A method and system for predicting and prospectively controlling production quality of a cigarette making machine based on operating condition parameters

Abstract

This invention relates to a method and system for predictive and proactive control of cigarette machine production quality based on operating condition parameters, belonging to the fields of tobacco machinery manufacturing, industrial automation control, and artificial intelligence. It includes: real-time acquisition of multi-dimensional operating condition parameters during cigarette machine operation; data preprocessing; acquisition of multi-dimensional time-series feature input sequences; construction of a time-series prediction model based on a long short-term memory network and incorporating an attention mechanism; training the time-series prediction model using the time-series feature input sequences; inputting the real-time operating condition feature sequences into the trained model to obtain key quality indicators; establishing a multi-index quality evaluation system using the analytic hierarchy process (AHP); comprehensively evaluating key quality indicators and other key quality indicators to obtain quantitative scores; and initiating graded quality early warning signals based on the degree to which the quantitative scores deviate from target values, using the early warning signals to drive parameter adjustments in the cigarette machine control system, thereby achieving proactive correction of quality deviations.

Description

Technical Field

[0001] This invention relates to a method and system for predicting and proactively controlling the production quality of cigarette machines based on operating condition parameters, belonging to the fields of tobacco machinery manufacturing, industrial automation control, and artificial intelligence technology. Background Technology

[0002] Cigarette product quality is a core element in tobacco companies' production management. Among these, the weight of a single cigarette is a key physical indicator affecting multiple downstream quality attributes such as draw resistance, hardness, and tar content. Its stability directly impacts the consistency and market competitiveness of cigarette products. To ensure controlled production processes, existing cigarette manufacturers generally employ Statistical Process Control (SPC) methods to monitor and analyze quality indicators such as the weight of a single cigarette.

[0003] However, SPC (Statistical Process Control) is essentially a post-production quality monitoring method based on outcome data, with its control based on the inspection results of already manufactured products. In the high-speed, continuous operation of cigarette production, this "manufacturing-inspection-judgment" control model inevitably has a time lag. When the SPC control chart triggers anomaly judgment or warning signals, quality deviations have often occurred continuously for some time, leading to the generation of batches of non-conforming products, thereby increasing the scrap rate and reducing the process capability index.

[0004] Furthermore, existing technologies have limited ability to utilize a large amount of early-stage information regarding the operation of cigarette making machines. Fluctuations in cigarette product quality are not isolated events, but rather the result of the combined effects of multiple factors, including the operating status of the cigarette making machine, material characteristics, and environmental conditions. For example, high-frequency vibration changes in key mechanical components, instantaneous shifts in the moisture content of tobacco, and subtle fluctuations in the feeding status are all direct causes and early-stage characteristics of quality changes. These operating parameters typically exhibit abnormal trends before significant changes in quality indicators occur, but traditional quality control methods struggle to systematically model and effectively utilize them.

[0005] Existing technologies generally lack a forward-looking analytical method that can establish an accurate mapping relationship between multi-dimensional operating parameters and macro-quality indicators of cigarette products. This makes it difficult to effectively predict the development trend of quality deviations before they actually occur, and even more difficult to implement early intervention and control based on this.

[0006] Therefore, there is an urgent need for a quality prediction and control method based on the operating parameters of cigarette machines. By analyzing and modeling multi-dimensional operating conditions in real time, potential risks can be identified before quality indicators deviate significantly, and a forward-looking quality control strategy can be provided to overcome the lag problem of traditional ex-post quality control methods. Summary of the Invention

[0007] To overcome the lag defects caused by the reliance on SPC in existing cigarette quality control and the aforementioned problems, this invention provides a method and system for predicting and proactively controlling the production quality of cigarette machines based on operating condition parameters. This invention can integrate and process multi-dimensional and heterogeneous operating condition parameters, and use deep learning models to achieve real-time prediction of the deviation trend of future key quality indicators.

[0008] The technical solution of this invention is: a method for predicting and proactively controlling the production quality of cigarette machines based on operating condition parameters, comprising:

[0009] S1. Real-time acquisition of multi-dimensional operating parameters during the operation of the cigarette machine;

[0010] S1 specifically includes: data acquisition and synchronization: establishing a multi-sensor network to collect three types of working condition parameters in real time: electromechanical, material, and environmental, and aligning them through timestamps to ensure the synchronization of feature sequences;

[0011] S2. Perform data cleaning, normalization, and data transformation on the collected multi-dimensional operating condition parameters, and extract the feature quantities of each operating condition parameter.

[0012] S2 specifically includes: data preprocessing: after completing the acquisition and synchronization of multi-dimensional operating condition parameters, data preprocessing and time-frequency feature extraction are performed on the multi-dimensional operating condition parameters.

[0013] First, the collected raw operating condition data is cleaned, including outlier removal, missing value correction, and noise suppression. Then, the cleaned operating condition parameters are normalized and necessary data transformations are performed to eliminate differences in the dimensions and numerical ranges of different operating condition parameters.

[0014] Based on the time-frequency analysis results, feature quantities reflecting the dynamic characteristics of the working condition are extracted, including time-frequency energy distribution characteristics, dominant frequency characteristics, and dominant frequency variation characteristics over time. These feature quantities are then fused with the normalized time-domain features to form a multidimensional feature dataset for subsequent modeling.

[0015] S3. Use grey relational analysis to analyze the correlation between the characteristics of each working condition parameter and the key quality indicators, and select key working condition characteristics according to the degree of correlation to form a multi-dimensional time series feature input sequence.

[0016] S3 specifically includes:

[0017] Feature Engineering and Correlation Analysis: After completing the data acquisition of multi-dimensional operating condition parameters, further feature engineering and correlation analysis are performed on these parameters. First, the acquired multi-dimensional operating condition parameters undergo data cleaning, normalization, and necessary data transformations to eliminate differences in dimensions, numerical ranges, and sampling scales among the different operating condition parameters. Based on this, time-frequency analysis is used to analyze the high-frequency dynamic signals contained in the operating condition parameters, extracting time-frequency feature quantities that characterize the changes in operating condition states. These feature quantities serve as the basis for subsequent analysis. Furthermore, Grey Relational Analysis (GRA) is introduced. Using key quality indicators of cigarette production as a reference sequence, the feature sequences of each operating condition parameter after feature extraction are used as comparison sequences. The grey correlation degree between the features of each operating condition parameter and the key quality indicators is calculated to quantify the degree of influence of different operating condition parameters on quality fluctuations.

[0018] Based on the results of the grey relational analysis, the working condition features are screened and reorganized to form a multi-dimensional time-series feature input sequence that reflects the key working condition-quality relationship, providing a unified and effective feature input basis for the construction of subsequent quality prediction models.

[0019] S4. Construct a time series prediction model based on a multi-layer neural network, and train the time series prediction model using the time series feature input sequence;

[0020] S4 specifically includes:

[0021] Construction and Training of the Feature Prediction Model: After completing the feature engineering and correlation analysis of multi-dimensional operating condition parameters, a feature prediction model for predicting cigarette production quality is constructed. This feature prediction model is a time-series prediction model based on a multi-layer neural network. The model structure sequentially includes a Long Short-Term Memory (LSTM) layer, an attention layer, and a projection layer. The LSTM layer is used to perform time-series modeling of multi-dimensional operating condition features to capture the dynamic characteristics of operating condition parameters evolving over time. The attention layer is used to weight the contribution of different features at different times. The projection layer maps the model output to predicted quality indicators. Using the multi-dimensional time-series feature input sequence formed during the feature engineering stage as the model input, the time-series prediction model is trained using historical production data. This allows the model to learn the complex nonlinear mapping relationship between cigarette machine operating condition parameters and key quality indicators of cigarette products. Through this training process, the model acquires the ability to predict future production states.

[0022] Based on the current operating condition characteristics input, the feature prediction model can predict the future... Predict key quality indicators of cigarette products within seconds and output the corresponding predicted values. This provides a predictive basis for subsequent quality assessment and control decisions.

[0023] S5. Input the real-time operating condition feature sequence into the trained model to obtain key quality indicators and evaluate the dynamic impact of key parameters on quality fluctuations.

[0024] S5 specifically includes:

[0025] Quality Advance Prediction: After the feature prediction model is built and trained, the feature prediction processing inputs the real-time collected and feature-engineered working condition feature sequences into the trained time-series prediction model to make advance predictions on the production quality of the cigarette machine. The prediction model outputs future... Predicted values ​​of key quality indicators for cigarette products within seconds By analyzing the prediction results, the system can gain a forward-looking understanding of future quality conditions before quality deviations actually occur. Furthermore, by combining the changes in operating characteristics over time, the predicted... Dynamic evaluation is conducted to analyze the impact of changes in key operating parameters on quality fluctuation trends, thereby enabling real-time prediction and evaluation of the production quality status of cigarette machines.

[0026] S6. A multi-index quality evaluation system is established through the analytic hierarchy process (AHP) to comprehensively evaluate the key quality indicators and other key quality indicators, obtain quantitative scores, and activate graded quality early warning signals based on the degree to which the quantitative scores deviate from the target values. The early warning signals are then used to drive the cigarette machine control system to adjust parameters, thereby achieving proactive correction of quality deviations.

[0027] S6 specifically includes:

[0028] Comprehensive Quality Evaluation and Early Warning Triggering: First, a multi-index comprehensive quality evaluation system is constructed based on the Analytic Hierarchy Process (AHP). Key quality indicators such as single cigarette weight, hardness, and circumference are used as evaluation factors. The weight coefficients of each quality indicator in the overall quality evaluation are determined through hierarchical structure modeling and consistency checks. Based on this, the predicted... The overall production quality is calculated by weighting and integrating other key quality indicators to obtain a quantitative comprehensive evaluation score reflecting the overall production quality status. This comprehensive evaluation score is then compared with a preset target quality level. Based on the degree to which the quantitative score deviates from the target value, a tiered quality early warning mechanism is activated. When the comprehensive evaluation score is within the allowable fluctuation range, the current production quality is considered to be in a normal state; when the comprehensive evaluation score exceeds a preset threshold, a corresponding level of quality early warning signal is triggered. After triggering the quality early warning signal, the early warning signal drives the cigarette machine control system to adjust the process parameters related to quality fluctuations. These parameter adjustments are based on a forward-looking correction of predicted quality deviations, allowing control actions to intervene before quality deviations actually occur, thereby suppressing quality deterioration trends and achieving proactive intervention and stable control of cigarette machine production quality.

[0029] As a further aspect of the present invention, S6 includes:

[0030] S6.1, Graded early warning decision-making and severe early warning determination: Based on the comprehensive quality evaluation results obtained in S6, determine whether the second threshold is met. The corresponding severe warning condition; the comprehensive quality evaluation result is lower than the second threshold. When the overall quality evaluation result is not lower than the second threshold, a severe quality warning signal is triggered, and an emergency intervention process is initiated; At that time, the mild warning determination step is initiated;

[0031] Specifically, S6.1 includes:

[0032] Based on the comprehensive quality evaluation score obtained in S6, a severe warning is issued for the production quality status of the cigarette machine. The comprehensive quality evaluation score is then compared with a second threshold. A comparison is made when the overall quality evaluation score is lower than the second threshold. When the current production quality status is determined to have entered a severely abnormal range, a severe quality warning signal is triggered, and an emergency intervention process is initiated; if the comprehensive quality evaluation score is not lower than the second threshold... If the conditions for triggering a severe warning are not met, the severe warning determination process ends and the process proceeds to the subsequent mild warning determination step. This severe warning determination mechanism enables the priority identification of serious quality risks, providing a basis for subsequent tiered warning decisions and the implementation of intervention strategies.

[0033] S6.2, Graded Early Warning Decision Judgment, Mild Early Warning Judgment: In the absence of triggering the severe quality early warning signal described in S6.1, a mild early warning judgment is made on the production quality status of the cigarette machine, and further judgment is made on whether the comprehensive quality evaluation result meets the first threshold. The corresponding mild warning conditions; the comprehensive quality evaluation results obtained in S6 are compared with the first threshold. A comparison is made, and when the overall quality evaluation result is lower than the first threshold... When a potential deviation risk is identified in the current production quality status, a mild quality warning signal is triggered, and a manual attention prompt is issued to the operator; when the comprehensive quality evaluation result is not lower than the first threshold... If the current production quality is determined to be within an acceptable range, the current operating parameters of the cigarette machine should be maintained unchanged.

[0034] The aforementioned mild early warning mechanism enables early warning of potential quality risks, providing a basis for decision-making for the stable operation of subsequent production processes.

[0035] S6.3 Control Execution and Closed-Loop Feedback: Based on the warning level triggered by S6.1 or S6.2, execute the corresponding control strategy to intervene in the cigarette machine production process;

[0036] Among them, when a severe warning signal is triggered, the control system automatically adjusts the operating parameters of the cigarette machine or implements emergency correction measures to suppress the further expansion of quality deviation;

[0037] When a mild warning signal is triggered, maintain the current operating parameters and monitor the system in conjunction with manual intervention;

[0038] After the control is executed, the multi-dimensional operating parameters of the cigarette machine are collected in real time, and the updated operating data is fed back to the data acquisition step of S1 to drive the subsequent data preprocessing, quality prediction and evaluation process, thereby forming a closed-loop quality forward-looking control process based on prediction-evaluation-decision-execution.

[0039] As a further aspect of the present invention, S3 specifically includes:

[0040] S3.1 Construction of reference and comparison sequences;

[0041] Let the normalized multidimensional operating parameters form a comparison sequence:

[0042] ;

[0043] in, Indicates the first The operating condition parameter at the first Normalized values ​​at each time sample This represents the total number of operating parameters. The sample length;

[0044] Use key quality indicators or their comprehensive quality characterization quantities in cigarette production as a reference sequence:

[0045] ;

[0046] Key quality indicators include the weight of a single cigarette. Cigarette hardness and circumference The reference sequence can be either a single quality index sequence or a quality target sequence obtained through comprehensive evaluation, to adapt to different modeling needs.

[0047] S3.2 Calculation of Grey Relational Coefficient;

[0048] Based on the concept of point association, the first... Each operating parameter at time The correlation coefficient between the location and the quality target is:

[0049]

[0050] in: The resolution coefficient is used to adjust the resolution of the correlation coefficient. The extreme value terms in the numerator and denominator represent the minimum and maximum absolute deviations between all operating parameters and the reference sequence, respectively.

[0051] S3.3 Definition of Grey Relational Degree;

[0052] The grey relational degree is obtained by averaging the correlation coefficients over time.

[0053]

[0054] in, Indicates the first The overall correlation strength between individual operating parameters and the quality target sequence;

[0055] S3.4 Feature filtering and weight guidance based on grey relational degree;

[0056] Based on the calculated grey relational degree sequence:

[0057]

[0058] The parameters of each working condition are sorted, and the working condition parameters with high gray correlation are selected according to the preset threshold or the sorting result as the key feature inputs for subsequent model training.

[0059] S3.5, Construction of multi-dimensional temporal features;

[0060] After completing data normalization and grey relational analysis, the operating parameters are filtered based on the grey relational degree results; let the set of operating parameter indices that meet the relational degree requirements be:

[0061]

[0062] in, The preset gray relational threshold;

[0063] For each selected key operating condition parameter While preserving its original time-series information, its time-domain features, frequency-domain features, and statistical features are further extracted, and these features are combined to form extended feature components. The extended feature components include mean, variance, rate of change, dominant frequency amplitude, and energy characteristics, which are used to characterize the dynamic change characteristics of the operating parameters.

[0064] The above are based on key operating condition parameters The derived extended features are aligned and fused along the time dimension to construct a multi-dimensional temporal feature input vector:

[0065]

[0066] in, Indicates the first Multidimensional temporal feature vectors at each time sample The feature dimension after fusion, and each feature component All parameters are derived from operating conditions filtered by grey relational analysis. and its feature transformation results.

[0067] As a further aspect of the present invention, in step S2, the feature quantities extracted from the mechanical vibration signal include at least the peak value, effective value, and main frequency characteristics of the vibration acceleration signal, which are used to characterize the dynamic stability and abnormal fluctuation characteristics of the key transmission components of the cigarette machine during operation.

[0068] As a further aspect of the present invention, in step S4, a time-series prediction model based on a multi-layer neural network is constructed. This model includes a long short-term memory network layer, an attention layer, and a projection layer. The time-series feature input sequence is used to train the model, enabling it to learn the nonlinear mapping relationship between multi-dimensional operating parameters in cigarette machine production data and key quality indicators of cigarette products. Furthermore, it can predict key quality indicators of cigarette products in the future time domain based on current operating conditions, thus obtaining predicted values. .

[0069] As a further aspect of the present invention, in step S4, the Long Short-Term Memory (LSTM) network layer updates its hidden state through the following gating mechanism. and cell state :

[0070]

[0071] in, For the Gate of Oblivion For input gate, For output gate, In cellular state, It is in a hidden state. For the current input features, It is the Sigmoid activation function. For Hadama accumulation, Candidate cell states are used to represent the current time step. Below, based on the hidden state vector from the previous time step... With current input features The newly generated memory information, through the combined action, performs a nonlinear mapping of the linear combination result using the hyperbolic tangent function, and is used at the input gate. Cell state update under control , , : These represent the LSTM at time 10:00 and 11:00 respectively. and time The hidden state vector. , : These represent the LSTM at time 10:00 and 11:00 respectively. and time The state of cells, : These are the weight matrices corresponding to the forget gate, input gate, candidate state, and output gate, respectively; : These are the bias vectors for the corresponding gating units. Hyperbolic tangent activation function Vector concatenation operator.

[0072] As a further aspect of the present invention, the training objective of the model in S4 is to minimize the mean squared error loss between the predicted value and the actual measured value. :

[0073] ;

[0074] in, Indicates time-based Operating condition information for future moments Prediction results of key quality indicators for cigarette products. Indicates future time The corresponding quality indicators obtained from actual measurements.

[0075] As a further aspect of the present invention, S5 includes:

[0076] S5.1, Construction of real-time operating condition feature sequences and model input;

[0077] During the operation of the cigarette rolling machine, the system uses a fixed sampling period. Multi-dimensional operating condition parameters are collected in real time, and the real-time data is synchronously cleaned, normalized, and processed using feature mapping according to the feature engineering process defined in S2 to construct a real-time operating condition feature vector.

[0078]

[0079] in, It includes key operating condition parameters and their time-domain, frequency-domain, and statistical characteristics after being filtered by grey relational analysis;

[0080] By using a sliding time window method, continuous The feature vectors at each time point are combined to form a real-time operating condition feature input sequence:

[0081]

[0082] The real-time operating condition feature input sequence serves as the real-time input to the trained prediction model;

[0083] S5.2 Advance prediction of key quality indicators;

[0084] Input the real-time operating condition features into the sequence Input the trained LSTM-Attention prediction model to obtain the future prediction time window. Predicted results of key quality indicators within the system:

[0085]

[0086] in, This represents the mapping relationship of the deep prediction model after training. This includes predicted values ​​for one or more key quality indicators such as single piece weight, circumference, and hardness.

[0087] S5.3 Assessment of quality fluctuation trends and dynamic change characteristics;

[0088] Introducing predicted quality increment and rate of change as dynamic evaluation indicators:

[0089]

[0090] in, The magnitude of the shift reflecting the prediction quality. Characterizes the rate of quality change and is used to identify the evolution trend of quality state;

[0091] S5.4 Assessment of the dynamic impact of key operating parameters on quality fluctuations;

[0092] In the real-time prediction process, the grey relational degree obtained in S2 is combined. and the feature importance weights output by the LightGBM model in S3 The impact of each operating condition parameter on the predicted quality fluctuation is dynamically evaluated.

[0093] Definition of the first Each operating parameter at time The comprehensive influencing factors on changes in quality forecasts are:

[0094]

[0095] in: Indicates the feature importance weights based on LightGBM; These are the operating parameters; This represents the degree of association obtained from grey relational analysis; This is the reference average value of the parameter under stable production conditions;

[0096] when When the preset threshold is exceeded, the operating condition parameter is determined to have a significant impact on the current quality fluctuation.

[0097] As a further aspect of the present invention, the first threshold Second threshold They are respectively:

[0098] ; ;

[0099] in, ,and Determined based on the company's quality control strategy or historical operating experience. , These are the mean and standard deviation of the comprehensive quantitative score, respectively.

[0100] This invention provides a predictive and forward-looking control system for the production quality of cigarette machines based on operating condition parameters. The system includes a module for executing the predictive and forward-looking control method for the production quality of cigarette machines based on operating condition parameters.

[0101] The beneficial effects of this invention are:

[0102] 1. This invention takes the multi-dimensional operating parameters formed during the operation of a cigarette machine as the research object. By modeling, analyzing and predicting the operating parameters, it realizes real-time evaluation, advanced prediction and active intervention control of key physical quality indicators of cigarette products (including but not limited to single cigarette weight, hardness and circumference).

[0103] 2. By constructing a working condition-quality mapping relationship and a hierarchical early warning control strategy, this invention transforms quality control from a passive approach that relies on post-event detection and experience-based adjustments to a proactive quality control method based on operational status prediction, providing an implementable technical solution for achieving stable operation and improved quality consistency in the cigarette production process.

[0104] 3. This invention identifies potential risks before quality indicators deviate significantly by analyzing and modeling multi-dimensional working condition information in real time, and provides a forward-looking quality control strategy to overcome the lag problem of traditional post-event quality control methods. Attached Figure Description

[0105] Figure 1 This is a block diagram of the real-time prediction system architecture for cigarette machine production quality based on multi-dimensional operating parameters, illustrating the overall process from data acquisition to closed-loop control. Detailed Implementation

[0106] Example 1: As Figure 1 As shown, a method for predictive and forward-looking control of cigarette machine production quality based on operating condition parameters includes:

[0107] S1. Real-time acquisition of multi-dimensional operating parameters during the operation of the cigarette machine;

[0108] The system establishes a multi-sensor network to collect and synchronize the following multi-dimensional operating parameters (input features) in real time. ):

[0109] Mechanical condition characteristics Vibration acceleration sensors and temperature sensors, derived from key transmission components, are used to detect abnormal machine conditions.

[0110] Material consistency characteristics include online moisture content of tobacco shreds, tobacco shred filling value (density), and cigarette paper air permeability, which are used to control the consistency of cigarette materials.

[0111] The operating parameters include at least electromechanical operating parameters, material operating parameters, and environmental operating parameters;

[0112] S2. Data preprocessing: The collected multi-dimensional operating condition parameters are cleaned, normalized, and transformed. Then, time-frequency analysis technology is used to analyze the signals collected in the operating condition parameters and extract the feature quantities of each operating condition parameter.

[0113] During the high-speed continuous production of cigarette machines, the collected multi-dimensional operating parameters exhibit significant differences in dimensions, numerical range, and fluctuation scale, and are inevitably affected by factors such as sensor noise, sampling jitter, and operating condition switching. To ensure the stability and comparability of subsequent feature analysis and model training, this invention first performs data preprocessing on the original collected data, including data cleaning, normalization, and necessary data transformations.

[0114] The original operating condition parameter sequence after outlier removal and missing value correction The dimensionless processing is performed using the Min–Max normalization method, and its formula is as follows:

[0115]

[0116] in, and These represent the minimum and maximum values ​​of the corresponding operating parameters within the historical stable production range.

[0117] Through the above processing, operating parameters with different dimensions and amplitude ranges are uniformly mapped to... The interval provides a unified scale basis for subsequent correlation analysis and feature fusion.

[0118] As a further aspect of the present invention, in step S2, the feature quantities extracted from the mechanical vibration signal include at least the peak value, effective value, and main frequency characteristics of the vibration acceleration signal, which are used to characterize the dynamic stability and abnormal fluctuation characteristics of the key transmission components of the cigarette machine during operation.

[0119] S3. Feature Engineering and Correlation Analysis: The grey relational analysis (GRA) method is used to analyze the correlation between the features of each working condition parameter and the key quality indicators, and the key working condition features are selected according to the correlation degree to form a multi-dimensional time series feature input sequence.

[0120] In industrial production systems, various operating parameters and quality indicators generally exhibit characteristics such as nonlinearity, weak coupling, and incomplete information. Traditional methods based on distance or statistical assumptions are difficult to accurately describe their inherent relationships.

[0121] Therefore, this invention introduces the Grey Relational Analysis (GRA) method from Grey System Theory to quantify the influence of different operating parameters on the fluctuation of cigarette machine production quality.

[0122] As a further aspect of the present invention, S3 specifically includes:

[0123] S3.1 Construction of reference and comparison sequences;

[0124] Let the normalized multidimensional operating parameters form a comparison sequence:

[0125] ;

[0126] in, Indicates the first The operating condition parameter at the first Normalized values ​​at each time sample This represents the total number of operating parameters. The sample length;

[0127] Use key quality indicators or their comprehensive quality characterization quantities in cigarette production as a reference sequence:

[0128] ;

[0129] Key quality indicators include the weight of a single cigarette. Cigarette hardness and circumference The reference sequence can be either a single quality index sequence or a quality target sequence obtained through comprehensive evaluation, to adapt to different modeling needs.

[0130] S3.2 Calculation of Grey Relational Coefficient;

[0131] Based on the concept of point association, the first... Each operating parameter at time The correlation coefficient between the location and the quality target is:

[0132]

[0133] in: The resolution coefficient is used to adjust the resolution of the correlation coefficient. The extreme value terms in the numerator and denominator represent the minimum and maximum absolute deviations between all operating parameters and the reference sequence, respectively.

[0134] Correlation coefficient The larger the value, the more consistent the trend of the working condition parameters with the trend of the quality target at that moment.

[0135] S3.3 Definition of Grey Relational Degree;

[0136] To comprehensively reflect the influence of operating parameters over the entire time series, the correlation coefficients are averaged over the time dimension to obtain the grey relational degree:

[0137]

[0138] in, Indicates the first The overall correlation strength between individual operating parameters and the quality target sequence;

[0139] The greater the grey relational degree, the higher the influence of the operating condition parameter on the fluctuation of the cigarette machine's production quality, and the stronger the indicative significance of its change on the quality status.

[0140] S3.4 Feature filtering and weight guidance based on grey relational degree;

[0141] Based on the calculated grey relational degree sequence:

[0142]

[0143] The parameters of each working condition are sorted, and the working condition parameters with high gray correlation are selected according to the preset threshold or the sorting result as the key feature inputs for subsequent model training.

[0144] Meanwhile, the gray relational degree is not only used for feature selection, but also serves as important prior information for feature weight allocation and attention mechanism guidance in subsequent models, thereby enhancing the model's sensitivity to key operating parameters and improving the quality prediction model's responsiveness to changes in production status.

[0145] S3.5, Construction of multi-dimensional temporal features;

[0146] After completing data normalization and grey relational analysis, the operating parameters are filtered based on the grey relational degree results; let the set of operating parameter indices that meet the relational degree requirements be:

[0147]

[0148] in, The preset gray relational threshold;

[0149] For each selected key operating condition parameter While preserving its original time-series information, its time-domain features, frequency-domain features, and statistical features are further extracted, and these features are combined to form extended feature components. The extended feature components include mean, variance, rate of change, dominant frequency amplitude, and energy characteristics, which are used to characterize the dynamic change characteristics of the operating parameters.

[0150] The above are based on key operating condition parameters The derived extended features are aligned and fused along the time dimension to construct a multi-dimensional temporal feature input vector:

[0151]

[0152] in, Indicates the first Multidimensional temporal feature vectors at each time sample The feature dimension after fusion, and each feature component All parameters are derived from operating conditions filtered by grey relational analysis. and its feature transformation results.

[0153] S4. Construct a time series prediction model based on a multi-layer neural network, and train the time series prediction model using the time series feature input sequence;

[0154] As a further aspect of the present invention, in step S4, a time-series prediction model based on a multi-layer neural network is constructed. This model includes a Long Short-Term Memory (LSTM) layer, an attention layer, and a projection layer. The time-series feature input sequence is used to train the model, enabling it to learn the nonlinear mapping relationship between multi-dimensional operating condition parameters and key quality indicators of cigarette products in cigarette machine production data, and to predict future trends based on current operating conditions. Key quality indicators of cigarette products were predicted within seconds. .

[0155] S4 includes:

[0156] S4.1 Quality Indicator Prediction Model Based on LSTM–Attention

[0157] (1) LSTM timing modeling structure

[0158] To capture the long-term dependencies and complex nonlinear mappings in the operating condition parameter sequence, this invention constructs a time series prediction model based on Long Short-Term Memory (LSTM) network.

[0159] LSTM networks selectively memorize historical information through a gating mechanism. The forget gate, input gate, output gate, and state update process are defined as follows:

[0160]

[0161] The meanings of each symbol are as follows: Time step index; :time The multidimensional working condition feature input vector is obtained from the aforementioned feature engineering steps; , : These represent the LSTM at time 10:00 and 11:00 respectively. and time The hidden state vector is used to characterize the short-term memory information of the system; , : These represent the LSTM at time 10:00 and 11:00 respectively. and time The cellular state used to store long-term memory information; Forget gate: Used to control the cell state at the previous moment. The degree to which information is retained; Input gate, used to control the current candidate state. The proportion of information written into the cell state; Candidate cell states are used to represent new information generated from the current input and historical states. Output gate: Used to control the proportion of information output from the cell state to the hidden state; : These are the weight matrices corresponding to the forget gate, input gate, candidate state, and output gate, respectively; : These are the bias vectors of the corresponding gating units; Sigmoid activation function; Hyperbolic tangent activation function; Hadamard product, representing element-wise multiplication of vectors; Vector concatenation operator.

[0162] (2) Introduction of attention mechanism

[0163] To further enhance the model's ability to model the differences in contributions of different time steps and features, an attention mechanism (Attention Layer) is introduced after the LSTM output layer.

[0164] Let the length be Within the sliding time window, the LSTM network models the continuously input sequence of operating conditions and outputs the hidden state sequence sequentially:

[0165]

[0166] in, Indicates that LSTM in the th The hidden state vector corresponding to each time step. Let be the number of historical time steps used for prediction. This represents the hidden state corresponding to the current prediction time, typically the hidden state at the end of the time window. , is used to characterize the overall system state at the current moment.

[0167] To describe the relative contribution of different historical hidden states to the current prediction task, an attention mechanism is introduced to weight the hidden state sequence. First, a correlation scoring function is used... Calculate the first A historical hidden state With the current hidden state Correlation score between them:

[0168]

[0169] in, It is a trainable correlation function used to measure the degree of matching between historical states and the current prediction target. Its specific form can be a vector dot product, a weighted linear mapping, or a scoring function based on a feedforward neural network.

[0170] Subsequently, the relevance scores were normalized using the Softmax function to obtain the attention weights corresponding to each time step:

[0171]

[0172] in, ,and , used to characterize the The relative importance of historical hidden states in current predictions.

[0173] Based on the attention weights, the historical hidden states are weighted and summed to construct a context vector:

[0174]

[0175] The context vector It integrates information from different historical time steps and can adaptively highlight key moment features that contribute most to the current prediction, thereby improving the model's ability to model the relationship between operating condition changes and quality evolution.

[0176] (3) Predicted output and objective function

[0177] Context vector A linear mapping is performed via a projection layer to output the future prediction time window. Predicted values ​​of key quality indicators within ,in Indicates time-based Operating condition information for future moments Predictive results of key quality indicators for cigarette products.

[0178] During model training, the corresponding quality indicators obtained from actual measurements are used. As a supervisory signal, the model parameters are optimized by minimizing the mean squared error (MSE) loss function between the predicted and actual values. The loss function is defined as:

[0179]

[0180] This model follows the principle of structural risk minimization, and can still maintain good generalization ability even when the sample size is limited or the data is noisy.

[0181] S4.2 Operating Condition Characteristic Impact Assessment Model Based on LightGBM

[0182] After completing the construction of multi-dimensional working condition features and training of the deep prediction model, in order to further quantify the influence of each working condition parameter on the fluctuation of the predicted quality index, this invention introduces a feature influence assessment model based on gradient boosting decision tree (Light Gradient Boosting Machine, LightGBM) to model the nonlinear contribution relationship between working condition features and quality index.

[0183] LightGBM employs an additive model structure composed of multiple progressively stacked regression decision trees, which at time... The predicted output of the quality indicator is defined as follows:

[0184]

[0185] in, Indicates time The multidimensional working condition feature input vector For the first The values ​​of each working condition parameter, The total dimension of the working condition features; Indicates the first The mapping function corresponding to each regression decision tree; This represents the total number of regression decision trees in the model; These are the predicted quality metrics output by the LightGBM model.

[0186] In the initial stage of model building, LightGBM uses a constant function as the base model to provide initial estimates for predictions:

[0187]

[0188] Where, constant Determined by minimizing the overall loss function between the predicted value and the actual quality indicator:

[0189]

[0190] In the formula, Indicates time The corresponding actual measured quality index value, For loss function, This represents the total number of training samples. From this, we obtain the initial prediction results:

[0191]

[0192] In the In this iteration, the model introduces a new regression decision tree based on the existing prediction results to fit the current prediction residuals. Its update form is as follows:

[0193]

[0194] in, Indicates the first The predicted output after the next iteration. For the first A regression decision tree model used to correct prediction errors.

[0195] No. Each regression decision tree is determined by minimizing the following objective function with a regularization term:

[0196] in, For the first The structural complexity constraint term of a decision tree is defined as follows:

[0197] In the formula, Indicates the first The number of leaf nodes in a decision tree. For the first The output weights corresponding to each leaf node and These are the regularization coefficients used to constrain the tree size and the leaf node weight magnitude, respectively.

[0198] To reduce the computational complexity of directly optimizing the objective function, LightGBM performs a second-order Taylor expansion of the loss function and incorporates gradient information for approximate solution. Defined in the... The first and second gradients in the next iteration are:

[0199]

[0200] in, and These represent the first and second derivatives of the loss function with respect to the predicted output, respectively.

[0201] At this time, the The optimization problem of decision trees can be transformed into:

[0202]

[0203] In LightGBM, each regression decision tree divides the feature space into several non-overlapping leaf node regions, and its mapping form is represented as:

[0204] in, The function representing the mapping from samples to leaf nodes. For the sample The output value of the leaf node.

[0205] For the first The optimal output weight of each leaf node can be analytically solved as follows:

[0206]

[0207] in, This represents the set of sample indices that fall into this leaf node.

[0208] Node splitting is based on maximizing the splitting gain, and its gain function is defined as:

[0209]

[0210] Among them, subscript and These represent the cumulative gradient values ​​of the samples within the left and right child nodes after the split. and This represents the cumulative gradient value of the samples within the current node.

[0211] Importance weights of each working condition feature in the model This feature is obtained by statistically analyzing the cumulative gain generated by its participation in node splitting across all regression decision trees, and is used to quantify the first... The relative influence of each operating condition parameter on the prediction results of quality indicators.

[0212] By iteratively constructing multiple regression decision trees, a complete LightGBM model is finally formed, and the set of feature importance weights corresponding to each working condition is obtained:

[0213]

[0214] The feature importance weights are used to characterize the influence of different operating parameters on the predicted changes in quality indicators, and serve as an important basis for subsequent quality fluctuation assessment and graded early warning decision-making.

[0215] S5. Input the real-time operating condition feature sequence into the trained model to obtain key quality indicators and evaluate the dynamic impact of key parameters on quality fluctuations.

[0216] As a further aspect of the present invention, S5 includes:

[0217] S5.1, Construction of real-time operating condition feature sequences and model input;

[0218] During the operation of the cigarette rolling machine, the system uses a fixed sampling period. Multi-dimensional operating condition parameters are collected in real time, and the real-time data is synchronously cleaned, normalized, and processed using feature mapping according to the feature engineering process defined in S2 to construct a real-time operating condition feature vector.

[0219]

[0220] in, It includes key operating condition parameters and their time-domain, frequency-domain, and statistical characteristics after being filtered by grey relational analysis;

[0221] By using a sliding time window method, continuous The feature vectors at each time point are combined to form a real-time operating condition feature input sequence:

[0222]

[0223] The real-time operating condition feature input sequence serves as the real-time input to the trained prediction model;

[0224] S5.2 Advance prediction of key quality indicators;

[0225] Input the real-time operating condition features into the sequence Input the trained LSTM-Attention prediction model to obtain the future prediction time window. Predicted results of key quality indicators within the system:

[0226]

[0227] in, This represents the mapping relationship of the deep prediction model after training. Predicted values ​​of one or more key quality indicators, including single piece weight, circumference, hardness, etc.

[0228] With this prediction result, the system can gain an advance perception of the future production quality status before quality deviations actually occur.

[0229] S5.3 Assessment of quality fluctuation trends and dynamic change characteristics;

[0230] To describe the changing trends of quality indicators, predicted quality increments and rates of change are introduced as dynamic evaluation indicators:

[0231]

[0232] in, The magnitude of the shift reflecting the prediction quality. Characterizes the rate of quality change and is used to identify the evolution trend of quality state;

[0233] S5.4 Assessment of the dynamic impact of key operating parameters on quality fluctuations;

[0234] In the real-time prediction process, the grey relational degree obtained in S2 is combined. and the feature importance weights output by the LightGBM model in S3 The impact of each operating condition parameter on the predicted quality fluctuation is dynamically evaluated.

[0235] Definition of the first Each operating parameter at time The comprehensive influencing factors on changes in quality forecasts are:

[0236]

[0237] in: Indicates the feature importance weights based on LightGBM; These are the operating parameters; This represents the degree of association obtained from grey relational analysis; This is the reference average value of the parameter under stable production conditions;

[0238] when When the preset threshold is exceeded, the operating condition parameter is determined to have a significant impact on the current quality fluctuation.

[0239] S6. A multi-index quality evaluation system is established through the Analytic Hierarchy Process (AHP) to comprehensively evaluate the key quality indicators and other key quality indicators, obtain quantitative scores, and activate graded quality early warning signals based on the degree to which the quantitative scores deviate from the target values. The early warning signals are then used to drive the cigarette machine control system to adjust parameters, thereby achieving proactive correction of quality deviations.

[0240] As a further aspect of the present invention, S6 includes:

[0241] S6.1, Graded early warning decision-making and severe early warning determination: Based on the comprehensive quality evaluation results obtained in S6, determine whether the second threshold is met. The corresponding severe warning condition; the comprehensive quality evaluation result is lower than the second threshold. When the overall quality evaluation result is not lower than the second threshold, a severe quality warning signal is triggered, and an emergency intervention process is initiated; At that time, the mild warning determination step is initiated;

[0242] S6.2, Graded early warning decision-making, mild early warning determination: If a severe early warning is not triggered, further determine whether the comprehensive quality evaluation result meets the first threshold. The corresponding mild warning condition; when the comprehensive quality evaluation result is lower than the first threshold. When the overall quality evaluation result is not lower than the first threshold, a mild quality warning signal is triggered, and a manual attention prompt is issued to the operator; At the same time, maintain the current operating parameters of the cigarette machine unchanged;

[0243] S6.3 Control Execution and Closed-Loop Feedback: Execute the corresponding control strategy according to the warning level triggered by S6.1 or S6.2.

[0244] Among them, when a severe warning signal is triggered, the control system automatically adjusts the operating parameters of the cigarette machine or implements emergency correction measures.

[0245] When a mild warning signal is triggered, maintain the current operating parameters and monitor the system in conjunction with manual intervention;

[0246] After the control is executed, multi-dimensional operating condition parameters are collected in real time and returned to S1, forming a closed-loop quality forward-looking control process based on prediction, evaluation, decision-making and execution.

[0247] Furthermore, S6 also includes the following:

[0248] (1) Construction of comprehensive evaluation system: In this invention, the comprehensive quality evaluation does not directly superimpose the original quality indicators linearly, but first performs predictive-driven efficacy mapping processing on different quality indicators.

[0249] This is because the quality attributes of cigarette products have the following significant characteristics:

[0250] 1. The sources of indicators are heterogeneous (physical quantities such as weight, hardness, and circumference);

[0251] 2. Significant differences exist in the dimensions and numerical ranges;

[0252] 3. Different indicators have different "optimal directions" (too large, too small, or have an optimal range).

[0253] 4. The evaluation object is the "predicted value" rather than the post-measured value.

[0254] Therefore, before comprehensive evaluation, all predicted quality indicators must be uniformly mapped to a dimensionless, comparable efficacy space to support subsequent weighted fusion and early warning decision-making using the Analytic Hierarchy Process (AHP).

[0255] Suppose the system obtains M sets of prediction samples at the prediction time, involving N key quality indicators, forming a prediction quality dataset: Each indicator The values ​​are all positive real numbers. This represents the predicted value of the nth quality indicator for the nth cigarette within the future prediction window; the predicted value is output by the LSTM-Attention deep model.

[0256] To facilitate subsequent fusion of multiple indicators, a predictive power score function is introduced. This is used to describe the degree to which a predictive indicator meets its ideal state.

[0257] S6.1.1 For indicators where larger predicted values ​​indicate better quality (denoted as skewed large indicators), the following predictive power function is constructed:

[0258]

[0259] in: Under historically stable production conditions, the first The statistical lower and upper limits of each indicator; the intervals can be obtained through SPC control limits or empirical process parameters; after mapping A larger value indicates a better predicted quality.

[0260] S6.1.2 For indicators where smaller predicted values ​​indicate better quality (denoted as smaller-than-average indicators).

[0261] For an indicator where a smaller predicted value signifies better quality, its power mapping function is defined as follows:

[0262]

[0263] S6.1.3 For quality indicators (such as single cigarette weight, circumference, etc.) that have a large number of "optimal process ranges" in cigarette production, this invention constructs a two-sided attenuation type prediction efficacy function:

[0264]

[0265] in: For the first The optimal target value of each indicator under stable operating conditions; when the predicted value deviates from the optimal point, the effectiveness score continuously decreases with the degree of deviation; the function in The maximum value is obtained at the location. .

[0266] S6.1.4 After the above power mapping, construct the predicted quality power matrix:

[0267]

[0268] This matrix serves as the direct input to the Analytic Hierarchy Process (AHP) comprehensive evaluation model in S5, and is used to calculate the predicted comprehensive quality score:

[0269]

[0270] in These are the indicator weights determined through AHP.

[0271] (2) Tiered early warning and closed-loop intervention;

[0272] Based on the principles of Statistical Process Control (SPC), this invention aims to predict comprehensive quality criterion variables. As a unified monitoring object, a hierarchical quality early warning mechanism based on prediction results is constructed.

[0273] S6.2.1 Statistical modeling of comprehensive quality criteria;

[0274] Under stable production conditions, the historical forecast comprehensive quality criterion sequence Perform statistical analysis and calculate its mean and standard deviation:

[0275]

[0276] in, Characterizes the center mass level under steady-state operating conditions. It reflects the intensity of predicted quality fluctuations.

[0277] S6.2.2 Definition of graded early warning thresholds;

[0278] Based on the SPC control concept, this study targets the comprehensive quality criterion variables. Set two levels of early warning thresholds:

[0279] First threshold (mild warning threshold):

[0280]

[0281] Second threshold (severe warning threshold):

[0282]

[0283] in, ,and It can be determined based on the company's quality control strategy or historical operating experience.

[0284] S6.2.3 based on The tiered early warning criteria;

[0285] Based on the relationship between the predictive comprehensive quality criterion variables and the early warning threshold, the following grading judgment rules are defined:

[0286]

[0287] Among them, mild warnings are used to identify potential deviations in prediction quality, while severe warnings are used to indicate that prediction quality has entered an abnormal range requiring immediate intervention.

[0288] The constraint relationship between the early warning threshold and the process capability index:

[0289] To ensure that the aforementioned early warning mechanism is consistent with industrial quality management standards, a process capability index is introduced. As a constraint indicator, it is defined as follows:

[0290]

[0291] in, and These are the upper and lower specification limits for the quality indicators. and These represent the statistical mean and standard deviation of the corresponding quality indicators.

[0292] S6.3 This invention effectively reduces the standard deviation of quality fluctuations by moving the early warning threshold forward to the prediction domain, allowing control actions to intervene before the actual occurrence of quality deviations. Achieve process capability index Continuous improvement.

[0293] S6.3.1 Definition of overall quality deviation;

[0294] Define the predicted overall quality deviation as:

[0295]

[0296] in, To determine the overall quality level, the expected value or design target value under stable operating conditions is usually taken.

[0297] S6.3.2 Construction of predictive-driven control laws;

[0298] Based on comprehensive quality deviation Construct a forward-looking regulatory control law:

[0299]

[0300] in: This represents the adjustment vector for key process parameters of the cigarette making machine; The control gain vector or gain matrix is ​​used; each control component corresponds to adjustable process parameters such as tobacco feeding speed, compaction mechanism position, and air separation pressure.

[0301] Severe warning determination;

[0302] Assuming at the prediction time The obtained overall quality score of the prediction is The second threshold is The severe warning determination rule is defined as follows:

[0303]

[0304] in, This indicates that a severe quality warning signal has been triggered, and the system has entered the emergency intervention process. If the conditions for triggering a severe warning are not met, the system will proceed to the mild warning determination step described in S8.

[0305] Mild warning assessment;

[0306] Without triggering a severe warning (i.e.) Under the premise that the first threshold is ), let it be . The criteria for determining a mild warning are defined as follows:

[0307]

[0308] in, This indicates that a mild quality warning signal has been triggered. This indicates that the overall predicted quality is within the acceptable range, and the current operating parameters of the cigarette machine will remain unchanged.

[0309] Control execution and closed-loop feedback;

[0310] Based on the judgment results of S7 and S8, the control execution strategy is defined as follows:

[0311]

[0312] in: This is the adjustment vector of key process parameters for cigarette making machines; This indicates the automatic emergency correction control quantity under severe warning conditions; This indicates the manual assistance or gradual adjustment control parameters under mild warning conditions. After the control is executed, the system continues to collect multi-dimensional operating condition parameters in the next sampling cycle.

[0313]

[0314] This results in the following closed-loop forward-looking control structure:

[0315]

[0316] Achieve closed-loop, forward-looking control of cigarette machine production quality, encompassing prediction, evaluation, decision-making, and execution.

[0317] This invention provides a predictive and forward-looking control system for the production quality of cigarette machines based on operating condition parameters. The system includes a module for executing the predictive and forward-looking control method for the production quality of cigarette machines based on operating condition parameters.

[0318] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for predictive and forward-looking control of cigarette machine production quality based on operating condition parameters, characterized in that, include: S1. Real-time acquisition of multi-dimensional operating parameters during the operation of the cigarette machine; S2. Perform data cleaning, normalization, and data transformation on the collected multi-dimensional operating condition parameters, and extract the feature quantities of each operating condition parameter. S3. Use grey relational analysis to analyze the correlation between the characteristics of each working condition parameter and the key quality indicators, and select key working condition characteristics according to the degree of correlation to form a multi-dimensional time series feature input sequence. S4. Construct a time series prediction model based on a multi-layer neural network, and train the time series prediction model using the time series feature input sequence; S5. Input the real-time operating condition feature sequence into the trained model to obtain key quality indicators and evaluate the dynamic impact of key parameters on quality fluctuations. S6. A multi-index quality evaluation system is established through the analytic hierarchy process (AHP) to comprehensively evaluate the key quality indicators and other key quality indicators, obtain quantitative scores, and activate graded quality early warning signals based on the degree to which the quantitative scores deviate from the target values. The early warning signals are then used to drive the cigarette machine control system to adjust parameters, thereby achieving proactive correction of quality deviations.

2. The method for predicting and proactively controlling the production quality of cigarette machines based on operating condition parameters according to claim 1, characterized in that: S6 includes: S6.1, Graded early warning decision-making and severe early warning determination: Based on the comprehensive quality evaluation results obtained in S6, determine whether the second threshold is met. The corresponding severe warning condition; the comprehensive quality evaluation result is lower than the second threshold. When the overall quality evaluation result is not lower than the second threshold, a severe quality warning signal is triggered, and an emergency intervention process is initiated; At that time, the mild warning determination step is initiated; S6.2, Graded early warning decision-making, mild early warning determination: If a severe early warning is not triggered, further determine whether the comprehensive quality evaluation result meets the first threshold. The corresponding mild warning condition; when the comprehensive quality evaluation result is lower than the first threshold. When the overall quality evaluation result is not lower than the first threshold, a mild quality warning signal is triggered, and a manual attention prompt is issued to the operator; At the same time, maintain the current operating parameters of the cigarette machine unchanged; S6.3 Control Execution and Closed-Loop Feedback: Execute the corresponding control strategy according to the warning level triggered by S6.1 or S6.

2. Among them, when a severe warning signal is triggered, the control system automatically adjusts the operating parameters of the cigarette machine or implements emergency correction measures. When a mild warning signal is triggered, maintain the current operating parameters and monitor the system in conjunction with manual intervention; After the control is executed, multi-dimensional operating condition parameters are collected in real time and returned to S1, forming a closed-loop quality forward-looking control process based on prediction, evaluation, decision-making and execution.

3. The method for predicting and proactively controlling the production quality of cigarette machines based on operating condition parameters according to claim 1, characterized in that: S3 specifically includes: S3.1 Construction of reference and comparison sequences; Let the normalized multidimensional operating parameters form a comparison sequence: ； in, Indicates the first The operating condition parameter at the first Normalized values ​​at each time sample This represents the total number of operating parameters. The sample length; Use key quality indicators or their comprehensive quality characterization quantities in cigarette production as a reference sequence: ； Key quality indicators include the weight of a single cigarette. Cigarette hardness and circumference The reference sequence can be either a single quality index sequence or a quality target sequence obtained through comprehensive evaluation, to adapt to different modeling needs. S3.2 Calculation of Grey Relational Coefficient; Based on the concept of point association, the first... Each operating parameter at time The correlation coefficient between the location and the quality target is: ； in: The resolution coefficient is used to adjust the resolution of the correlation coefficient. The extreme value terms in the numerator and denominator represent the minimum and maximum absolute deviations between all operating parameters and the reference sequence, respectively. S3.3 Definition of Grey Relational Degree; The grey relational degree is obtained by averaging the correlation coefficients over time. ； in, Indicates the first The overall correlation strength between individual operating parameters and the quality target sequence; S3.4 Feature filtering and weight guidance based on grey relational degree; Based on the calculated grey relational degree sequence: ； The parameters of each working condition are sorted, and the working condition parameters with high gray correlation are selected according to the preset threshold or the sorting result as the key feature inputs for subsequent model training. S3.5, Construction of multi-dimensional temporal features; After completing data normalization and grey relational analysis, the operating parameters are filtered based on the grey relational degree results; let the set of operating parameter indices that meet the relational degree requirements be: ； in, The preset gray relational threshold; For each selected key operating condition parameter While preserving its original time-series information, its time-domain features, frequency-domain features, and statistical features are further extracted, and these features are combined to form extended feature components. The extended feature components include mean, variance, rate of change, dominant frequency amplitude, and energy characteristics, which are used to characterize the dynamic change characteristics of the operating parameters. The above are based on key operating condition parameters The derived extended features are aligned and fused along the time dimension to construct a multi-dimensional temporal feature input vector: ； in, Indicates the first Multidimensional temporal feature vectors at each time sample The feature dimension after fusion, and each feature component All parameters are derived from operating conditions filtered by grey relational analysis. and its feature transformation results.

4. The method for predicting and proactively controlling the production quality of cigarette machines based on operating condition parameters according to claim 1, characterized in that: In step S2, the feature quantities extracted from the mechanical vibration signal include at least the peak value, effective value, and dominant frequency characteristics of the vibration acceleration signal, which are used to characterize the dynamic stability and abnormal fluctuation characteristics of the key transmission components of the cigarette machine during operation.

5. The method for predicting and proactively controlling the production quality of cigarette machines based on operating condition parameters according to claim 1, characterized in that: In step S4, a time-series prediction model based on a multi-layer neural network is constructed. This model includes a long short-term memory network layer, an attention layer, and a projection layer. The model is trained using the time-series feature input sequence, enabling it to learn the nonlinear mapping relationship between multi-dimensional operating parameters in cigarette machine production data and key quality indicators of cigarette products. Furthermore, it can predict key quality indicators of cigarette products in the future time domain based on current operating conditions, thus obtaining predicted values. .

6. The method for predicting and proactively controlling the production quality of cigarette machines based on operating condition parameters according to claim 5, characterized in that: In step S4, the Long Short-Term Memory (LSTM) network layer updates its hidden state through the following gating mechanism. and cell state : ； in, For the Gate of Oblivion For input gate, For output gate, In cellular state, It is in a hidden state. For the current input features, It is the Sigmoid activation function. For Hadama accumulation, Candidate cell states are used to represent the current time step. Below, based on the hidden state vector from the previous time step... With current input features The newly generated memory information, through the combined action, performs a nonlinear mapping of the linear combination result using the hyperbolic tangent function, and is used at the input gate. Cell state update under control , , : These represent the LSTM at time 10:00 and 11:00 respectively. and time The hidden state vector. , : These represent the LSTM at time 10:00 and 11:00 respectively. and time The state of cells, : These are the weight matrices corresponding to the forget gate, input gate, candidate state, and output gate, respectively; : These are the bias vectors for the corresponding gating units. Hyperbolic tangent activation function Vector concatenation operator.

7. The method for predicting and proactively controlling the production quality of cigarette machines based on operating condition parameters according to claim 1, characterized in that: The training objective of the model in S4 is to minimize the mean squared error loss between the predicted and actual measured values. : ； in, Indicates time-based Operating condition information for future moments Prediction results of key quality indicators for cigarette products. Indicates future time The corresponding quality indicators obtained from actual measurements.

8. The method for predicting and proactively controlling the production quality of cigarette machines based on operating condition parameters according to claim 1, characterized in that: S5 includes: S5.1, Construction of real-time operating condition feature sequences and model input; During the operation of the cigarette rolling machine, the system uses a fixed sampling period. Multi-dimensional operating condition parameters are collected in real time, and the real-time data is synchronously cleaned, normalized, and processed using feature mapping according to the feature engineering process defined in S2 to construct a real-time operating condition feature vector. ； in, It includes key operating condition parameters and their time-domain, frequency-domain, and statistical characteristics after being filtered by grey relational analysis; By using a sliding time window method, continuous The feature vectors at each time point are combined to form a real-time operating condition feature input sequence: ； The real-time operating condition feature input sequence serves as the real-time input to the trained prediction model; S5.2 Advance prediction of key quality indicators; Input the real-time operating condition features into the sequence Input the trained LSTM-Attention prediction model to obtain the future prediction time window. Predicted results of key quality indicators within the system: ； in, This represents the mapping relationship of the deep prediction model after training. This includes predicted values ​​for one or more key quality indicators such as single piece weight, circumference, and hardness. S5.3 Assessment of quality fluctuation trends and dynamic change characteristics; Introducing predicted quality increment and rate of change as dynamic evaluation indicators: ； in, The magnitude of the shift reflecting the prediction quality. Characterizes the rate of quality change and is used to identify the evolution trend of quality state; S5.4 Assessment of the dynamic impact of key operating parameters on quality fluctuations; In the real-time prediction process, the grey relational degree obtained in S2 is combined. and the feature importance weights output by the LightGBM model in S3 The impact of each operating condition parameter on the predicted quality fluctuation is dynamically evaluated. Definition of the first Each operating parameter at time The comprehensive influencing factors on changes in quality forecasts are: ； in: Indicates the feature importance weights based on LightGBM; These are the operating parameters; This represents the degree of association obtained from grey relational analysis; This is the reference average value of the parameter under stable production conditions; when When the preset threshold is exceeded, the operating condition parameter is determined to have a significant impact on the current quality fluctuation.

9. The method for predicting and proactively controlling the production quality of cigarette machines based on operating condition parameters according to claim 2, characterized in that: The first threshold Second threshold They are respectively: ； ； in, ,and Determined based on the company's quality control strategy or historical operating experience. , These are the mean and standard deviation of the comprehensive quantitative score, respectively.

10. A predictive and forward-looking control system for cigarette machine production quality based on operating condition parameters, characterized in that, The system includes a module for executing the cigarette machine production quality prediction and forward-looking control method based on operating condition parameters as described in any one of claims 1 to 9.

Images