A multi-process quality prediction method based on attention mechanism

Through the feature encoding network based on attention mechanism processing of multi-process quality prediction data, the problem that the existing technology cannot capture complex relationships between processes is solved, intelligent optimization of process parameters and quality prediction are realized, and the practicality of product quality control is improved.

CN118552092BActive Publication Date: 2024-11-05NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202411022464.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-05
Estimated Expiration
2044-07-29

AI Technical Summary

Technical Problem

Existing multi-process quality prediction methods cannot effectively capture the potential mutual influence and complex dynamic changes between processes, resulting in a decrease in the applicability and prediction ability of the prediction model.

Method used

Using a feature coding network based on attention mechanism, the process parameters and multi-source sensor timing feature data are characterized by feature coding and multi-process quality prediction training in historical product processing, and the quality prediction model is generated, and the data is processed using the multi-head attention network to capture Complex relationships between processes.

Benefits of technology

It realizes the mapping relationship acquisition between process parameters and quality indicators, supports intelligent optimization of process parameters, discovers processing problems in advance, and improves the practicality of product quality control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118552092B_ABST
    Figure CN118552092B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of quality prediction technology, and solves the technical problem that the existing methods cannot capture the potential mutual influence and complex dynamic changes between processes, thereby reducing the applicability and prediction ability of the prediction model, and in particular relates to a multi-process quality prediction method based on an attention mechanism, the method comprising the following steps: obtaining the process parameters of each process in the historical product processing and the time series feature data of multi-source sensors to construct a training set; using the training set to perform multi-process quality prediction training on the feature encoding network based on the attention mechanism to obtain a quality prediction model; using the trained quality prediction model to perform quality prediction on the target data. The present invention realizes the acquisition of the mapping relationship between process parameters and quality indicators, provides support for the intelligent optimization of process parameters, discovers possible problems in processing in advance, and controls quality in advance, and has good practicality.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of quality prediction technology, and in particular to a multi-process quality prediction method based on an attention mechanism. Background Art

[0002] Multi-process quality prediction is an important production management method used to predict the quality status of products during the manufacturing process. In the manufacturing industry, product manufacturing often requires multiple processes or steps, and the quality status of each process affects the quality of the final product. Therefore, by predicting and monitoring the quality of each process, potential quality issues can be identified early, avoiding quality defects in subsequent processes, thereby improving product consistency and overall quality. The amount of data involved in the manufacturing process is large and complex, including sensor data, process parameters, equipment status, etc., and how to effectively collect, integrate, and process this data is a key issue.

[0003] Multi-process quality prediction is typically based on statistical methods. However, these methods often assume independence and identical distribution of data. In actual multi-process manufacturing, data between processes may have complex correlations and nonlinear relationships. The limitations of this assumption can lead to the model failing to capture the potential interactions and complex dynamic changes between processes, reducing the applicability and predictive power of the prediction model. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a multi-process quality prediction method based on the attention mechanism, which solves the technical problem that the existing methods are unable to capture the potential mutual influence and complex dynamic changes between processes, thereby reducing the applicability and prediction ability of the prediction model.

[0005] To solve the above technical problems, the present invention provides the following technical solution: a multi-process quality prediction method based on an attention mechanism, the method comprising the following steps:

[0006] Obtain the process parameters of each step in historical product processing and the time series feature data of multi-source sensors to build a training set;

[0007] The training set is used to train the feature encoding network based on the attention mechanism for multi-process quality prediction to obtain a quality prediction model;

[0008] Use the trained quality prediction model to predict the quality of the target data.

[0009] Furthermore, the process of constructing the training set specifically includes:

[0010] Arrange the process sequence according to the processing order, obtain the historical product in N processing steps, and the process parameters of each processing step constitute the process parameter set ,The process parameters include equipment power, spindle speed, feed speed and other process parameters;

[0011] Obtain the historical product in N processing steps, and the time series features obtained by multi-source sensor monitoring of each processing step constitute the time series feature data set ,The timing characteristics include servo load, spindle vibration, spindle load and other timing characteristics.

[0012] Furthermore, the training of the feature encoding network to obtain the quality prediction model includes the following steps:

[0013] S21, the time series feature data set of the t-th processing step The elements in are reorganized into time series , where T is the duration of the t-th processing step and C is the number of time series data;

[0014] S22, time series The input uses the residual method to organize the temporal coding module of L Temp Conv layers to generate the encoded temporal features , where L=2, that is, the temporal encoding module organizes two Temp Conv layers, and the expression is: ;

[0015] in, is the number of TempConv layers;

[0016] S23, set the process parameters of the t-th processing step Each element is concatenated and then combined with the encoded temporal features Splicing is performed to obtain the coding features of the tth processing step ,Right now:

[0017] ;

[0018] in, Represents the set of process parameters for the tth processing step The nth element of ; Represents a splicing operation;

[0019] S24, each coding feature Input into the LSTM unit in sequence, and the LSTM unit calculates the output of the current processing step , and update the cell state of the unit ;

[0020] S25, output the current processing step of N processing steps As input and produce output feature matrix B in the multi-head attention network;

[0021] S26. Input the feature matrix B into the MLP to obtain the processing error z, and iteratively complete the multi-process quality prediction training of the feature encoding network to obtain a quality prediction model.

[0022] Furthermore, in step S22, the temporal coding module organizes two TempConv layers.

[0023] Furthermore, in step S22, the specific process includes the following steps:

[0024] S221, Use Fast Fourier Transform (FFT) to analyze time series in the frequency domain Get some amplitude values, namely:

[0025] ;

[0026] in, and Respectively represent the calculation of FFT and amplitude value; Indicates the TempConv layers, represents the calculated amplitude for each frequency, averaged from the C dimension ; Represents the amplitude of the periodic basis function with frequency j, and the corresponding period length is ; T is the processing time; is the cycle length; is the most significant frequency;

[0027] S222, select the first k amplitude values, and use the non-normalized amplitude value Get the most significant frequency , the period length corresponding to the most significant frequency is ;

[0028] S223, based on the most significant frequency selected and the corresponding period length , convert the 1D time series Reconstructed into multiple two-dimensional tensors to represent different two-dimensional changes in time derived from different periods , the formula is as follows:

[0029] ;

[0030] in, It is to zero-extend the time series along the time dimension so that it is consistent with Compatibility; cycle length and the most significant frequency Respectively represent the number of rows and columns of the transformed two-dimensional tensor; Indicates the most significant frequency No. A reconstructed time series, where the columns and rows represent the corresponding period lengths intra-cycle and inter-cycle variations under

[0031] S224, use two-dimensional convolution kernel to two-dimensional tensor Perform convolution separately to obtain the two-dimensional tensor after convolution , expressed as:

[0032] ;

[0033] Among them, Conv2D represents a 3x3 two-dimensional convolution kernel;

[0034] S225, the convolved two-dimensional tensor After rearrangement, the length is The one-dimensional sequence of is truncated to the original length T to obtain a one-dimensional representation, which is:

[0035] ;

[0036] Wherein, Trunc represents the Trunc function; Represents a two-dimensional tensor After rearrangement, the length is One-dimensional sequence of

[0037] S226, based on the amplitude aggregation of the one-dimensional representation of the layer to obtain the temporal characteristics of the layer output ,Right now:

[0038] ;

[0039] in, ; is the number of TempConv layers; k is the number of amplitude values ​​selected; is the amplitude value; Represents the Softmax function; Indicates the The frequency calculated by the layer TempConv is The amplitude of the periodic basis function.

[0040] Furthermore, in step S24, the specific process includes the following steps:

[0041] S241, the coding feature and the output of a process Input the forget gate to get the output ,Right now:

[0042] ;

[0043] in, is the weight matrix of the forget gate; is the forget gate bias term; For data splicing; is the sigmoid function;

[0044] S242, the coding feature , the output of the previous process , cell status and output Input the update gate layer to get the current cell state , the expression is:

[0045] ;

[0046] in, Both are weight matrices for updating gate layers; Both are bias items for updating gate layers; For data splicing; is the sigmoid function;

[0047] S243. Use the updated cell state Output the calculation result of the LSTM unit, which is the output of the current processing step , the expression is:

[0048] ;

[0049] in, To update the weight matrix of the gate layer; To update the bias term of the gate layer, is the sigmoid function.

[0050] Furthermore, in step S25, the specific process includes the following steps:

[0051] S251, output the current processing step of N processing steps Perform preliminary coding to obtain , the expression is:

[0052] ;

[0053] Where W is the preliminary encoding weight matrix;

[0054] S252, use Weight matrix Respectively Multiply to get , and Denoted as a matrix ,Right now:

[0055] ;

[0056] in, , , is the number of attention heads;

[0057] S253, according to Calculate the attention value of each attention head , calculated as follows:

[0058] ;

[0059] in, represents the dimension of the attention head;

[0060] S254, according to the attention value Calculate the output B of the multi-head attention module, that is:

[0061] ;

[0062] in, is the weight matrix.

[0063] By means of the above technical solution, the present invention provides a multi-process quality prediction method based on the attention mechanism, which has at least the following beneficial effects:

[0064] 1. The present invention realizes the acquisition of the mapping relationship between process parameters and quality indicators by feature encoding and multi-process quality prediction training of process parameters and monitoring data in the product production process, provides support for intelligent optimization of process parameters, discovers possible problems in processing in advance, controls quality in advance, and has good practicality.

[0065] 2. This invention utilizes the process parameters recorded in the historical processing process and the time series feature data monitored by multi-source sensors to propose a feature encoding network based on the attention mechanism to process and analyze the data, thereby realizing the quality prediction of multi-process products and the identification of key processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0067] Figure 1 This is a flow chart of the multi-process quality prediction method of the present invention;

[0068] Figure 2 This is a diagram of the feature encoding network structure based on the attention mechanism of the present invention;

[0069] Figure 3 This is an example diagram of the spindle motor current data of the present invention;

[0070] Figure 4 is an example diagram of servo load data of the present invention;

[0071] Figure 5 This is an example diagram of the spindle vibration data of the present invention;

[0072] Figure 6 This is an example diagram of the spindle load data of the present invention;

[0073] Figure 7 This is an example of processing error output by the quality prediction model of the present invention. DETAILED DESCRIPTION

[0074] To make the above-mentioned objectives, features, and advantages of the present invention more clearly understood, the present invention is further described below in detail with reference to the accompanying drawings and specific embodiments. This will enable a full understanding of how this application uses technical means to solve technical problems and achieve technical effects, and to implement the invention accordingly.

[0075] In order to solve the technical problem that in the actual multi-process manufacturing process, the data between each process may have complex correlations and nonlinear relationships, which makes the model unable to capture the potential mutual influence and complex dynamic changes between processes, thus reducing the applicability and prediction ability of the prediction model. Figure 1-Figure 7 This embodiment proposes a multi-process quality prediction method based on an attention mechanism. By utilizing the process parameters recorded during historical processing and the time series feature data monitored by multi-source sensors, this embodiment proposes a feature encoding network based on the attention mechanism to process and analyze the data, thereby achieving quality prediction of multi-process products and identifying key processes. The method includes the following steps:

[0076] S1. Obtain the process parameters of each step in the historical product processing and the time series feature data of multi-source sensors to build a training set. The process of building the training set specifically includes:

[0077] Arrange the process sequence according to the processing order, obtain the historical product in N processing steps, and the process parameters of each processing step constitute the process parameter set , process parameters include equipment power, spindle speed, feed speed and other process parameters; obtain the historical product in N processing steps, and the time series features obtained by multi-source sensor monitoring of each processing step constitute the time series feature data set , ,The timing characteristics include servo load, spindle vibration, spindle load and other timing characteristics.

[0078] S2. Use the training set to train the feature encoding network based on the attention mechanism for multi-process quality prediction to obtain a quality prediction model; Figure 2 As shown, the feature encoding network is trained to obtain a quality prediction model, which includes the following steps:

[0079] S21, the time series feature data set of the t-th processing step The elements in are reorganized into time series , where T is the duration of the t-th processing step and C is the number of time series data;

[0080] S22, time series The input uses the residual method to organize the temporal coding module of L Temp Conv layers to generate the encoded temporal features , where L=2, that is, the temporal encoding module organizes two Temp Conv layers, and the expression is: ;

[0081] in, is the number of layers of TempConv; in step S22, the specific process includes the following steps:

[0082] S221, Use Fast Fourier Transform (FFT) to analyze time series in the frequency domain Get some amplitude values, namely:

[0083] ;

[0084] in, and Respectively represent the calculation of FFT and amplitude value; Indicates the TempConv layers, represents the calculated amplitude for each frequency, averaged from the C dimension ; Represents the amplitude of the periodic basis function with frequency j, and the corresponding period length is ; T is the processing time; is the cycle length; The most significant frequency.

[0085] S222, select the first k amplitude values, and use the non-normalized amplitude value Get the most significant frequency , the period length corresponding to the most significant frequency is ; The most significant frequencies of these choices also correspond to Cycle length , here we take .

[0086] S223, based on the most significant frequency selected and the corresponding period length , convert the 1D time series Reconstructed into multiple two-dimensional tensors to represent different two-dimensional changes in time derived from different periods , the formula is as follows:

[0087] ;

[0088] in, It is to zero-extend the time series along the time dimension so that it is consistent with Compatible; where the cycle length and the most significant frequency Respectively represent the number of rows and columns of the transformed two-dimensional tensor; Indicates the most significant frequency No. A reconstructed time series, where the columns and rows represent the corresponding period lengths Therefore, according to the selected frequency and estimated period, we get a set of two-dimensional tensors: , which represents different temporal two-dimensional changes derived from different periods.

[0089] S224, use 3x3 two-dimensional convolution kernel Conv2D to two-dimensional tensor Perform convolution separately to obtain the two-dimensional tensor after convolution , expressed as:

[0090] ;

[0091] Among them, Conv2D represents a 3x3 two-dimensional convolution kernel;

[0092] S225, use the Trunc function to convolve the two-dimensional tensor After rearrangement, the length is The one-dimensional sequence of is truncated to the original length T to obtain a one-dimensional representation, which is:

[0093] ;

[0094] Wherein, Trunc represents the Trunc function; Represents a two-dimensional tensor After rearrangement, the length is , that is, the two-dimensional tensors are connected end to end in rows to form a one-dimensional sequence.

[0095] S226, based on the amplitude aggregation of the one-dimensional representation of the layer to obtain the temporal characteristics of the layer output ,Right now:

[0096] ;

[0097] in, ; is the number of TempConv layers; k is the number of amplitude values ​​selected; is the amplitude value; Represents the Softmax function; Indicates the first The frequency calculated by the layer TempConv is The amplitude of the periodic basis function.

[0098] S23, set the process parameters of the t-th processing step Each element is concatenated and then combined with the encoded temporal features Splicing is performed to obtain the coding features of the tth processing step ,Right now:

[0099] ;

[0100] in, Represents the set of process parameters for the tth processing step The nth element of ; Represents a splicing operation;

[0101] S24, each coding feature Input into the LSTM unit in sequence, and the LSTM unit calculates the output of the current processing step , and update the cell state of the unit In step S24, the specific process includes the following steps:

[0102] S241, the coding feature and the output of the previous process Input the forget gate to get the output ,Right now:

[0103] ;

[0104] in, is the weight matrix of the forget gate; is the forget gate bias term; For data splicing; is the sigmoid function;

[0105] S242, the coding feature , the output of the previous process , cell status and output Input the update gate layer to get the current cell state , the expression is:

[0106] ;

[0107] in, Both are weight matrices for updating gate layers; Both are bias items for updating gate layers; For data splicing; is the sigmoid function;

[0108] S243. Use the updated cell state Output the calculation result of the LSTM unit, which is the output of the current processing step , the expression is:

[0109] ;

[0110] in, To update the weight matrix of the gate layer; To update the bias term of the gate layer, is the sigmoid function.

[0111] S25, output the current processing step of N processing steps As input and generate output feature matrix B in the multi-head attention network; in step S25, the specific process includes the following steps:

[0112] S251, output the current processing step of N processing steps Perform preliminary coding to obtain , the expression is:

[0113] ;

[0114] in, is the preliminary encoding weight matrix;

[0115] S252, use Weight matrix Respectively Multiply to get , and Denoted as a matrix ,Right now:

[0116] ;

[0117] in, , , is the number of attention heads, which is set to 5 here;

[0118] S253, according to Calculate the attention value of each attention head , calculated as follows:

[0119] ;

[0120] in, represents the dimension of the attention head;

[0121] S254, according to the attention value Calculate the output B of the multi-head attention module, that is:

[0122] ;

[0123] in, is the weight matrix.

[0124] S26. Input the feature matrix B into the MLP to obtain the processing error z, and iteratively complete the multi-process quality prediction training of the feature encoding network to obtain a quality prediction model.

[0125] S3, using the trained quality prediction model to predict the quality of the target data; Figure 3-6 As shown in the figure, the time series feature data input into the prediction model are examples, namely spindle motor current, servo load, spindle vibration and spindle load. By inputting the process parameter set of the target data and the time series feature data set into the trained prediction model, the output real number z can be obtained, which is the predicted product processing error. The output result is as follows: Figure 7 shown.

[0126] The present invention achieves the acquisition of the mapping relationship between process parameters and quality indicators by performing feature encoding and multi-process quality prediction training on process parameters and monitoring data in the product production process, provides support for intelligent optimization of process parameters, discovers possible problems in processing in advance, controls quality in advance, and has good practicality.

[0127] Those skilled in the art will appreciate that all or part of the steps in the above-mentioned embodiment methods can be accomplished by instructing the relevant hardware through a program. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented 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.

[0128] Each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to in detail. For the above embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For relevant parts, please refer to the partial description of the method embodiments.

[0129] The above embodiments provide a detailed introduction to the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A multi-process quality prediction method based on attention mechanism, characterized in that: The method comprises the following steps: Obtain the process parameters of each process in the historical product processing and the time series feature data of multi-source sensors to build a training set; The training set is used to train the feature encoding network based on the attention mechanism for multi-process quality prediction to obtain a quality prediction model, which includes the following steps: S21, the time series feature data set of the tth processing step The elements in are reorganized into time series , where T is the duration of the tth processing step, and C is the number of time series data; S22. Time series The input uses the residual method to organize the temporal coding module of L Temp Conv layers to generate the encoded temporal features , where L=2, that is, the temporal encoding module organizes 2 Temp Conv layers, and the expression is: ; in, is the number of layers of TempConv; S23, set the process parameters of the tth processing step Each element is concatenated and then combined with the encoded time series features Splice to get the coding features of the tth processing step ,Right now: ; in, Represents the set of process parameters for the tth processing step The nth element of ; Represents a splicing operation; S24, each coding feature Input into the LSTM unit in sequence, and the LSTM unit calculates the output of the current processing step , and update the cell state of the unit ; S25, output the current processing step of N processing steps As input and produce output feature matrix B in the multi-head attention network; S26, inputting the feature matrix B into the MLP to obtain the processing error z, iteratively completing the multi-process quality prediction training of the feature encoding network to obtain a quality prediction model; Use the trained quality prediction model to predict the quality of the target data.

2. The multi-process quality prediction method according to claim 1, characterized in that: The constructing of the training set comprises: Arrange the process sequence according to the processing order, obtain the process parameters of each process in the N processing steps of the historical product, and form a process parameter set ,The process parameters include equipment power, spindle speed, and feed speed process parameters; Obtain the time series features of N processing steps of historical products, and the time series features obtained by multi-source sensor monitoring of each processing step constitute the time series feature data set ,The timing characteristics include servo load, spindle vibration, and spindle load timing characteristics.

3. The multi-process quality prediction method according to claim 1, characterized in that: In step S22, the specific process includes the following steps: S221, Use Fast Fourier Transform (FFT) to analyze time series in the frequency domain Get some amplitude values, namely: ; in, and Respectively represent the calculation of FFT and amplitude value; Indicates TempConv layers, represents the calculated amplitude for each frequency, averaged from the C dimension ; represents the amplitude of the periodic basis function with frequency j, and the corresponding period length is ; T is the processing time; Represents the TempConv layer as The length of the cycle when is the most significant frequency; S222, select the first k amplitude values, and use the non-normalized amplitude values Get the most significant frequency , the period length corresponding to the most significant frequency is ; S223, based on the most significant frequency selected and the corresponding period length , convert the 1D time series Reconstructed into multiple two-dimensional tensors to represent different two-dimensional changes in time derived from different periods , the formula is as follows: ; in, It is to zero-expand the time series along the time dimension so that it is consistent with Compatibility; Cycle Length and the most significant frequency Respectively represent the number of rows and columns of the transformed two-dimensional tensor; The most significant frequency No. A reconstructed time series, where the columns and rows represent the corresponding period lengths intra-cycle and inter-cycle variations under S224, use two-dimensional convolution kernel to two-dimensional tensor Perform convolution separately to obtain the two-dimensional tensor after convolution , expressed as: ; Among them, Conv2D represents a 3x3 two-dimensional convolution kernel; S225, the convolved two-dimensional tensor After rearrangement, the length is The one-dimensional sequence of is truncated to the original length T to obtain a one-dimensional representation, which is: ; Wherein, Trunc represents the Trunc function; Represents a two-dimensional tensor After rearrangement, the length is One-dimensional sequence of S226, based on the amplitude aggregation of the one-dimensional representation of the layer, obtain the temporal characteristics of the output of the layer ,Right now: ; in, ; is the number of TempConv layers; k is the number of amplitude values ​​selected; is the amplitude value; Represents the Softmax function; Indicates The frequency calculated by the layer TempConv is The amplitude of the periodic basis function.

4. The multi-process quality prediction method according to claim 1, characterized in that: In step S24, the specific process includes the following steps: S241, the coding feature and the output of the previous process Input into the forget gate to get the output ,Right now: ; in, is the weight matrix of the forget gate; is the forget gate bias term; For data splicing; is the sigmoid function; S242, the coding feature , the output of the previous process , cell status And the output Input the update gate layer to get the cell state at the current moment , the expression is: ; in, They are all weight matrices for updating gate layers; They are all bias items for updating the gate layer; For data splicing; is the sigmoid function; S243. Use the updated cell state Output the calculation result of the LSTM unit, which is the output of the current processing step , the expression is: ; in, To update the weight matrix of the gate layer; To update the bias term of the gate layer, is the sigmoid function.

5. The multi-process quality prediction method according to claim 1, characterized in that: In step S25, the specific process includes the following steps: S251, output of the current processing step of N processing steps Perform preliminary coding to obtain , the expression is: ; Where W is the preliminary encoding weight matrix; S252, Use Weight Matrix Respectively Multiply to get , and Denoted in matrix form ,Right now: ; in, , , is the number of attention heads; S253, according to Calculate the attention value of each attention head , calculated as follows: ; in, represents the dimension of the attention head; S254, according to the attention value Calculate the output B of the multi-head attention module, that is: ; in, is the weight matrix.

Citation Information

Patent Citations

  • Process quality prediction method based on CNN-LSTM hybrid neural network model

    CN115169737A

  • Multi-quality index output prediction method and system for shred-making loosening and moisture-regaining process

    CN115456460A

  • Cut tobacco production workshop process quality online prediction method

    CN115545321A

  • Product quality production online monitoring method

    CN116820040A

  • LSTM and TPA mechanism-based cut tobacco flavoring process quality prediction method and system

    CN116843225A