Cigarette end shred falling amount characterization method and system and storage medium

Abstract

The invention relates to a tobacco production data feature screening technology, in particular to a method and a system for characterizing cigarette end shred falling quantity based on prior process knowledge and a storage medium, and the method comprises the following steps: obtaining cigarette production process data and carrying out dynamic region division; performing multi-dimensional feature extraction on the data after region division to obtain a multi-dimensional feature set; performing edge feature pre-screening according to the multi-dimensional feature set to obtain an effective feature set; obtaining a comprehensive score feature set according to the effective feature set; performing intra-region feature screening on the comprehensive score feature set to obtain an initial feature set and a candidate feature set; and performing improved sequence forward selection fine screening based on the initial feature set and the candidate feature set to obtain an optimal feature subset. According to the characterization method for the tobacco rod end portion shred falling amount, higher characterization precision, higher model robustness, better process interpretability and wider production scene adaptability are achieved.

Description

Technical Field

[0001] This invention relates to tobacco production data feature screening technology, specifically to a method, system, and storage medium for characterizing the amount of tobacco strands at the end of a cigarette based on prior process knowledge. Background Technology

[0002] In cigarette production, the amount of tobacco falling at the cigarette tip is a core indicator of cigarette quality, and its stability directly affects the smoking experience, combustion performance, and production costs. Traditional methods for detecting tobacco falling at the cigarette tip rely on offline sampling and weighing, which suffers from detection lag and the inability to monitor in real time. Microwave density data from 32 segments of the cigarette tip is a significant factor influencing tobacco falling at the tip, and this density data is measured online in real time, providing a data foundation for the digital representation of tobacco falling at the tip. However, accurately selecting features related to tobacco falling at the tip from the 32 density segments faces multiple challenges: the interplay of noise and effective information in high-dimensional data, the difficulty in characterizing the nonlinear relationship between features and tobacco falling amount, insufficient integration of process knowledge and data-driven methods, and the decreased model generalization ability due to feature redundancy. These issues have become technical bottlenecks restricting the accurate representation of tobacco falling at the cigarette tip.

[0003] Existing feature selection methods have significant limitations in the application of industrial quality control. Traditional methods often use a single indicator (such as the Pearson correlation coefficient) to measure the correlation between features and target variables, which can only capture linear associations and is difficult to characterize the nonlinear mapping relationship between density data and the amount of tobacco scrap. For example, when dealing with the correlation between density gradient abrupt changes and the amount of tobacco scrap, linear correlation analysis often leads to feature selection bias due to the asymmetry of data distribution. In addition, most feature selection algorithms lack a systematic integration of process knowledge: on the one hand, fixed region division methods (such as uniformly dividing cigarette segments) do not consider the process characteristics of density distribution (such as the gradient difference between the ends and the middle), resulting in feature extraction that does not match actual process requirements; on the other hand, because features are not grouped and managed according to process knowledge, it is difficult to effectively control feature redundancy within groups, resulting in increased model complexity and a tendency to overfit.

[0004] The shortcomings of existing technologies are more pronounced in the feature pre-screening and fine-screening stages. Statistical screening methods (such as coefficient of variation and significance tests) typically process each feature independently, ignoring the temporal correlation and spatial distribution characteristics between features. Search algorithms such as sequence forward selection are prone to getting trapped in local optima because they lack dynamic redundancy verification mechanisms. In the cigarette production scenario, the 32 density data segments exhibit strong spatial correlation, with density fluctuations in adjacent segments influencing each other. Traditional methods struggle to retain key features while eliminating redundant information. Furthermore, existing technologies lack a dynamic weight adjustment mechanism for feature importance, failing to balance the contributions of linear and nonlinear correlations to the characterization of end-burn cigarette weight, resulting in insufficient robustness of feature screening results.

[0005] Therefore, how to deeply integrate prior process knowledge with data-driven feature selection methods to achieve dynamic region division of cigarette density data, nonlinear correlation measurement of multi-dimensional features, and redundancy control based on process knowledge has become a core technical challenge for improving the accuracy of cigarette burnout characterization. This invention aims to solve the problems of linear correlation dependence, poor process adaptability, and insufficient redundancy control in existing methods during the feature selection process by integrating process knowledge with a multi-dimensional feature evaluation system. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, and storage medium for characterizing the amount of tobacco residue at the end of a cigarette based on prior process knowledge, in order to solve the problems of linear correlation dependence, poor process adaptability, and insufficient redundancy control in the feature selection process of existing technologies.

[0007] To achieve the above objectives, embodiments of the present invention provide a method for characterizing the amount of tobacco residue at the cigarette tip, comprising: Acquire data on the cigarette production process and dynamically divide the region; Multi-dimensional feature extraction is performed on the data after regional division to obtain a multi-dimensional feature set; Based on the multi-dimensional feature set, edge feature pre-screening is performed to obtain an effective feature set; Obtain the comprehensive scoring feature set based on the effective feature set; The comprehensive scoring feature set is subjected to regional feature filtering to obtain an initial feature set and a candidate feature set; Based on the initial feature set and candidate feature set, an improved sequence forward selection fine screening is performed to obtain the optimal feature subset.

[0008] Optionally, acquiring cigarette production process data and dynamically dividing the area includes: Acquire production data on cigarette burnout amount and discrete microwave density data of multiple segments along the cigarette axis; The density gradient is calculated based on the discrete microwave density data of the cigarette along the axial direction of the multi-segment to obtain the gradient value; Potential mutation points are obtained based on the gradient values; Valid mutation points are selected based on the potential mutation points; The regions are divided based on the effective mutation points.

[0009] Optionally, multi-dimensional feature extraction is performed on the data after region segmentation to obtain a multi-dimensional feature set including: Features were extracted from the partitioned data from the dimensions of central tendency, dispersion, distribution pattern, and time series characteristics. Central tendency includes mean, median, and mode; dispersion includes standard deviation, variance, range, and interquartile range; distribution pattern includes skewness and kurtosis; and time series characteristics include difference mean, difference standard deviation, time trend, and zero crossover rate.

[0010] Optionally, edge feature pre-screening is performed based on the multi-dimensional feature set to obtain an effective feature set, including: The coefficient of variation is used to filter the multidimensional feature set to obtain a first filtered dataset; The multi-dimensional feature set is subjected to monotonicity-saliency joint screening to obtain a second screening dataset; The intersection of the first and second filtered datasets is used to obtain the effective feature set.

[0011] Optionally, obtaining the comprehensive scoring feature set based on the effective feature set includes: Based on the effective feature set, calculate the nonlinear correlation value and the linear correlation value respectively; Dynamic weights are set based on the nonlinear and linear correlation values; The comprehensive score for each effective feature is calculated based on the dynamic weights. Based on the comprehensive score of each valid feature, obtain the comprehensive score feature set.

[0012] Optionally, calculating the comprehensive score for each valid feature based on the dynamic weights includes: Calculate the comprehensive score for each valid feature according to formula (1). (1) in, For comprehensive scoring, For dynamic weights, The normalized linear correlation value, This is the normalized nonlinear correlation value.

[0013] Optionally, improving the sequence forward selection based on the initial feature set and candidate feature set to obtain the optimal feature subset includes: Based on the current initial feature set and the current candidate feature set, the best feature set is selected for inclusion to obtain the current feature subset; Based on the current feature subset, calculate the correlation coefficient and remove redundant features; Update the current initial feature set and the current candidate feature set based on the current feature subset after removing redundant features; Determine if the stopping condition is met; If the stopping condition is not met, return to the step of selecting the best feature set based on the current initial feature set and the current candidate feature set to obtain the current feature subset; If the stopping condition is met, select the feature subset corresponding to the lowest prediction error as the optimal feature subset.

[0014] Optionally, the current feature subset is obtained by selectively adding features based on the current initial feature set and the current candidate feature set, including: Set the initial feature set to The candidate feature set is ; For the candidate feature set Each feature in Construct an extended feature set ; Based on each extended feature set Train random forest models separately and calculate the corresponding prediction errors on the validation set; Select the feature that minimizes the prediction error from all candidate features. ; Selected features Add the current initial feature set as the current feature subset.

[0015] On the other hand, the present invention also provides a cigarette end-burning quantity characterization system, the system including a processor configured to perform any of the methods described above.

[0016] In another aspect, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement any of the methods described above.

[0017] The beneficial effects of this invention are: The embodiments of this invention utilize a dynamic region division method that integrates the spatial distribution characteristics of density data with process knowledge. First, it adaptively divides the process regions of the cigarette end, middle, and end based on density gradient mutation points. This overcomes the technical limitations of traditional fixed region division methods that are disconnected from process characteristics, achieves precise adaptation to the process characteristics of cigarette density distribution, and solves the common problem of mismatch between feature extraction and process requirements.

[0018] The embodiments of this invention construct a multivariate pre-screening framework based on the coefficient of variation, monotonicity test, and significance analysis. Combined with a feature redundancy control strategy within process groups, it effectively separates noise information and effective features in high-dimensional data while preserving the spatiotemporal correlation of density data. This overcomes the shortcomings of traditional statistical methods that ignore the spatiotemporal correlation between features, which leads to screening bias. It provides a high-quality, low-redundancy initial feature set for subsequent fine feature screening.

[0019] The embodiments of this invention, by designing a comprehensive evaluation mechanism that combines linear and nonlinear correlation metrics, and employing a dual evaluation system that combines Pearson correlation coefficient with feature importance based on random forest, achieve a comprehensive characterization of the complex mapping relationship between density data and the amount of fiber shedding. This solves the technical problem that traditional methods cannot effectively characterize nonlinear process correlations due to their reliance on linear correlation analysis alone.

[0020] The embodiments of this invention introduce an improved sequence forward selection algorithm with dynamic redundancy check. During the iterative addition of features, the feature correlation coefficient matrix is ​​calculated in real time and highly redundant features are eliminated. Combined with a model performance-oriented search strategy, this invention overcomes the technical bottleneck of traditional search algorithms that are prone to getting trapped in local optima and cannot dynamically control feature redundancy. It achieves the optimization goal of maximizing representation accuracy while ensuring the model's generalization ability.

[0021] The embodiments of this invention establish a feature grouping management and weight balancing mechanism guided by process knowledge, integrate prior process knowledge into the entire feature screening process in a structured manner, and achieve deep integration of data-driven methods and domain knowledge by dynamically balancing the contribution of linear and nonlinear correlations through parameters. This solves the core problems of poor process adaptability and weak feature interpretability in the prior art.

[0022] Through the above-mentioned technical means, the cigarette end-fiber quantity characterization method finally constructed by this invention maintains the advantages of online real-time monitoring while achieving higher characterization accuracy, stronger model robustness, better process interpretability, and wider adaptability to production scenarios, providing reliable technical support for precise quality control in the cigarette production process.

[0023] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0024] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart of a method for characterizing the amount of tobacco residue at the end of a cigarette according to an embodiment of the present invention; Figure 2 A flowchart illustrating a method for acquiring cigarette production process data and dynamically dividing regions according to an embodiment of the present invention; Figure 3 A flowchart illustrating a method for pre-screening edge features to obtain an effective feature set according to an embodiment of the present invention; Figure 4A flowchart of a method for obtaining a comprehensive scoring feature set according to an embodiment of the present invention; Figure 5 A flowchart illustrating a method for improving sequence forward selection and fine screening to obtain an optimal feature subset according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the dynamic region division result according to one embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the mutual information values ​​between the features and the amount of yarn discarded according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the Pearson correlation coefficient between a feature and the amount of fiber shedding according to an embodiment of the present invention; Figure 9 This is a line graph illustrating the changes in the number of process features according to one embodiment of the present invention. Figure 10 This is a bar chart illustrating the changes in the number of process features according to one embodiment of the present invention; Figure 11 This is a schematic diagram showing the order of importance of the final process features according to one embodiment of the present invention. Detailed Implementation

[0025] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0026] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0027] like Figure 1 The diagram shows a flowchart of a method for characterizing the amount of tobacco residue at the cigarette tip according to an embodiment of the present invention. Figure 1 The detection method may include the following steps: In step S10, data on the cigarette production process is acquired and dynamic area division is performed; In step S11, multi-dimensional feature extraction is performed on the data after region division to obtain a multi-dimensional feature set; In step S12, edge features are pre-screened based on the multi-dimensional feature set to obtain an effective feature set; In step S13, a comprehensive scoring feature set is obtained based on the effective feature set; In step S14, the comprehensive scoring feature set is filtered within the region to obtain an initial feature set and a candidate feature set; In step S15, an improved sequence forward selection fine screening is performed based on the initial feature set and candidate feature set to obtain the optimal feature subset.

[0028] In such Figure 1 In the illustrated method for characterizing the amount of tobacco strands at the branch end, step S10 is used to acquire tobacco production process data and perform dynamic region division. In this embodiment, the specific method for acquiring tobacco production process data and performing dynamic region division in step S10 can be of various forms known to those skilled in the art. In one example of the present invention, step S10 may include, for example... Figure 2 The steps shown are described in this. Figure 2 In this context, step S10 may include: In step S20, production data on the amount of cigarette shreds and discrete microwave density data of multiple segments along the axial direction of the cigarette are obtained. In step S21, the density gradient is calculated based on the discrete microwave density data of multiple segments along the cigarette axis to obtain the gradient value; In step S22, potential mutation points are obtained based on gradient values; In step S23, valid mutation points are screened based on potential mutation points; In step S24, the region is divided according to the effective mutation points.

[0029] In such Figure 2In the method shown, step S20 is used to acquire tobacco shred production data and discrete microwave density data of multiple segments along the cigarette axis. The tobacco shred production data represents the overall weight or filling amount of tobacco shreds within the cigarette, macroscopically reflecting the weight of a cigarette. In this example, this could involve reading the tobacco shred production data and microwave density data for 32 segments of the cigarette. Using microwave technology, a cigarette is cut into 32 segments along its length (axial direction) at equal (or approximately equal) intervals, and the density of each segment is measured. This results in a sequence of 32 density values, the order of which corresponds to the position of the cigarette from one end to the other. Step S21 is used to calculate the density gradient. In this example, this could involve sorting the density sequence along the cigarette axis and then calculating the density gradient between adjacent segments using the central difference method. The density difference between each segment and its preceding and following segments is calculated, resulting in a value reflecting the degree of change. The greater the change, the larger the absolute value of the gradient, ultimately yielding a sequence of 31 absolute gradient values. Step S22 is used to obtain potential mutation points based on gradient values. In this example, the 75th quantile of the absolute gradient value sequence can be used as a threshold to locate potential mutation points: sort the 31 absolute gradient values ​​from smallest to largest, find their 75th quantile (i.e., the value larger than 75% of the data), and set this value as the threshold. Then, check all gradient values; any point exceeding this threshold is considered to have a sudden change in density, and this point is marked as a potential mutation point. Step S23 is used to filter effective mutation points based on potential mutation points. Since the two ends of the cigarette (the beginning and the end) may have large density changes due to physical reasons (such as the filter tip, mouthpiece, etc.), these mutation points are not internal region boundaries. Therefore, it is necessary to remove the boundary mutation points located at the very beginning and the very end of the sequence, and what remains are the effective density mutation points inside the cigarette. Usually, 1-2 effective mutation points remain. Step S24 is used to divide the region based on the effective mutation points. If two effective mutation points are found, they divide the cigarette density sequence into three segments. Therefore, the first segment (the area with lower density / greater variation) can be defined as the end region, the middle segment (the area with higher density and stability) can be defined as the middle region, and the last segment (the area where the density decreases / changes again) can be defined as the tail region. If only one abrupt change point is found, the cigarette may be simply divided into two parts (such as the end and the middle and tail).

[0030] Step S11 is used to extract multi-dimensional features from the data after region division to obtain a multi-dimensional feature set. In this embodiment, the specific method for multi-dimensional feature extraction can be of various forms known to those skilled in the art. In one example of the present invention, step S11 can be based on the ends, middle, and tails in step S10, performing multi-dimensional feature extraction for each region and global feature extraction, with the extraction processes for the ends, middle, and tails being consistent. Features are extracted from four dimensions: central tendency, dispersion, distribution pattern, and temporal characteristics. Among them, central tendency includes mean, median, and mode; dispersion includes standard deviation, variance, range, and interquartile range; distribution pattern includes skewness and kurtosis; and temporal characteristics include difference mean, difference standard deviation, time trend, and zero crossover rate. Global features are extracted synchronously on the entire density data according to the same dimensions, forming a multi-dimensional feature set covering both local and global aspects.

[0031] Step S12 is used to perform edge feature pre-screening based on the multi-dimensional feature set to obtain an effective feature set. In this embodiment, the specific method for performing edge feature pre-screening to obtain an effective feature set in step S12 can be of various forms known to those skilled in the art. In one example of the present invention, step S12 may include, for example... Figure 3 The steps shown are described in this. Figure 3 In this context, step S12 may include: In step S30, the coefficient of variation of the multi-dimensional feature set is filtered to obtain the first filtered dataset; In step S31, the multi-dimensional feature set is subjected to monotonicity-significance joint screening to obtain the second screening dataset; In step S32, the intersection of the first filtered dataset and the second filtered dataset is taken to obtain the effective feature set.

[0032] In such Figure 3In the method shown, step S30 is used to perform coefficient of variation (CV) filtering on the multi-dimensional feature set, removing features that show almost no change across different samples. By calculating the CV for each feature (CV = standard deviation / mean), the relative volatility of the data is measured, eliminating the influence of units of measurement. Then, a threshold is set to filter features with extremely low volatility. In this example, features with CV > 0.01 may be retained. Step S31 is used to perform monotonicity-significance joint filtering on the multi-dimensional feature set, finding features that have a statistically significant and clearly directional association with the target variable to be predicted (such as cigarette smoke resistance or sensory rating). The monotonicity test (Spearman correlation coefficient) measures the strength and direction of the monotonic relationship between two variables, i.e., whether if one increases, the other tends to increase or decrease. The significance test (p-value test) is used to determine whether the calculated correlation is real or merely due to random sampling error. A threshold is set and filtering is performed. In this example, features with p-value < 0.1 may be retained. Step S32 is used to select the intersection of the two filtering steps to obtain the pre-filtered effective features.

[0033] Step S13 is used to obtain a comprehensive score feature set based on the effective feature set. In this embodiment, the specific method for obtaining the comprehensive score feature set in step S13 can be of various forms known to those skilled in the art. In one example of the present invention, step S13 may include, for example... Figure 4 The steps shown are described in this. Figure 4 In this context, step S13 may include: In step S40, the nonlinear correlation value and the linear correlation value are calculated respectively based on the effective feature set; In step S41, dynamic weights are set based on the nonlinear correlation value and the linear correlation value; In step S42, a comprehensive score for each valid feature is calculated based on the dynamic weights; In step S43, a comprehensive score feature set is obtained based on the comprehensive score of each valid feature.

[0034] In such Figure 4In the method shown, step S40 is used to calculate the nonlinear correlation value and the linear correlation value. In this example, the calculation of the nonlinear correlation value can be based on the features selected in step S12, paired with the target variable of silk loss, and the mutual information value is calculated using the KNN mutual information estimation method (parameters alpha=0.6, c=15). In this example, the calculation of the linear correlation value can be based on the features selected in step S12, paired with the target variable of silk loss, and the degree of linear correlation between the two is calculated using the Pearson correlation coefficient formula. The absolute value of the correlation coefficient is used to quantify the strength of the linear correlation, and the Pearson correlation coefficient is calculated. Step S41 is used to perform dynamic weight adjustment based on the nonlinear correlation value and the linear correlation value. In this example, it can be done by comparing the nonlinear correlation value and the linear correlation value in step S40 to determine whether the linear relationship is dominant or the nonlinear relationship is dominant. If the linear relationship is determined to be dominant, the weight α=0.6 is set; otherwise, the nonlinear relationship is dominant, and α=0.8 is set. Step S42 is used to calculate the comprehensive score of each effective feature based on the dynamic weight. In this example, the overall score for each valid feature could be calculated using, for example, formula (1): (1) in, For comprehensive scoring, For dynamic weights, The normalized linear correlation value, This is the normalized nonlinear correlation value.

[0035] Step S43 is used to obtain a comprehensive score feature set based on the comprehensive score of each valid feature. In this example, the features pre-screened in step S12 can be sorted from high to low according to the calculated comprehensive score, and the top 70% of the features after sorting can be selected to form a comprehensive score feature set, prioritizing the retention of core features with a stronger correlation to the amount of silk lost, thus obtaining the comprehensive score feature set.

[0036] Step S14 is used to define feature categories based on process knowledge and perform regional feature filtering on the comprehensive scoring feature set to obtain an initial feature set and a candidate feature set. Specifically, based on the features filtered in step S13, feature categories are defined and grouped. Within each group, each feature is paired with the target variable of yarn loss, the Pearson correlation coefficient is calculated, and the absolute value is used to quantify the correlation strength. Feature filtering is then performed within each group, with a maximum of one feature with the highest correlation selected from each group. These features are then integrated to obtain the initial feature set. The unselected features are then integrated to obtain the candidate feature set.

[0037] Step S15 is used to perform improved sequence forward selection and fine screening based on the initial feature set and candidate feature set to obtain the optimal feature subset. In this embodiment, the specific method for performing improved sequence forward selection and fine screening in step S15 to obtain the optimal feature subset can be of various forms known to those skilled in the art. In one example of the present invention, step S15 may include, for example... Figure 5 The steps shown are described in this. Figure 5 In this context, step S15 may include: In step S50, the current feature set and the current candidate feature set are selectively added to obtain the current feature subset; In step S51, based on the current feature subset, the correlation coefficient is calculated and redundant features are removed; In step S52, the current initial feature set and the current candidate feature set are updated based on the current feature subset after removing redundant features; In step S53, it is determined whether the stopping condition is met; In step S54, if it is determined that the stopping condition is not met, the process returns to the step of selectively adding features based on the current initial feature set and the current candidate feature set to obtain the current feature subset. If the stopping condition is met, select the feature subset corresponding to the lowest prediction error as the optimal feature subset.

[0038] In such Figure 5 In the method shown, step S50 is used to selectively add features based on the current initial feature set and the current candidate feature set. In this embodiment, the specific method for selectively adding features in step S50 can be of various forms known to those skilled in the art. In one example of the present invention, step S50 may include: In step S60, the initial feature set is set to... The candidate feature set is ; In step S61, for the candidate feature set Each feature in Construct an extended feature set ; In step S62, based on each extended feature set Train random forest models separately and calculate the corresponding prediction errors on the validation set; In step S63, the feature that minimizes the prediction error is selected from all candidate features. ; In step S64, the selected feature Add the current initial feature set as the current feature subset.

[0039] Steps S60 to S64 are used to traverse each unselected feature in the candidate feature set, temporarily add it to the initial feature set, construct an expanded feature set, calculate its prediction error RMSE on the validation set, select the feature that can reduce the RMSE to the lowest level, this feature is the one that contributes the most to the improvement of model performance in this round, and add it to the initial feature set.

[0040] Step S51 is used to perform internal redundancy removal based on the current feature subset. Specifically, in this example, it may involve first calculating the correlation coefficient matrix of the current feature subset, then identifying all feature pairs with a correlation coefficient greater than 0.85, indicating that the two features are highly correlated and have significant information overlap. For each pair of highly correlated features, their feature importance scores in the random forest are compared, and the feature with the lower importance score is discarded.

[0041] Step S52 is used to update the current initial feature set and the current candidate feature set based on the current feature subset after removing redundant features.

[0042] Step S53 is used to determine whether the stopping condition is met. Specifically, in this example, the stopping condition can be that after three consecutive iterations, the improvement in RMSE after adding new features is less than or equal to 0.01, indicating that performance has saturated and it is time to stop. The stopping condition can also be that the number of features reaches a pre-set threshold, such as a maximum of 15 features, to prevent the model from becoming too complex.

[0043] Step S54 is used to perform corresponding operations based on the judgment result: If the stopping condition is not met, return to the step of selectively adding features based on the current initial feature set and the current candidate feature set to obtain the current feature subset. If the stopping condition is met, select the feature subset corresponding to the lowest prediction error as the optimal feature subset. After the algorithm stops, it does not directly output the feature set of the last round, but reviews the entire iteration process and outputs the feature subset that makes the model's RMSE reach the lowest value (optimal performance) throughout the process, ensuring that the final result is the best combination.

[0044] Example 1: This embodiment presents a method for characterizing the amount of tobacco residue at the end of a cigarette based on prior process knowledge. This method is used to achieve dynamic region division and multi-dimensional feature filtering of cigarette density data. The specific method is as follows: S1. Dynamic Region Division: Based on the tobacco shedding data and microwave density data of 32 segments of the tobacco, the cigarette is adaptively divided into ends, middle, and tail sections according to the density gradient abrupt change points. Segments 0-4 are designated as the ends, segments 5-26 as the middle, and segments 27-31 as the tail. The division results are as follows: Figure 6 As shown.

[0045] S2. Multi-dimensional feature extraction: Based on the end, middle and tail parts in step S1, multi-dimensional feature extraction and global feature extraction are performed for each region, and a total of 52 features are extracted. The initial RMSE value is 10.7755.

[0046] S3. Edge feature pre-screening: Based on the features extracted in step S2, the coefficient of variation and monotonicity-significance are screened together, leaving 16 features after screening.

[0047] S4. Calculate the nonlinear correlation value: Based on the features selected in step S3, calculate their KNN mutual information estimates, such as... Figure 7 As shown.

[0048] S5. Calculate the linear correlation value of the features: Based on the features selected in step S3, calculate their Pearson correlation coefficient, such as... Figure 8 As shown.

[0049] S6. Dynamic weight adjustment: Based on the nonlinear correlation value in step S4 and the linear correlation value in step S5, dynamic weight adjustment is performed. From the calculated mutual information value and Pearson correlation coefficient, it can be seen that the nonlinear relationship is currently dominant, α=0.8.

[0050] S7. Calculate the comprehensive feature score: Normalize and weight the nonlinear correlation values ​​in step S4 and the linear correlation values ​​in step S5, and calculate the comprehensive score for each feature.

[0051] S8. Obtain the comprehensive score feature set: Based on the comprehensive score in step S7, sort the features filtered in step S3, and filter from the sorted features to obtain the comprehensive score feature set. After filtering, there are 11 features, and the RMSE value at this time is 8.7028.

[0052] S9. Redundancy control of feature grouping based on process knowledge: Based on the feature category grouping defined after screening in step S8, the features are divided into central tendency group (containing 4 features), dispersion group (containing 4 features), distribution pattern group (containing 2 features), and time series group (containing 1 feature). The correlation between the features and the target variable is calculated, and feature screening within the group is performed. At most one feature with the highest correlation is selected in each group.

[0053] S10. Obtain the initial feature set: Based on the features selected in each feature group in S9, obtain the initial feature set. For the central tendency group, select the mean difference feature; for the dispersion group, select the standard deviation of the middle difference feature; for the distribution pattern group, select the skewness feature; and for the time series group, select the time trend feature. Since the mean difference feature and the time trend feature are highly correlated, the mean difference feature is removed. Finally, the standard deviation of the middle difference, the skewness, and the time trend feature are retained in the initial feature set.

[0054] S11. Obtain candidate feature set: Based on the features that were not selected in step S9 in step S8, obtain a candidate feature set, which consists of 8 features. At this time, the RMSE value is 7.5540.

[0055] S12. Improved forward selection and fine screening of sequences: Based on the initial feature set in step S10 and the candidate feature set in step S11, improve the forward selection and fine screening of sequences.

[0056] S13. Obtaining the final feature set: Based on the improved sequence forward selection algorithm in step S12, feature refinement is performed, and the optimal feature subset is determined by minimizing the root mean square error (RMSE). This feature subset can effectively characterize the variation characteristics of the amount of silk lost. The final feature set includes five features: mid-section difference standard deviation, end-section skewness, end-section time trend, end-section difference standard deviation, and end-section standard deviation. The optimized RMSE value is 6.4400, an improvement of 4.3355 compared to the initial model's 10.7755. The trend of feature quantity changes during the feature selection process is as follows: Figure 9 and Figure 10 As shown, the final feature importance ranking results are as follows: Figure 11 As shown.

[0057] On the other hand, the present invention also provides a cigarette end-burning quantity characterization system, the system including a processor configured to perform any of the methods for characterizing cigarette end-burning quantity.

[0058] In another aspect, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement any of the methods described in the method for characterizing the amount of cigarette residue at the cigarette end.

[0059] The beneficial effects of this invention are: The embodiments of this invention utilize a dynamic region division method that integrates the spatial distribution characteristics of density data with process knowledge. First, it adaptively divides the process regions of the cigarette end, middle, and end based on density gradient mutation points. This overcomes the technical limitations of traditional fixed region division methods that are disconnected from process characteristics, achieves precise adaptation to the process characteristics of cigarette density distribution, and solves the common problem of mismatch between feature extraction and process requirements.

[0060] The embodiments of this invention construct a multivariate pre-screening framework based on the coefficient of variation, monotonicity test, and significance analysis. Combined with a feature redundancy control strategy within process groups, it effectively separates noise information and effective features in high-dimensional data while preserving the spatiotemporal correlation of density data. This overcomes the shortcomings of traditional statistical methods that ignore the spatiotemporal correlation between features, which leads to screening bias. It provides a high-quality, low-redundancy initial feature set for subsequent fine feature screening.

[0061] The embodiments of this invention, by designing a comprehensive evaluation mechanism that combines linear and nonlinear correlation metrics, and employing a dual evaluation system that combines Pearson correlation coefficient with feature importance based on random forest, achieve a comprehensive characterization of the complex mapping relationship between density data and the amount of fiber shedding. This solves the technical problem that traditional methods cannot effectively characterize nonlinear process correlations due to their reliance on linear correlation analysis alone.

[0062] The embodiments of this invention introduce an improved sequence forward selection algorithm with dynamic redundancy check. During the iterative addition of features, the feature correlation coefficient matrix is ​​calculated in real time and highly redundant features are eliminated. Combined with a model performance-oriented search strategy, this invention overcomes the technical bottleneck of traditional search algorithms that are prone to getting trapped in local optima and cannot dynamically control feature redundancy. It achieves the optimization goal of maximizing representation accuracy while ensuring the model's generalization ability.

[0063] The embodiments of this invention establish a feature grouping management and weight balancing mechanism guided by process knowledge, integrate prior process knowledge into the entire feature screening process in a structured manner, and achieve deep integration of data-driven methods and domain knowledge by dynamically balancing the contribution of linear and nonlinear correlations through parameters. This solves the core problems of poor process adaptability and weak feature interpretability in the prior art.

[0064] Through the above-mentioned technical means, the cigarette end-fiber quantity characterization method finally constructed by this invention maintains the advantages of online real-time monitoring while achieving higher characterization accuracy, stronger model robustness, better process interpretability, and wider adaptability to production scenarios, providing reliable technical support for precise quality control in the cigarette production process.

[0065] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0066] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0069] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0070] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0071] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0072] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0073] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for characterizing the amount of tobacco residue at the end of a cigarette, characterized in that, The characterization method includes: Acquire data on the cigarette production process and dynamically divide the region; Multi-dimensional feature extraction is performed on the data after regional division to obtain a multi-dimensional feature set; Based on the multi-dimensional feature set, edge feature pre-screening is performed to obtain an effective feature set; Obtain the comprehensive scoring feature set based on the effective feature set; The comprehensive scoring feature set is subjected to regional feature filtering to obtain an initial feature set and a candidate feature set; Based on the initial feature set and candidate feature set, an improved sequence forward selection fine screening is performed to obtain the optimal feature subset.

2. The characterization method according to claim 1, characterized in that, Acquiring data on the cigarette production process and dynamically dividing it into regions includes: Acquire production data on cigarette burnout amount and discrete microwave density data of multiple segments along the cigarette axis; The density gradient is calculated based on the discrete microwave density data of the cigarette along the axial direction of the multi-segment to obtain the gradient value; Potential mutation points are obtained based on the gradient values; Valid mutation points are selected based on the potential mutation points; The regions are divided based on the effective mutation points.

3. The characterization method according to claim 1, characterized in that, Multi-dimensional feature extraction is performed on the data after regional segmentation to obtain a multi-dimensional feature set, including: Features were extracted from the partitioned data from the dimensions of central tendency, dispersion, distribution pattern, and time series characteristics. Central tendency includes mean, median, and mode; dispersion includes standard deviation, variance, range, and interquartile range; distribution pattern includes skewness and kurtosis; and time series characteristics include difference mean, difference standard deviation, time trend, and zero crossover rate.

4. The characterization method according to claim 1, characterized in that, Based on the multi-dimensional feature set, edge feature pre-screening is performed to obtain an effective feature set, including: The coefficient of variation is used to filter the multidimensional feature set to obtain a first filtered dataset; The multi-dimensional feature set is subjected to monotonicity-saliency joint screening to obtain a second screening dataset; The intersection of the first and second filtered datasets is used to obtain the effective feature set.

5. The characterization method according to claim 1, characterized in that, The comprehensive scoring feature set obtained from the effective feature set includes: Based on the effective feature set, calculate the nonlinear correlation value and the linear correlation value respectively; Dynamic weights are set based on the nonlinear and linear correlation values; The comprehensive score for each effective feature is calculated based on the dynamic weights. Based on the comprehensive score of each valid feature, obtain the comprehensive score feature set.

6. The characterization method according to claim 5, characterized in that, The comprehensive score for each valid feature is calculated based on the dynamic weights, including: Calculate the comprehensive score for each valid feature according to formula (1). ，（1） in, For comprehensive scoring, For dynamic weights, The normalized linear correlation value, This is the normalized nonlinear correlation value.

7. The characterization method according to claim 1, characterized in that, Based on the initial feature set and candidate feature set, an improved sequence forward selection process is performed to obtain the optimal feature subset, including: Based on the current initial feature set and the current candidate feature set, the best feature set is selected for inclusion to obtain the current feature subset; Based on the current feature subset, calculate the correlation coefficient and remove redundant features; Update the current initial feature set and the current candidate feature set based on the current feature subset after removing redundant features; Determine if the stopping condition is met; If the stopping condition is not met, return to the step of selecting the best feature set based on the current initial feature set and the current candidate feature set to obtain the current feature subset; If the stopping condition is met, select the feature subset corresponding to the lowest prediction error as the optimal feature subset.

8. The characterization method according to claim 1, characterized in that, Based on the current initial feature set and the current candidate feature set, the best feature set is selected for inclusion to obtain the current feature subset, which includes: Set the initial feature set to The candidate feature set is ; For the candidate feature set Each feature in Construct an extended feature set ; Based on each extended feature set Train random forest models separately and calculate the corresponding prediction errors on the validation set; Select the feature that minimizes the prediction error from all candidate features. ; Selected features Add the current initial feature set as the current feature subset.

9. A system for characterizing the amount of tobacco residue at the end of a cigarette, characterized in that, The system includes a processor configured to perform the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a processor, implement the method as described in any one of claims 1 to 8.

Images