CN-121998190-A - Cigarette blank head value prediction method, system and storage medium

Abstract

The invention relates to the technical field of cigarette process quality control, in particular to a cigarette empty head value prediction method, a system and a storage medium, which comprise the steps of obtaining multidimensional cigarette process parameter data; the method comprises the steps of preprocessing the multi-dimensional cigarette process parameter data, constructing a cigarette blank head value prediction model, constructing a loss function based on physical information, training the cigarette blank head value prediction model by adopting the preprocessed multi-dimensional cigarette process parameter data according to the loss function based on the physical information to obtain an optimal feature set, and predicting cigarettes to be predicted by adopting the cigarette blank head value prediction model corresponding to the optimal feature set to obtain a prediction result. According to the embodiment of the invention, by providing the tobacco hollow head value prediction method based on the physical information neural network, physical priori knowledge of the tobacco industry can be deeply integrated into model learning, so that a prediction result accords with a data statistics rule and conforms to a physical mechanism, and the prediction precision and the model reliability are remarkably improved.

Inventors

• QIU YUCAN
• ZHU LIMING
• FAN HU
• JIA QIAODONG
• ZHU QIANG
• XIA CHEN
• SHEN KAI
• XU XUE
• QIN TING

Assignees

• 浙江中烟工业有限责任公司

Dates

Publication Date

20260508

Application Date

20260127

Claims (10)

1. 1. A method for predicting a cigarette loose end value, the method comprising: Acquiring multi-dimensional cigarette process parameter data; Preprocessing the multi-dimensional cigarette process parameter data; Constructing a cigarette empty head value prediction model; Constructing a loss function based on physical information; Training the cigarette blank head value prediction model by adopting preprocessed multidimensional cigarette process parameter data according to the loss function based on the physical information so as to obtain an optimal feature set; And predicting the cigarettes to be predicted by adopting a cigarette blank head value prediction model corresponding to the optimal feature set so as to obtain a prediction result.
2. 2. The method of claim 1, wherein preprocessing the multi-dimensional cigarette process parameter data comprises: Detecting an abnormal value and removing; Performing data consistency verification; filling the missing technological parameter characteristics by adopting a multiple interpolation method; and (3) carrying out data standardization by adopting standard deviation, and eliminating dimension differences.
3. 3. The prediction method according to claim 1, wherein constructing a loss function based on physical information comprises: constructing a loss function based on the physical information according to formula (1), ,(1) Wherein, the In order to be based on the loss function of the physical information, The loss term is predicted on the basis of the prediction, As a port density loss term, In order to compact the end distance loss term, Is that Is used for the weight coefficient of the (c), Is that Weight coefficient of (c) in the above-mentioned formula (c).
4. 4. A prediction method according to claim 3, wherein the base prediction loss term is obtained according to formula (2): ,(2) Wherein, the The loss term is predicted on the basis of the prediction, Is the true value of the empty cigarette head value, Is the predicted value of the null head value, The threshold is segmented for the loss function.
5. 5. A prediction method according to claim 3, wherein the port density loss term is obtained according to formula (3): ,(3) Wherein, the As a port density loss term, Is the density average value of the front n sections of the combustion end in the multi-dimensional cigarette process parameter data, Is the value of the standard density which is the value of the standard density, Is the predicted value of the null head value, For the maximum allowed null value, Is a proportionality coefficient.
6. 6. A prediction method according to claim 3, wherein the compaction end distance loss term is obtained according to equation (4): ,(4) Wherein, the In order to compact the end distance loss term, For the value of the compaction tip distance offset, In order to lose the magnitude coefficient, As a non-linear influence coefficient, Is the predicted value of the null head value, Is the maximum allowed null value.
7. 7. The prediction method according to claim 1, wherein training the tobacco rod empty head value prediction model with preprocessed multidimensional tobacco rod process parameter data according to the loss function based on physical information to obtain an optimal feature set comprises: Inputting a multidimensional cigarette technological parameter data set into the cigarette empty head value prediction model; Calculating the global importance score of each feature in the multi-dimensional cigarette process parameter dataset; removing the characteristics of which the global importance score is lower than a preset threshold value, updating the multi-dimensional cigarette process parameter data set and calculating the quantity of the residual characteristics; Judging whether the residual feature quantity is smaller than or equal to a preset minimum value or not; Under the condition that the residual feature quantity is judged to be smaller than or equal to a preset minimum value, model training is completed; Returning to execute the step of inputting the multi-dimensional cigarette process parameter data set into the cigarette blank head value prediction model under the condition that the residual characteristic quantity is judged to be larger than a preset minimum value; And evaluating the prediction capability of each model by using model evaluation indexes to obtain an optimal feature set.
8. 8. The method of claim 7, wherein calculating a global importance score for each feature in the multi-dimensional cigarette process parameter dataset comprises: a global importance score is calculated according to equation (5), ,(5) Wherein, the As a global importance score, the number of the nodes, Is characterized by In the sample Upper part of the cylinder The value of the sum of the values, Is desired for the sample.
9. 9. A cigarette butt value prediction system, characterized in that the system comprises a processor configured to perform the method of any one of claims 1 to 8.
10. 10. A computer readable storage medium having instructions stored thereon which, when executed by a processor, implement the method of any of claims 1 to 8.

Description

Cigarette blank head value prediction method, system and storage medium Technical Field The invention relates to the technical field of cigarette process quality control, in particular to a cigarette empty head value prediction method, a system and a storage medium. Background In the cigarette production process, the empty cigarette head value is one of the core indexes for measuring the quality of the cigarette process. The existing empty head of the cigarette expenditure is a common quality defect in the cigarette production process, namely, the phenomenon that the cut tobacco at the end part of the cigarette is not filled enough and obvious gaps exist. From the aspect of use, the condition that the burning speed of tobacco shreds is accelerated and the smoking resistance of cigarettes is reduced can appear at the empty part, the stability and the consistency of the taste in the smoking process are further damaged, and the smoking experience of consumers is seriously influenced. Therefore, accurate and efficient prediction and control of the void fraction is of great importance in the tobacco industry. Current methods of cigarette loose end detection rely mainly on manual sampling inspection and automatic optical detection based on machine vision. The manual method has low efficiency, strong subjectivity and easy omission. The machine vision method improves the efficiency, is easily interfered by factors such as ambient light, cigarette placement posture and the like, and is difficult to predict the unoccupied internal empty trend. If the empty head value prediction technology can be combined, the position and probability of empty head cigarettes can be predicted in advance before the manual sampling and machine vision technology, so that a detector can be helped to manage equipment more pertinently, the aim of eliminating unqualified products in advance is fulfilled, and the detection efficiency and accuracy are improved. Meanwhile, the empty value prediction result can be used as a basis for production optimization by engineers, and waiting for shutdown maintenance and waste of defective product production are reduced. The existing cigarette process quality prediction method has significant limitation, most of the existing cigarette process quality prediction methods adopt statistical and machine learning technologies, and data driving models such as a Support Vector Machine (SVM), a random forest (RandomForest) and a deep learning network are widely applied to cigarette process quality prediction. Although these methods improve the predictive ability to some extent, they are essentially black box models, which rely purely on statistical rules in the data, with obvious limitations. For example, the prediction may violate known physical laws and industry knowledge, resulting in models that are not trusted in practical applications. In addition, part of the model performance is severely dependent on a large amount of high-quality annotation data, and in an actual industrial scene, high-quality defect samples are rare, so that the model generalization capability is insufficient. On the other hand, the traditional machine learning model has poor interpretability, engineers have difficulty in understanding the decision basis of the model, and cannot relate the prediction result with specific production process parameters to guide production optimization. The Physical Information Neural Network (PINN) is a new paradigm that has been raised in recent years to combine physical laws with deep learning. The method guides the learning process of the model by introducing a control equation, boundary conditions or known physical relations as constraints into the construction of the neural network, thereby solving the problems of missing physical information consistency, data dependence, poor interpretability and the like of the traditional machine learning technology. However, the prior art lacks the ability to efficiently encode specific process knowledge and physical mechanisms of the tobacco industry into the PINN framework and design a complete, engineered set of cigarette end value prediction methods. Therefore, how to deeply integrate prior tobacco process knowledge with a quality prediction method of a physical information neural network, effectively construct the physical information neural network related to the cigarette blank head value, efficiently screen multi-dimensional process parameters and control parameters based on process logic becomes a core technical problem for improving the prediction precision of the cigarette blank head value. Disclosure of Invention The embodiment of the invention aims to provide a cigarette loose end value prediction method, a system and a storage medium, which aim to solve the problems of missing physical consistency, strong data dependence and poor interpretability of the existing method in a quality prediction link by integrating process knowledge and a physical infor