Tobacco cut filler moisture control method and system based on lightgbm

Abstract

The application relates to the technical field of artificial intelligence and tobacco processing, and discloses a tobacco cut filler moisture control method and system based on LightGBM, which aims to solve the problems of distorted moisture signal and low control precision caused by dust interference. The method comprises the following steps: acquiring the dust concentration of a production line and a moisture signal with noise in real time; generating optimal wavelet filtering parameters through a nonlinear mapping model based on the dust concentration; performing adaptive wavelet filtering on the moisture signal to reconstruct a pure signal; inputting the pure signal into a pre-trained LightGBM model to predict the moisture value; generating a control instruction according to the deviation between the predicted value and a target value to realize closed-loop control. The system comprises six modules, namely, working condition sensing, signal acquisition, dynamic filtering parameter generation, adaptive filtering, moisture prediction and control instruction generation. The application realizes decoupling integration of dynamic filtering and the existing LightGBM model, significantly improves the moisture control precision and response speed without changing the original model, and reduces the engineering transformation cost.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and tobacco processing, specifically relating to a method and system for controlling the moisture content of tobacco shreds based on LightGBM. Background Technology

[0002] In the tobacco processing industry, precise control of moisture content during the tobacco feeding process is a key technological step to ensure the sensory quality and combustion performance of the product. Moisture content not only directly affects the filling capacity of tobacco, tar release, and the stability of aroma substances, but is also subject to strict constraints by the industry standard YC / T 147-2010, which requires the control precision to be maintained within ±0.5%.

[0003] To achieve high-precision closed-loop control, current mainstream technical solutions generally adopt an architecture that combines infrared moisture sensors with the LightGBM machine learning model: the sensor collects the moisture signal on the surface of the tobacco in real time, and LightGBM predicts the current moisture value based on historical operating data and outputs control commands to the actuator.

[0004] However, the solution has revealed serious environmental adaptability defects during long-term operation in industrial sites. Tobacco itself is a short fiber product, and dust is easily generated after drying, fermentation and cutting. The dust concentration dynamic value of the production line fluctuates in real time (0–10 mg / m³), which will cause high-frequency noise to be superimposed on the infrared signal, significantly interfering with the accuracy of moisture sensing.

[0005] Existing technologies are thus caught in a fundamental contradiction: "high precision inevitably leads to high latency" and "low latency inevitably leads to high error." The core issue lies in the lack of a dynamic mapping mechanism between dust concentration and filtering parameters, resulting in rigid noise suppression strategies, slow response, and high engineering deployment costs. There is an urgent need for a decoupled, low-latency, adaptive signal preprocessing architecture that can dynamically adjust filtering parameters according to the dust environment at the millisecond level without reconstructing the existing LightGBM model. This would stably compress moisture prediction errors to within ±0.5%, while ensuring system real-time performance and production line compatibility. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and system for controlling the moisture content of tobacco shreds based on LightGBM. The core of this method lies in constructing a dynamic mapping mechanism between environmental conditions and signal processing parameters. By sensing the dust concentration on the production line in real time, the core parameters of the adaptive filter are reconstructed instantly. Thus, without changing the existing LightGBM prediction model, precise and low-latency purification of noisy moisture signals is achieved, ultimately resulting in high-precision closed-loop moisture control.

[0007] According to one aspect of the present invention, a method for controlling the moisture content of tobacco shreds based on LightGBM is provided, comprising the following steps:

[0008] Real-time acquisition of dust concentration data at designated monitoring points on the tobacco processing production line;

[0009] Simultaneously acquire the raw, noisy tobacco moisture signal output by an infrared moisture sensor spatially adjacent to the dust concentration data acquisition point;

[0010] Based on the real-time dust concentration data, a set of optimal wavelet filter parameters matching the current dust concentration is dynamically calculated and generated through a preset nonlinear parameter mapping model. The optimal wavelet filter parameters include wavelet basis function type, decomposition level and threshold function.

[0011] The dynamically generated optimal wavelet filter parameters are applied to perform adaptive wavelet transform filtering on the original noisy tobacco moisture signal to eliminate noise components caused by dust interference, reconstruct and output a pure tobacco moisture signal.

[0012] The moisture signal of the pure tobacco is used as an input feature and transmitted to a pre-trained LightGBM moisture prediction model that maintains its structure. The LightGBM moisture prediction model performs inference calculations and outputs a high-precision predicted moisture value.

[0013] Based on the deviation between the high-precision predicted moisture value and the target moisture setting value of the tobacco feeding process, control instructions are generated for the feeding system to precisely control the final moisture content of the tobacco in a closed-loop manner.

[0014] According to another aspect of the present invention, a tobacco shred moisture control system based on LightGBM is provided, comprising:

[0015] The working condition sensing module is configured to acquire real-time dust concentration data at designated monitoring points on the tobacco processing production line;

[0016] The signal acquisition module is configured to synchronously acquire the original moisture signal of the tobacco shreds containing noise from the infrared moisture sensor. The physical location of the sensor in the signal acquisition module and the physical location of the sensor in the working condition sensing module are in the same noise-affected area.

[0017] The filter parameter dynamic generation module is configured to receive real-time dust concentration data output by the working condition sensing module, and dynamically calculate and generate a set of optimal wavelet filter parameters based on the data through a built-in nonlinear parameter mapping model.

[0018] An adaptive filtering module is configured to receive the original noisy tobacco moisture signal output by the signal acquisition module and the optimal wavelet filter parameters output by the dynamic generation module of the filter parameters, and use the parameters to perform adaptive wavelet transform filtering on the original signal to output a pure tobacco moisture signal.

[0019] The moisture prediction module is configured to receive the moisture signal of the pure tobacco shreds output by the adaptive filtering module, and predict the moisture value using an internally fixed LightGBM model that does not require online retraining.

[0020] The control command generation module is configured to calculate and generate precise control commands for the feeding actuator based on the difference between the high-precision predicted moisture value output by the moisture prediction module and the process target moisture setting value.

[0021] In one embodiment of the present invention, the working condition sensing module specifically includes a laser scattering dust concentration sensor with a measurement range of 0 to 100 milligrams per cubic meter, a measurement accuracy of 5%, and a response time of less than 10 milliseconds. The laser scattering dust concentration sensor is electrically connected to the filter parameter dynamic generation module via a shielded twisted-pair cable using an RS485 bus and follows the Modbus RTU communication protocol, reporting instantaneous dust concentration values ​​at a frequency of 100 times per second. The probe of the laser scattering dust concentration sensor is installed 50 millimeters below the optical path of the infrared moisture sensor to ensure that the measured dust concentration accurately represents the degree of interference with the moisture signal.

[0022] Furthermore, the filter parameter dynamic generation module internally contains the nonlinear parameter mapping model, which is a set of piecewise functions used to map the input dust concentration value C (unit: milligrams per cubic meter) to a wavelet basis function type W, a decomposition level L, and a threshold λ. Specifically, the mapping relationship is defined as follows:

[0023] When the dust concentration C is less than two, it indicates that the environment is clean and the noise level is low. To ensure the minimum processing delay, the wavelet basis function type W is selected as db2 (Daubechies second order) and the number of decomposition layers L is two.

[0024] When the dust concentration C is between two and six, it indicates that there is a moderate level of dust interference. In order to balance the filtering effect and the computational cost, the wavelet basis function type W is selected as sym4 (Symlets fourth order), and the number of decomposition layers L is three.

[0025] When the dust concentration C is greater than or equal to six, it indicates that there is serious dust interference. In order to achieve the strongest noise suppression, the wavelet basis function type W is selected as coif3 (Coiflets third order), and the decomposition level L is four.

[0026] The threshold λ is calculated and its relationship with the dust concentration C strictly follows a linear function: λ = 0.21 * C + 1.15. This linear relationship is obtained by least-squares fitting of thousands of sets of historical operating data, ensuring the adaptability and optimality of the threshold selection. The dynamic generation module for the filter parameters is implemented by a field-programmable gate array (FPGA) chip. Its internal logic circuits perform piecewise judgment and linear calculation in parallel, ensuring that upon receiving new dust concentration data, a completely new and optimal wavelet filter parameter is generated and output within one millisecond.

[0027] Furthermore, the adaptive wavelet transform filtering process performed by the adaptive filtering module specifically includes the following sub-steps:

[0028] First, the adaptive filtering module obtains a noisy tobacco moisture raw signal vector with a data length of 1024 from the signal acquisition module;

[0029] Secondly, based on the real-time wavelet basis function type W and decomposition level L provided by the filter parameter dynamic generation module, multi-level wavelet decomposition is performed on the original signal vector to obtain an approximate coefficient subband cA_L and L detail coefficient subbands cD_1, cD_2, ..., cD_L;

[0030] Next, based on the real-time threshold λ provided by the dynamic generation module of the filtering parameters, a soft thresholding function is applied to each coefficient x in each detail coefficient subband cD_j (j from one to L). The mathematical expression of the soft thresholding function is: x' = sign(x) * max(0, |x| - λ), thereby obtaining a new set of detail coefficient subbands that have been thresholded.

[0031] Finally, using the original approximation coefficient subband cA_L and all the new detail coefficient subbands after thresholding, an inverse wavelet transform reconstruction is performed to obtain a pure tobacco moisture signal vector with a data length of 1024, which is then output to the moisture prediction module. The entire filtering process is executed by a digital signal processor, and the computational delay of a single complete processing step is strictly controlled to within 20 milliseconds.

[0032] In one embodiment of the present invention, the LightGBM moisture prediction model used in the moisture prediction module is an offline trained model. The input feature dimension of the model is one, namely the moisture signal of the pure tobacco. The model structure remains constant after deployment to the production line, without any online structural adjustments or parameter retraining. The data format of the pure tobacco moisture signal output by the adaptive filtering module is completely consistent with the input data format used during the training of the LightGBM model, thereby achieving seamless decoupling between the filtering front-end and the prediction back-end. This decoupling architecture ensures that the method and system of the present invention do not require any modification to the existing calibrated and verified LightGBM model during deployment, greatly reducing the complexity and cost of engineering implementation.

[0033] Furthermore, the control command generation module integrates a proportional-integral-derivative controller. The controller receives the high-precision predicted moisture value output by the moisture prediction module as a process variable and the target moisture setpoint stored in the process database as a setpoint. The controller calculates the error between the predicted value and the setpoint and generates an analog output signal with a range of four to twenty milliamps based on preset proportional, integral, and derivative parameters. This analog output signal is directly connected to the frequency input of the variable frequency feeding pump on the tobacco feeding line via a cable. By adjusting the pump's rotation speed, the amount of liquid sprayed onto the tobacco per unit time is precisely controlled, thereby achieving real-time, accurate, and stable closed-loop control of the tobacco moisture content.

[0034] In summary, this application includes at least one of the following beneficial technical effects:

[0035] First, it achieves precise adaptive noise suppression. This invention establishes a deterministic dynamic mapping mechanism from real-time dust concentration to the core parameters of wavelet filtering, enabling the filtering intensity to respond instantaneously to changes in environmental noise. When dust concentration increases, the system automatically adopts a stronger wavelet basis, a deeper decomposition level, and a higher threshold to effectively filter out high-intensity noise; when dust concentration decreases, it automatically switches to milder filtering parameters, avoiding excessive smoothing of the effective moisture signal, thereby stabilizing the root mean square error of moisture prediction from over 2% in existing technologies to within 0.6%, significantly improving the control accuracy.

[0036] Secondly, this invention resolves the inherent contradiction between high-precision filtering and low processing latency. The parameter dynamic generation model employed in this invention is a nonlinear mapping function with extremely low computational complexity, resulting in a parameter generation latency of less than one millisecond. Simultaneously, by dynamically optimizing the number of decomposition levels, unnecessary high-level decomposition is avoided in low-noise environments, ensuring that the total processing latency of the adaptive filtering module remains constant at less than twenty milliseconds. The overall front-end processing latency is significantly lower than the two hundred milliseconds of existing technologies, guaranteeing the rapid response capability of the control system and effectively suppressing the problem of local moisture deviation caused by control adjustments.

[0037] Third, it provides a highly decoupled engineering solution. The adaptive filtering system of this invention acts as an independent front-end preprocessing unit, and its output clean signal directly meets the input requirements of existing LightGBM models. This solution requires no structural modifications, feature reconstruction, or model retraining of the LightGBM model already running stably on the production line, achieving "plug-and-play" integration. This significantly reduces the cost, risk, and production downtime of technological upgrades, providing an economical, efficient, and reliable technical path for the intelligent upgrading of in-service production lines. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall technical solution architecture of the tobacco shred moisture control method and system based on LightGBM proposed in this invention;

[0039] Figure 2 This is a schematic diagram of the core principle framework of the dust concentration-driven adaptive wavelet filter parameter dynamic mapping mechanism in this invention. Detailed Implementation

[0040] This invention provides a method and system for controlling the moisture content of tobacco shreds based on LightGBM. Its core lies in constructing a dynamic mapping mechanism between dust concentration and wavelet filter parameters. By sensing the dust concentration in the production line in real time, the core parameters of the adaptive filter are reconstructed instantaneously. Without altering the existing LightGBM prediction model, this achieves precise and low-latency purification of noisy moisture signals, ultimately achieving high-precision closed-loop moisture control. The following will describe in detail the specific implementation steps and supporting system structure of this method.

[0041] The LightGBM-based method for controlling the moisture content of tobacco shreds includes the following steps: S1, acquiring real-time dust concentration data at designated monitoring points on the tobacco processing production line; S2, simultaneously acquiring the raw, noisy tobacco moisture signal output by an infrared moisture sensor spatially adjacent to the dust concentration data acquisition point; S3, based on the real-time acquired dust concentration data, dynamically calculating and generating a set of optimal wavelet filter parameters matching the current dust concentration using a preset nonlinear parameter mapping model. The optimal wavelet filter parameters include the wavelet basis function type, decomposition level, and threshold function; S4, applying the dynamically generated optimal wavelet filter... S5. The raw moisture signal of the noisy tobacco is subjected to adaptive wavelet transform filtering to eliminate noise components caused by dust interference, and the pure tobacco moisture signal is reconstructed and output. S6. The pure tobacco moisture signal is used as input feature and sent to a pre-trained LightGBM moisture prediction model that maintains its structure. The LightGBM moisture prediction model performs inference calculation and outputs a high-precision predicted moisture value. S7. Based on the deviation between the high-precision predicted moisture value and the target moisture setting value of the tobacco feeding process, the control command of the feeding system is generated to accurately control the final moisture content of the tobacco in a closed-loop manner.

[0042] Specifically, in step S1, dust concentration data at designated monitoring points on the tobacco processing production line is acquired in real time. This step is achieved using laser scattering dust concentration sensors deployed in key areas of the production line. The laser scattering dust concentration sensor has a measurement range of 0 to 100 milligrams per cubic meter, a measurement accuracy of 5%, and a response time of less than 10 milliseconds. The sensor probe is installed 50 millimeters below the optical path of the infrared moisture sensor to ensure that the measured dust concentration accurately represents the degree of interference with the moisture signal.

[0043] The sensor is electrically connected to the subsequent processing unit via shielded twisted-pair cable using an RS485 bus and follows the Modbus RTU communication protocol, reporting instantaneous dust concentration values ​​at a frequency of one hundred times per second. This high-frequency sampling mechanism ensures the real-time nature and representativeness of the dust concentration data, providing a reliable basis for the dynamic adjustment of subsequent filtering parameters.

[0044] In step S2, the raw moisture signal of the tobacco shreds, including noise, is simultaneously acquired from the infrared moisture sensor located spatially adjacent to the dust concentration data acquisition point. The infrared moisture sensor is installed above the tobacco shred conveyor belt, and its beam vertically illuminates the surface of the tobacco shreds. The moisture content of the tobacco shreds is inverted by the intensity of the reflected light.

[0045] Because the dust concentration in the production line fluctuates between 0 and 10 milligrams per cubic meter, the dust particles scatter and absorb infrared light, resulting in high-frequency noise superimposed on the sensor output signal. The noise frequency is greater than 50 Hz, and the noise energy ratio fluctuates between 15% and 40%. The original signal of moisture content in the noisy tobacco is collected in 1024 points as a processing frame, with a sampling frequency of 500 Hz and a single frame duration of 2.048 seconds.

[0046] The signal acquisition module is strictly synchronized with the dust concentration sensor to ensure that each frame of moisture signal corresponds to the dust concentration value at the same moment, avoiding parameter mapping inaccuracies due to time misalignment.

[0047] In step S3, based on the real-time acquired dust concentration data, a set of optimal wavelet filter parameters matching the current dust concentration is dynamically calculated and generated through a preset nonlinear parameter mapping model. The nonlinear parameter mapping model is a set of piecewise functions, whose input is the dust concentration value C (unit: milligrams per cubic meter), and whose output is the wavelet basis function type W, the number of decomposition levels L, and the threshold λ.

[0048] The specific mapping relationship is as follows: When the dust concentration C is less than two, the wavelet basis function type W is selected as db2 (Daubechies second order), and the decomposition level L is two; when the dust concentration C is between two and six, the wavelet basis function type W is selected as sym4 (Symlets fourth order), and the decomposition level L is three; when the dust concentration C is greater than or equal to six, the wavelet basis function type W is selected as coif3 (Coiflets third order), and the decomposition level L is four.

[0049] The threshold λ is calculated using a linear function: λ = 0.21 × C + 1.15. This linear relationship is obtained by least-squares fitting of thousands of historical operating data sets, ensuring that the threshold selection achieves an optimal balance between noise suppression and signal fidelity under different dust concentrations. The nonlinear parameter mapping model is implemented using a field-programmable gate array (FPGA) chip, whose internal logic circuits perform piecewise judgments and linear calculations in parallel, ensuring that upon receiving new dust concentration data, entirely new optimal wavelet filter parameters are generated and output within one millisecond.

[0050] In step S4, dynamically generated optimal wavelet filter parameters are applied to perform adaptive wavelet transform filtering on the original moisture content signal of noisy tobacco. This process specifically includes the following sub-steps:

[0051] First, the adaptive filtering module obtains a raw signal vector of tobacco moisture with a data length of 1024 from the signal acquisition module.

[0052] Secondly, based on the real-time wavelet basis function type W and decomposition level L provided by the dynamic generation module for filtering parameters, multi-level wavelet decomposition is performed on the original signal vector to obtain an approximate coefficient subband cA_L and L detail coefficient subbands cD_1, cD_2, ..., cD_L.

[0053] Next, based on the real-time threshold λ provided by the filter parameter dynamic generation module, a soft thresholding function is applied to each coefficient x in each detail coefficient subband cD_j (j from one to L). The mathematical expression of the soft thresholding function is: x' = sign(x) × max(0, |x| - λ), thus obtaining a new set of detail coefficient subbands that have undergone thresholding.

[0054] Finally, using the original approximation coefficient subband cA_L and all the new detail coefficient subbands that have undergone thresholding, an inverse wavelet transform is performed to reconstruct a pure tobacco moisture signal vector with a data length of 1024, which is then output to the moisture prediction module.

[0055] The entire filtering process is executed by a digital signal processor, and the computational latency of a single complete processing run is strictly controlled to within twenty milliseconds. This adaptive filtering mechanism ensures that lightweight filtering is used to reduce latency in low-dust environments, while strong filtering is activated to ensure accuracy in high-dust environments, thereby achieving the optimal balance between noise suppression and real-time performance across the entire operating range.

[0056] In step S5, the moisture signal of the pure tobacco is used as the input feature and transmitted to a pre-trained LightGBM moisture prediction model that maintains its structural integrity. The LightGBM moisture prediction model performs inference calculations and outputs a high-precision predicted moisture value. The LightGBM moisture prediction model is an offline trained model with a single input feature dimension, namely the moisture signal of the pure tobacco.

[0057] The model structure remains constant after deployment to the production line, without any online structural adjustments or parameter retraining. The data format of the pure tobacco moisture signal output by the adaptive filtering module is completely consistent with the input data format used during the training of the LightGBM model, including data length, sampling frequency, numerical range, and normalization method, thus achieving seamless decoupling between the filtering front-end and the prediction back-end.

[0058] This decoupled architecture ensures that the method of this invention can be deployed without any modification to the existing calibrated and verified LightGBM model, greatly reducing the complexity and cost of engineering implementation. The inference time of the LightGBM model is optimized to less than twelve milliseconds, ensuring that the overall front-end processing latency (dust sensing + filtering + prediction) does not exceed thirty-three milliseconds.

[0059] In step S6, based on the deviation between the high-precision predicted moisture value and the target moisture setting value of the tobacco feeding process, a control command for the feeding system is generated to precisely control the final moisture content of the tobacco in a closed-loop manner. The control command generation module integrates a proportional-integral-derivative controller.

[0060] The controller receives the high-precision predicted moisture value from the moisture prediction module as a process variable and the target moisture setpoint stored in the process database as the setpoint. The controller calculates the error between the predicted value and the setpoint and generates an analog output signal with a range of four to twenty milliamps based on preset proportional, integral, and derivative parameters.

[0061] The analog output signal is directly connected to the frequency input of the variable frequency feeding pump on the tobacco feeding line via a cable. By adjusting the speed of the feeding pump, the amount of liquid sprayed onto the tobacco per unit time is precisely controlled. This closed-loop control mechanism ensures that the moisture content of the tobacco remains stable within the target value of ±0.5%, meeting industry standard requirements.

[0062] Corresponding to the above method, the present invention also provides a tobacco shred moisture control system based on LightGBM, which includes a working condition sensing module, a signal acquisition module, a filter parameter dynamic generation module, an adaptive filtering module, a moisture prediction module, and a control command generation module.

[0063] Specifically, the operating condition sensing module is configured to acquire real-time dust concentration data at designated monitoring points on the tobacco processing production line. This module consists of a laser scattering dust concentration sensor, whose probe is installed in the same noise-affected area as the infrared moisture sensor to ensure the representativeness of the measurement data. The sensor outputs data via an RS485 bus using the Modbus RTU protocol, with a communication latency of less than one millisecond.

[0064] The signal acquisition module is configured to synchronously acquire the raw, noisy tobacco moisture signal output from the infrared moisture sensor. This module includes a high-precision analog-to-digital converter and a time synchronization unit to ensure strict alignment of the moisture signal and dust concentration data on the time axis. The signal is buffered frame by frame, with 1024 points per frame, for subsequent filtering and processing.

[0065] The dynamic filter parameter generation module is configured to receive real-time dust concentration data output from the operating condition sensing module and dynamically calculate and generate a set of optimal wavelet filter parameters based on the data using a built-in nonlinear parameter mapping model. This module is implemented using a field-programmable gate array (FPGA) chip, which internally incorporates segmented judgment logic and linear calculation units, with a parameter generation delay of no more than one millisecond.

[0066] The adaptive filtering module is configured to receive the noisy raw tobacco moisture signal from the signal acquisition module and the optimal wavelet filter parameters from the dynamic filter parameter generation module. It then applies these parameters to perform adaptive wavelet transform filtering on the raw signal to output a clean tobacco moisture signal. This module is implemented using a dedicated digital signal processor and incorporates wavelet decomposition, thresholding, and reconstruction algorithms, with a single-frame processing delay of no more than twenty milliseconds.

[0067] The moisture prediction module receives the moisture signal of the pure tobacco from the adaptive filtering module and performs prediction using an internally embedded LightGBM model that does not require online retraining, outputting a high-precision predicted moisture value. This module is an embedded computing unit that runs a LightGBM model trained offline with fixed parameters, achieving an inference latency of no more than twelve milliseconds.

[0068] The control command generation module is configured to calculate and generate precise control commands for the feeding actuator based on the difference between the high-precision predicted moisture value output by the moisture prediction module and the process target moisture setpoint. This module integrates a proportional-integral-derivative control algorithm, outputting a 4-20 mA analog signal to directly drive the variable frequency feeding pump, achieving closed-loop moisture control.

[0069] This invention, through the aforementioned method and system, achieves dust concentration-driven adaptive filtering, resolving the dilemma of "high precision → high latency" or "low latency → high error" in existing technologies. In actual testing on a production line of Yunnan Tobacco, this invention reduced the root mean square error of moisture prediction from 2.1% to 0.58%, compressed the total front-end latency from 200 milliseconds to 33 milliseconds, shortened the control lag distance from 40 centimeters to 6.6 centimeters, and increased the moisture qualification rate from 78% to 99.3%. Furthermore, since no modification to the existing LightGBM model is required, production line upgrades only require the addition of a front-end filtering unit, with downtime of less than two hours and a cost reduction of over 95%. This solution provides the tobacco industry with a new paradigm for intelligent moisture control characterized by high precision, low latency, and easy deployment.

[0070] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0071] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A LightGBM-based cut filler moisture regulation method, characterized in that, include: S1: Real-time acquisition of dust concentration data at designated monitoring points on the tobacco processing production line; S2: Simultaneously acquire the raw moisture signal of tobacco shreds with noise output by the infrared moisture sensor near the dust concentration data monitoring point; S3: Based on the real-time dust concentration data, a set of wavelet filter parameters matching the current dust concentration is dynamically calculated and generated through a preset nonlinear parameter mapping model. The wavelet filter parameters include the wavelet basis function type, the number of decomposition layers, and the threshold. S4: Apply dynamically generated wavelet filter parameters to perform adaptive wavelet transform filtering on the original noisy tobacco moisture signal to eliminate noise components caused by dust interference, reconstruct and output a pure tobacco moisture signal; S5: The moisture signal of pure tobacco is used as the input feature and transmitted to a pre-trained LightGBM moisture prediction model that maintains its structure. The LightGBM moisture prediction model performs inference calculations and outputs a high-precision predicted moisture value. S6: Based on the deviation between the high-precision predicted moisture value and the target moisture setting value of the tobacco feeding process, the control command of the feeding system is generated to accurately control the final moisture content of the tobacco in a closed-loop manner. In step S3, a set of wavelet filter parameters matching the current dust concentration is dynamically calculated and generated, specifically including: When the dust concentration data C is less than two milligrams per cubic meter, select the wavelet basis function type as db2 and the number of decomposition levels as two; When the dust concentration data C is greater than or equal to two and less than six milligrams per cubic meter, the wavelet basis function type is selected as sym4 and the number of decomposition levels is three. When the dust concentration data C is greater than or equal to six milligrams per cubic meter, select the wavelet basis function type as coif3 and the number of decomposition layers as four; The threshold λ is calculated based on the linear function relationship λ=0.21×C+1.15, where C is the dust concentration data in milligrams per cubic meter.

2. The LightGBM-based cut filler moisture regulation method according to claim 1, wherein, In step S1, dust concentration data at designated monitoring points on the tobacco processing production line is acquired in real time, specifically including: Instantaneous dust concentration values ​​are obtained using a laser scattering dust concentration sensor; The probe of the laser scattering dust concentration sensor is installed 50 mm below the optical path of the infrared moisture sensor and is connected to the subsequent processing unit via shielded twisted pair cable in RS485 bus mode. It follows the Modbus RTU communication protocol and reports dust concentration data at a frequency of 100 times per second.

3. The LightGBM-based cut filler moisture regulation method of claim 1, wherein, In step S4 above, dynamically generated wavelet filter parameters are applied to perform adaptive wavelet transform filtering on the original moisture content signal of noisy tobacco shreds, specifically including: Obtain a raw moisture signal vector of noisy tobacco with a data length of 1024; Based on the wavelet basis function type and the number of decomposition levels, multi-level wavelet decomposition is performed on the original signal vector of moisture content in noisy tobacco to obtain an approximate coefficient subband cA_L and L detail coefficient subbands cD_1 to cD_L. Based on the threshold λ, apply the soft thresholding function x'=sign(x)×max(0,|x|-λ) to each coefficient x in each detail coefficient subband cD_j to obtain a new detail coefficient subband; By performing inverse wavelet transform reconstruction using the approximate coefficient subband cA_L and all new detail coefficient subbands, a pure tobacco moisture signal vector with a data length of 1024 is obtained.

4. The LightGBM-based cut filler moisture regulation method of claim 3, wherein, The moisture content signal of pure tobacco is used as input feature and fed into a pre-trained LightGBM moisture prediction model that maintains its structural invariance. The LightGBM moisture prediction model performs inference calculations and outputs high-precision predicted moisture values, including: The moisture signal vector of pure tobacco is input into the LightGBM moisture prediction model in the same data format as when the LightGBM model was trained. The LightGBM moisture prediction model is a single-input model trained offline. Its structure remains constant after deployment and no online structure adjustment or parameter retraining is performed. The LightGBM moisture prediction model completes inference and outputs a high-precision predicted moisture value within 12ms after receiving input.

5. The LightGBM-based cut filler moisture regulation method of claim 1, wherein, In step S6, based on the deviation between the high-precision predicted moisture value and the target moisture setting value of the tobacco feeding process, a control command for the feeding system is generated to precisely control the final moisture content of the tobacco in a closed-loop manner. The specific steps include: The high-precision predicted moisture value is used as a process variable, and the target moisture value of tobacco stored in the existing tobacco production process database is used as the set point. The error between the process variable and the setpoint is calculated by a proportional-integral-derivative controller, and an analog output signal of 4 to 20mA is generated according to the preset proportional, integral and derivative parameters. The analog output signal is transmitted to the frequency input terminal of the variable frequency feeding pump in the existing production process, and the amount of liquid sprayed onto the tobacco per unit time is controlled by adjusting the speed of the feeding pump.

6. The LightGBM-based cut filler moisture regulation method of claim 1, wherein, The original moisture signal of the noisy tobacco shreds was strictly synchronized with the dust concentration data on the time axis. The original moisture signal of the noisy tobacco shreds was collected in frames of 1024 points each, with a sampling frequency of 500 Hz and a single frame duration of 2.048 s.

7. A LightGBM-based tobacco shred moisture control system, applied to the LightGBM-based tobacco shred moisture control method as described in any one of claims 1-6, characterized in that, include: The working condition sensing module is used to acquire real-time dust concentration data at designated monitoring points in the tobacco processing production line; The signal acquisition module is used to synchronously acquire the original moisture signal of the tobacco shreds with noise output by the infrared moisture sensor. The physical location of the sensor of the signal acquisition module and the physical location of the sensor of the working condition sensing module are in the same noise-affected area. The filter parameter dynamic generation module is used to receive the real-time dust concentration data output by the working condition sensing module, and dynamically calculate and generate a set of wavelet filter parameters that match the current dust concentration based on the data through a built-in nonlinear parameter mapping model. An adaptive filtering module is used to receive the original noisy tobacco moisture signal output by the signal acquisition module and the wavelet filtering parameters output by the dynamic generation module of the filtering parameters, and to use the parameters to perform adaptive wavelet transform filtering on the original signal to output a pure tobacco moisture signal. The moisture prediction module is used to receive the moisture signal of the pure tobacco shreds output by the adaptive filtering module, and to predict the moisture value by using an internally fixed LightGBM model that does not require online retraining. The control command generation module is used to calculate and generate precise control commands for the feeding actuator based on the difference between the high-precision predicted moisture value output by the moisture prediction module and the process target moisture setting value.

Images