Intelligent regulation and control method for cigarette weight of cigarette making machine

Abstract

The invention relates to the technical field of cigarette manufacturing, and discloses an intelligent regulation and control method for cigarette weight of a cigarette making machine, and the method comprises the following steps: S1, obtaining production parameters of the cigarette making machine, and carrying out data preprocessing and feature selection; s2, constructing a hybrid prediction model based on TCN-BiLSTM and training the hybrid prediction model; s3, forming a closed-loop feedback mechanism by taking the prediction model constructed in S2 as a predictor and taking an MPC controller as a center node; according to the invention, data closed-loop feedback is effectively realized; a mixed structure of a time sequence convolutional network and a long and short memory network is adopted, the cigarette suction resistance value is predicted in real time, and the cigarette suction resistance predicted value is output in real time; a prediction model and a real-time data stream are integrated, millisecond-level rolling optimization of the cigarette weight is achieved, an optimal weight control value is output, and dependence on artificial experience is effectively eliminated.

Description

Technical Field

[0001] The present application belongs to the technical field of cigarette manufacturing, and in particular relates to an intelligent control method for cigarette weight of a cigarette making machine. Background Art

[0002] Cigarette draw resistance significantly impacts sensory quality and chemical indicators such as tar content. It is a key factor influencing consumer perception and a crucial indicator of cigarette physical quality. Numerous factors influence draw resistance, including raw materials, cigarette-making machines, and filter rods. However, in practice, this is often controlled solely by a single parameter: cigarette weight. This control method is entirely manual, subject to blind spots and reliance on experience. Existing manual control methods have the following technical features: discrete puff testing. Operators conduct quality checks at three fixed times daily, weighing and analyzing 20 short cigarette samples from each of eight production units. Weight parameter compensation is performed when the measured weight fluctuates beyond a ±5mg threshold. A dual-track calibration system includes operator-performed 5mg weight adjustments and monthly equipment slope calibration. The slope calibration threshold is a system-displayed coefficient of deviation of 1.5; if this threshold is exceeded, professional maintenance personnel intervene. Experience-based decision-making. A flexible control range of [-3, +1] is set for the cigarette resistance index. When the test value exceeds the limit, the re-inspection frequency and compensation amount are determined based on the operator's experience. The compensation data is generated into a quality tracking log through the MES system.

[0003] The existing cigarette weight control technology is mainly manual control, which has the following shortcomings: excessive dependence on manual intervention: the sampling interval is as long as 8 hours, weight compensation lags behind actual production fluctuations, and the re-inspection strategy lacks quantitative standards; the calibration system fails to coordinate: weight calibration and slope calibration belong to different operating levels, and the equipment status data does not form a closed-loop feedback; data transmission fault: laboratory test data is physically isolated from the production line control system, and the MES system only realizes the post-recording function. Summary of the Invention

[0004] The technical problem to be solved by the present application is to overcome the deficiencies of the prior art. The present application provides an intelligent control method for the weight of cigarettes in a cigarette-making machine.

[0005] To achieve the above objectives, the present application provides a method for intelligently controlling cigarette weight in a cigarette making machine, comprising the following steps:

[0006] S1. Obtain the production parameters of the cigarette making machine and perform data preprocessing and feature selection:

[0007] Cigarette machine operating parameters and QTM physical measurement data were obtained, and the importance metric of the gradient boosting decision tree was used to select features with high correlation with cigarette draw resistance.

[0008] S2. Build a hybrid prediction model based on TCN-BiLSTM:

[0009] The features selected by S1 are used as input to the prediction model. TCN-BiLSTM adopts a multimodal time series fusion architecture to extract local features and global time series features through parallel processing paths, ultimately fusing and outputting the predicted value of cigarette draw resistance. TCN uses a multi-layer residual causal dilation convolution structure to extract local features, while BiLSTM uses a two-layer bidirectional structure to capture global time series features for feature modeling. The local features represent local fluctuations in cigarette production parameters, while the global time series features represent long-term trends in the quality of different batches of cigarettes.

[0010] S3, based on the prediction model built in S2, builds an MPC-based closed-loop control system:

[0011] The current state and historical data are used as the training set input into the TCN-BiLSTM hybrid prediction model. The prediction results output by the hybrid prediction model are optimized in the rolling time domain through an optimization process. The optimal control parameters of cigarette weight are calculated by minimizing the objective function. The optimal control parameters are sent to the cigarette making machine, and the closed-loop error compensation of the prediction model is performed using real-time resistance to draw detection data.

[0012] Optionally, the TCN-BiLSTM hybrid prediction model constructed in S2 includes:

[0013] The TCN network adopts a 4-layer residual causal expansion convolution structure. The output of the previous layer is used as the input of the next layer. The output of each layer is expressed as:

[0014] h (l) =ReLU(BN(CausalConv(x (l-1) ,dilation=2 l-1 )))+x (l-1) (4)

[0015] Among them, h (l) ReLU represents the hidden layer output after the calculation of the lth layer, ReLU represents the activation function, BN represents batch normalization, CausalConv represents causal expansion convolution, dilation represents the expansion rate, 2 l-1 is the specific expansion rate value, x (l-1) Indicates that the input data of the l-1 layer of the TCN network is used as the basic input for the current layer calculation;

[0016] By exponentially increasing the expansion coefficient (1, 2, 4, 8) to expand the receptive field, the final TCN output is expressed as:

[0017] z TCN =ReLU(BN(CausalConv(tcn3)))∈RB×T×64 (5)

[0018] Among them, tcn3 is the output of the residual block of the third layer of the TCN network, R represents the shape description of the multidimensional real tensor, B represents the batch size, and T represents the time step.

[0019] Optionally, the TCN-BiLSTM hybrid prediction model constructed in S2 includes:

[0020] The architecture of the BiLSTM network adopts a two-layer bidirectional structure, which calculates the forward LSTM and backward LSTM respectively, expressed as:

[0021]

[0022] Concatenating the last time step output captures the bidirectional temporal dynamics, expressed as:

[0023] Among them, σ represents the sigmoid activation function, Denote the hidden states of the forward and backward LSTM at time step t, respectively, W xh represents the weight matrix input to the hidden layer; x t represents the input vector at time step t; Represents the weight matrices from the forward and backward hidden layers to the hidden layers respectively; Represent the hidden states of the forward and backward LSTM at time steps t-1 and t+1 respectively; represents the hidden state sequence of the forward and backward LSTM from time step 1 to T.

[0024] Optionally, the TCN-BiLSTM hybrid prediction model constructed in S2 includes:

[0025] The features output by the TCN and BiLSTM networks are concatenated and expressed as:

[0026] z concat =[FC T (z TCN W proj );FC L (z LSTM )]∈R B×160 (9)

[0027] After passing through the timing decoder, the final output sequence is obtained, which is expressed as:

[0028] Among them, B represents the batch size, N represents the prediction target dimension, and W proj is the projection matrix, FCT 、FC L Represents the fully connected operations acting on TCN and BiLSTM respectively, W out Represents the output layer weight matrix, W1∈R 160→128 ,W2∈R 128→N Represents the weight matrix, Conv1D represents the one-dimensional convolution operation, and ELU represents the activation function to alleviate gradient diffusion.

[0029] Optionally, the TCN-BiLSTM hybrid prediction model constructed in S2 includes:

[0030] The TCN-BiLSTM hybrid prediction model uses layered Dropout and L2 regularization constraint weights, and the loss function uses Smooth L1 Loss, which is expressed as:

[0031] Among them, y i 、 Represent the true value and predicted value of the i-th sample respectively, and i represents the index of the number of samples in the same batch.

[0032] Optionally, the optimization process in S3 includes predicting the resistance to draw of multiple cigarettes:

[0033] Based on current status and historical data X k ∈R B×T×24 (The input data at time k includes B samples, each sample has T time steps, and the 24-dimensional features selected by S1). The predicted resistance to draw of the next N steps of cigarettes is used as the prediction input of MPC and is expressed as:

[0034] in, for represents the predicted sequence of cigarette resistance values ​​from step k+1 to step k+N output by TCN-BiLSTM, f TCN-BiLST Represents the TCN-BiLSTM hybrid prediction model function.

[0035] Optionally, the optimization process in S3 includes constructing an MPC optimization problem:

[0036] In the control time domain M, the cigarette weight adjustment sequence is optimized to minimize the objective function, which minimizes the prediction error and the objective function of the control variable change, and is expressed as:

[0037]

[0038] Where ΔW k:k+M-1 =(ΔW k ,ΔW K+1 ,…,ΔWK+M-1 ) represents the cigarette weight adjustment sequence, ΔW k , ΔW k+1 , ΔW k+M-1 Respectively represent the cigarette weight adjustment sequence at time k, k+1, and k+M-1, 1≤i≤N, 1≤j≤M, M<N, i and j represent the i-th and j-th moments, respectively, Y target Indicates the target value of cigarette draw resistance, represents the cigarette resistance value at time k+i, γ∈(0,1] is the discount factor of the prediction error, and γ i-1 The further back the prediction step is, the error weight decays in proportion to γ. It is expressed as the square value of the cigarette weight adjustment at time k+j-1, and λ is the balance factor, which controls the weight of "smoothness" and "error";

[0039] The constraints are expressed as:

[0040]

[0041] Abnormal processing conditions include the cigarette draw resistance exceeding limit alarm and the cigarette weight setting exceeding limit cutoff, which are expressed as:

[0042]

[0043] W set =clip(W set ,W min ,W max );

[0044] Among them, W set,k+j W represents the cigarette weight setting value at the k+jth moment. set,k+j-1 Indicates the cigarette weight setting value at the k+j-1th moment, ΔW k+j-1 W is the weight adjustment of the cigarette at time k+j-1; set Indicates weight setting value; W min W is the lower limit of the safe range of cigarette weight; max ΔW is the upper limit of the safe range of cigarette weight; max For single-step cigarette weight adjustment limit, Y min Y is the lower limit of the standard process for cigarette draw resistance, max is the upper limit of the standard process for cigarette draw resistance, Y represents the actual detected cigarette draw resistance value, and clip represents the cutoff function.

[0045] Optionally, the optimization process in S3 includes rolling optimization and execution:

[0046] Solve the MPC optimization problem and get the optimal sequence in represents the optimal cigarette weight adjustment at the corresponding time i. MPC adopts the "optimize multiple steps, execute one step" strategy. Although it calculates the adjustment sequence of the next M steps, it only executes the first step to avoid control deviation caused by long-term prediction uncertainty.

[0047] The optimized solution is verified by constraints, including: weight setting range constraint; single-step adjustment limit; if all constraints are met, the cigarette machine only performs the first step adjustment and updates the weight setting value as follows: If the optimized solution does not meet any constraint conditions (e.g., the weight setting exceeds the limit or the adjustment amount is too large), the exception handling mechanism is triggered (e.g., execution of over-limit truncation or alarm), and the optimization process is re-entered;

[0048] Update the system status, slide the prediction window, and repeat the optimization process.

[0049] Optionally, in each update step of the optimization process in S3, the current true state s is estimated by Kalman filtering k =[W real,k ,Y real,k ] T , providing accurate initial conditions for the prediction model and updating X k Value, where W real,k represents the actual value of the cigarette weight setting at time k; Y real,k Represents the true value of the cigarette draw resistance at time k.

[0050] After adopting the above technical solution, this application has the following beneficial effects compared with the prior art:

[0051] This application realizes data closed-loop feedback by obtaining cigarette machine operating parameters and QTM physical measurement data; this application constructs a TCN-BiLSTM hybrid prediction model by combining TCN and BiLSTM, establishes a nonlinear system model of the cigarette machine based on historical data, and outputs the predicted value of cigarette draw resistance in real time; this application uses the MPC optimizer to integrate the prediction model and real-time data stream to realize millisecond-level rolling optimization of cigarette weight, output the optimal weight control value, and effectively eliminate dependence on manual experience.

[0052] The specific implementation methods of the present application are further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings are part of this application and are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application but do not constitute an undue limitation of this application. Obviously, the drawings described below are only some embodiments. For those of ordinary skill in the art, other drawings can be derived from these drawings without inventive effort.

[0054] In the attached figure:

[0055] Figure 1 Schematic diagram of the flow of the method for intelligently controlling cigarette weight of a cigarette making machine in this specific embodiment;

[0056] Figure 2 This is a system structure diagram of the method for intelligently controlling cigarette weight in a cigarette-making machine according to this specific embodiment;

[0057] Figure 3 This is a schematic diagram of the system structure of the TCN-BiLSTM hybrid prediction model in this specific implementation;

[0058] Figure 4 Schematic diagram of the TCN network architecture in this specific implementation;

[0059] Figure 5 This is a framework diagram of the intelligent control system for cigarette weight based on MPC in this specific embodiment;

[0060] Figure 6 This is a schematic diagram of the results of M01 intelligent control in this specific implementation;

[0061] Figure 7 This is a schematic diagram of the results of manual control of M01 in this specific implementation method;

[0062] Figure 8 This is a schematic diagram of the results of M02 intelligent control in this specific implementation method;

[0063] Figure 9 This is a schematic diagram of the results of manual control of M02 in this specific implementation method;

[0064] Figure 10 This is a schematic diagram of the results of M03 intelligent control in this specific implementation method;

[0065] Figure 11 This is a schematic diagram of the results of manual control of M03 in this specific implementation method;

[0066] Figure 12 This is a schematic diagram of the results of M04 intelligent control in this specific implementation;

[0067] Figure 13 This is a schematic diagram of the results of manual control of M04 in this specific implementation method;

[0068] Figure 14 This is a schematic diagram of the results of M05 intelligent control in this specific implementation method;

[0069] Figure 15 This is a schematic diagram of the results of manual control of M05 in this specific implementation method;

[0070] Figure 16 This is a schematic diagram of the weight setting of A01 intelligent cigarette control in this specific implementation;

[0071] Figure 17 This is a schematic diagram of the weight setting of A02 intelligent cigarette control in this specific implementation;

[0072] Figure 18 This is a schematic diagram of the weight setting of A03 intelligent cigarette control in this specific implementation;

[0073] Figure 19 This is a schematic diagram of the weight setting of A04 intelligent cigarette control in this specific implementation;

[0074] Figure 20 This is a schematic diagram of the weight setting of A05 intelligent cigarette control in this specific implementation;

[0075] Figure 21 This is a schematic diagram of the weight setting of the A06 intelligent cigarette control in this specific embodiment;

[0076] Figure 22 This is a schematic diagram of the weight setting of A07 intelligent cigarette control in this specific implementation. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. The following embodiments are used to illustrate the present application but are not used to limit the scope of the present application.

[0078] See Figure 1-Figure 5 The present application provides a method for intelligently controlling the weight of cigarettes in a cigarette-making machine, comprising the following steps:

[0079] S1. Obtain the production parameters of the cigarette making machine and perform data preprocessing and feature selection:

[0080] Cigarette machine operating parameters and QTM physical measurement data were obtained, and the importance metric of the gradient boosting decision tree was used to select features with high correlation with cigarette draw resistance.

[0081] S2. Build a hybrid prediction model based on TCN-BiLSTM:

[0082] The features selected by S1 are used as input to the prediction model. TCN-BiLSTM adopts a multimodal time series fusion architecture to extract local features and global time series features through parallel processing paths, ultimately fusing and outputting the predicted value of cigarette draw resistance. TCN uses a multi-layer residual causal dilation convolution structure to extract local features, while BiLSTM uses a two-layer bidirectional structure to capture global time series features for feature modeling. The local features represent local fluctuations in cigarette production parameters, while the global time series features represent long-term trends in the quality of different batches of cigarettes.

[0083] S3, based on the prediction model built in S2, builds an MPC-based closed-loop control system:

[0084] The current state and historical data are used as the training set input into the TCN-BiLSTM hybrid prediction model. The prediction results output by the hybrid prediction model are optimized in the rolling time domain through an optimization process. The optimal control parameters of cigarette weight are calculated by minimizing the objective function. The optimal control parameters are sent to the cigarette making machine, and the closed-loop error compensation of the prediction model is performed using real-time resistance to draw detection data.

[0085] It should be noted that the execution subject of the method in this embodiment is an intelligent control device for the weight of cigarettes in a cigarette making machine, which can be an electronic device, a component in an electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. For example, the mobile electronic device can be a mobile phone, a tablet computer, a laptop computer, a PDA, an in-vehicle electronic device, a wearable device, etc., and the non-mobile electronic device can be a server and a personal computer, etc., which are not specifically limited in this application. The following describes the intelligent control method for the weight of cigarettes in a cigarette making machine in this embodiment, taking the execution subject as an example of a server.

[0086] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of this application, the meaning of "plurality" is two or more, unless otherwise clearly specified.

[0087] See Figure 1 and Figure 2Cigarette-making machine operating parameters and QTM physical measurement data are uploaded to the Industrial Internet of Things (IIoTs) platform via the communication platform's OPC protocol. The IIoT platform transmits this real-time production data to the intelligent controller, which uses a data rectifier module to preprocess and feature-engineer the data. It then constructs a TCN-BiLSTM hybrid prediction model to predict cigarette draw resistance. This hybrid prediction model serves as a predictor for a closed-loop control system based on Model Predictive Control (MPC). This system performs rolling optimization to control cigarette weight and outputs the optimal target weight to the supervisory control system. Upon receiving the target weight, the supervisory control system verifies and monitors it, uploads it to the IIoT platform, and then distributes it to the corresponding cigarette-making machine, thereby controlling cigarette draw resistance based on cigarette weight.

[0088] Specifically, the cigarette-making machine collects data on individual cigarettes and receives cigarette weight values ​​from the supervisory control system to the IoT platform. The QTM physical measurement system collects QTM physical measurement data. The communication platform is responsible for data collection and management, ensuring data accuracy and availability. Specifically, it collects and uploads physical measurement data from the cigarette-making machine and QTM via the OPC protocol to the IoT platform. The supervisory control system monitors and evaluates target weight values ​​issued by the intelligent controller, providing outlier warnings. This enables multi-level evaluation and supervision of cigarette weight settings, ensuring stable and efficient production operations. The IoT platform visualizes physical measurement data from the cigarette-making machine and QTM, connects to and monitors supervisory equipment, ensures compliance with relevant regulations and standards, and records and displays supervisory control system operations for feedback to production departments. The intelligent controller then uploads physical measurement data from the cigarette-making machine and QTM to the intelligent controller in real time. The intelligent controller stores a TCN-BiLSTM hybrid prediction model to accurately predict draw resistance. The MPC-based closed-loop control system outputs the optimal weight setting through rolling optimization.

[0089] In practical applications, data preprocessing and feature selection are required before inputting data into the TCN-BiLSTM hybrid prediction model. Feature selection primarily considers two aspects: feature divergence: If a feature does not diverge, meaning its variance is 0, indicating that the samples have essentially no difference in this feature, then this feature is useless for distinguishing samples and should be eliminated; and correlation between each feature and the target cigarette draw resistance: Features with a high correlation with the target cigarette draw resistance are selected, using the Gradient Boosting Decision Tree (GBDT) importance metric for feature selection. The specific hyperparameter settings for GBDT are shown in Table 1, and the final optimized input data is shown in Table 2.

[0090] Table 1 GBDT hyperparameter values:

[0091]

[0092] Table 2 Model input features after feature importance selection

[0093]

[0094] As a specific implementation, production data, after data preprocessing and characterization, is used to train a TCN-BiLSTM hybrid prediction model, which outputs predicted values ​​for cigarette draw resistance. Its input feature dimension is 24 (excluding the current value of cigarette draw resistance), as shown in Table 2. As the predictor of the predictive controller, it provides the intelligent controller with accurate estimates of future system performance and provides critical cigarette draw resistance prediction information for optimizing cigarette weight.

[0095] It should be noted that the key algorithms in the TCN-BiLSTM hybrid prediction model are TCN and BiLSTM. Both are widely used in time series, weather, and traffic forecasting, but their combined architecture has not been applied in the tobacco industry. To address the large volume of tobacco data and the high computational resource consumption of traditional methods, TCN's weight-shared convolution reduces computational overhead. Combined with BiLSTM's ability to capture forward and backward information in the sequence, this model efficiently processes large amounts of data while ensuring the modeling of complex temporal relationships. Regarding feature modeling, TCN utilizes causal convolution to extract local features, such as local fluctuations in cigarette production parameters, while BiLSTM bidirectionally captures long-term dependencies, such as the long-term quality trends of different batches of cigarettes. This combination of the two enables customized modeling of complex "local + long-term" feature patterns in tobacco data, with superior representational capabilities compared to a single model alone. Furthermore, the TCN convolution kernel can learn local patterns, while the BiLSTM can adapt to dynamic changes in the sequence. This makes the model more adaptable to dynamic data characteristics caused by factors such as time and equipment status in tobacco production, such as fluctuations in production parameters. It can accurately capture and model these features in a timely manner, improving tobacco data processing efficiency and modeling accuracy.

[0096] See Figure 3 and Figure 4 In practical applications, the TCN-BiLSTM hybrid prediction model is a multimodal time series fusion architecture designed specifically for the regression prediction task of cigarette draw resistance. The TCN-BiLSTM hybrid prediction model takes 24-dimensional time series data as input, extracts local features and global time series dependencies through parallel processing paths, and finally fuses them to output the prediction results. The input data is first fed simultaneously into the improved TCN and bidirectional BiLSTM. The TCN network uses a four-layer residual causal dilation convolutional structure, with the output of the previous layer serving as the input to the next layer. The output of each layer is represented as:

[0097] h (l) =ReLU(BN(CausalConv(x (l-1) ,dilation=2 l-1 )))+x (l-1) (4)

[0098] Among them, h (l) ReLU represents the hidden layer output after the calculation of the lth layer, ReLU represents the activation function, BN represents batch normalization, CausalConv represents causal expansion convolution, dilation represents the expansion rate, 2 l-1 is the specific expansion rate value, x (l-1) Indicates that the input data of the l-1 layer of the TCN network is used as the basic input for the current layer calculation;

[0099] By exponentially increasing the expansion coefficient (1, 2, 4, 8) to expand the receptive field, the final TCN output is expressed as:

[0100] z TCN = ReLU(BN(CausalConv(tcn3))) e R B×T×64 (5)

[0101] where tcn3 is the output of the 3rd residual block of the TCN network, R represents the shape description of the multi-dimensional real tensor, B represents the batch size, and T represents the time step.

[0102] Please continue to see Figure 3 , the architecture of the BiLSTM network adopts a double-layer bidirectional structure, which respectively calculates the forward LSTM and the backward LSTM, denoted as:

[0103]

[0104] Concatenate the output of the last time step to capture the bidirectional time series dynamics, denoted as:

[0105]

[0106] where σ represents the sigmoid activation function, respectively represent the hidden states of the forward and backward LSTMs at time step t, W xh represents the weight matrix of the input to the hidden layer; x t represents the input vector at time step t; respectively represent the weight matrices of the forward and backward hidden layers to the hidden layer; respectively represent the hidden states of the forward and backward LSTMs at time steps t-1 and t+1; represents the hidden state sequence of the forward and backward LSTMs from time step 1 to T.

[0107] Please continue to see Figure 3 , the features output by the TCN and BiLSTM networks are concatenated, denoted as:

[0108] z concat = [FC T (z TCN W proj ); FC L (z LSTM )] e R B×160 (9)

[0109] The final output sequence is obtained after the time series decoder, denoted as:

[0110]

[0111] where B represents the batch size, N represents the prediction target dimension, and Wproj is the projection matrix, FC T 、FC L Represents the fully connected operations acting on TCN and BiLSTM respectively, W out Represents the output layer weight matrix, W1∈R 160→128 ,W2∈R 128→N Represents the weight matrix, Conv1D represents the one-dimensional convolution operation, and ELU represents the activation function to alleviate gradient diffusion.

[0112] Please continue to see Figure 3 and Figure 4 , the TCN-BiLSTM hybrid prediction model adopts layered Dropout, TCN is 0.2, LSTM is 0.3, FC is 0.5 and L2 regular constraint weight, and the loss function adopts Smooth L1 Loss, which is expressed as:

[0113]

[0114] Among them, y i 、 Represent the true value and predicted value of the i-th sample respectively, and i represents the index of the number of samples in the same batch.

[0115] The regression accuracy is improved by integrating local pattern perception and global time series modeling. The definitions of other hyperparameters are shown in Table 3, and the changes in the input data dimensions of each network are shown in Table 4.

[0116] Table 3 Hyperparameter settings

[0117] Parameter name Parameter value Optimizer AdamW Learning rate 1e-3 Weight decay 1e-4 Batch size 64 Maximum number of rounds 200 Loss Function SmoothL1Loss

[0118] Table 4 Changes in neural network input dimensions

[0119]

[0120] See Figure 1 、 Figure 3 and Figure 5To precisely control cigarette weight, this embodiment proposes a closed-loop control system based on Model Predictive Control (MPC). The system uses the MPC controller as the central node and integrates a TCN-BiLSTM hybrid prediction model to perform multi-step predictions on the draw resistance parameters of cigarettes during cigarette making machine production. Historical data is used as the model input, and the prediction results are optimized through the MPC training set. The MPC optimizer performs rolling time domain optimization and calculates the optimal control parameters by minimizing the objective function. The decision node uses "yes / no" logic to determine whether the current optimized solution meets the constraints, forming a closed-loop feedback mechanism that ultimately outputs the adjusted control instructions to the cigarette making machine actuator. The entire system achieves dynamic optimization control of the draw resistance parameters during cigarette production through an iterative cycle of "prediction-optimization-correction," reflecting the deep integration of data-driven and model-driven approaches, ensuring that cigarette weight and draw resistance values ​​remain stable within the process standard range.

[0121] See Figure 1 and Figure 5 The optimization process in S3 includes predicting the resistance to smoking of multi-step cigarettes:

[0122] Input data X selected based on S1 k ∈R B×T×24 The input data refers to the current state and the historical state. The input data contains B samples and T time steps. The predicted resistance value of the cigarette for the next N steps is used as the prediction input of MPC and is expressed as:

[0123]

[0124] in, for represents the predicted sequence of cigarette draw resistance values ​​from step k+1 to step k+N output by TCN-BiLSTM, f TCN-BiLSTM Represents the TCN-BiLSTM hybrid prediction model function.

[0125] See Figure 1 and Figure 5 , the optimization process in S3 includes constructing the MPC optimization problem:

[0126] In the control time domain M, the cigarette weight adjustment sequence is optimized to minimize the objective function, which minimizes the prediction error and the objective function of the control variable change, and is expressed as:

[0127] Where ΔW k:k+M-1 =(ΔW k ,ΔW k+1 ,…,ΔW k+M-1 ) represents the cigarette weight adjustment sequence, ΔW k , ΔWk+1 , ΔW k+M-1 Represents the cigarette weight adjustment sequence at time k, k+1, and k+M-1, 1≤i≤N, 1≤j≤M, M<N, i and j represent the i-th and j-th moments respectively, Y target Indicates the target value of cigarette draw resistance, represents the cigarette resistance value at time k+i, γ∈(0,1] is the discount factor of the prediction error, γ i-1 The further back the prediction step is, the error weight decays in proportion to γ. It is expressed as the square value of the cigarette weight adjustment at time k+j-1, and λ is the balance factor, which controls the weight of "smoothness" and "error";

[0128] The constraints are expressed as:

[0129]

[0130] Abnormal processing conditions include the cigarette draw resistance exceeding limit alarm and the cigarette weight setting exceeding limit cutoff, which are expressed as:

[0131]

[0132] W set =clip(W set ,W min ,W max );

[0133] Among them, W set,k+j W represents the cigarette weight setting value at the k+jth moment. set,k+j-1 Indicates the cigarette weight setting value at the k+j-1th moment, ΔW k+j-1 W is the weight adjustment of the cigarette at time k+j-1; set Indicates weight setting value; W min W is the lower limit of the safe range of cigarette weight; max ΔW is the upper limit of the safe range of cigarette weight; max For single-step cigarette weight adjustment limit, Y min Y is the lower limit of the standard process for cigarette draw resistance, max is the upper limit of the standard process for cigarette draw resistance, Y represents the actual detected cigarette draw resistance value, and clip represents the cutoff function.

[0134] Please continue to see Figure 1 and Figure 5 The optimization process in S3 includes rolling optimization and execution:

[0135] Solve the MPC optimization problem and get the optimal sequence in represents the optimal cigarette weight adjustment at the corresponding time i. MPC adopts the "optimize multiple steps, execute one step" strategy. Although it calculates the adjustment sequence of the next M steps, it only executes the first step to avoid control deviation caused by long-term prediction uncertainty.

[0136] The optimized solution is verified by constraints, including: weight setting range constraint; single-step adjustment limit; if all constraints are met, the cigarette machine only performs the first step adjustment and updates the weight setting value as follows: If the optimized solution does not meet any constraint condition (such as the weight setting exceeds the limit or the adjustment amount is too large), the exception handling mechanism is triggered (such as executing over-limit truncation or alarm), and the optimization process is re-entered; the system status is updated, the prediction window is slid, and the optimization process is repeated.

[0137] Please continue to see Figure 1 and Figure 5 In each update step of the optimization process in S3, the current true state s is estimated by Kalman filtering k =[W real,k ,Y real,k ] T , providing accurate initial conditions for the prediction model and updating X k Value, where W real,k represents the actual value of the cigarette weight setting at time k; Y real,k Represents the true value of the cigarette draw resistance at time k.

[0138] The experimental cigarette making machine is an E09 cigarette making machine from a Zhejiang cigarette factory. It uses intelligent control in the morning and manual control in the afternoon. Because the machine needs to warm up in the morning, there may be large fluctuations at the beginning, requiring more adjustments, which can better test the effectiveness of the control. At the same time, considering factors such as cutter position and circumferential fluctuation, a narrow range control is adopted, with 106 as the center and a fluctuation of 3mmH2O. The number of cigarettes outside this range is counted. Figure 6 It can be seen that in the first two days, the intelligent control system was basically close to manual control, but in the last three days, manual control caused the number of cigarettes whose suction resistance was not in the optimal range to increase significantly, by about 1 million more.

[0139] Figure 6-Figure 15 The control results of M01-M05 intelligent control and manual control are shown. Overall, manual control tends not to make adjustments, and even if adjustments are made, they are very delayed. However, the intelligent control system will perform timely control when an abnormality occurs or even when an upward or downward trend is predicted.

[0140] Since the cigarette machine's cigarette draw resistance detection cap has an attenuation problem, C2 calibration was performed during this process, and the measurement accuracy of the cigarette draw resistance was verified and corrected in combination with the QTM test results. Then, in order to verify the effectiveness of the intelligent control system, a full-day intelligent control experiment was conducted. It can be seen from Table 5 that the average SD deviation of the draw resistance per minute is less than 1, and the overall cigarette draw resistance can achieve a control effect centered at 106. After correction, the measured value of the cigarette machine's draw resistance can reflect the control of the cigarette draw resistance throughout the entire process. The local target weight is dynamically adjusted to mostly below 630 mg, which can significantly reduce the weight of short cigarettes and tobacco consumption, with an average daily reduction of 20.1 kg. The specific control situation is as follows. Figure 16-Figure 22 shown.

[0141] Table 5: All-day intelligent control results statistics

[0142] Test Number Suction resistance per minute SD Corrected average suction resistance (mmH2O) Reduce cigarette weight (g) A01 0.93 106.0 11998 A02 0.97 105.4 15174 A03 0.77 106.3 18675 A04 0.76 106.4 22420 A05 0.99 106.2 21534 A06 0.88 106.1 30717

[0143] The above is only a preferred embodiment of the present application and does not constitute any form of limitation to the present application. Although the present application has been disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technician familiar with the present application can make some changes or modifications to equivalent embodiments with equivalent changes using the technical content suggested above without departing from the scope of the technical solution of the present application. The implementation schemes in the above embodiments can also be further combined or replaced. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present application that do not depart from the content of the technical solution of the present application still fall within the scope of the solution of the present application.

Claims

1. An intelligent method for controlling cigarette weight in a cigarette making machine, characterized in that: The following steps are involved: S1. Obtain the production parameters of the cigarette making machine and perform data preprocessing and feature selection: Cigarette machine operating parameters and QTM physical measurement data were obtained, and the importance metric of the gradient boosting decision tree was used to select features with high correlation with cigarette draw resistance. S2. Build a hybrid prediction model based on TCN-BiLSTM: The features selected by S1 are used as input to the prediction model. TCN-BiLSTM adopts a multimodal time series fusion architecture to extract local features and global time series features through parallel processing paths, ultimately fusing and outputting the predicted value of cigarette draw resistance. TCN uses a multi-layer residual causal dilation convolution structure to extract local features, while BiLSTM uses a two-layer bidirectional structure to capture global time series features for feature modeling. The local features represent local fluctuations in cigarette production parameters, while the global time series features represent long-term trends in the quality of different batches of cigarettes. S3, based on the prediction model built in S2, builds an MPC-based closed-loop control system: The current state and historical data are used as the training set input into the TCN-BiLSTM hybrid prediction model. The prediction results output by the hybrid prediction model are optimized in the rolling time domain through an optimization process. The optimal control parameters of cigarette weight are calculated by minimizing the objective function. The optimal control parameters are sent to the cigarette making machine, and the closed-loop error compensation of the prediction model is performed using real-time resistance to draw detection data.

2. The method according to claim 1, characterized in that The TCN-BiLSTM hybrid prediction model built in S2 includes: The TCN network adopts a 4-layer residual causal expansion convolution structure. The output of the previous layer is used as the input of the next layer. The output of each layer is expressed as: h (l) =ReLU(BN(CausalConv(x (l-1) ,dilation=2 l-1 )))+x (l-1) (4) Among them, h (l) ReLU represents the hidden layer output after the calculation of the lth layer, ReLU represents the activation function, BN represents batch normalization, CausalConv represents causal expansion convolution, dilation represents the expansion rate, 2 l-1 is the specific expansion rate value, x (l-1) Indicates that the input data of the l-1 layer of the TCN network is used as the basic input for the current layer calculation; By exponentially increasing the expansion coefficient (1, 2, 4, 8) to expand the receptive field, the final TCN output is expressed as: z TCN =ReLU(BN(CausalConv(tcn3)))∈R B×T×64 (5) Among them, tcn3 is the output of the residual block of the third layer of the TCN network, R represents the shape description of the multidimensional real tensor, B represents the batch size, and T represents the time step.

3. The method according to claim 2, characterized in that The TCN-BiLSTM hybrid prediction model built in S2 includes: The architecture of the BiLSTM network adopts a two-layer bidirectional structure, which calculates the forward LSTM and backward LSTM respectively, expressed as: Concatenating the last time step output captures the bidirectional temporal dynamics, expressed as: Among them, σ represents the sigmoid activation function, Represent the hidden states of the forward and backward LSTM at time step t, W xh represents the weight matrix input to the hidden layer; x t represents the input vector at time step t; Represents the weight matrices from the forward and backward hidden layers to the hidden layers respectively; Represent the hidden states of the forward and backward LSTM at time steps t-1 and t+1 respectively; represents the hidden state sequence of the forward and backward LSTM from time step 1 to T.

4. The method according to claim 3, characterized in that The TCN-BiLSTM hybrid prediction model built in S2 includes: The features output by the TCN and BiLSTM networks are concatenated and expressed as: from concat =[FC T (from TCN W proj );FC L (from LSTM )]∈R B×160 (9) After passing through the timing decoder, the final output sequence is obtained, which is expressed as: Among them, B represents the batch size, N represents the prediction target dimension, and W proj is the projection matrix, FC T 、FC L Represents the fully connected operations acting on TCN and BiLSTM respectively, W out Represents the output layer weight matrix, W1∈R 160→128 ,W2∈R 128→N Represents the weight matrix, Conv1D represents the one-dimensional convolution operation, and ELU represents the activation function to alleviate gradient diffusion.

5. The method according to claim 2, characterized in that The TCN-BiLSTM hybrid prediction model built in S2 includes: The TCN-BiLSTM hybrid prediction model uses layered Dropout and L2 regularization constraint weights, and the loss function uses SmoothL1 Loss, which is expressed as: Among them, y i 、 Represent the true value and predicted value of the i-th sample respectively, and i represents the index of the number of samples in the same batch.

6. The method according to claim 3, characterized in that The optimization process in S3 includes predicting the resistance to draw of multiple cigarettes: Input data X selected based on S1 k ∈R B×T×24 , predict the future N steps of cigarette resistance, as the prediction input of MPC, expressed as: in, for represents the predicted sequence of cigarette draw resistance values ​​from step k+1 to step k+N output by TCN-BiLSTM, f TCN-BiLST Represents the TCN-BiLSTM hybrid prediction model function.

7. The method according to claim 6, characterized in that The optimization process in S3 involves constructing an MPC optimization problem: In the control time domain M, the cigarette weight adjustment sequence is optimized to minimize the objective function, which minimizes the prediction error and the objective function of the control variable change, and is expressed as: Where ΔW k:k+M-1 =(ΔW k ,ΔW k+1 ,…,ΔW k+M-1 ) represents the cigarette weight adjustment sequence, ΔW k , ΔW k+1 , ΔW k+M-1 Respectively represent the cigarette weight adjustment sequence at time k, k+1, and k+M-1, 1≤i≤N, 1≤j≤M, M<N, i and j represent the i-th and j-th moments, respectively, Y target Indicates the target value of cigarette draw resistance, represents the cigarette resistance value at time k+i, γ∈(0,1] is the discount factor of the prediction error, and γ i-1 The further back the prediction step is, the error weight decays in proportion to γ. It is expressed as the square value of the cigarette weight adjustment at time k+j-1, and λ is the balance factor, which controls the weight of "smoothness" and "error". The constraints are expressed as: Abnormal processing conditions include the cigarette draw resistance exceeding limit alarm and the cigarette weight setting exceeding limit cutoff, which are expressed as: W set =clip(W set ,W min ,W max ); Among them, W set,k+j W represents the cigarette weight setting value at the k+jth moment. set,k+j-1 Indicates the cigarette weight setting value at the k+j-1th moment, ΔW k+j-1 W is the weight adjustment of the cigarette at time k+j-1; set Indicates weight setting value; W min W is the lower limit of the safe range of cigarette weight; max ΔW is the upper limit of the safe range of cigarette weight; max For single-step cigarette weight adjustment limit, Y min Y is the lower limit of the standard process for cigarette draw resistance, max is the upper limit of the standard process for cigarette draw resistance, Y represents the actual detected cigarette draw resistance value, and clip represents the cutoff function.

8. The method according to claim 7, characterized in that The optimization process in S3 includes rolling optimization and execution: solving and constructing MPC optimization problems to obtain the optimal sequence in represents the optimal cigarette weight adjustment at the corresponding time i. MPC adopts an "optimize multiple steps, execute one step" strategy. Although it calculates the adjustment sequence for the next M steps, it only executes the first step to avoid control bias caused by long-term prediction uncertainty. The constraints of the optimized solution are verified, including: the constraints include the weight setting range constraint and the single-step adjustment limit constraint; if all constraints are met, the cigarette machine only performs the first step adjustment and updates the weight setting value as follows: If the optimized solution does not satisfy any constraint, the exception handling mechanism is triggered and the optimization process is re-entered; Update the system status, slide the prediction window, and repeat the optimization process.

9. The method according to claim 8, characterized in that In each update step of the optimization process in S3, the current true state s is estimated by Kalman filtering k =[W real,k ,Y real,k ] T , providing accurate initial conditions for the prediction model and updating X k Value, where W real,k represents the actual value of the cigarette weight setting at time k; Y real,k Represents the true value of the cigarette draw resistance at time k.

Images