to form a complete input feature set. Using the input feature set as the independent variable and the machine adaptation result as the dependent variable, machine learning models are trained separately for different machines. The core of the training process is to allow the model to repeatedly learn the mapping relationship between input features and output results until the model can accurately predict the corresponding adaptive result based on the input features. Since the operating characteristics of different machines differ, machine-specific modeling allows the model to more accurately capture the response patterns of each machine to the physical properties of the label paper. After training, the model internally stores the mapping relationship from physical parameters to machine performance—i.e., the judgment rules.
[0054] In step S3, for newly arrived trademark paper to be tested, its multidimensional physical parameters are measured in the same manner as in step S1, and its interactive features are constructed using the same mathematical transformation method as in step S2. These features are then input into the machine models trained in step S2. Since the model has learned the mapping rules for predicting machine performance from physical parameters, when the features of the paper to be tested are input, the model can automatically calculate the possible machine performance of the paper on each machine and output it in the form of adaptive prediction results (including potential fault types). Based on the prediction results, the system automatically recommends the machine with the highest suitability, achieving precise matching of trademarks "one batch, one policy".
[0055] Traditional prediction methods for the adaptability of cigarette labels on machines rely on past production experience and fixed binding strategies. This approach is highly dependent on operator subjectivity and easily affected by accidental factors such as equipment status at specific times and batch fluctuations in raw materials. It struggles to accurately reflect the intrinsic relationship between key parameters of the label paper and equipment performance, resulting in poor prediction accuracy. This application constructs a quantitative mapping relationship between the physical properties of the label paper and its machine performance by collecting multidimensional physical parameters of historical batches of label paper and their corresponding machine adaptability results. This relationship is then used to train a machine learning prediction model. In subsequent predictions, the physical parameters of the label paper to be tested are input into the trained model, which outputs its adaptability prediction results on different machines. This method transforms traditional experience-based judgment into model prediction based on historical data training, more accurately reflecting the intrinsic relationship between key parameters of the label paper and equipment performance, while reducing uncertainty caused by subjective factors, thereby improving prediction accuracy.
[0056] Example 2:
[0057] Based on Embodiment 1, optionally, in step S1, the multidimensional physical parameters include indentation height, indentation width, indentation parallelism, static friction coefficient, dynamic friction coefficient, and indentation stiffness;
[0058] The indentation height, indentation width, and indentation parallelism are obtained using a high-precision laser measuring instrument; the static friction coefficient and dynamic friction coefficient are measured using a friction measuring instrument; and the indentation stiffness is obtained using a stiffness meter.
[0059] Specifically, indentation height refers to the depth of each indentation on the label paper, measured in millimeters (mm); indentation width refers to the lateral dimension of each indentation on the label paper, measured in millimeters (mm); indentation parallelism refers to the degree of parallelism between indentations or between an indentation and a reference edge, reflecting the consistency of indentation positions; static friction coefficient refers to the coefficient of friction between the label paper surface and the contact material in a static state, reflecting the anti-slip performance of the paper during transport; dynamic friction coefficient refers to the coefficient of friction between the label paper surface and the contact material in a relative motion state, reflecting the smoothness of the paper during high-speed transport; and indentation stiffness refers to the bending stiffness of the label paper at the indentation point, reflecting the stability and resilience of the indented area during folding.
[0060] The above parameters were obtained using the following specialized measuring equipment: Indentation height, indentation width, and indentation parallelism were measured non-contactly using a high-precision laser measuring instrument, which accurately obtained the microscopic geometric dimensions of the indentation with a measurement accuracy down to the micrometer level; static and dynamic friction coefficients were measured using a tribometer, obtaining the frictional characteristics between the paper surface and the specific material through standard testing methods; indentation stiffness was measured using a stiffness meter, obtaining the bending stiffness value at the indentation point by applying a standard bending force. These multidimensional physical parameters collectively constitute the original attribute data of the label paper, serving as the basic input for subsequent construction of interactive features and training of predictive models.
[0061] Optionally, in step S1, the on-machine adaptability result is obtained in the following way:
[0062] Obtain the actual production indicators of historical batches of trademark paper on the corresponding machine, classify the actual production indicators based on preset indicator thresholds, and obtain the machine adaptability label as the machine adaptability result.
[0063] The machine adaptability labels include good, molding defect, easy to clog, and surface whitening.
[0064] Specifically, firstly, the actual production indicators of historical batches of trademark paper on the corresponding machines are obtained. In this embodiment, the actual production indicators include, but are not limited to, one or more of the following:
[0065] ①Equipment effective operating rate: refers to the effective operating rate of the cigarette machine during operation, reflecting the frequency of machine downtime caused by label paper issues;
[0066] ②Small box forming quality: The forming effect of cigarette packs is automatically judged by a visual inspection system, including whether the folds are neat, whether there is any glue separation, and whether the surface is flat.
[0067] ③ Number of cigarette packs rejected by the equipment: refers to the number of cigarette packs automatically rejected by the equipment per unit time due to problems with the label paper (such as paper jams, blockages, poor forming, etc.), and the unit is times / 10,000 cigarettes or times / hour.
[0068] These actual production indicators are automatically recorded during the production process by the equipment's built-in sensors, vision inspection systems, or production management (MES) systems, and are objective and quantifiable production data.
[0069] Secondly, the actual production indicators are categorized based on preset threshold values. Specifically, for each actual production indicator, several classification thresholds are pre-set, mapping continuous production indicator values to discrete classification labels. For example, for the number of rejections per unit (in ten thousand pieces), multiple classification thresholds can be preset according to actual production conditions, dividing the rejection count into different intervals, with each interval corresponding to a machine adaptability label. Further, a first threshold, a second threshold, and a third threshold can be set, classifying rejections less than the first threshold as "good," those between the first and second thresholds as "forming defects," those between the second and third thresholds as "easy to clog," and those greater than the third threshold as "surface whitening." Similar classification thresholds can be set for equipment effective operating rate and small box forming quality, i.e., pre-setting several thresholds for each indicator, mapping continuous indicator values to discrete adaptability labels.
[0070] Through the above classification process, the actual production performance of each historical batch of trademark paper on each machine is transformed into a discrete classification label, namely, the machine adaptability label. In this embodiment, the machine adaptability label includes the following four categories:
[0071] Good: The label paper runs smoothly after being loaded onto the machine, with no obvious quality problems; Forming defects: The label paper has slight creases, alignment deviations and other defects during the folding and forming process; Easy to clog: The label paper frequently jams or clogs during the conveying or folding process; Surface scratches: The label paper surface has scratches, wear and other appearance damage during the conveying process.
[0072] The machine adaptability label serves as the machine adaptability result and, together with the multidimensional physical parameters of the corresponding batch of trademark paper, constitutes the pairing data required for subsequent model training.
[0073] In this embodiment, by acquiring the actual production indicators of historical batches of trademark paper on the corresponding machines, and classifying these continuous production data based on preset indicator thresholds, the data is transformed into discrete machine adaptability labels (such as good, forming defect, easy to clog, surface whitening), thus providing a clear and objective "standard answer" for model training. This process not only quantifies the vague "machine performance" into specific labels that the model can learn, but also replaces the traditional judgment method that relies on subjective experience with objective rules; at the same time, the label generation rules also provide a unified quantitative basis for feedback optimization after the subsequent batches of trademark paper are machined, ensuring the consistency between training data and feedback data.
[0074] Example 3:
[0075] Based on the above embodiments, optionally, in step S2, the interaction feature is constructed in the following way:
[0076] The height and width of each indentation on the historical batch of trademark paper are obtained. Based on the indentation height and width, the interaction features are determined. The interaction features include the aspect ratio of each indentation, the height difference and width difference between adjacent indentations, the mean and standard deviation of all indentation heights, and the mean and standard deviation of all indentation widths.
[0077] Specifically, for each historical batch of trademark paper, a high-precision measuring device (such as a laser measuring instrument) is used to detect the geometric dimensions of each indentation in the key indentation area, recording the height Hi and width Wi of each indentation, where the subscript i represents the indentation number (for example, common trademark paper designs have 12 key indentations, then i=1, 2, ..., 12). Then, based on the obtained height and width of each indentation, interactive features that reflect the correlation between indentations and the overall distribution characteristics are calculated. In this embodiment, the interactive features include the indentation width-to-height ratio WH_divi (i=1, 2, ..., 12), the indentation height difference H_deltai (i=2, 3, ..., 12), the width difference W_deltai (i=2, 3, ..., 12), the mean indentation height H_mean and standard deviation H_std, and the mean indentation width W_mean and standard deviation W_std. These interactive feature variables are added as explanatory factors. The specific calculation method is as follows:
[0078] ① Aspect ratio of each indentation:
[0079]
[0080] This feature reflects the cross-sectional shape of each indentation, and the aspect ratio may affect the stress concentration and deformation behavior of the indentation during the folding process.
[0081] ② Height difference between adjacent indentations:
[0082]
[0083] This feature characterizes the variation trend of indentation height along the arrangement direction. Abrupt changes in the height of adjacent indentations may lead to uneven stress during folding, thus affecting the molding quality.
[0084] ③ Width difference between adjacent indentations:
[0085]
[0086] This feature is used to quantify the local fluctuations in the indentation width. Excessive width differences may cause local stress concentration in the paper at the indentation point.
[0087] ④ Average height of all indentations:
[0088]
[0089] This feature characterizes the overall level of indentation depth for this batch of trademark paper.
[0090] ⑤ Standard deviation of height for all indentations:
[0091]
[0092] This feature measures the dispersion between the heights of each indentation. A larger standard deviation indicates that the indentation depth is more uneven, which may lead to poorer folding consistency.
[0093] ⑥ Average width of all indentations:
[0094]
[0095] This feature reflects the overall level of the indentation width.
[0096] ⑦ Standard deviation of the width of all indentations:
[0097]
[0098] This feature is used to measure the uniformity of the indentation width. A large standard deviation means that the width of each indentation is significantly different, which may affect the consistency of folding.
[0099] The aforementioned interactive features, along with the original indentation height, width, and other multidimensional physical parameters (such as coefficient of friction and stiffness), constitute the input feature set required for subsequent model training. By introducing these interactive features, the geometric characteristics of the indentation area of the label paper can be more comprehensively characterized, especially the interrelationships between indentations and the overall uniformity. This provides the machine learning model with more effective information related to machine adaptability, helping to improve prediction accuracy.
[0100] Example 4:
[0101] Based on the above embodiments, optionally, after the interaction features are constructed, the multidimensional physical parameters and interaction features are subjected to data standardization processing to eliminate the dimensional differences between the features in the multidimensional physical parameters and interaction features, so that the features are on the same scale.
[0102] Specifically, after constructing the interaction features, a complete set of input features is now obtained, including:
[0103] Original multidimensional physical parameters: such as indentation height, indentation width, indentation parallelism, static friction coefficient, dynamic friction coefficient, indentation stiffness, etc. These parameters have different physical units and dimensions (for example, indentation height is in millimeters, friction coefficient is a dimensionless value, and stiffness may have a special unit of measurement).
[0104] The constructed interactive features include aspects such as the aspect ratio of each indentation, the height and width differences between adjacent indentations, the mean and standard deviation of all indentation heights, and the mean and standard deviation of all indentation widths. These features also have different numerical ranges and dimensions (for example, the aspect ratio is usually a dimensionless ratio, while the standard deviation has the same unit as the original height / width).
[0105] Due to their different units of measurement, the numerical ranges of these features can vary significantly. For example, indentation height may range from 0.2 to 0.5 mm, while aspect ratio may range from 1.6 to 6, and the standard deviation may be as small as 0.01 to 0.1. Without standardization, features with larger numerical ranges (such as aspect ratio) may dominate the optimization process during subsequent model training, while features with smaller numerical ranges (such as the standard deviation of indentation height) may be ignored by the model. This would prevent the model from fairly learning the true contribution of each feature to the prediction results.
[0106] Therefore, this embodiment performs data standardization on the above features, specifically using the Z-score standardization method (also known as zero-mean standardization). The calculation formula is as follows:
[0107]
[0108] in, These are the original eigenvalues. The mean of this feature on the training set. The standard deviation of this feature on the training set. These are the standardized feature values.
[0109] Through the above transformation, each feature is converted into a standardized feature with a mean of 0 and a standard deviation of 1. At this point, all features are on the same scale and are no longer affected by the original units or numerical ranges, allowing the model to fairly learn the importance of each feature. After standardization, a standardized feature set is obtained, which will serve as input for subsequent feature selection and model training. This step provides the model with feature data of uniform scale and is a preprocessing step that ensures the model's training effectiveness and predictive stability.
[0110] Example 5:
[0111] Based on the above embodiments, optionally, after the data standardization process, the features of the standardized data are filtered, and the filtering includes:
[0112] Calculate the correlation between each feature and the on-machine adaptability result, and remove redundant features and features with low correlation;
[0113] Principal component analysis was performed on the remaining features after correlation screening to screen the original features that made significant contributions to the principal components;
[0114] Importance scores of the original features are extracted based on a random forest model. The top few original features whose cumulative importance reaches a preset importance threshold are selected as the key parameter set of the historical batch of trademark paper.
[0115] Specifically, after the data standardization process in Example 5, a standardized feature set with uniform scale is obtained. However, these features may include noisy features unrelated to the machine adaptation results, highly correlated redundant features, and minor features with small contributions. Directly inputting all of them into the model will not only increase computational complexity but may also introduce the risk of overfitting, affecting the model's generalization ability. Therefore, it is necessary to perform multi-level screening on the standardized features to extract the most representative and predictive key feature subset as input for subsequent model training.
[0116] In this embodiment, the screening process specifically includes the following three steps executed sequentially:
[0117] Step 1: Relevance Screening – Eliminating Redundant and Low-Relevance Features
[0118] First, the correlation between each standardized feature and the computer adaptation result is calculated. Here, the computer adaptation result is the classification label defined in Example 2 (e.g., good, molding defect, easy to clog, surface whitening). The correlation measure can be the Spearman rank correlation coefficient, which effectively measures the degree of monotonic association between features and classification labels without making assumptions about data distribution, making it suitable for feature selection in classification problems.
[0119] For each feature, calculate the absolute value of its correlation coefficient with the computer-based adaptability results. If the absolute value is less than a preset correlation threshold (e.g., 0.2), it indicates that the feature has a weak correlation with the computer-based performance and is considered a low-correlation feature, which should be removed. Meanwhile, for features with excessively high correlation coefficients (e.g., absolute values greater than 0.9), it indicates that they carry highly repetitive information and are considered redundant features; one or both can be retained or removed depending on the actual situation.
[0120] The first step of screening removed a large number of irrelevant and redundant features, retaining a subset of features that are strongly correlated with the computer-based adaptability results.
[0121] Step 2: Principal Component Analysis – Screening Primitive Features That Significantly Contribute to Principal Components
[0122] The remaining features after the first screening step, while showing some correlation with the computer-based adaptability results, may still exhibit complex linear correlations. To further explore the intrinsic structure of the features and identify the core features that contribute most to data variation, principal component analysis (PCA) was performed on the remaining features.
[0123] Principal component analysis (PCA) combines original features into several uncorrelated principal components through linear transformation. Each principal component is a linear combination of the original features and can reflect the main variance information in the original data. For the first few principal components (e.g., the first 3-4, whose cumulative variance contribution rate usually reaches a high level), the factor loadings (i.e., the correlation coefficients between the original features and the principal components) of each original feature on these principal components are examined. The larger the absolute value of the factor loading, the more significant the contribution of the original feature to the principal component, meaning that the feature plays an important role in explaining the data variance.
[0124] Set a factor loading threshold (e.g., 0.6) to filter out the original features whose absolute factor loadings are greater than this threshold in the first few principal components. These features are considered the main contributors to the formation of the principal components and deserve special attention. This step further focuses on those original features that dominate the data structure.
[0125] Step 3: Random Forest Feature Importance Screening – Selecting Original Features with Cumulative Importance Metrics
[0126] After the first two steps of screening, a set of features that are significant in both relevance and data structure have been obtained. To finally determine the key parameter set for model training, a random forest model is used to evaluate feature importance.
[0127] Based on the currently selected feature set, a preliminary random forest classification model is trained. During training, the random forest model automatically calculates the importance score of each feature based on the purity improvement (such as the reduction in the Gini coefficient) brought about by each feature splitting at a decision tree node. After training, the importance scores of each feature are extracted and sorted from highest to lowest. Then, starting with the feature with the highest importance, the importance scores are accumulated sequentially to calculate the cumulative importance. A preset importance threshold (e.g., 60%) is set, and the top few features whose cumulative importance reaches this threshold are selected as the final set of key parameters.
[0128] These selected features are the original features (including the original multidimensional physical parameters and the constructed interaction features) that excel in all three dimensions: correlation strength, data structure contribution, and predictive ability. They constitute the core input for subsequent model training.
[0129] Through the above three steps of screening, the key parameter set of the historical batches of trademark paper is finally obtained. This key parameter set eliminates noise and redundancy in the original features, retains the most representative and predictive subset of features, and provides high-quality, low-dimensional input data for subsequent model training, which helps to improve the training efficiency and prediction accuracy of the model.
[0130] Example 6:
[0131] Based on the above embodiments, optionally, the training of the machine learning prediction model includes:
[0132] A training dataset is constructed using the aforementioned key parameter set and the corresponding on-machine adaptation labels;
[0133] The training dataset is divided into a training set and a test set according to a preset ratio, and stratified sampling is performed according to the on-machine adaptability label to ensure that the sample ratio of each category in the training set and the test set is consistent with that in the training dataset.
[0134] Within the training set, cross-validation is used to tune the hyperparameters of various machine learning algorithms, and the optimal combination of hyperparameters for each algorithm is determined by the average performance score of cross-validation.
[0135] The various machine learning algorithms mentioned include support vector machines, decision trees, random forests, neural networks, and XGBoost.
[0136] Specifically, after obtaining the key parameter set of historical batches of trademark paper through multi-level screening in Example 5, it is necessary to construct a dataset for model training based on this key parameter set and its corresponding machine adaptability labels, and complete the model training and hyperparameter optimization. Specifically:
[0137] Step 1: Construct the training dataset
[0138] Using the key parameter set determined in Example 5 as input features and the machine adaptability labels corresponding to each historical batch as defined in Example 2 as output targets, the two are paired and combined to construct a training dataset. Each data point in this training dataset contains a key parameter vector of a historical batch of trademark paper and its actual machine adaptability label on that machine. This dataset serves as the foundational material for subsequent model training.
[0139] Step 2: Divide the dataset into training and test sets (including stratified sampling).
[0140] The completed training dataset is divided into a training set and a test set according to a preset ratio. For example, 80% of the sample data can be used as the training set for model training and hyperparameter tuning; the remaining 20% of the sample data can be used as the test set to evaluate the generalization ability of the finally trained model.
[0141] During the partitioning process, stratified sampling is performed based on the aforementioned adaptive labels. Specifically, firstly, the proportion of samples for each category label (e.g., good, molding defect, prone to clogging, surface whitening) in the training dataset is statistically analyzed. Then, when partitioning the training and test sets, the proportion of samples for each category in the training and test sets is ensured to remain consistent with that in the original training dataset. The purpose of stratified sampling is to avoid a particular category having an excessively high or low proportion in the training or test set due to random partitioning, thereby ensuring the class balance of the model training and the reliability of the evaluation results.
[0142] Step 3: Cross-validation hyperparameter tuning
[0143] Within the training set, cross-validation is used to fine-tune the hyperparameters of various machine learning algorithms. In this embodiment, K-fold cross-validation can be used, for example, dividing the training set into 5 equal parts (K=5) and performing 5 training and validation iterations. In each iteration, 4 parts are selected as the training subset, and the remaining part is used as the validation set, and this process is repeated until every data set has been used as the validation set. The average of the 5 validation results is taken as the model performance score under this set of hyperparameter configurations, thereby reducing the evaluation variance caused by a single random partition.
[0144] For each machine learning algorithm, a set of candidate hyperparameter combinations is pre-defined (e.g., the number of trees and tree depth in a random forest, the kernel function type and penalty coefficient in a support vector machine). For each hyperparameter combination, the aforementioned K-fold cross-validation is performed on the training set to obtain an average performance score. The hyperparameter combination with the highest average performance score is selected as the optimal hyperparameter combination for the algorithm.
[0145] Through the cross-validation tuning described above, we can find the most suitable hyperparameter configuration for each algorithm based on the current data features without overfitting.
[0146] Step 4: Determine the various algorithms to be used in training
[0147] In this embodiment, the various machine learning algorithms involved in the training include, but are not limited to, the following five:
[0148] Support Vector Machine (SVM): It maps input features to a high-dimensional space through a kernel function to find the optimal hyperplane that maximizes the class margin. It is suitable for handling high-dimensional small sample data and non-linear classification problems.
[0149] Decision tree: It simulates the decision-making process with a tree structure. It recursively divides the data into subsets with high purity through feature splitting, and the model has strong interpretability.
[0150] Random Forest: Based on the idea of Bagging ensemble learning, it classifies multiple decorrelational decision trees by combining their voting results, which can effectively reduce the risk of overfitting and improve generalization ability.
[0151] Neural Network: By simulating the structure of biological neurons, it automatically learns the complex mapping relationship between features and labels using multi-layer nonlinear transformations, and has the ability to fit highly nonlinear functions;
[0152] XGBoost: Based on the idea of Boosting ensemble learning, it generates multiple decision trees in sequence. Each tree fits the residuals based on the previous tree, and introduces second-order Taylor expansion and regularization terms into the objective function, resulting in high optimization efficiency and generalization ability.
[0153] Using multiple algorithms for training and optimization can provide diverse candidates for subsequent model selection, which helps to select the prediction model that is most suitable for the current machine characteristics.
[0154] Through the above steps, the model training preparation, data partitioning, cross-validation optimization, and training of various algorithms for historical batches of trademark paper data were completed, laying the foundation for subsequent model selection.
[0155] Example 7:
[0156] Based on the above embodiments, optionally, after the cross-validation is completed, the classification accuracy on the test set is used as the primary indicator to select the optimal machine learning prediction model for each machine from multiple algorithms.
[0157] When classification accuracy is similar, the machine learning prediction model with stronger interpretability should be selected as the final model for the machine.
[0158] Specifically, after the cross-validation and hyperparameter tuning in Example 6, a training model with its optimal hyperparameter combination was obtained for each machine learning algorithm (such as support vector machine, decision tree, random forest, neural network, XGBoost, etc.). At this point, it is necessary to select the final prediction model that is most suitable for the current machine from these candidate models. The selection process follows the following two levels of evaluation criteria.
[0159] (a) Primary metric: Classification accuracy on the test set
[0160] Using the reserved test set (not used for training and cross-validation) in Example 6 as the evaluation benchmark, the classification accuracy of each candidate model on the test set was calculated. Classification accuracy is defined as the proportion of samples correctly predicted by the model to the total number of samples in the test set, and is a core indicator for measuring the model's ability to predict unknown future data.
[0161] The reason for using classification accuracy on the test set as the primary screening criterion is:
[0162] The samples in the test set were not involved in the model training and hyperparameter tuning process, which can objectively reflect the model's generalization ability in real application scenarios; the higher the accuracy, the more reliable the model's prediction results for unknown data, and the more in line with the actual application needs of this application.
[0163] For each machine, calculate the test set accuracy of all candidate algorithms on that machine and sort them from highest to lowest accuracy.
[0164] (ii) Secondary indicator: Model interpretability
[0165] In some cases, the classification accuracy of different algorithms on the test set may be very close, for example, the difference is less than a preset threshold (such as within 1%). In this case, it is difficult to distinguish the quality of the models based on accuracy alone, and a secondary evaluation criterion needs to be introduced—the interpretability of the model.
[0166] Model interpretability refers to the degree to which humans can understand and explain the model's decision-making process. In the application scenario of this application, interpretability has the following practical significance: it helps process engineers understand the specific basis for the model's judgment that "a certain batch of trademark paper is prone to clogging"; it helps analyze the reasons for predicted failures and guide subsequent process improvements or data supplementation; and it provides directional reference for model iterative optimization.
[0167] Based on the above considerations, when classification accuracy is similar, the machine learning prediction model with stronger interpretability is preferred as the final model for this machine. Among common algorithms, decision trees and random forests are generally considered to have better interpretability (e.g., they can output feature importance and visualize decision paths), while "black box" models such as neural networks have relatively weaker interpretability.
[0168] (III) Result of the selection
[0169] Through the above two steps of selection, a machine learning prediction model with the best overall performance was finally determined for each machine. This model ensures high prediction accuracy while taking into account the need for interpretability in practical applications, and can serve as the core engine for subsequent machine-based adaptive prediction of the trademark paper to be tested.
[0170] Example 8:
[0171] Based on the above embodiments, optionally, after step S3, the method further includes:
[0172] Based on the machine adaptability prediction results, the machine with the highest matching degree with the trademark paper to be tested is recommended;
[0173] Obtain the actual production indicators of the trademark paper to be tested in actual production, update the training dataset based on the actual production indicators, and iteratively optimize the machine learning prediction model based on the updated training dataset.
[0174] Specifically, after completing the machine adaptability prediction for the test label paper, this application further applies the prediction results to actual production decisions and forms a self-optimizing closed loop through feedback from actual production data to continuously improve the model's prediction accuracy and adaptability. This includes:
[0175] (i) Recommending the optimal matching machine based on the prediction results
[0176] In step S3 of Example 1, after inputting the features of the trademark paper to be tested into the trained machine learning prediction model of each machine, the machine adaptability prediction results of the trademark paper to be tested on each machine are obtained. These prediction results can be specific classification labels (such as good, forming defect, easy to clog, surface whitening) or prediction probabilities of each category.
[0177] Based on these prediction results, the system automatically assesses the degree of matching between the test label paper and each machine. The criteria for judging the degree of matching can be set according to actual production needs. For example, machines with a prediction result of "good" are given priority; if multiple machines are predicted to be "good", the prediction probabilities can be further compared; if there is no "good" option, the machine with the highest prediction result level is selected (e.g., "forming defects" is better than "easy to clog").
[0178] Based on the aforementioned evaluation rules, one or more machines with the highest matching degree with the label paper to be tested are selected from all available machines and output as recommendations. This recommendation information can be used to guide production scheduling, allocating the batch of label paper to the most suitable machines for production, thereby reducing production failures, efficiency reductions, and material losses caused by mismatch between auxiliary materials and equipment, achieving precise matching of "one batch, one policy".
[0179] (ii) Feedback and model iterative optimization based on actual production data
[0180] After the test label paper is recommended for production, real operational data will be generated during the actual production process. This step obtains these actual production indicators and feeds them into the model training stage, forming a data-driven self-learning loop.
[0181] Furthermore, after the production of the trademark paper to be tested is completed, the actual production indicators of that batch of trademark paper on the corresponding machine are obtained from the equipment sensors, vision inspection system, or production management system (MES). The actual production indicators are consistent with the types defined in claim 3, and may include, but are not limited to: equipment effective operating rate; small box forming quality; equipment rejection count.
[0182] Following the same method as in Example 2, these actual production indicators are classified based on preset indicator thresholds to obtain the actual machine adaptability labels for this batch of trademark paper on the machine (e.g., good, forming defect, easy to clog, surface whitening). These actual labels, together with the key parameter set of this batch of trademark paper (i.e., the features of this batch of data after interactive feature construction, data standardization, and feature filtering), constitute a new set of sample data.
[0183] This new set of sample data is then added to the training dataset, making it contain more and newer production instances. Subsequently, based on the updated training dataset, the model training process described in Example 6 (including data partitioning, cross-validation, hyperparameter tuning, etc.) is re-executed to iteratively optimize the machine learning prediction model. Iterative optimization can be triggered periodically (e.g., weekly, monthly) or after a certain number of new samples have been added.
[0184] Through this feedback optimization mechanism, the model can continuously absorb new production data and learn the impact of dynamic factors such as changes in equipment status and fluctuations in raw material batches on adaptability, thereby continuously improving prediction accuracy and system adaptability. This closed-loop design of "prediction → production → feedback → optimization" enables the prediction method of this application to have the ability to learn and continuously improve itself, and to better adapt to changes in the actual production environment.
[0185] It should be noted that the overall process of Examples 3 to 8 can be referred to Figure 3The detailed flowchart shown is implemented.
[0186] Example 9:
[0187] See Figure 3 Based on the above embodiments, this embodiment provides a cigarette label machine adaptability prediction system, applied to the cigarette label machine adaptability prediction method described above, including:
[0188] The data import module is used to import the multidimensional physical parameters of historical batches of label paper and the corresponding machine adaptability results, as well as the multidimensional physical parameters of the label paper to be tested;
[0189] The data processing and storage module is used to construct interactive features based on the multidimensional physical parameters of the historical batches of trademark paper and store the processed data; and to construct interactive features based on the multidimensional physical parameters of the trademark paper to be tested using the same method.
[0190] The model prediction module is used to train a machine learning prediction model based on the stored processed data, and input the interactive features of the trademark paper to be tested into the trained model, and output the machine adaptive prediction result.
[0191] The result output module is used to output the prediction results and recommend the optimal matching machine.
[0192] The visualization module is used to visualize the prediction results, recommendation information, and system status.
[0193] The result feedback module is used to obtain the actual production indicators after the test trademark paper is put into the machine, process the actual production indicators and store them in the database to update the processed data, and iteratively optimize the machine learning prediction model.
[0194] Specifically, the data import module is responsible for receiving and importing various types of raw data required for system operation, including:
[0195] Multidimensional physical parameters of historical batches of trademark paper, such as indentation height, indentation width, indentation parallelism, static friction coefficient, dynamic friction coefficient, and indentation stiffness, can be obtained through offline measurement or imported from a historical database.
[0196] Machine adaptability results corresponding to historical batches: that is, the actual production performance of each historical batch of trademark paper on the corresponding machine, and the labels obtained after classification (such as good, forming defect, easy to clog, surface whitening).
[0197] Multidimensional physical parameters of the label paper to be tested: Physical parameters of the same type as newly arrived batches of label paper, used for subsequent prediction. Imported data, after format verification, is transmitted to the data processing and storage module for further processing.
[0198] The data processing and storage module is responsible for data processing, feature construction, and storage, including the separate processing of historical data and data to be tested, specifically:
[0199] Processing of historical batch data: Based on the multidimensional physical parameters of the imported historical batch trademark paper, interactive features are constructed using the same method as subsequent predictions. These interactive features enhance the expressive power of the original parameters, for example, deriving features such as aspect ratio, adjacent difference, mean, and standard deviation from the indentation height and width. The original multidimensional physical parameters are merged with the constructed interactive features to form a complete feature set, and necessary preprocessing (such as data cleaning and standardization) is performed to obtain the processed historical data. The processed historical data and its corresponding on-machine adaptation results are associated and stored in the system database as the training dataset for subsequent model training.
[0200] Processing of the test batch data: Based on the imported multidimensional physical parameters of the test label paper, interactive features are constructed using the same method as historical batches, resulting in an interactive feature set for the test label paper. This feature set is either temporarily stored or directly transferred to the model prediction module for prediction. The data processing and storage module ensures consistency in feature construction methods between historical data and test data, which is crucial for guaranteeing the effectiveness of model prediction.
[0201] The model prediction module is the core computing unit of the system, and it mainly performs two tasks:
[0202] Model Training: Processed historical data and their corresponding on-machine adaptation results are read from the database to form a training sample set. Based on this sample set, machine learning prediction models are trained separately for different cigarette-making machines. During training, various algorithms (such as support vector machines, decision trees, random forests, neural networks, XGBoost, etc.) are used for model construction, and model parameters are optimized through methods such as cross-validation. After training, the optimal model parameters and structure for each machine are saved to the model library for subsequent prediction calls.
[0203] Online prediction: Receives the interactive feature set of the trademark paper to be tested from the data processing and storage module. Inputs this feature set into the pre-trained prediction models of each machine, runs the model inference, and obtains the machine adaptability prediction results (e.g., classification label or probability) of the trademark paper to be tested on each machine. Outputs the prediction results to the result output module.
[0204] The results output module is responsible for transforming the prediction results generated by the model prediction module into usable decision information, specifically including:
[0205] Output the machine adaptability prediction results: display the predicted category (good, forming defect, easy to clog, surface whitening) of the test label paper on each machine in the form of text, table or graphic.
[0206] Recommended Optimal Matching Machine: Based on the prediction results, the system automatically calculates and recommends the machine with the highest matching degree to the test label paper according to preset matching rules (e.g., prioritizing machines predicted as "good," or selecting the highest-level machine if none is available), for production scheduling reference. Output information can be simultaneously transmitted to the visualization module and the external production management system.
[0207] The visualization module provides users with an intuitive system operation interface, which mainly includes:
[0208] Visualization of prediction results: The distribution and probability of the prediction results of the trademark paper to be tested on each machine are displayed in the form of charts.
[0209] Recommended information display: Highlights the recommended machines and the reasons for their matching.
[0210] System status monitoring: Real-time display of data import status, model running status, database storage status, etc., making it easy for operation and maintenance personnel to grasp the health of the system.
[0211] The visualization module helps users quickly understand the prediction results and improves the system's usability.
[0212] The results feedback module enables the system's self-learning and continuous optimization loop. The specific workflow is as follows:
[0213] After the test label paper is recommended for production, its actual production indicators (such as equipment effective operating rate, small box forming quality, equipment rejection rate, etc.) are obtained from the production site. These actual production indicators are converted into real machine adaptability labels using the same method as historical data. The real labels are combined with the key parameter set corresponding to this batch of label paper (i.e., features after interactive feature construction, standardization, and feature selection) to form new sample data. The new sample data is stored in the database to update the original training dataset. Based on the updated dataset, the model prediction module is triggered periodically or as needed to retrain and iteratively optimize the machine learning model, enabling the model to adapt to changes in production conditions and continuously improve prediction accuracy.
[0214] Through the results feedback module, the system forms a complete closed loop of "data → training → prediction → production → feedback → optimization", which has the ability to dynamically adapt and continuously improve.
[0215] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for predicting the adaptability of cigarette labels for machine use, characterized in that, include: S1, obtain the multi-dimensional physical parameters of historical batches of trademark paper and the corresponding machine adaptability results; S2, construct interactive features based on the multidimensional physical parameters, and train machine learning prediction models for different machines by taking the multidimensional physical parameters and interactive features as inputs and the machine adaptability results as outputs. S3. Obtain the multidimensional physical parameters of the trademark paper to be tested, construct interactive features in the same way as in step S2, input the multidimensional physical parameters and interactive features into the trained machine learning prediction model, and output the machine adaptive prediction result.
2. The method for predicting the adaptability of cigarette trademarks for machine use according to claim 1, characterized in that, In step S1, the multidimensional physical parameters include indentation height, indentation width, indentation parallelism, static friction coefficient, dynamic friction coefficient, and indentation stiffness. The indentation height, indentation width, and indentation parallelism are obtained using a high-precision laser measuring instrument; the static friction coefficient and dynamic friction coefficient are measured using a friction measuring instrument; and the indentation stiffness is obtained using a stiffness meter.
3. The method for predicting the adaptability of cigarette trademarks for machine use according to claim 2, characterized in that, In step S1, the onboard adaptability result is obtained in the following way: Obtain the actual production indicators of historical batches of trademark paper on the corresponding machine, classify the actual production indicators based on preset indicator thresholds, and obtain the machine adaptability label as the machine adaptability result. The machine adaptability labels include good, molding defect, easy to clog, and surface whitening.
4. The method for predicting the adaptability of cigarette trademarks for machine use according to claim 1, characterized in that, In step S2, the interaction features are constructed in the following way: The height and width of each indentation on the historical batch of trademark paper are obtained. Based on the indentation height and width, the interaction features are determined. The interaction features include the aspect ratio of each indentation, the height difference and width difference between adjacent indentations, the mean and standard deviation of all indentation heights, and the mean and standard deviation of all indentation widths.
5. The method for predicting the adaptability of cigarette trademarks for machine use according to claim 4, characterized in that, After the interaction features are constructed, the multidimensional physical parameters and interaction features are subjected to data standardization processing to eliminate the dimensional differences between the features in the multidimensional physical parameters and interaction features, so that the features are on the same scale.
6. The method for predicting the adaptability of cigarette trademarks for machine use according to claim 5, characterized in that, After the data standardization process, the standardized features are filtered, and the filtering includes: Calculate the correlation between each feature and the on-machine adaptability result, and remove redundant features and features with low correlation; Principal component analysis was performed on the remaining features after correlation screening to screen the original features that made significant contributions to the principal components; Importance scores of the original features are extracted based on a random forest model. The top few original features whose cumulative importance reaches a preset importance threshold are selected as the key parameter set of the historical batch of trademark paper.
7. The method for predicting the adaptability of cigarette trademarks for machine use according to claim 6, characterized in that, The trained machine learning prediction model includes: A training dataset is constructed using the aforementioned key parameter set and the corresponding on-machine adaptation labels; The training dataset is divided into a training set and a test set according to a preset ratio, and stratified sampling is performed according to the on-machine adaptability label to ensure that the sample ratio of each category in the training set and the test set is consistent with that in the training dataset. Within the training set, cross-validation is used to tune the hyperparameters of various machine learning algorithms, and the optimal combination of hyperparameters for each algorithm is determined by the average performance score of cross-validation. The various machine learning algorithms include support vector machines, decision trees, random forests, neural networks, and XGBoost.
8. The method for predicting the adaptability of cigarette trademarks for machine use according to claim 7, characterized in that, After the cross-validation is completed, the classification accuracy on the test set is used as the primary indicator to select the optimal machine learning prediction model for each machine from multiple algorithms. When classification accuracy is similar, the machine learning prediction model with stronger interpretability should be selected as the final model for the machine.
9. The method for predicting the adaptability of cigarette trademarks for machine use according to claim 7, characterized in that, The step S3 is followed by: Based on the machine adaptability prediction results, the machine with the highest matching degree with the trademark paper to be tested is recommended; Obtain the actual production indicators of the trademark paper to be tested in actual production, update the training dataset based on the actual production indicators, and iteratively optimize the machine learning prediction model based on the updated training dataset.
10. A predictive system for the adaptability of cigarette labels for machine use, characterized in that, The method for predicting the adaptability of cigarette labels on a machine, as described in any one of claims 1 to 9, includes: The data import module is used to import the multidimensional physical parameters of historical batches of label paper and the corresponding machine adaptability results, as well as the multidimensional physical parameters of the label paper to be tested; The data processing and storage module is used to construct interactive features based on the multidimensional physical parameters of the historical batches of trademark paper and store the processed data; and to construct interactive features based on the multidimensional physical parameters of the trademark paper to be tested using the same method. The model prediction module is used to train a machine learning prediction model based on the stored processed data, and input the interactive features of the trademark paper to be tested into the trained model, and output the machine adaptive prediction result. The result output module is used to output the prediction results and recommend the optimal matching machine. The visualization module is used to visualize the prediction results, recommendation information, and system status. The result feedback module is used to obtain the actual production indicators after the test trademark paper is put into the machine, process the actual production indicators and store them in the database to update the processed data, and iteratively optimize the machine learning prediction model.