An adaptive predictive control method, system, and storage medium for tobacco leaf drying moisture content.

Abstract

This invention discloses an adaptive predictive control method, system, and storage medium for tobacco leaf drying moisture content. The method includes: acquiring historical tobacco leaf drying process data and denoising it; standardizing the denoised historical tobacco leaf drying process data and constructing a sample dataset using a multi-scale segmentation strategy based on a sliding window mechanism; constructing a differential recursive neural network prediction model architecture; training the differential recursive neural network prediction model using the sample dataset; collecting tobacco leaf drying process data online, standardizing it, constructing input features, and inputting them into the trained differential recursive neural network prediction model to output the predicted vectors for the instantaneous moisture content of the outlet material, the hot air valve opening, and the exhaust damper opening for the future prediction step; and generating and issuing tobacco leaf drying control commands based on the model prediction results. This invention can solve the problem of unstable tobacco leaf drying quality in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of tobacco processing technology, and in particular to an adaptive predictive control method, system and storage medium for moisture content in tobacco leaves during drying. Background Technology

[0002] In tobacco processing lines, the drying process is a core thermal process that determines the physical properties (such as filling value and whole tobacco yield) and sensory quality of cigarettes. The main technological objective of this process is to precisely reduce the moisture content of the tobacco shreds from a relatively high level at the inlet (usually 20%–22%) to the required outlet level (usually 12%–13%) while ensuring the tobacco processing temperature, and to maintain minimal fluctuations in outlet moisture content (the smaller the standard deviation SD value, the better).

[0003] Current yarn drying equipment (such as airflow yarn dryers or drum yarn dryers) is typically equipped with a basic process control system. The most common and well-known control technology is PID feedback control, and its basic implementation scheme is as follows:

[0004] (1) Data acquisition: The system collects the instantaneous moisture content of the material at the outlet in real time through an infrared moisture meter installed at the outlet. ;

[0005] (2) Deviation calculation: Calculate the difference between the actual measured value and the target value set by the process. Deviation between ;

[0006] (3) Feedback adjustment: The PID controller adjusts according to the deviation. Proportional, integral, and differential terms are used to calculate control quantities. In the yarn drying machine, This typically corresponds to the hot air temperature setpoint or the opening of the exhaust damper. The actuator adjusts accordingly. Adjust the hot air valve or damper to change the drying intensity.

[0007] However, the wire drying process is a typical thermodynamic system with large time lag, strong coupling, and nonlinearity. There is often a pure time lag of tens of seconds to several minutes between adjusting the hot air temperature and the output moisture content. Simple PID control mainly relies on retrospective deviations for adjustment, lacking the ability to predict future trends. When faced with fluctuations in inlet flow rate or sudden changes in incoming material moisture content, it is prone to overshoot or untimely adjustment, resulting in a large amount of material run-off and waste.

[0008] To address the issue of strong lag in traditional PID control, data-driven time series forecasting techniques have been increasingly introduced. Current mainstream long-sequence forecasting methods based on piecewise deep learning architectures include piecewise forecasting models such as Transformer and its variants (e.g., PatchTST, Autoformer). However, existing long-sequence forecasting methods based on piecewise deep learning architectures suffer from the following problems:

[0009] (1) The discrete segmentation mechanism cuts off the continuity of the physical process and cannot characterize the large hysteresis decay characteristics of thermal inertia.

[0010] Specific shortcomings analysis: To reduce computational complexity, existing technologies typically divide continuous time series into multiple independent time segments as input tokens. While this mechanism improves parallel computing efficiency, it severs the direct interaction channel between implicit states of the segments. The model treats each segment as an independent semantic unit, lacking a recursive mechanism across segments to pass on the system's historical state.

[0011] Technical challenges in tobacco drying scenarios: Tobacco drying is a thermodynamic process with significant time lag. The current outlet moisture content depends not only on the current control parameters but also on the continuous influence of the equipment's thermal state from minutes earlier or even earlier (such as the deep heat storage in the cylinder wall metal and the historical enthalpy accumulation in the hot air ducts). Furthermore, this thermal influence is not constant but exhibits physical decay over time. Existing segmented mechanisms cause the model to lose the implicit thermal potential energy and its decay patterns from the previous segment when processing the current segment. This loss of cross-segment physical context prevents the model from establishing long-distance temporal causal dependencies. Consequently, in long-period predictions, the prediction error amplifies dramatically with increasing prediction step size, making it difficult to guide feedforward control of tobacco drying machines with large time lag characteristics.

[0012] (2) Simple temporal attention mechanism has weak noise resistance and cannot separate mechanical noise from process details in frequency domain aliasing.

[0013] Specific Defect Analysis: Existing attention mechanisms primarily allocate weights based on numerical dot product operations in the time domain. However, from a time-domain waveform perspective, high-frequency random mechanical noise and effective high-frequency detail signals often exhibit similar fluctuation patterns, resulting in severe time-domain waveform aliasing, making it extremely difficult to distinguish them solely based on time-domain amplitude.

[0014] Technical challenges in the wire drying scenario: Sensor data collected on-site (especially exhaust air volume, instantaneous flow rate of electronic scales, etc.) contains a large amount of high-frequency random noise caused by high-power fan vibration and airflow turbulence. Existing technologies, when calculating attention in the time domain, are prone to misjudging large amounts of random noise as key features, thus assigning incorrect attention weights, leading to model overfitting and unnecessary, severe jitter in the output control commands. Simultaneously, high-frequency details that truly reflect minute changes in moisture (such as step signals in the feed head moisture content) are often submerged by strong noise, making them impossible for the model to extract effectively, resulting in insufficient sensitivity and robustness of the control system to changes in real-world operating conditions.

[0015] (3) Single time scale modeling leads to insufficient extraction of multivariate dynamic coupling features.

[0016] Specific shortcomings analysis: Existing models typically extract features from all input variables on a single time scale, forcing all variables to adapt to the same time receptive field, ignoring the essential differences in response frequency and rate of change of different physical quantities.

[0017] Technical challenges in the wire drying scenario: The wire drying process involves nonlinear coupling of multiple variables, and the dynamic characteristics of each variable vary greatly, making it a typical multi-timescale system. For example, fast variables (such as the opening of the exhaust damper and steam pressure) fluctuate drastically on the order of seconds, while slow variables (such as cylinder wall temperature and ambient temperature and humidity) drift slowly on the order of minutes. Existing single-scale techniques struggle to simultaneously capture the instantaneous response of fast variables and the long-term trend of slow variables, resulting in models that cannot accurately characterize this multi-frequency heterogeneous coupling relationship, thus limiting the prediction accuracy of outlet moisture under complex operating conditions.

[0018] (4) Static models lack online adaptive capabilities and cannot cope with the drift in operating conditions caused by equipment aging and raw material fluctuations.

[0019] Specific shortcomings: Existing technologies mostly adopt a mode of fixing parameters after offline training, lacking an online parameter correction mechanism during operation. The model cannot perceive changes in environmental distribution.

[0020] Technical challenges in tobacco drying scenarios: Significant conceptual drift exists in actual production. On the one hand, as equipment operating time increases, heat exchange efficiency changes due to pipe fouling and mechanical aging; on the other hand, different batches of tobacco raw materials (origin, grade, moisture content) also have different characteristics. Existing static models cannot perceive and adapt to these slow changes in physical characteristics, causing the predictive performance to gradually degrade over time after the model has been online for a period of time, resulting in model failure. This fails to meet the precise control requirements of long-term continuous production, forcing factories to frequently perform manual intervention or retraining.

[0021] The aforementioned issues result in poor accuracy of long-sequence prediction methods based on segmented deep learning architecture when applied to adaptive prediction and control of tobacco leaf drying moisture, thus affecting the stability of the final drying quality. Summary of the Invention

[0022] This invention provides an adaptive prediction and control method, system, and storage medium for tobacco leaf drying moisture, in order to solve the problem that the existing technology has poor accuracy in predicting the moisture content at the outlet of the drying process, which leads to unstable drying quality.

[0023] Firstly, an adaptive predictive control method for moisture content in tobacco leaf drying is provided, comprising the following steps:

[0024] S1: Obtain historical tobacco drying process data and remove non-steady-state noise;

[0025] S2: Standardize the denoised historical tobacco drying process data and construct a sample dataset using a multi-scale segmentation strategy based on a sliding window mechanism. Each sample includes input features consisting of tobacco drying process data sequences at different time scales and a label vector for the prediction step size. The label vector includes the instantaneous moisture vector of the outlet material, the hot air valve opening vector, and the dehumidification damper opening vector.

[0026] S3: Construct a neural network prediction model architecture based on differential recursion, including a feature extraction module, a feature fusion module, and a multi-objective regression head connected in sequence; the feature extraction module includes multiple branches that process input features at different time scales, and each branch includes a spatiotemporal feature embedding layer and a differential recursive unit connected in sequence; the feature fusion module is used to fuse the outputs of the multiple branches of the feature extraction module.

[0027] S4: Train the differential recursive neural network prediction model using the sample dataset;

[0028] S5: Collect tobacco drying process data online and standardize it. Then extract tobacco drying process data sequences at different time scales as input features and input them into a trained differential recursive neural network prediction model. Output the instantaneous moisture content prediction vector of the outlet material, the hot air valve opening prediction vector, and the dehumidification damper opening prediction vector for the future prediction step.

[0029] S6: Generate and issue tobacco drying control instructions based on the hot air valve opening prediction vector and the dehumidification damper opening prediction vector.

[0030] Furthermore, step S1 specifically includes:

[0031] Obtain historical tobacco drying process data and construct the original feature matrix. ,in, The number of sampling points. For feature dimensions;

[0032] In high-dimensional space, the conditional probability between two sample points is calculated using the Gaussian kernel function of dynamic variance to represent the local similarity between the two sample points, and then the high-dimensional joint probability distribution is calculated.

[0033] Set the target low-dimensional space, and assume high-dimensional sample points. The corresponding mapping point in the target low-dimensional space is: In the target low-dimensional space, the Student-t distribution with 1 degree of freedom in the long tail is used to calculate the low-dimensional joint probability distribution between two low-dimensional mapping points.

[0034] Kullback-Leibler divergence is used as the loss function to solve for the optimal set of low-dimensional mapping point coordinates Y, so that the low-dimensional joint probability distribution approximates the high-dimensional joint probability distribution.

[0035] For the optimal set of low-dimensional mapping point coordinates Y, perform a clustering operation with a cluster size of 2, retaining only the data indices belonging to the steady-state clusters, and then extracting the data from the original feature matrix. Extract the corresponding original data to complete the noise reduction.

[0036] Furthermore, step S2 specifically includes:

[0037] For the denoised historical tobacco drying process data, Z-Score transformation is performed on each feature dimension to make it conform to a standard normal distribution.

[0038] The standardized historical tobacco drying process data is segmented using a sliding window mechanism to obtain an initial sample dataset: the basic time window length is set. , The original feature step size, To predict the step size; the data feature dimensions extracted from the first L steps of the base time window are: Data corresponding to the tobacco drying process for all categories; the subsequent basic time window The data feature dimensions extracted in each step are 3, corresponding to the three feature dimensions of instantaneous moisture content of the outlet material, hot air valve opening, and exhaust damper opening, which are used to construct the label vector;

[0039] Multi-scale processing: Multi-scale feature extraction is performed on the first L steps of data of each initial sample. The feature vectors of different scales extracted together constitute the input features, thereby constructing the sample dataset.

[0040] Furthermore, the calculation logic of the differential recursive unit is as follows:

[0041] Thermal potential injection and adaptive decay: The input of the differential recursive unit at time t is represented as follows:

[0042] ;

[0043] In the formula, The input features of the differential recursive unit at time t; The hidden state vector at time t-1 is used to characterize the historical thermal potential energy; It is a learnable adaptive decay coefficient. ; Input features of the differential recursive unit that injects historical thermal potential energy at time t;

[0044] Deep feature interaction: The input is fed into a K-layer stacked nonlinear transformation subunit, each layer comprising a spectral-temporal dual-domain attention layer, a first normalization layer, a feedforward neural network, and a second normalization layer connected in sequence; the output features of the nonlinear transformation subunit are denoted as... ;

[0045] Two-stream differential separation mechanism: A differential stream separation mechanism is set up: Predictive stream: Explicit predictive features for water regression at the current time are generated using a multilayer perceptron. , The prediction flow represents the multilayer perceptron operation; the state flow involves extracting latent state features that are not currently consumed but need to be passed on to the future using the multilayer perceptron. , Represents state flow multilayer perceptron operation;

[0046] Gated memory update: Using the Sigmoid activation function as the gating unit, the hidden state for the next time step is generated. and the output at the current moment :

[0047] ;

[0048] ;

[0049] in, As a renewed historical thermal energy transfer to time.

[0050] Furthermore, the spectral-temporal dual-domain attention layer includes parallel time-domain and frequency-domain branches:

[0051] Temporal branch:

[0052] Global average pooling is performed on the tensor X of the dual-domain attention layer at the input spectrum to generate channel context descriptors and form a global context vector. This global context vector is then used to generate a temporal attention mask through a multilayer perceptron. Identify the main trend channels:

[0053] ;

[0054] In the formula, This represents the output tensor of the time-domain branch. This represents element-wise multiplication;

[0055] Frequency domain branch:

[0056] Orthogonal frequency projection: The tensor X of the dual-domain attention layer at input spectrum is projected onto the orthogonal frequency domain space using discrete cosine transform to obtain the spectral tensor. ;

[0057] Spectrum mask learning: Learning a frequency domain filter mask through a neural network. The mask automatically suppresses high-frequency diffuse components representing mechanical noise while retaining specific frequency bands representing process abrupt changes.

[0058] Inverse projection restoration: restoring the weighted clean spectrum back to the time domain, expressed as: , This represents the frequency domain branch output tensor. Represents the inverse discrete cosine transform;

[0059] Bi-domain orthogonal fusion: The output tensor of the time-domain branch is added to the output tensor of the frequency-domain branch, and the enhanced feature representation is output. .

[0060] Furthermore, in step S6, when generating the tobacco drying control command, process constraint verification is performed. If the predicted hot air valve opening or dehumidification damper opening exceeds the corresponding set process red line, it is forcibly limited to the corresponding safe range.

[0061] Furthermore, it also includes:

[0062] Drift detection is performed on the online differential recursive neural network prediction model. If a substantial drift is determined to have occurred, the online update process of the differential recursive neural network prediction model is automatically triggered.

[0063] The latest samples and the remembered samples are mixed in proportion to construct a hybrid fine-tuning training set;

[0064] Based on a hybrid fine-tuning training set, a hierarchical parameter freezing and gradient fine-tuning strategy is used to fine-tune and update the differential recursive neural network prediction model.

[0065] A shadow evaluation mechanism is adopted to compare the prediction accuracy of the updated new model with that of the current model in real time within the evaluation window; only when the accuracy of the updated new model is better than that of the current model, the parameters of the inference engine are pointed to the updated new model in milliseconds through memory pointer swapping technology, so as to achieve seamless model iteration in the production process.

[0066] Secondly, an adaptive predictive control system for tobacco leaf drying moisture is provided, including:

[0067] The noise reduction module is used to acquire historical tobacco drying process data and remove non-steady-state noise;

[0068] The sample dataset construction module is used to standardize the denoised historical tobacco drying process data and construct the sample dataset using a multi-scale segmentation strategy based on a sliding window mechanism. Each sample includes input features consisting of tobacco drying process data sequences at different time scales and a label vector for the prediction step size. The label vector includes the instantaneous moisture vector of the outlet material, the hot air valve opening vector, and the dehumidification damper opening vector.

[0069] The model building module is used to build a neural network prediction model architecture based on differential recursion, including a feature extraction module, a feature fusion module, and a multi-objective regression head connected in sequence. The feature extraction module includes multiple branches that process input features at different time scales, and each branch includes a spatiotemporal feature embedding layer and a differential recursive unit connected in sequence. The feature fusion module is used to fuse the outputs of the multiple branches of the feature extraction module.

[0070] The model training module is used to train a differential recursive neural network prediction model using a sample dataset.

[0071] The online prediction module is used to collect tobacco drying process data online, standardize it, extract tobacco drying process data sequences at different time scales as input features, and input them into a trained differential recursive neural network prediction model to output the instantaneous moisture content prediction vector of the outlet material, the hot air valve opening prediction vector, and the dehumidification damper opening prediction vector for the future prediction step.

[0072] The instruction generation and issuance module is used to generate and issue tobacco drying control instructions based on the hot air valve opening prediction vector and the dehumidification damper opening prediction vector.

[0073] Thirdly, an adaptive predictive control system for tobacco leaf drying moisture is provided, including:

[0074] A memory on which computer programs are stored;

[0075] The processor is used to execute the computer program to implement the adaptive predictive control method for tobacco leaf drying moisture as described above.

[0076] Fourthly, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the adaptive predictive control method for tobacco leaf drying moisture as described above.

[0077] This invention proposes an adaptive predictive control method, system, and storage medium for the moisture content of tobacco leaves during drying. Compared with existing technologies, it has the following advantages:

[0078] (1) Overcoming the thermal inertia discontinuity caused by discrete slicing and improving the long-period prediction accuracy of large-time-lag systems. This invention adopts a differential recursive unit architecture and introduces an adaptive thermal inertia decay mechanism in state transfer. By establishing recursive channels between discrete time segments, the thermal potential energy hidden state vector is transferred. This method seamlessly transfers the implicit thermodynamic state from the previous moment (such as deep heat storage in the cylinder wall and the hysteretic response of moisture) to the current moment. A learnable adaptive decay coefficient is introduced to simulate the natural dissipation of physical heat over time. The model can automatically adjust according to the current operating conditions. The value retains long-term memory in steady state and quickly forgets the old state when the operating conditions change abruptly, thereby accurately capturing the dynamic evolution law of large time lag systems.

[0079] (2) Frequency domain signal decoupling under strong noise conditions is achieved, ensuring the stability and robustness of control commands. This invention constructs a spectrum-time dual-domain attention mechanism and introduces adaptive soft threshold denoising of the frequency domain branch. Orthogonal frequency projection is used to map features to the frequency domain. Since random mechanical noise usually manifests as diffuse components across the entire frequency band, while effective moisture mutation signals manifest as specific spectral structures, the two are significantly separable in the frequency domain. In the frequency domain, a spectral mask is generated through a neural network, which is equivalent to an adaptive bandpass filter to accurately suppress noise bands; at the same time, global context embedding is used in the time domain to lock the macroscopic trend. This dual mechanism of "frequency domain purification + time domain fidelity" ensures that the model can still output stable and accurate control parameters under harsh noise conditions.

[0080] (3) This invention solves the problem of insufficient feature extraction caused by multivariate dynamic coupling. The invention processes feature data streams at different time scales (different granularities) in parallel. The fine-grained stream focuses on capturing the instantaneous changes of fast variables such as flow rate at the second level, while the coarse-grained stream focuses on capturing the thermal inertia curves of slow variables such as temperature at the minute level. Through cross-channel convolution fusion at the end, physical features with different time resolutions are mapped to a unified semantic state space, thereby accurately representing the complex "fast-slow" dynamic coupling relationship in the wire drying process.

[0081] (4) It possesses full lifecycle adaptability to operating conditions, overcoming the problem of model performance degradation. This invention can perceive the statistical significance changes of the prediction residuals in real time and automatically trigger incremental fine-tuning. During the update process, through a hierarchical parameter freezing strategy and a hybrid memory sample pool (mixing new and old operating condition data), the model can quickly adapt to the current new equipment status or new raw material characteristics, while avoiding the catastrophic forgetting of the general rules of the old operating conditions. This ensures that the control system maintains optimal performance throughout long-term operation, reduces reliance on manual operation and maintenance, and improves production efficiency. Attached Figure Description

[0082] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0083] Figure 1 This is a flowchart of an adaptive predictive control method for tobacco leaf drying moisture provided in an embodiment of the present invention;

[0084] Figure 2 This is a diagram of the neural network prediction model architecture based on differential recursion provided in an embodiment of the present invention;

[0085] Figure 3 This is a differential recursive unit architecture diagram provided in an embodiment of the present invention;

[0086] Figure 4 This is a schematic diagram of the spectral-temporal dual-domain attention layer provided in an embodiment of the present invention. Detailed Implementation

[0087] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0088] like Figure 1 As shown in the figure, this invention provides an adaptive predictive control method for the moisture content of tobacco leaves during drying, comprising the following steps:

[0089] S1: Obtain historical tobacco drying process data and remove non-steady-state noise.

[0090] The main task of this step is to remove non-steady-state noise from high-dimensional, non-linear raw historical process data. Traditional fixed-threshold cleaning methods can only identify outliers exceeding the measurement range, but they are difficult to identify complex operating conditions such as equipment warm-up, material shortage, and idling. This step maps the hidden non-steady-state data distribution in high-dimensional space to intuitive physical distances in low-dimensional space, and then achieves accurate isolation and cleaning of operating conditions through clustering.

[0091] In this embodiment, step S1 specifically includes:

[0092] S11: Construction of high-dimensional feature space.

[0093] Obtain historical tobacco drying process data and construct the original feature matrix. ,in, The number of sampling points. For feature dimension; the first The high-dimensional sample points are denoted as . It includes three types of key parameters:

[0094] (1) State variable: actual hot air temperature Actual cylinder wall temperature , exhaust air temperature , entrance moisture Electronic scale flow rate ;

[0095] (2) Controlled variable: hot air valve opening Dehumidification damper opening ;

[0096] (3) Target variable: instantaneous moisture content of the outlet material .

[0097] S12: High-dimensional probability distribution modeling.

[0098] In high-dimensional space, with conditional probability To represent sample points and The local similarity between samples is measured using a Gaussian kernel function of dynamic variance, as shown in the following formula:

[0099] ;

[0100] In the formula, For The variance of the Gaussian function centered at the center; to adapt to the differences in data density under different working conditions of the red silk process (such as dense steady-state data and sparse non-steady-state data), a binary search method is used for each Dynamically determined This ensures that the perplexity of all Gaussian distributions meets the preset value.

[0101] To facilitate subsequent optimization, the joint probability distribution is calculated. Construct a symmetric high-dimensional similarity matrix:

[0102] ;

[0103] In the formula, Represents a given sample point At that time, sample points As the conditional probability of its neighboring points; Represents a given sample point At that time, sample points N represents the number of sampling points, which serves as the conditional probability of its neighboring points.

[0104] S13: Low-dimensional mapping and target space optimization.

[0105] Set a target low-dimensional space (usually two-dimensional, for easy visualization and clustering), and assume high-dimensional sample points. and The corresponding mapping points (coordinates) in two-dimensional space are respectively and During initialization, and Assign a random initial value.

[0106] In the target low-dimensional space, to alleviate the data congestion problem during dimensionality reduction, a Student-t distribution with 1 degree of freedom and a long tail is used to calculate the low-dimensional joint probability distribution between two low-dimensional mapping points. :

[0107] ;

[0108] In the formula, and These represent the mapping points of the k-th and l-th sample points in the target low-dimensional space, respectively.

[0109] Using Kullback-Leibler divergence as the loss function, we solve for the optimal set of low-dimensional mapping point coordinates Y, so that the low-dimensional joint probability distribution approximates the high-dimensional joint probability distribution to the maximum extent.

[0110] S14: Cluster cleaning.

[0111] For the optimal low-dimensional (two-dimensional) set of mapping point coordinates Y, K-Means clustering is performed in the two-dimensional manifold space, with the number of cluster centers set to K=2 (corresponding to both steady-state and non-steady-state core states). During clustering, based on intra-cluster inertia and sample density, the system automatically identifies the steady-state production clusters with large sample sizes and compact distribution, as well as the non-steady-state clusters with discrete banded distributions. Only the data indices belonging to the steady-state clusters are retained, and the data is extracted from the original feature matrix. Extract the corresponding original data to generate a high-quality cleaned dataset. This completes the noise reduction process.

[0112] Equipment warm-up and idling due to material shortage generally occur at the beginning and end of the equipment's operation. Therefore, the denoising data obtained through the aforementioned data processing is generally continuous time-series data. If sample points are missing at certain points in the middle, they can be filled by interpolation (such as interpolating the mean of adjacent sample points) to obtain continuous time-series data.

[0113] S2: Standardize the denoised historical tobacco drying process data and construct a sample dataset using a multi-scale segmentation strategy based on a sliding window mechanism. Each sample includes input features consisting of tobacco drying process data sequences at different time scales and a label vector for the prediction step size. The label vector includes the instantaneous moisture vector of the outlet material, the hot air valve opening vector, and the dehumidification damper opening vector.

[0114] To address the challenge of dynamically coupling fast variables (such as flow rate) and slow variables (such as temperature) at a single time scale, this invention constructs a multi-scale data stream input architecture.

[0115] In this embodiment, step S2 specifically includes:

[0116] S21: Dynamic Z-Score Standardization.

[0117] To address the issue of significant differences in different physical dimensions in the extracted steady-state operating data (for example, temperature values ​​are typically around 130℃, while outlet moisture content is around 12%), a Z-Score transformation is performed on each feature dimension to make it conform to a standard normal distribution, thus eliminating the influence of dimensions.

[0118] S22: Multi-scale segmentation strategy:

[0119] The standardized historical tobacco drying process data is segmented using a sliding window mechanism to obtain an initial sample dataset: the basic time window length is set. , The original feature step size, To predict the step size; the data feature dimensions extracted from the first L steps of the base time window are: Data corresponding to the tobacco drying process for all categories; the subsequent basic time window The data feature dimensions extracted in each step are 3, corresponding to the three feature dimensions of instantaneous moisture content of the outlet material, hot air valve opening, and exhaust damper opening, which are used to construct the label vector;

[0120] Multi-scale processing: Multi-scale feature extraction is performed on the first L steps of data of each initial sample. The feature vectors of different scales extracted together constitute the input features, thereby constructing the sample dataset.

[0121] Taking the extraction of feature vectors from two scales to form the input features as an example, it includes fine-grained flow and coarse-grained flow:

[0122] (1) Fine-grained flow: Maintain the original sampling frequency (e.g., 1Hz), and denote the input tensor as ( (This refers to the batch size). This stream is used to capture instantaneous, second-level mutations in fast variables such as exhaust airflow.

[0123] (2) Coarse-grained flow: Perform one-dimensional average pooling (AvgPool) on the original sequence, with a pooling kernel size of (In other embodiments, other values ​​such as 2, 3, 5, and 6 can be selected as needed. For multi-scale scenarios, multiple different values ​​can be selected to construct flows of different granularities.) The input tensor is denoted as... This flow acts as a low-pass filter, used to capture minute-level thermal inertia trends of slow variables such as cylinder wall temperature.

[0124] S3: Construct a neural network prediction model architecture based on differential recursion, such as... Figure 2 As shown, it includes a feature extraction module, a feature fusion module, and a multi-objective regression head connected in sequence; the feature extraction module includes multiple branches that process input features at different time scales, and each branch includes a spatiotemporal feature embedding layer and a differential recursive unit connected in sequence; the feature fusion module is used to fuse the outputs of the multiple branches of the feature extraction module.

[0125] To address the issues of interrupting time continuity and failing to characterize thermal inertia decay in existing Transformer models, this invention designs a differential recursive unit to replace the traditional self-attention block.

[0126] In this embodiment, the spatiotemporal feature embedding layer uses a linear layer to embed the input tensor of this branch (e.g., or Mapped to a unified hidden dimension And superimpose learnable positional encodings to obtain embedded features (such as...) or ).

[0127] In this embodiment, the differential recursive unit is an improved recurrent neural network architecture, as shown in the following diagram. Figure 3 As shown, the differential recursive unit is in Receive input features of the current segment at any time and the hidden state vector of the previous time step The specific calculation logic is as follows:

[0128] (1) Thermal potential energy injection and adaptive decay:

[0129] A learnable adaptive decay coefficient is introduced. Used to simulate the natural dissipation of physical thermal inertia over time:

[0130] ;

[0131] In the formula, The input features of the differential recursive unit at time t; The hidden state vector at time t-1 is used to characterize the historical thermal potential energy; Input features of the differential recursive unit that injects historical thermal potential energy at time t;

[0132] This mechanism enables the model to automatically adjust its reliance on historical information based on different operating conditions (e.g., in steady state). Larger, maintaining inertia; during operating condition switching Smaller size, faster response time;

[0133] (2) Deep feature interaction:

[0134] Will The input is fed into a K-layer stacked nonlinear transformation subunit, each layer comprising a spectral-temporal dual-domain attention layer, a first normalization layer, a feedforward neural network, and a second normalization layer connected in sequence; the output features of the nonlinear transformation subunit are denoted as... ;

[0135] (3) Two-stream differential separation mechanism:

[0136] To mathematically enforce the separation of transient output from steady-state inertia, a differential current shunt mechanism was designed:

[0137] Predictive Stream: Generates explicit predictive features for moisture regression at the current time using a multilayer perceptron. , This represents the prediction flow multilayer perceptron operation;

[0138] State flow: Extracting latent state features that are not currently consumed but need to be passed on to the future using a multilayer perceptron. , Represents state flow multilayer perceptron operation;

[0139] This mechanism forces the model to store implicit information (such as slow thermal decay);

[0140] (4) Gating memory update:

[0141] Using the Sigmoid activation function as a gate unit, the hidden state for the next time step is generated. and the output at the current moment :

[0142] ;

[0143] ;

[0144] in, As a renewed historical thermal energy transfer to At any given moment, the physical state is continuously transferred across time steps.

[0145] To address the high-frequency noise interference caused by fan vibration in industrial settings, this invention designs a spectral-temporal dual-domain attention layer within the differential recursive unit, such as... Figure 4 As shown, in this embodiment, the spectral-temporal dual-domain attention layer includes parallel time-domain and frequency-domain branches:

[0146] (1) Time-domain branch:

[0147] Global Context Embedding: Global average pooling (GAP) is performed on the tensor X of the dual-domain attention layer when the input spectrum is applied to generate channel context descriptors. Specifically, assuming the time step of the input tensor X is L and the feature dimension is D, the features are compressed along the time dimension, and the channel context descriptor of the d-th channel is generated. The calculation formula is:

[0148] ;

[0149] In the formula, t represents the time step index. This represents the eigenvalue of the input tensor at time step t and channel d.

[0150] Channel dependency mapping and temporal gating: A global context vector C is constructed from the context descriptors of all D channels. This global context vector C is then input into a channel dependency mapping network (implemented using a multilayer perceptron MLP) and processed through temporal gating (using a sigmoid activation function) to generate a temporal attention mask. The calculation formula is as follows:

[0151] In the formula, This indicates multilayer perceptron processing. Represents the activation function for time-domain gating;

[0152] Using the generated temporal attention mask Identify the main trend channels:

[0153] ;

[0154] In the formula, This represents the output tensor of the time-domain branch. This represents element-wise multiplication;

[0155] (2) Frequency domain branch:

[0156] Orthogonal frequency projection: The tensor X of the dual-domain attention layer at input spectrum is projected to the orthogonal frequency domain space using Discrete Cosine Transform (DCT) to obtain the spectral tensor. :

[0157] Spectrum mask learning: Learning a frequency domain filter mask through a neural network. This mask is equivalent to an adaptive soft threshold gate, which automatically suppresses high-frequency diffuse components representing mechanical noise while retaining specific frequency bands representing process abrupt changes.

[0158] ;

[0159] In the formula, This represents the sigmoid activation function. This indicates multilayer perceptron processing;

[0160] Inverse projection restoration: restoring the weighted clean spectrum back to the time domain, expressed as: , This represents the frequency domain branch output tensor. It represents the inverse discrete cosine transform.

[0161] Frequency domain branching utilizes the energy concentration characteristics of the frequency domain to separate noise from effective signals;

[0162] (3) Dual-domain orthogonal fusion:

[0163] Add the output tensor of the time-domain branch to the output tensor of the frequency-domain branch to output the enhanced feature representation: By superimposing time-domain trend features and frequency-domain cleanliness features, a high-fidelity process state characterization is output.

[0164] In this embodiment, the feature fusion module includes a cross-stream semantic alignment and fusion layer and a global pooling layer connected in sequence; wherein, the outputs of multiple branches of the cross-stream semantic alignment and fusion layer feature extraction module are concatenated in the channel dimension and then processed through... Convolutional layers integrate cross-channel information, unifying microscopic instantaneous responses and macroscopic long-term trends into a semantic space, generating a global context vector. .

[0165] In this embodiment, the multi-objective regression head uses a fully connected layer to output a prediction vector for the future prediction time step. This vector contains key guidance information in three dimensions:

[0166] (1) Instantaneous moisture prediction vector for exported materials: used for product quality early warning;

[0167] (2) Hot air valve opening prediction vector: used for feedforward control to intervene in heat supply in advance;

[0168] (3) Prediction vector of dehumidification damper opening: used for coordinated control to balance the humidity environment inside the cylinder.

[0169] S4: Train the differential recursive neural network prediction model using the sample dataset.

[0170] In this embodiment, since the prediction model needs to simultaneously output key guidance information in three dimensions—instantaneous moisture content of the outlet material, hot air valve opening, and exhaust damper opening—the following joint loss function and training strategy are set to ensure the effectiveness of multi-objective collaborative optimization and prevent model overfitting:

[0171] S41: Construction of Multi-Objective Joint Loss: Using mean squared error (MSE) as the basic evaluation index, a multi-objective joint loss function is constructed, consisting of the weighted sum of the mean squared errors of instantaneous moisture content of the outlet material, hot air valve opening, and exhaust damper opening. Since the instantaneous moisture content of the outlet material is the core quality control indicator, it is assigned the highest weight. The specific weight ratio can be selected according to the actual situation.

[0172] S42: Training parameter settings: The Adam optimizer is used to update the gradient of network parameters, and the initial learning rate and learning rate decay strategy are set to ensure convergence stability in the later stages of training; at the same time, the batch size and the maximum number of iterations are set.

[0173] S43: Model Iteration and Early Stopping: Divide the sample dataset into a training set and a validation set; during iterative training, if the joint loss value on the validation set no longer decreases within a set number of consecutive rounds, trigger the early stopping mechanism to terminate the training in advance, and save the model weight with the smallest loss value on the validation set as the final trained differential recursive neural network prediction model.

[0174] S5: Collect tobacco drying process data online and standardize it. Then, extract tobacco drying process data sequences at different time scales as input features and input them into a trained differential recursive neural network prediction model. Output the instantaneous moisture content prediction vector of the outlet material, the hot air valve opening prediction vector, and the dehumidification damper opening prediction vector for the future prediction step.

[0175] During standardization, the online-collected tobacco drying process data is standardized using the standardization parameters from step S2. The specific process of extracting tobacco drying process data sequences at different time scales as input features can be referenced from step S2 and will not be elaborated further here. After the model outputs the predicted vectors for the instantaneous moisture content of the outlet material, the hot air valve opening, and the exhaust damper opening for future prediction steps, the model output needs to be de-standardized to obtain data that can be used to generate control commands.

[0176] S6: Generate and issue tobacco drying control instructions based on the hot air valve opening prediction vector and the dehumidification damper opening prediction vector.

[0177] In this embodiment, when generating tobacco drying control instructions, process constraint verification is also required. If the predicted opening degree of the hot air valve or the opening degree of the dehumidification damper exceeds the corresponding set process red line, it is forcibly limited to the corresponding safe range.

[0178] To address the problem of model concept drift caused by changes in the origin of tobacco raw materials, equipment aging (such as thermal efficiency decline), or changes in environmental temperature and humidity, in some preferred embodiments, an adaptive predictive control method for tobacco leaf drying moisture further includes:

[0179] S7: Perform drift detection on the online differential recursive neural network prediction model. If a substantial drift is determined to have occurred, the online update process of the differential recursive neural network prediction model will be automatically triggered.

[0180] In this embodiment, to avoid accidental model updates due to occasional sensor malfunctions or transient process fluctuations, a dual "instantaneous-trend" judgment logic is designed:

[0181] S71: Real-time residual monitoring: Real-time calculation of the absolute percentage error (APE) of the instantaneous moisture prediction value of the exported material.

[0182] S72: Sliding statistical test: Maintain a length of... Using a sliding window, calculate the mean error. and standard deviation Define the drift index , This is the drift parameter, typically set to 2 or 3.

[0183] S73: Dual threshold trigger: Sets an alert threshold and a drift threshold. When An alert is triggered when the warning threshold is exceeded; when When the drift threshold is exceeded for a continuous period of p time steps, a substantial concept drift is determined to have occurred, and the online update process is automatically triggered.

[0184] S8: Mix the latest samples with the memory samples in proportion to construct a mixed fine-tuning training set.

[0185] In this embodiment, to prevent the model from forgetting the general rules of old operating conditions during the fine-tuning process (catastrophic forgetting), a hybrid fine-tuning training set is constructed.

[0186] S81: Latest Sample Pool: Extracts new data before the current batch triggers drift, characterizing the characteristics of the current new operating conditions.

[0187] S82: Memory Sample Pool: A memory pool containing historical typical operating conditions (different grades, seasons) is maintained using a reservoir sampling algorithm.

[0188] S83: Hybrid Strategy: Mix the latest samples with the memory samples in a ratio of 7:3 to construct a hybrid fine-tuning training set.

[0189] S9: Based on the hybrid fine-tuning training set, a hierarchical parameter freezing and gradient fine-tuning strategy is used to fine-tune and update the neural network prediction model based on differential recursion.

[0190] To achieve rapid updates under the limited computing power of industrial control computers, a hierarchical parameter freezing and gradient fine-tuning strategy is adopted for efficient parameter fine-tuning.

[0191] S91: Parameter Freeze: Completely freeze the general physical feature extraction layer at the bottom of the differential recursive unit (i.e., the first K-1 layers of deep nonlinear interactive network, including its internal dual-domain attention mask generation parameters). This part of the network extracts general time-frequency physical laws, which do not need to be frequently changed.

[0192] S92: Unfreeze Update: Only the top layer of the differential recurrent unit, the multi-stream fusion convolutional layer, and the prediction fully connected layer are unfrozen. Using mixed incremental data, gradient descent updates the unfrozen parameters with a small learning rate, allowing the model to quickly adapt to changes in the thermal efficiency of the current equipment.

[0193] S10: Employs a shadow evaluation mechanism to compare the prediction accuracy of the updated new model with that of the current model in real time within the evaluation window; only when the accuracy of the updated new model is better than that of the current model, the parameters of the inference engine are pointed to the updated new model within milliseconds through memory pointer swapping technology, thus achieving seamless model iteration in the production process.

[0194] In this embodiment, to ensure production safety, a shadow evaluation mechanism and hot-switching are designed:

[0195] S101: Shadow Parallel Operation: The updated new model first runs in the background in shadow mode, receiving real-time input but not outputting control commands.

[0196] S102: Winning determination: In the evaluation window, the prediction accuracy of the updated new model is compared with that of the current model in real time; the updated new model is determined to win only when the accuracy of the updated new model is better than that of the current model.

[0197] S103: Parameter Hot Switching: Through memory pointer swapping technology, the parameters of the inference engine are switched to the updated new model within milliseconds, enabling seamless model iteration during the production process.

[0198] Compared with the prior art, the adaptive predictive control method for tobacco leaf drying moisture provided in the above embodiments has the following advantages:

[0199] (1) Overcome the thermal inertia fault caused by discrete slicing and improve the long-period prediction accuracy of large-lag systems.

[0200] Existing technical challenges: To reduce computational load, existing long-sequence deep prediction models generally employ a sequence patching strategy. This mechanism severs the physical connection between time segments, causing the model to be unable to remember the accumulated thermodynamic state of the equipment from minutes prior. When dealing with systems like the wire drying process, which exhibits significant hysteresis, the prediction error diverges sharply as the prediction step size increases due to the loss of historical thermal potential information.

[0201] Improvements of this invention: This invention adopts a differential recursive unit architecture and introduces an adaptive thermal inertia decay mechanism in state transmission.

[0202] Technical principle: By establishing recursive channels between discrete time segments, the hidden state vector of thermal potential energy is transferred. This allows for the lossless transfer of implicit thermodynamic states from the previous moment (such as deep heat storage in the cylinder wall and the hysteretic response of moisture) to the current moment. A learnable adaptive attenuation coefficient is introduced. This model simulates the natural dissipation of physical heat over time. It can automatically adjust according to the current operating conditions. Values ​​that retain long-term memory (large) in steady state. ), quickly forget the old state when operating conditions change abruptly (small This allows for the precise capture of the dynamic evolution patterns of systems with large time lags.

[0203] (2) Frequency domain signal decoupling under strong noise conditions was achieved, ensuring the stability and robustness of control commands.

[0204] Existing technical challenges: Sensor data from the drying process (especially exhaust air volume and instantaneous flow rate) contains a large amount of high-frequency noise caused by mechanical vibration or airflow turbulence. Existing time-domain correlation analysis mechanisms only calculate correlation on the time waveform, which is prone to time-domain waveform aliasing, misjudging high-amplitude random noise as key features, leading to incorrect weight allocation, and causing the output control commands to jitter severely due to noise interference.

[0205] Improvements of this invention: This invention constructs a spectral-temporal dual-domain attention mechanism and introduces adaptive soft thresholding denoising for frequency domain branching.

[0206] Technical Principle: Features are mapped to the frequency domain using orthogonal frequency projection. Since random mechanical noise typically manifests as diffuse components across the entire frequency band, while effective moisture mutation signals exhibit a specific spectral structure, the two are significantly separable in the frequency domain. A spectral mask is generated in the frequency domain using a neural network, essentially acting as an adaptive bandpass filter to precisely suppress noise bands; simultaneously, global context embedding is used in the time domain to lock in macroscopic trends. This dual mechanism of "frequency domain cleanup + time domain fidelity preservation" ensures that the model can still output stable and accurate control parameters even in harsh noise environments.

[0207] (3) It solves the problem of insufficient feature extraction caused by multivariate dynamic coupling.

[0208] Existing technical challenges: The wire drying process involves multiple variables with vastly different frequencies of change. Fast variables (such as the dehumidification damper) fluctuate dramatically on a second-by-second basis, while slow variables (such as the cylinder temperature) drift slowly on a minute-by-minute basis. Existing models typically use a single temporal resolution (such as a fixed convolutional kernel or a single window) to process all variables, making it difficult to simultaneously account for both instantaneous response and long-term trends.

[0209] Improvements of this invention: This invention constructs a multi-scale input architecture and a cross-stream semantic alignment mechanism.

[0210] Technical Principle: The system processes fine-grained streams that retain the original sampling rate and coarse-grained streams that have been downsampled in parallel. The fine-grained stream focuses on capturing the instantaneous changes of fast variables such as flow rate at the second level, while the coarse-grained stream focuses on capturing the thermal inertia curves of slow variables such as temperature at the minute level. Through cross-channel convolutional fusion at the end, physical features with different time resolutions are mapped to a unified semantic state space, thereby accurately representing the complex "fast-slow" dynamic coupling relationship in the wire drying process.

[0211] (4) It has the ability to adapt to working conditions throughout the entire life cycle, thus overcoming the problem of model performance degradation.

[0212] Existing technical challenges: Most existing industrial prediction models are static models trained offline. As equipment operating time increases (heat exchange efficiency decreases due to fouling) or the origin of tobacco raw materials changes (hygroscopic properties change), the prediction accuracy of static models gradually decreases over time, resulting in conceptual drift and requiring frequent manual retraining and downtime maintenance.

[0213] The improvement method of this application: This application establishes an online closed-loop update mechanism based on dual drift detection and experience playback.

[0214] Technical Principle: The system can detect statistically significant changes in the prediction residuals in real time and automatically trigger incremental fine-tuning. During the update process, through a hierarchical parameter freezing strategy (freezing the underlying general features and updating only the top decision layer) and a hybrid memory sample pool (mixing new and old operating condition data), the model can quickly adapt to the current new equipment status or new raw material characteristics while avoiding the catastrophic forgetting of general rules from old operating conditions. This ensures that the control system maintains optimal performance throughout long-term operation, reduces reliance on manual maintenance, and improves production efficiency.

[0215] This invention also provides an adaptive predictive control system for tobacco leaf drying moisture content, comprising:

[0216] The noise reduction module is used to acquire historical tobacco drying process data and remove non-steady-state noise;

[0217] The sample dataset construction module is used to standardize the denoised historical tobacco drying process data and construct the sample dataset using a multi-scale segmentation strategy based on a sliding window mechanism. Each sample includes input features consisting of tobacco drying process data sequences at different time scales and a label vector for the prediction step size. The label vector includes the instantaneous moisture vector of the outlet material, the hot air valve opening vector, and the dehumidification damper opening vector.

[0218] The model building module is used to construct a neural network prediction model architecture based on differential recursion, including a feature extraction module, a feature fusion module, and a multi-objective regression head connected in sequence. The feature extraction module includes multiple branches that process input features at different time scales, and each branch includes a spatiotemporal feature embedding layer and a differential recursive unit connected in sequence. The feature fusion module is used to fuse the outputs of the multiple branches of the feature extraction module.

[0219] The model training module is used to train a differential recursive neural network prediction model using a sample dataset.

[0220] The online prediction module is used to collect tobacco drying process data online, standardize it, extract tobacco drying process data sequences at different time scales as input features, and input them into a trained differential recursive neural network prediction model to output the instantaneous moisture content prediction vector of the outlet material, the hot air valve opening prediction vector, and the dehumidification damper opening prediction vector for the future prediction step.

[0221] The instruction generation and issuance module is used to generate and issue tobacco drying control instructions based on the hot air valve opening prediction vector and the dehumidification damper opening prediction vector.

[0222] It should be understood that the functional unit modules in the various embodiments of the present invention can be concentrated in one processing unit, or each unit module can exist physically separately, or two or more unit modules can be integrated into one unit module, and can be implemented in hardware or software.

[0223] This invention also provides an adaptive predictive control system for tobacco leaf drying moisture content, comprising:

[0224] A memory on which computer programs are stored;

[0225] The processor is used to execute the computer program to implement the adaptive predictive control method for tobacco leaf drying moisture as described above.

[0226] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the adaptive predictive control method for tobacco leaf drying moisture as described above.

[0227] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0228] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0229] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0230] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0231] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0232] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for adaptive predictive control of moisture content in tobacco leaf drying, characterized in that, Includes the following steps: S1: Obtain historical tobacco drying process data and remove non-steady-state noise; S2: Standardize the denoised historical tobacco drying process data and construct a sample dataset using a multi-scale segmentation strategy based on a sliding window mechanism. Each sample includes input features consisting of tobacco drying process data sequences at different time scales and a label vector for the prediction step size. The label vector includes the instantaneous moisture vector of the outlet material, the hot air valve opening vector, and the dehumidification damper opening vector. S3: Construct a neural network prediction model architecture based on differential recursion, including a feature extraction module, a feature fusion module, and a multi-objective regression head connected in sequence; the feature extraction module includes multiple branches that process input features at different time scales, and each branch includes a spatiotemporal feature embedding layer and a differential recursive unit connected in sequence; the feature fusion module is used to fuse the outputs of the multiple branches of the feature extraction module. The calculation logic of the differential recursive unit is as follows: Thermal potential injection and adaptive decay: The input of the differential recursive unit at time t is represented as follows: ； In the formula, The input features of the differential recursive unit at time t; The hidden state vector at time t-1 is used to characterize the historical thermal potential energy; It is a learnable adaptive decay coefficient. ; Input features of the differential recursive unit that injects historical thermal potential energy at time t; Deep feature interaction: The input is fed into a K-layer stacked nonlinear transformation subunit, each layer comprising a spectral-temporal dual-domain attention layer, a first normalization layer, a feedforward neural network, and a second normalization layer connected in sequence; the output features of the nonlinear transformation subunit are denoted as... ; Two-stream differential separation mechanism: A differential stream separation mechanism is set up: Predictive stream: Explicit predictive features for water regression at the current time are generated using a multilayer perceptron. , The prediction flow represents the multilayer perceptron operation; the state flow involves extracting latent state features that are not currently consumed but need to be passed on to the future using the multilayer perceptron. , Represents state flow multilayer perceptron operation; Gated memory update: Using the Sigmoid activation function as the gating unit, the hidden state for the next time step is generated. and the output at the current moment : ； ； in, As a renewed historical thermal energy transfer to time; S4: Train the differential recursive neural network prediction model using the sample dataset; S5: Collect tobacco drying process data online and standardize it. Then extract tobacco drying process data sequences at different time scales as input features and input them into a trained differential recursive neural network prediction model. Output the instantaneous moisture content prediction vector of the outlet material, the hot air valve opening prediction vector, and the dehumidification damper opening prediction vector for the future prediction step. S6: Generate and issue tobacco drying control instructions based on the hot air valve opening prediction vector and the dehumidification damper opening prediction vector.

2. The adaptive predictive control method for tobacco leaf drying moisture content according to claim 1, characterized in that, Step S1 specifically includes: Obtain historical tobacco drying process data and construct the original feature matrix. ,in, The number of sampling points. For feature dimensions; The conditional probability between two sample points is calculated using the Gaussian kernel function of dynamic variance to represent the local similarity between the two sample points, and then the high-dimensional joint probability distribution is calculated. Set the target low-dimensional space, and assume high-dimensional sample points. The corresponding mapping point in the target low-dimensional space is: In the target low-dimensional space, the Student-t distribution with 1 degree of freedom in the long tail is used to calculate the low-dimensional joint probability distribution between two low-dimensional mapping points. Kullback-Leibler divergence is used as the loss function to solve for the optimal set of low-dimensional mapping point coordinates Y, so that the low-dimensional joint probability distribution approximates the high-dimensional joint probability distribution. For the optimal set of low-dimensional mapping point coordinates Y, perform a clustering operation with a cluster size of 2, retaining only the data indices belonging to the steady-state clusters, and then extracting the data from the original feature matrix. Extract the corresponding original data to complete the noise reduction.

3. The adaptive predictive control method for tobacco leaf drying moisture content according to claim 1, characterized in that, Step S2 specifically includes: For the denoised historical tobacco drying process data, Z-Score transformation is performed on each feature dimension to make it conform to a standard normal distribution. The standardized historical tobacco drying process data is segmented using a sliding window mechanism to obtain an initial sample dataset: the basic time window length is set. , The original feature step size, To predict the step size; the data feature dimensions extracted from the first L steps of the base time window are: Data corresponding to the tobacco drying process for all categories; the subsequent basic time window The data feature dimensions extracted in each step are 3, corresponding to the three feature dimensions of instantaneous moisture content of the outlet material, hot air valve opening, and exhaust damper opening, which are used to construct the label vector; Multi-scale processing: Multi-scale feature extraction is performed on the first L steps of data of each initial sample. The feature vectors of different scales extracted together constitute the input features, thereby constructing the sample dataset.

4. The adaptive predictive control method for tobacco leaf drying moisture content according to claim 1, characterized in that, The spectral-time dual-domain attention layer includes parallel time-domain branches and frequency-domain branches: Temporal branch: Global average pooling is performed on the tensor X of the dual-domain attention layer at the input spectrum to generate channel context descriptors and form a global context vector. This global context vector is then used to generate a temporal attention mask through a multilayer perceptron. Identify the main trend channels: ； In the formula, This represents the output tensor of the time-domain branch. This represents element-wise multiplication; Frequency domain branch: Orthogonal frequency projection: The tensor X of the dual-domain attention layer at input spectrum is projected onto the orthogonal frequency domain space using discrete cosine transform to obtain the spectral tensor. ; Spectrum mask learning: Learning a frequency domain filter mask through a neural network. The mask automatically suppresses high-frequency diffuse components representing mechanical noise while retaining frequency bands representing process abrupt changes. Inverse projection restoration: restoring the weighted clean spectrum back to the time domain, expressed as: , This represents the frequency domain branch output tensor. Represents the inverse discrete cosine transform; Bi-domain orthogonal fusion: The output tensor of the time-domain branch is added to the output tensor of the frequency-domain branch, and the enhanced feature representation is output. .

5. The adaptive predictive control method for tobacco leaf drying moisture content according to claim 1, characterized in that, In step S6, when generating the tobacco drying control command, process constraint verification is performed. If the predicted hot air valve opening or dehumidification damper opening exceeds the corresponding set process red line, it is forcibly limited to the corresponding safe range.

6. The adaptive predictive control method for tobacco leaf drying moisture content according to claim 1, characterized in that, Also includes: Drift detection is performed on the online differential recursive neural network prediction model. If a substantial drift is determined to have occurred, the online update process of the differential recursive neural network prediction model is automatically triggered. The latest samples and the remembered samples are mixed in proportion to construct a hybrid fine-tuning training set; Based on a hybrid fine-tuning training set, a hierarchical parameter freezing and gradient fine-tuning strategy is used to fine-tune and update the differential recursive neural network prediction model. A shadow evaluation mechanism is adopted to compare the prediction accuracy of the updated new model with that of the current model in real time within the evaluation window; only when the accuracy of the updated new model is better than that of the current model, the parameters of the inference engine are pointed to the updated new model in milliseconds through memory pointer swapping technology, so as to achieve seamless model iteration in the production process.

7. An adaptive predictive control system for tobacco leaf drying moisture content, characterized in that, include: The noise reduction module is used to acquire historical tobacco drying process data and remove non-steady-state noise; The sample dataset construction module is used to standardize the denoised historical tobacco drying process data and construct the sample dataset using a multi-scale segmentation strategy based on a sliding window mechanism. Each sample includes input features consisting of tobacco drying process data sequences at different time scales and a label vector for the prediction step size. The label vector includes the instantaneous moisture vector of the outlet material, the hot air valve opening vector, and the dehumidification damper opening vector. The model building module is used to build a neural network prediction model architecture based on differential recursion, including a feature extraction module, a feature fusion module, and a multi-objective regression head connected in sequence. The feature extraction module includes multiple branches that process input features at different time scales, and each branch includes a spatiotemporal feature embedding layer and a differential recursive unit connected in sequence. The feature fusion module is used to fuse the outputs of the multiple branches of the feature extraction module. The calculation logic of the differential recursive unit is as follows: Thermal potential injection and adaptive decay: The input of the differential recursive unit at time t is represented as follows: ； In the formula, The input features of the differential recursive unit at time t; The hidden state vector at time t-1 is used to characterize the historical thermal potential energy; It is a learnable adaptive decay coefficient. ; Input features of the differential recursive unit that injects historical thermal potential energy at time t; Deep feature interaction: The input is fed into a K-layer stacked nonlinear transformation subunit, each layer comprising a spectral-temporal dual-domain attention layer, a first normalization layer, a feedforward neural network, and a second normalization layer connected in sequence; the output features of the nonlinear transformation subunit are denoted as... ; Two-stream differential separation mechanism: A differential stream separation mechanism is set up: Predictive stream: Explicit predictive features for water regression at the current time are generated using a multilayer perceptron. , The prediction flow represents the multilayer perceptron operation; the state flow involves extracting latent state features that are not currently consumed but need to be passed on to the future using the multilayer perceptron. , Represents state flow multilayer perceptron operation; Gated memory update: Using the Sigmoid activation function as the gating unit, the hidden state for the next time step is generated. and the output at the current moment : ； ； in, As a renewed historical thermal energy transfer to time; The model training module is used to train a differential recursive neural network prediction model using a sample dataset. The online prediction module is used to collect tobacco drying process data online, standardize it, extract tobacco drying process data sequences at different time scales as input features, and input them into a trained differential recursive neural network prediction model to output the instantaneous moisture content prediction vector of the outlet material, the hot air valve opening prediction vector, and the dehumidification damper opening prediction vector for the future prediction step. The instruction generation and issuance module is used to generate and issue tobacco drying control instructions based on the hot air valve opening prediction vector and the dehumidification damper opening prediction vector.

8. An adaptive predictive control system for tobacco leaf drying moisture content, characterized in that, include: A memory on which computer programs are stored; A processor, when executing the computer program, implements the adaptive predictive control method for tobacco leaf drying moisture as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the adaptive predictive control method for tobacco leaf drying moisture as described in any one of claims 1 to 6.

Images